harmonisation standard UQO WP (2026-10-17)

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Simon-Pierre Boucher committed 1 h ago (Oct 17, 2026)

61 changed files +7,212 −0

added .env.example +2 −0

		`@@ -0,0 +1,2 @@`
	1	+OPENAI_API_KEY=
	2	+ANTHROPIC_API_KEY=

added .gitignore +42 −0

		`@@ -0,0 +1,42 @@`
	1	+# Raw & processed data — never committed (see data/raw/README.md for provenance)
	2	+data/raw/*
	3	+!data/raw/README.md
	4	+data/processed/*
	5	+!data/processed/.gitkeep
	6	+
	7	+# Python
	8	+__pycache__/
	9	+*.pyc
	10	+.pytest_cache/
	11	+.venv/
	12	+*.egg-info/
	13	+
	14	+# LaTeX build artifacts (paper/main.pdf IS committed)
	15	+paper/*.aux
	16	+paper/*.log
	17	+paper/*.out
	18	+paper/*.toc
	19	+paper/*.lof
	20	+paper/*.lot
	21	+paper/*.fls
	22	+paper/*.fdb_latexmk
	23	+paper/*.bbl
	24	+paper/*.blg
	25	+paper/*.spl
	26	+paper/*.synctex.gz
	27	+paper/*.run.xml
	28	+paper/*.bcf
	29	+
	30	+# Logs
	31	+logs/
	32	+*.log
	33	+
	34	+# macOS
	35	+.DS_Store
	36	+
	37	+# Project-specific
	38	+# OPENAI_API_KEY / ANTHROPIC_API_KEY — never commit
	39	+.env
	40	+# LLM/vision caches (specs.jsonl, facade_specs.jsonl), merged datasets, OOS predictions
	41	+data/processed/*
	42	+!data/processed/README.md

added README.md +164 −0

		`@@ -0,0 +1,164 @@`
	1	+<!-- Author: Simon-Pierre Boucher — simon-pierre.boucher@uqo.ca -->
	2	+# 🏠 Hedonic House Pricing with Language-Model and Vision Features
	3	+
	4	+UQO Working Paper No. 14 — Evidence from Quebec Real-Estate Listings
	5	+
	6	+[![Paper](https://img.shields.io/badge/paper-PDF%20(33%20p.)-b31b1b?logo=latex&logoColor=white)](paper/main.pdf)
	7	+[![Python](https://img.shields.io/badge/python-3.11%2B-3776AB?logo=python&logoColor=white)](requirements.txt)
	8	+[![Data](https://img.shields.io/badge/data-19%2C123%20house%20listings%20·%20Realtor.ca%20%2B%20DuProprio-4051b5)](data/raw/README.md)
	9	+[![Reproducible](https://img.shields.io/badge/reproducible-end--to--end%20pipeline-2e7d32)](scripts/)
	10	+[![Institution](https://img.shields.io/badge/UQO-D%C3%A9pt.%20des%20sciences%20administratives-16365c)](https://uqo.ca)
	11	+[![Author](https://img.shields.io/badge/author-Simon--Pierre%20Boucher-16365c)](mailto:simon-pierre.boucher@uqo.ca)
	12	+
	13	+> TL;DR — Much of a home's value is recorded only in the free text and the photographs of its listing. We use a
	14	+> large language model to turn each description into structured specifications and a vision–language model to turn the
	15	+> façade photo into objective visual attributes, both under strict JSON schemas. On 19,123 Quebec house listings the two
	16	+> AI layers lift the R² of an already rich log-linear hedonic model from 0.717 to 0.755; the vision layer is the largest
	17	+> out-of-sample gain of any feature block for every estimator (OLS, gradient boosting, neural nets); the best ensemble
	18	+> reaches out-of-sample R² 0.779 with a 14.0% median absolute error. Implicit prices are economically sensible (+2.3% per
	19	+> point of curb appeal, +3–4% for brick or stone) and survive municipality-clustered inference.
	20	+
	21	+---
	22	+
	23	+## 📌 Status
	24	+
	25	+\| Item \| Value \|
	26	+\|---\|---\|
	27	+\| Version \| 2.0 (July 2026) — revised draft after an internal review round (`scripts/13_revision_extras.py`) \|
	28	+\| SSRN \| not yet submitted \|
	29	+\| Journal \| Elsevier-format highlights prepared (`docs/highlights_elsevier.tex`); target to be decided \|
	30	+\| Code & data \| reproducible from `scripts/` run from the repository root; raw crawls and LLM caches not committed \|
	31	+
	32	+## 🎯 What this paper does
	33	+
	34	+Hedonic models are bounded by the characteristics an analyst can measure. We treat general-purpose AI models as a
	35	+feature-engineering technology: each listing description is parsed by an LLM into a fixed schema of property
	36	+specifications (renovations, finishes, systems, extras), and each façade photograph is parsed by a vision–language model
	37	+into objective attributes (cladding, storeys, garage, curb appeal). The outputs are ordinary econometric variables and
	38	+enter a log-linear hedonic regression with city fixed effects:
	39	+
	40	+```
	41	+ln P_i = α_city(i) + X_i'β + S_i'γ (LLM text layer) + V_i'δ (vision layer) + ε_i
	42	+```
	43	+
	44	+We then measure, under identical 5-fold splits, how much each layer adds for OLS, histogram gradient boosting, a
	45	+residual MLP with city/type embeddings, and a stacked blend.
	46	+
	47	+## 🏆 Headline results
	48	+
	49	+\| Object \| Estimate \|
	50	+\|---\|---\|
	51	+\| Base hedonic OLS (structural + location), in-sample R² \| 0.717 \|
	52	+\| + LLM text layer + vision layer \| 0.755 \|
	53	+\| Out-of-sample R², OLS: base → +text → +text+vision (5-fold) \| 0.52 → 0.62 → … (see `paper/tables/ablation.tex`) \|
	54	+\| Out-of-sample R², GBM: base → +text \| 0.69 → 0.73 \|
	55	+\| Best ensemble (robust NN + GBM blend) \| R²_OOS = 0.779, median absolute error 14.0% \|
	56	+\| Curb-appeal score (1–5) \| +2.3% per point \|
	57	+\| Brick / stone façade \| +3–4% \|
	58	+\| Inference \| HC3 and municipality-clustered standard errors agree \|
	59	+
	60	+## 🗂 Repository layout
	61	+
	62	+```
	63	+hedonic-llm-vision-features/
	64	+├── data/
	65	+│ ├── raw/ # two Apify crawls (Realtor.ca, DuProprio), 2026-07-03 — not committed
	66	+│ └── processed/ # LLM/vision caches, merged datasets, OOS predictions — not committed
	67	+├── scripts/ # run from the repository root, in order
	68	+│ ├── 00a_test_openai_extraction.py / 00b_test_vision_claude.py # API smoke tests
	69	+│ ├── 01_extract_specs_llm.py # description → JSON specs (parallel, resumable)
	70	+│ ├── 02_facade_vision.py # façade photo → visual attributes
	71	+│ ├── 03_merge_llm_features.py
	72	+│ ├── 04_hedonic_ols_base.py / 05_hedonic_ols_v4_text.py
	73	+│ ├── 06_hedonic_ml_cv.py / 07_torch_nn.py / 08_torch_nn_robust.py
	74	+│ ├── 09_v5_ml_text_vision.py / 10_final_blend.py
	75	+│ ├── 11_extras_v5_robustness.py / 12_ablation_extras.py / 13_revision_extras.py
	76	+│ └── 14_figures_v5.py / 15_make_tables.py / 16_make_tables_v5.py / 17_gen_regression_table.py
	77	+├── results/ # hedonic summaries (txt), ablation_extras.json, revision_extras.json
	78	+├── docs/ # METHODOLOGIE.md + methodologie.tex/pdf (French data & method note), Elsevier highlights
	79	+└── paper/ # main.tex + sections/ + tables/ + figures/ + references.bib → main.pdf
	80	+```
	81	+
	82	+## 🚀 Quick start
	83	+
	84	+```bash
	85	+python3 -m pip install -r requirements.txt
	86	+cp .env.example .env # OPENAI_API_KEY=…, ANTHROPIC_API_KEY=… (only steps 00–02 call the APIs)
	87	+# place the two Apify crawls in data/raw/ and build data/processed/maisons_avec_prix.json (docs/METHODOLOGIE.md §2–4), then:
	88	+WORKERS=24 python3 scripts/01_extract_specs_llm.py
	89	+python3 scripts/02_facade_vision.py
	90	+python3 scripts/03_merge_llm_features.py
	91	+python3 scripts/05_hedonic_ols_v4_text.py
	92	+python3 scripts/08_torch_nn_robust.py && python3 scripts/09_v5_ml_text_vision.py && python3 scripts/10_final_blend.py
	93	+python3 scripts/11_extras_v5_robustness.py && python3 scripts/12_ablation_extras.py && python3 scripts/13_revision_extras.py
	94	+python3 scripts/14_figures_v5.py && python3 scripts/16_make_tables_v5.py && python3 scripts/17_gen_regression_table.py
	95	+cd paper && make # latexmk → main.pdf
	96	+```
	97	+
	98	+LLM extraction ≈ 600 tokens in / 110 out per listing (24 threads); vision ≈ 2,470 tokens in per image (16 threads);
	99	+both are resumable. Econometrics and ML ≈ 20 minutes on an Apple-silicon laptop (PyTorch on MPS).
	100	+
	101	+## 🗃 Data
	102	+
	103	+27,302 Realtor.ca and 7,693 DuProprio listings crawled on 2026-07-03; houses (single-family and townhouse) with a valid
	104	+price retained, outliers removed by the rules in `docs/METHODOLOGIE.md` §3, giving 19,123 observations. Structural
	105	+variables (living area, lot, rooms, year built, type, city), text-derived specifications and façade attributes. See
	106	+`data/raw/README.md` and `data/processed/README.md`.
	107	+
	108	+## 🔬 Methodology
	109	+
	110	+1. AI feature layers with strict JSON schemas and retry/backoff — `scripts/01_…`, `scripts/02_…`.
	111	+2. Log-linear hedonic OLS with city fixed effects, HC3 and municipality-clustered SEs — `scripts/05_…`, `scripts/13_…`.
	112	+3. Ablation by feature block × estimator under identical folds (OLS, GBM, NN) — `scripts/09_…`, `scripts/12_…`.
	113	+4. Stacked blend of the robust NN and HistGradientBoosting — `scripts/10_final_blend.py`.
	114	+5. Robustness by segment, extreme city effects, calibration and importance plots — `scripts/11_…`, `scripts/14_…`.
	115	+
	116	+## 🖼 Figures & tables inventory
	117	+
	118	+\| Figure \| Content \|
	119	+\|---\|---\|
	120	+\| `fig_ablation` \| out-of-sample gain by feature block and estimator \|
	121	+\| `fig_calibration` \| predicted vs actual, blend \|
	122	+\| `fig_importance` \| feature importance \|
	123	+\| `fig_premia` \| location premia \|
	124	+\| `fig_validation` \| validation of AI-extracted attributes \|
	125	+\| `nn_diagram` (TikZ) \| architecture of the robust NN \|
	126	+
	127	+Tables (`paper/tables/*.tex`): summary stats, regression (A/B panels), models, ablation, robustness, AI summary.
	128	+
	129	+## ⚠️ Limitations
	130	+
	131	+- Asking prices from listings, single cross-section (July 2026).
	132	+- Realtor.ca listings lack coordinates; location enters through city fixed effects and postal codes.
	133	+- LLM/vision outputs are model-dependent; schemas and prompts are versioned in the scripts.
	134	+- API costs make the extraction layers the expensive part of replication.
	135	+
	136	+## 📝 Citation
	137	+
	138	+```bibtex
	139	+@techreport{boucher2026llmvisionhedonic,
	140	+ author = {Boucher, Simon-Pierre},
	141	+ title = {Hedonic House Pricing with Language-Model and Vision Features: Evidence from Quebec Real-Estate Listings},
	142	+ institution = {Universit\'e du Qu\'ebec en Outaouais, D\'epartement des sciences administratives},
	143	+ type = {Working Paper},
	144	+ number = {14},
	145	+ year = {2026},
	146	+ month = {July}
	147	+}
	148	+```
	149	+
	150	+## 👤 Author & contact
	151	+
	152	+Simon-Pierre Boucher — Département des sciences administratives,
	153	+Université du Québec en Outaouais (UQO), Gatineau, QC.
	154	+📧 [simon-pierre.boucher@uqo.ca](mailto:simon-pierre.boucher@uqo.ca)
	155	+
	156	+## 🗒 Notes
	157	+
	158	+- Numbering: this paper was first issued as "Working Paper No. 11" (July 2026); number 11 belongs to *Half a Million
	159	+ Prices, Twenty Models. Renumbered 14* on 2026-10-17; cover date fixed to July 2026 (was `\today`).
	160	+- Harmonised 2026-10-17: flat scripts moved to `scripts/` (numbered), data to `data/`, logs removed, summaries to
	161	+ `results/`, methodology note to `docs/`; file paths inside the scripts now point to `data/…` and `results/…`
	162	+ (run them from the repository root). Bibliography switched from `plainnat` to `aer`.
	163	+- The script `13_revision_extras.py` was written for a Journal of Housing Economics revision round (clustered SEs,
	164	+ in-sample ablation, extreme city effects).

added data/processed/README.md +10 −0

		`@@ -0,0 +1,10 @@`
	1	+# Processed data (not committed)
	2	+
	3	+\| File \| Produced by \| Content \|
	4	+\|---\|---\|---\|
	5	+\| `maisons_avec_prix.json` \| house filter + cleaning described in `docs/METHODOLOGIE.md` §2–4 (combined Realtor + DuProprio houses with price) \| 19,123 house listings \|
	6	+\| `specs.jsonl` \| `scripts/01_extract_specs_llm.py` (GPT-4o-mini, strict JSON schema, resumable cache keyed by uid) \| structured specifications extracted from descriptions \|
	7	+\| `facade_specs.jsonl` \| `scripts/02_facade_vision.py` (vision–language model on the façade photo) \| objective visual attributes (curb appeal, cladding, …) \|
	8	+\| `maisons_specs.json` / `.csv` \| `scripts/03_merge_llm_features.py` \| listings merged with LLM features \|
	9	+\| `maisons_hedonic.csv`, `maisons_hedonic_v4.csv` \| `scripts/04_…`, `scripts/05_…` \| final datasets with fitted price and `resid_pct` \|
	10	+\| `nn_robust_oos.npy`, `v5_blend_oos.npy` \| `scripts/08_…`, `scripts/09_…` \| out-of-sample predictions (5-fold, seed 42) reused by steps 10–14 \|

added data/raw/README.md +14 −0

		`@@ -0,0 +1,14 @@`
	1	+# Raw data — Realtor.ca and DuProprio listings (Apify crawlers, 2026-07-03)
	2	+
	3	+\| File \| Source \| Records \| Size \|
	4	+\|---\|---\|---\|---\|
	5	+\| `dataset_realtor-property-crawler_2026-07-03_10-36-26-001.json` \| Realtor.ca (MLS) via Apify `realtor-property-crawler` \| 27,302 listings \| 161 MB \|
	6	+\| `dataset_duproprio-property-crawler_2026-07-03_00-37-49-449.json` \| DuProprio via Apify `duproprio-property-crawler` \| 7,693 listings \| 87 MB \|
	7	+
	8	+Cross-sectional snapshot extracted on 2026-07-03 (no time dimension). Common fields: `source`, `url`, `listingId`,
	9	+`price`, `address{…}`, `property{…}`, `features[]`, `description`, `images[]`. Realtor adds `rooms[]` (room dimensions)
	10	+but no coordinates; DuProprio adds `coordinates{lat,lon}` and captioned images. Unique key: `uid = source:listingId`.
	11	+Houses (single-family + townhouse) with a valid price: 19,123 listings used in the paper.
	12	+
	13	+Not committed (size and crawler terms). Re-run the two Apify actors or request the snapshot from the author.
	14	+Full data dictionary and cleaning rules: `docs/METHODOLOGIE.md` (French).

added docs/METHODOLOGIE.md +172 −0

		`@@ -0,0 +1,172 @@`
	1	+# Méthodologie — Modèle hédonique de prix des maisons (Québec)
	2	+
	3	+Pipeline : Realtor.ca + DuProprio → nettoyage → feature engineering (règles + LLM + vision) → régression hédonique log-linéaire à effets fixes.
	4	+
	5	+---
	6	+
	7	+## 1. Origine des données
	8	+
	9	+\| Source \| Fichier \| Nb enreg. \| Nature \|
	10	+\|---\|---\|---\|---\|
	11	+\| Realtor.ca \| `dataset_realtor-property-crawler_...json` (161 Mo) \| 27 302 \| Annonces MLS via crawler Realtor (Apify) \|
	12	+\| DuProprio \| `dataset_duproprio-property-crawler_...json` (87 Mo) \| 7 693 \| Annonces vente sans intermédiaire (Apify) \|
	13	+
	14	+- Collecte : scrapers Apify, extraction le 2026-07-03. Chaque enregistrement est un objet JSON structuré.
	15	+- Champs communs : `source`, `url`, `listingId`, `title`, `price`/`priceNumber`, `currency`, `address{full,street,city,province,postalCode,...}`, `property{...}`, `features[]`, `description`, `images[]`, `scrapedAt`.
	16	+- Spécificités :
	17	+ - Realtor : `property.buildingType`, `storeys`, `yearBuilt`, `parkingSpaces`, `rooms[]` (dimensions par pièce), pas de coordonnées GPS, images sans caption utile.
	18	+ - DuProprio : `property.type`, `levels`, `coordinates{lat,lon}` (7 691 présentes), images avec captions (`Kitchen`, etc.), pas de `rooms`.
	19	+- Instantané transversal (cross-section) : un seul point dans le temps → pas de dimension temporelle, donc pas d'effet fixe temporel nécessaire.
	20	+
	21	+Identifiant unique construit : `uid = source + ":" + listingId` (évite toute collision entre les deux espaces d'ID).
	22	+
	23	+---
	24	+
	25	+## 2. Construction du dataset combiné « maisons »
	26	+
	27	+1. Fusion des deux tableaux JSON (schéma harmonisé).
	28	+2. Définition de « maison » (choix utilisateur : unifamiliales + townhouses) :
	29	+ - Realtor : `buildingType ∈ {House, Row / Townhouse}`
	30	+ - DuProprio : `type ∈ {Home, Bungalow, Semi Detached, Intergenerational, Country Home, Cottage, Villa, Townhouse}`
	31	+ - Exclus : condos, plex (duplex→quintuplex), terrains/lots, fermes, commercial, mobile homes.
	32	+3. Filtre prix présent : `priceNumber` non nul et > 0.
	33	+
	34	+→ 19 273 maisons avec prix (15 470 Realtor + 3 803 DuProprio).
	35	+
	36	+---
	37	+
	38	+## 3. Nettoyage des aberrations
	39	+
	40	+Les prix scrappés contiennent des erreurs de parsing (min 899 $, max 368 G$). Règles appliquées :
	41	+
	42	+\| Variable \| Règle \| Justification \|
	43	+\|---\|---\|---\|
	44	+\| Prix \| garder 50 k$ – 15 M$ \| plage plausible d'une maison au QC ; retire loyers mal parsés et concaténations \|
	45	+\| Superficie habitable \| 300 – 20 000 ft² (sinon → manquant) \| valeurs physiquement possibles \|
	46	+\| Terrain \| 200 – 5 000 000 ft² (sinon → manquant) \| idem \|
	47	+\| Chambres / sdb \| 0–15 / 0–12 (sinon → manquant) \| plafonds réalistes \|
	48	+\| Prix/pied² \| garder P0,5 – P99,5 (≈ 73–1 276 $/ft²) sur les biens à surface réelle \| retire les erreurs résiduelles de parsing prix ou surface \|
	49	+
	50	+→ 19 115 maisons conservées (158 retirées).
	51	+
	52	+Principe : on borne plutôt que de supprimer quand une seule variable est aberrante (on met la variable à manquant et on l'impute), pour ne pas perdre l'observation entière.
	53	+
	54	+---
	55	+
	56	+## 4. Construction des variables
	57	+
	58	+### 4.1 Variable cible
	59	+`ln_price = ln(priceNumber)` — transformation logarithmique (rend la distribution symétrique, linéarise les effets multiplicatifs, coefficients ≈ % d'impact sur le prix).
	60	+
	61	+### 4.2 Variables structurelles (règles déterministes)
	62	+- Surface & terrain : parsing d'unités hétérogènes → tout en pieds carrés.
	63	+ - `"1535 sqft"`, `"1,333.2 ft²"`, `"11 561.52 ft²"` (séparateurs de milliers variés), `"961.6 m2"` (m² × 10,7639).
	64	+- Reconstruction de surface via les pièces (Realtor) : dimensions `"23 ft ,4 in x 13 ft ,8 in"` → pieds-décimaux → aire ; somme sur toutes les pièces. Calibration sur le recouvrement (ratio médian surface réelle / somme pièces ≈ 0,98) → sert à combler `squareFeet` manquant.
	65	+ - Constat : les annonces à pièces détaillées ≈ celles qui ont déjà `squareFeet` → gain de couverture faible, mais les comptes/types de pièces restent informatifs.
	66	+- Autres : `beds`, `baths`, `parking` (Realtor), `storeys`/`levels`, `age = 2026 − yearBuilt` (+ `age²` pour la courbure), drapeaux `has_*` marquant les valeurs imputées.
	67	+
	68	+### 4.3 Aménités (règles par mots-clés bilingues)
	69	+17 dummies détectés dans `features[] + description + title` (FR/EN) : piscine, bord de l'eau, garage, foyer, climatisation/thermopompe, sous-sol fini, cul-de-sac, golf, ski, proximité parc/école/autoroute/transport, absence de voisin arrière.
	70	+
	71	+### 4.4 Détail des pièces (Realtor, `rooms[]`)
	72	+Nombre de pièces, nombre de niveaux, chambre principale, sous-sol, bureau, salle familiale, atelier, cellier à vin, solarium, mezzanine, bibliothèque, buanderie.
	73	+
	74	+### 4.5 Signaux de listing
	75	+Nombre de photos (`n_images`, écrêté à 60), présence de visite virtuelle — proxies de gamme/qualité de la mise en marché.
	76	+
	77	+### 4.6 Variables géographiques (effets fixes)
	78	+- Ville nettoyée (Realtor : on retire le quartier entre parenthèses ; minuscules ; normalisation). Villes avec ≥ 20 annonces → dummy propre ; les rares → `__other__`. ≈ 194 villes en effets fixes.
	79	+- Type de propriété harmonisé, regroupé si < 30 annonces.
	80	+- Source (Realtor/DuProprio) → dummy captant les différences systématiques entre plateformes.
	81	+
	82	+### 4.7 Imputation
	83	+Valeurs manquantes après nettoyage : imputées par médiane par type de propriété (puis médiane globale), *avec un drapeau `has_`** pour que le modèle distingue « valeur observée » de « valeur imputée ».
	84	+
	85	+---
	86	+
	87	+## 5. Mécanismes IA pour le feature engineering
	88	+
	89	+Deux couches IA extraient des variables impossibles à obtenir par règles — l'une du texte, l'autre des images.
	90	+
	91	+### 5.1 Extraction de specs depuis la description — GPT-4o-mini (OpenAI)
	92	+- But : transformer la description libre en 20 attributs structurés (état : à rénover→neuf, clé en main, cuisine/sdb rénovée, revêtements de sol, comptoirs quartz/granit, chauffage, vue sur l'eau, revenus locatifs, grand terrain, secteur tranquille, + mots-clés de vente).
	93	+- Mécanisme : appel Chat Completions avec `response_format = json_schema` strict (schéma imposé, énumérations fermées) → sortie JSON valide garantie, pas de parsing fragile. `temperature = 0` (déterminisme). Système : « extracteur factuel, uniquement d'après le texte ; absence → false/inconnu/[] ».
	94	+- Ingénierie : ~600 tokens in / ~110 out par annonce ; parallèle (24 threads), résumable (cache `specs.jsonl` clé = uid), retry exponentiel sur 429/5xx.
	95	+- Coût/temps : ~$3 / ~19 k annonces.
	96	+
	97	+### 5.2 Attributs visuels de façade — Claude Haiku 4.5 (Anthropic, vision)
	98	+- But : lire ce que le texte ne dit pas — état visuel réel, attrait extérieur (curb appeal), niveau de rénovation, matériaux, aménagement paysager, cadre (urbain/banlieue/rural/bord d'eau/boisé), indice de luxe, qualité de la photo.
	99	+- Sélection d'image : 1 photo/annonce = la façade. Realtor → 1re photo. DuProprio → photo dont la caption est l'adresse (pas une pièce), via exclusion de mots-clés de pièces.
	100	+- Mécanisme : Messages API, image passée par URL (`source.type = url`, pas de téléchargement), `output_config.format = json_schema` (12 attributs, énumérations fermées). Système : « évaluateur objectif ; scores fondés sur critères concrets (surfaces, matériaux, aménagement), pas sur un goût personnel ; si non-extérieur → is_exterior=false ».
	101	+- Ingénierie : ~2 470 tokens in / ~100 out par image ; parallèle (16 threads), résumable (cache `facade_specs.jsonl`), retry sur 429/5xx/529.
	102	+- Coût/temps : ~$57 / ~19 k images.
	103	+- Pourquoi Haiku : modèle vision le moins cher et rapide, largement suffisant pour de la classification d'attributs concrets.
	104	+
	105	+### 5.3 Pourquoi le JSON structuré strict est central
	106	+Le schéma imposé (types + `enum` + champs requis + `additionalProperties:false`) fait valider la sortie au niveau de l'API : le modèle est forcé de renvoyer exactement les champs attendus, dans le bon domaine de valeurs. On obtient des variables directement tabulables, reproductibles, sans post-traitement de texte libre.
	107	+
	108	+### 5.4 Intégration au dataset
	109	+Les sorties LLM (clé `uid`) sont converties en variables numériques : booléens → 0/1, `condition` → score ordinal 0-5, revêtements → dummies (bois franc, quartz/granit), vue → dummy eau, etc., avec un drapeau `has_gpt` / `is_exterior` pour marquer la disponibilité.
	110	+
	111	+---
	112	+
	113	+## 6. Modèle économétrique
	114	+
	115	+### 6.1 Spécification — hédonique log-linéaire à effets fixes
	116	+$$\ln(P_i) = \alpha + \beta\,\ln(\text{surface}_i) + \gamma\,\ln(\text{terrain}_i) + \sum_k \delta_k X_{ki} + \sum_c \lambda_c \mathbb{1}[\text{ville}_i=c] + \varepsilon_i$$
	117	+
	118	+- Log-log sur surface et terrain → coefficients = élasticités (% prix par % surface).
	119	+- Log-niveau sur le reste → un coefficient $b$ ≈ $(e^b-1)\times100$ % d'impact sur le prix.
	120	+- Effets fixes ville ($\lambda_c$) : absorbent tout le différentiel de localisation (le plus gros déterminant du prix). Chaque coefficient = prime/décote de la ville vs la ville de référence.
	121	+- Effets fixes type de propriété et source.
	122	+
	123	+### 6.2 Double spécification (couverture vs richesse)
	124	+Realtor ne rapporte `squareFeet` que pour ~38 % des annonces → deux modèles, prédictions combinées :
	125	+- Modèle A (riche) : inclut `ln(surface)`, estimé sur les biens à surface mesurée (réelle ou reconstruite via pièces). Meilleur ajustement.
	126	+- Modèle B (large) : sans surface, estimé sur toutes les maisons → sert à pricer celles sans surface (via beds/baths/terrain/pièces + specs).
	127	+- Prédiction finale : A si surface mesurée, sinon B.
	128	+
	129	+### 6.3 Estimation & inférence
	130	+- MCO (OLS) sur `ln_price`.
	131	+- Erreurs-types robustes à l'hétéroscédasticité HC3 (les résidus immobiliers sont hétéroscédastiques ; HC3 est conservateur en petit/moyen échantillon).
	132	+- Correction du biais de rétrotransformation : $\hat{P} = e^{\hat{\ln P}} \times e^{\hat\sigma^2/2}$ (correction normale de Duan) — sinon $e^{\hat{\ln P}}$ sous-estime systématiquement le niveau.
	133	+
	134	+### 6.4 Interprétation
	135	+Coefficients rapportés en % d'impact prix + significativité (p-values sur SE HC3). Effets fixes ville présentés comme primes de localisation.
	136	+
	137	+---
	138	+
	139	+## 7. Statistiques & diagnostics
	140	+
	141	+\| Indicateur \| Signification \|
	142	+\|---\|---\|
	143	+\| R² / R²-ajusté \| part de variance de `ln_price` expliquée \|
	144	+\| Écart-type des résidus (log) \| → erreur typique en % = $(e^{\sigma}-1)\times100$ \|
	145	+\| Erreur absolue médiane \| \\|prix réel / prix prédit − 1\\| médian sur toutes les prédictions \|
	146	+\| % à ±15 % / ±25 % \| part des prédictions dans la fourchette \|
	147	+\| `resid_pct` par annonce \| surprix/sous-prix vs modèle → repère aubaines & surévaluations \|
	148	+
	149	+Progression observée : v2 (structurel) R²≈0,70 → v3 (+ pièces/photos) R²≈0,72 → v4 (+ specs GPT) → v5 (+ vision) attendu.
	150	+
	151	+---
	152	+
	153	+## 8. Limites & biais (transparence)
	154	+
	155	+- Endogénéité / non-causalité : de belles photos ou des specs riches corrèlent avec des courtiers haut de gamme et des biens chers → coefficients prédictifs, pas causaux. À ne pas lire comme « ajouter une piscine ⇒ +X % ».
	156	+- Biais de sélection : uniquement des biens en vente (prix demandé, pas prix de transaction) ; les biens sans surface/pièces sont non aléatoires (annonces moins complètes).
	157	+- Imputation : les médianes par type réduisent la variance des variables imputées (atténuation possible des coefficients) — d'où les drapeaux `has_*`.
	158	+- Extraction LLM : erreurs/hallucinations possibles ; `temperature=0` + schéma strict + énumérations fermées les limitent, mais pas de validation humaine systématique. La vision juge sur une seule photo.
	159	+- Instantané unique : pas de dynamique de marché ; primes de localisation figées au 2026-07-03.
	160	+- Couverture géographique : villes < 20 annonces regroupées en `__other__` (moins de finesse spatiale là où les données sont rares).
	161	+
	162	+---
	163	+
	164	+## 9. Fichiers produits
	165	+
	166	+\| Fichier \| Contenu \|
	167	+\|---\|---\|
	168	+\| `hedonic.py`, `hedonic_v4.py` \| pipelines reproductibles (nettoyage + modèles) \|
	169	+\| `extract_specs.py`, `facade_vision.py` \| extracteurs LLM (texte / vision), parallèles & résumables \|
	170	+\| `specs.jsonl`, `facade_specs.jsonl` \| sorties LLM brutes (clé uid) \|
	171	+\| `maisons_hedonic*.csv` \| dataset final + prix prédit + `resid_pct` \|
	172	+\| `hedonic_*_summary.txt` \| sorties statistiques complètes \|

added docs/highlights_elsevier.tex +17 −0

		`@@ -0,0 +1,17 @@`
	1	+% ============================================================================
	2	+% Highlights — separate file required at Elsevier submission
	3	+% (3-5 bullets, max 85 characters each, results-oriented)
	4	+% ============================================================================
	5	+\documentclass[12pt]{article}
	6	+\usepackage[utf8]{inputenc}
	7	+\usepackage[T1]{fontenc}
	8	+\begin{document}
	9	+\section*{Highlights}
	10	+\begin{itemize}
	11	+ \item LLM and vision models turn listing text and photos into hedonic regressors
	12	+ \item Schema-constrained decoding makes AI outputs reproducible econometric variables
	13	+ \item Two AI feature layers add four points of $R^2$ to a rich hedonic model
	14	+ \item Façade-vision features give the largest out-of-sample gain of any block
	15	+ \item Curb appeal, materials, and luxury cues carry sensible implicit prices
	16	+\end{itemize}
	17	+\end{document}

added docs/methodologie.pdf +0 −0

Binary file not shown.

added docs/methodologie.tex +372 −0

		`@@ -0,0 +1,372 @@`
	1	+\documentclass[11pt,a4paper]{article}
	2	+
	3	+\usepackage[utf8]{inputenc}
	4	+\usepackage[T1]{fontenc}
	5	+\usepackage[french]{babel}
	6	+\usepackage{lmodern}
	7	+\usepackage{amsmath,amssymb}
	8	+\usepackage{booktabs}
	9	+\usepackage{longtable}
	10	+\usepackage{array}
	11	+\usepackage{geometry}
	12	+\usepackage{xcolor}
	13	+\usepackage{enumitem}
	14	+\usepackage{hyperref}
	15	+\usepackage{fancyhdr}
	16	+
	17	+\geometry{margin=2.4cm}
	18	+\hypersetup{colorlinks=true, linkcolor=blue!55!black, urlcolor=blue!55!black,
	19	+ citecolor=blue!55!black}
	20	+
	21	+\newcommand{\code}[1]{\texttt{\small #1}}
	22	+\renewcommand{\arraystretch}{1.25}
	23	+
	24	+\pagestyle{fancy}
	25	+\fancyhf{}
	26	+\lhead{\small Modèle hédonique de prix — Québec}
	27	+\rhead{\small\thepage}
	28	+\renewcommand{\headrulewidth}{0.4pt}
	29	+
	30	+\title{\textbf{Méthodologie}\\[2mm]
	31	+\large Modèle hédonique de prix des maisons (Québec)\\[1mm]
	32	+\normalsize Realtor.ca $+$ DuProprio $\rightarrow$ nettoyage $\rightarrow$ ingénierie
	33	+de variables (règles $+$ LLM $+$ vision) $\rightarrow$ régression hédonique
	34	+log-linéaire à effets fixes}
	35	+\author{}
	36	+\date{Données extraites le 3 juillet 2026}
	37	+
	38	+\begin{document}
	39	+\maketitle
	40	+\thispagestyle{fancy}
	41	+
	42	+\begin{abstract}
	43	+\noindent Ce document décrit la méthodologie complète de construction d'un modèle
	44	+hédonique de prix pour les maisons unifamiliales et maisons en rangée du Québec.
	45	+Le pipeline combine deux jeux de données d'annonces immobilières, un nettoyage des
	46	+valeurs aberrantes fondé sur des règles, une ingénierie de variables à quatre
	47	+origines (règles déterministes, extraction de texte par grand modèle de langage,
	48	+analyse d'images par modèle de vision, effets fixes géographiques) et une
	49	+estimation par moindres carrés ordinaires du logarithme du prix avec erreurs-types
	50	+robustes à l'hétéroscédasticité.
	51	+\end{abstract}
	52	+
	53	+\tableofcontents
	54	+\bigskip
	55	+
	56	+%======================================================================
	57	+\section{Origine des données}
	58	+
	59	+\begin{longtable}{@{}p{2.4cm}p{5.2cm}rp{4.6cm}@{}}
	60	+\toprule
	61	+\textbf{Source} & \textbf{Fichier} & \textbf{Nb} & \textbf{Nature}\\
	62	+\midrule
	63	+\endhead
	64	+Realtor.ca & \code{dataset\_realtor-\dots.json} (161 Mo) & 27\,302 &
	65	+Annonces MLS via crawler Apify\\
	66	+DuProprio & \code{dataset\_duproprio-\dots.json} (87 Mo) & 7\,693 &
	67	+Annonces de vente sans intermédiaire (Apify)\\
	68	+\bottomrule
	69	+\end{longtable}
	70	+
	71	+\begin{itemize}[leftmargin=1.2em]
	72	+ \item \textbf{Collecte} : scrapers Apify, extraction le \textbf{2026-07-03}. Chaque
	73	+ enregistrement est un objet JSON structuré.
	74	+ \item \textbf{Champs communs} : \code{source}, \code{url}, \code{listingId},
	75	+ \code{title}, \code{price}/\code{priceNumber}, \code{currency},
	76	+ \code{address\{...\}}, \code{property\{...\}}, \code{features[]},
	77	+ \code{description}, \code{images[]}, \code{scrapedAt}.
	78	+ \item \textbf{Spécificités} :
	79	+ \begin{itemize}[leftmargin=1.2em]
	80	+ \item Realtor : \code{buildingType}, \code{storeys}, \code{yearBuilt},
	81	+ \code{parkingSpaces}, \textbf{\code{rooms[]}} (dimensions par pièce) ;
	82	+ pas de coordonnées GPS ; images sans légende utile.
	83	+ \item DuProprio : \code{type}, \code{levels},
	84	+ \textbf{\code{coordinates}} (7\,691 présentes) ; images \textbf{avec légendes}
	85	+ (\emph{Kitchen}, etc.) ; pas de \code{rooms}.
	86	+ \end{itemize}
	87	+ \item \textbf{Instantané transversal} (une seule date) : pas de dimension
	88	+ temporelle, donc pas d'effet fixe temporel requis.
	89	+\end{itemize}
	90	+
	91	+\noindent\textbf{Identifiant unique} : \code{uid = source + ":" + listingId}
	92	+(évite toute collision entre les deux espaces d'identifiants).
	93	+
	94	+%======================================================================
	95	+\section{Construction du dataset «~maisons~»}
	96	+
	97	+\begin{enumerate}[leftmargin=1.4em]
	98	+ \item \textbf{Fusion} des deux tableaux JSON (schéma harmonisé).
	99	+ \item \textbf{Définition de «~maison~»} (unifamiliales $+$ maisons en rangée) :
	100	+ \begin{itemize}[leftmargin=1.2em]
	101	+ \item Realtor : \code{buildingType} $\in \{$House, Row/Townhouse$\}$ ;
	102	+ \item DuProprio : \code{type} $\in \{$Home, Bungalow, Semi Detached,
	103	+ Intergenerational, Country Home, Cottage, Villa, Townhouse$\}$ ;
	104	+ \item \emph{Exclus} : condos, plex (duplex$\rightarrow$quintuplex),
	105	+ terrains/lots, fermes, commercial, maisons mobiles.
	106	+ \end{itemize}
	107	+ \item \textbf{Filtre prix présent} : \code{priceNumber} non nul et $>0$.
	108	+\end{enumerate}
	109	+
	110	+\noindent$\rightarrow$ \textbf{19\,273 maisons avec prix} (15\,470 Realtor
	111	+$+$ 3\,803 DuProprio).
	112	+
	113	+%======================================================================
	114	+\section{Nettoyage des valeurs aberrantes}
	115	+
	116	+Les prix extraits contiennent des erreurs d'analyse syntaxique (minimum
	117	+observé 899~\$, maximum 368~G\$). Règles appliquées :
	118	+
	119	+\begin{longtable}{@{}p{3.0cm}p{6.2cm}p{5.2cm}@{}}
	120	+\toprule
	121	+\textbf{Variable} & \textbf{Règle} & \textbf{Justification}\\
	122	+\midrule
	123	+\endhead
	124	+Prix & garder \textbf{50\,k\$ -- 15\,M\$} & plage plausible au QC ; retire loyers
	125	+mal analysés et concaténations\\
	126	+Surface habitable & 300 -- 20\,000 ft$^2$ (sinon $\rightarrow$ manquant) &
	127	+valeurs physiquement possibles\\
	128	+Terrain & 200 -- 5\,000\,000 ft$^2$ (sinon $\rightarrow$ manquant) & idem\\
	129	+Chambres / sdb & 0--15 / 0--12 (sinon $\rightarrow$ manquant) & plafonds réalistes\\
	130	+Prix/pied$^2$ & garder \textbf{P0{,}5 -- P99{,}5} ($\approx 73$--$1276~\$$/ft$^2$) &
	131	+retire les erreurs résiduelles de prix ou surface\\
	132	+\bottomrule
	133	+\end{longtable}
	134	+
	135	+\noindent$\rightarrow$ \textbf{19\,115 maisons} conservées (158 retirées).
	136	+
	137	+\paragraph{Principe.} On \emph{borne} plutôt que de supprimer quand une seule
	138	+variable est aberrante : la variable est mise à \emph{manquant} puis imputée, pour
	139	+ne pas perdre l'observation entière.
	140	+
	141	+%======================================================================
	142	+\section{Construction des variables}
	143	+
	144	+\subsection{Variable cible}
	145	+\[
	146	+\code{ln\_price} = \ln(\code{priceNumber})
	147	+\]
	148	+La transformation logarithmique rend la distribution symétrique, linéarise les
	149	+effets multiplicatifs et donne des coefficients interprétables en pourcentage
	150	+d'impact sur le prix.
	151	+
	152	+\subsection{Variables structurelles (règles déterministes)}
	153	+\begin{itemize}[leftmargin=1.2em]
	154	+ \item \textbf{Surface \& terrain} : analyse d'unités hétérogènes, tout converti en
	155	+ \textbf{pieds carrés} : \code{"1535 sqft"}, \code{"1,333.2 ft²"},
	156	+ \code{"11 561.52 ft²"} (séparateurs de milliers variés),
	157	+ \code{"961.6 m2"} ($\times 10{,}7639$).
	158	+ \item \textbf{Reconstruction de surface via les pièces (Realtor)} : dimensions
	159	+ \code{"23 ft ,4 in x 13 ft ,8 in"} $\rightarrow$ pieds décimaux $\rightarrow$ aire,
	160	+ sommées sur toutes les pièces. Calibration sur le recouvrement (ratio médian
	161	+ surface réelle / somme des pièces $\approx 0{,}98$) pour combler
	162	+ \code{squareFeet} manquant.
	163	+ \item \textbf{Autres} : \code{beds}, \code{baths}, \code{parking},
	164	+ \code{storeys}/\code{levels}, \code{age} $= 2026 - \code{yearBuilt}$ (avec
	165	+ \code{age}$^2$ pour la courbure), drapeaux \code{has\_*} marquant les valeurs
	166	+ imputées.
	167	+\end{itemize}
	168	+
	169	+\subsection{Aménités (règles par mots-clés bilingues)}
	170	+17 variables indicatrices détectées dans \code{features[]} $+$ \code{description}
	171	+$+$ \code{title} (FR/EN) : piscine, bord de l'eau, garage, foyer, climatisation/
	172	+thermopompe, sous-sol fini, cul-de-sac, golf, ski, proximité parc/école/autoroute/
	173	+transport, absence de voisin arrière.
	174	+
	175	+\subsection{Détail des pièces (Realtor)}
	176	+Nombre de pièces, nombre de niveaux, chambre principale, sous-sol, bureau, salle
	177	+familiale, atelier, cellier à vin, solarium, mezzanine, bibliothèque, buanderie.
	178	+
	179	+\subsection{Signaux de mise en marché}
	180	+Nombre de photos (\code{n\_images}, écrêté à 60), présence de visite virtuelle —
	181	+indicateurs indirects de gamme et de qualité de l'annonce.
	182	+
	183	+\subsection{Variables géographiques (effets fixes)}
	184	+\begin{itemize}[leftmargin=1.2em]
	185	+ \item \textbf{Ville} nettoyée (Realtor : retrait du quartier entre parenthèses ;
	186	+ minuscules ; normalisation). Villes avec $\ge 20$ annonces $\rightarrow$
	187	+ indicatrice propre, les rares $\rightarrow$ \code{\_\_other\_\_}
	188	+ ($\approx 194$ villes en effets fixes).
	189	+ \item \textbf{Type de propriété} harmonisé, regroupé si $<30$ annonces.
	190	+ \item \textbf{Source} (Realtor/DuProprio) $\rightarrow$ indicatrice captant les
	191	+ différences systématiques entre plateformes.
	192	+\end{itemize}
	193	+
	194	+\subsection{Imputation}
	195	+Valeurs manquantes après nettoyage : imputées par \textbf{médiane par type de
	196	+propriété} (puis médiane globale), \textbf{avec un drapeau \code{has\_*}} pour que
	197	+le modèle distingue valeur observée et valeur imputée.
	198	+
	199	+%======================================================================
	200	+\section{Mécanismes d'IA pour l'ingénierie de variables}
	201	+
	202	+Deux couches d'IA extraient des variables impossibles à obtenir par règles~: l'une
	203	+depuis le texte, l'autre depuis les images.
	204	+
	205	+\subsection{Extraction de specs depuis la description --- GPT-4o-mini}
	206	+\begin{itemize}[leftmargin=1.2em]
	207	+ \item \textbf{But} : transformer la description libre en 20 attributs structurés
	208	+ (état~: à rénover$\rightarrow$neuf ; clé en main ; cuisine/sdb rénovée ;
	209	+ revêtements de sol ; comptoirs quartz/granit ; chauffage ; vue sur l'eau ;
	210	+ revenus locatifs ; grand terrain ; secteur tranquille ; mots-clés de vente).
	211	+ \item \textbf{Mécanisme} : appel \emph{Chat Completions} avec
	212	+ \code{response\_format = json\_schema} \textbf{strict} (schéma imposé,
	213	+ énumérations fermées) $\rightarrow$ sortie JSON valide garantie, sans analyse
	214	+ fragile. \code{temperature = 0} (déterminisme). Instruction système : extracteur
	215	+ factuel, uniquement d'après le texte ; absence $\rightarrow$ false/inconnu/[].
	216	+ \item \textbf{Ingénierie} : $\sim$600 tokens en entrée / $\sim$110 en sortie par
	217	+ annonce ; \textbf{parallèle} (24 fils), \textbf{résumable} (cache
	218	+ \code{specs.jsonl}, clé \code{uid}), \textbf{réessais} exponentiels sur 429/5xx.
	219	+ \item \textbf{Coût / temps} : $\sim$\$3 pour $\sim$19\,k annonces.
	220	+\end{itemize}
	221	+
	222	+\subsection{Attributs visuels de façade --- Claude Haiku 4.5 (vision)}
	223	+\begin{itemize}[leftmargin=1.2em]
	224	+ \item \textbf{But} : lire ce que le texte ne dit pas --- état visuel réel,
	225	+ attrait extérieur (\emph{curb appeal}), niveau de rénovation, matériaux,
	226	+ aménagement paysager, cadre (urbain/banlieue/rural/bord d'eau/boisé), indice de
	227	+ luxe, qualité de la photo.
	228	+ \item \textbf{Sélection d'image} : une photo/annonce $=$ la façade. Realtor
	229	+ $\rightarrow$ 1\up{re} photo ; DuProprio $\rightarrow$ photo dont la légende est
	230	+ l'adresse (pas une pièce), par exclusion de mots-clés de pièces.
	231	+ \item \textbf{Mécanisme} : \emph{Messages API}, image passée par \textbf{URL}
	232	+ (\code{source.type = url}, sans téléchargement),
	233	+ \code{output\_config.format = json\_schema} (12 attributs, énumérations fermées).
	234	+ Instruction système : évaluateur objectif ; scores fondés sur des critères
	235	+ concrets (surfaces, matériaux, aménagement), pas sur un goût personnel.
	236	+ \item \textbf{Ingénierie} : $\sim$2\,470 tokens en entrée / $\sim$100 en sortie
	237	+ par image ; \textbf{parallèle} (16 fils), \textbf{résumable} (cache
	238	+ \code{facade\_specs.jsonl}), \textbf{réessais} sur 429/5xx/529.
	239	+ \item \textbf{Coût / temps} : $\sim$\$57 pour $\sim$19\,k images.
	240	+ \item \textbf{Choix de Haiku} : modèle de vision le moins cher et rapide,
	241	+ suffisant pour de la classification d'attributs concrets.
	242	+\end{itemize}
	243	+
	244	+\subsection{Rôle central du JSON structuré strict}
	245	+Le schéma imposé (types $+$ \code{enum} $+$ champs requis $+$
	246	+\code{additionalProperties:false}) fait valider la sortie \textbf{au niveau de
	247	+l'API}~: le modèle renvoie exactement les champs attendus, dans le bon domaine de
	248	+valeurs. On obtient des variables directement tabulables et reproductibles, sans
	249	+post-traitement de texte libre.
	250	+
	251	+\subsection{Intégration au dataset}
	252	+Les sorties des modèles (clé \code{uid}) sont converties en variables numériques~:
	253	+booléens $\rightarrow$ 0/1, état $\rightarrow$ score ordinal 0--5, revêtements
	254	+$\rightarrow$ indicatrices (bois franc, quartz/granit), vue $\rightarrow$
	255	+indicatrice eau, etc., avec un drapeau \code{has\_gpt} / \code{is\_exterior}
	256	+marquant la disponibilité.
	257	+
	258	+%======================================================================
	259	+\section{Modèle économétrique}
	260	+
	261	+\subsection{Spécification --- hédonique log-linéaire à effets fixes}
	262	+\begin{equation}
	263	+\ln(P_i) = \alpha + \beta\,\ln(\text{surface}_i) + \gamma\,\ln(\text{terrain}_i)
	264	++ \sum_k \delta_k X_{ki}
	265	++ \sum_c \lambda_c\,\mathbb{1}[\text{ville}_i=c]
	266	++ \varepsilon_i
	267	+\end{equation}
	268	+
	269	+\begin{itemize}[leftmargin=1.2em]
	270	+ \item \textbf{Log--log} sur surface et terrain $\rightarrow$ coefficients $=$
	271	+ \textbf{élasticités} (\% de prix par \% de surface).
	272	+ \item \textbf{Log--niveau} sur le reste $\rightarrow$ un coefficient $b$ vaut
	273	+ $\approx (e^{b}-1)\times 100$~\% d'impact sur le prix.
	274	+ \item \textbf{Effets fixes ville} ($\lambda_c$) : absorbent le différentiel de
	275	+ localisation (premier déterminant du prix) ; chaque coefficient $=$ prime/décote
	276	+ de la ville vs la référence.
	277	+ \item \textbf{Effets fixes} type de propriété et source.
	278	+\end{itemize}
	279	+
	280	+\subsection{Double spécification (couverture vs richesse)}
	281	+Realtor ne rapporte \code{squareFeet} que pour $\sim$38~\% des annonces, d'où deux
	282	+modèles et des prédictions combinées~:
	283	+\begin{itemize}[leftmargin=1.2em]
	284	+ \item \textbf{Modèle A (riche)} : inclut $\ln(\text{surface})$, estimé sur les
	285	+ biens à surface \emph{mesurée} (réelle ou reconstruite via les pièces). Meilleur
	286	+ ajustement.
	287	+ \item \textbf{Modèle B (large)} : sans surface, estimé sur \emph{toutes} les
	288	+ maisons $\rightarrow$ price celles sans surface (beds/baths/terrain/pièces $+$
	289	+ specs).
	290	+ \item \textbf{Prédiction finale} : A si surface mesurée, sinon B.
	291	+\end{itemize}
	292	+
	293	+\subsection{Estimation et inférence}
	294	+\begin{itemize}[leftmargin=1.2em]
	295	+ \item \textbf{Moindres carrés ordinaires} sur \code{ln\_price}.
	296	+ \item \textbf{Erreurs-types robustes à l'hétéroscédasticité HC3} (les résidus
	297	+ immobiliers sont hétéroscédastiques ; HC3 est conservateur en échantillon
	298	+ moyen).
	299	+ \item \textbf{Correction du biais de rétrotransformation} :
	300	+ \[
	301	+ \hat{P} = e^{\widehat{\ln P}} \times e^{\hat\sigma^2/2}
	302	+ \]
	303	+ (correction normale de Duan) --- sinon $e^{\widehat{\ln P}}$ sous-estime
	304	+ systématiquement le niveau de prix.
	305	+\end{itemize}
	306	+
	307	+\subsection{Interprétation}
	308	+Coefficients rapportés en \textbf{\% d'impact prix} $+$ significativité (p-values
	309	+sur SE HC3). Effets fixes ville présentés comme primes de localisation.
	310	+
	311	+%======================================================================
	312	+\section{Statistiques et diagnostics}
	313	+
	314	+\begin{longtable}{@{}p{5.2cm}p{9.4cm}@{}}
	315	+\toprule
	316	+\textbf{Indicateur} & \textbf{Signification}\\
	317	+\midrule
	318	+\endhead
	319	+$R^2$ / $R^2$ ajusté & part de variance de \code{ln\_price} expliquée\\
	320	+Écart-type des résidus (log) & $\rightarrow$ erreur typique en \% $=(e^{\sigma}-1)\times100$\\
	321	+Erreur absolue médiane & $\|\,$prix réel $/$ prix prédit $-1\,\|$ médian\\
	322	+\% à $\pm15$~\% / $\pm25$~\% & part des prédictions dans la fourchette\\
	323	+\code{resid\_pct} par annonce & surprix/sous-prix vs modèle
	324	+$\rightarrow$ repère aubaines et surévaluations\\
	325	+\bottomrule
	326	+\end{longtable}
	327	+
	328	+\noindent Progression observée : v2 (structurel) $R^2\approx0{,}70$
	329	+$\rightarrow$ v3 ($+$ pièces/photos) $R^2\approx0{,}72$
	330	+$\rightarrow$ v4 ($+$ specs GPT) $\rightarrow$ v5 ($+$ vision) attendu.
	331	+
	332	+%======================================================================
	333	+\section{Limites et biais}
	334	+
	335	+\begin{itemize}[leftmargin=1.2em]
	336	+ \item \textbf{Endogénéité / non-causalité} : de belles photos ou des specs riches
	337	+ corrèlent avec des courtiers haut de gamme et des biens chers $\rightarrow$
	338	+ coefficients \emph{prédictifs}, pas causaux.
	339	+ \item \textbf{Biais de sélection} : uniquement des biens \emph{en vente} (prix
	340	+ demandé, pas prix de transaction) ; les biens sans surface/pièces sont non
	341	+ aléatoires (annonces moins complètes).
	342	+ \item \textbf{Imputation} : les médianes par type réduisent la variance des
	343	+ variables imputées (atténuation possible des coefficients), d'où les drapeaux
	344	+ \code{has\_*}.
	345	+ \item \textbf{Extraction LLM} : erreurs/hallucinations possibles ;
	346	+ \code{temperature=0} $+$ schéma strict $+$ énumérations fermées les limitent,
	347	+ sans validation humaine systématique. La vision juge sur \emph{une seule} photo.
	348	+ \item \textbf{Instantané unique} : pas de dynamique de marché ; primes de
	349	+ localisation figées au 2026-07-03.
	350	+ \item \textbf{Couverture géographique} : villes $<20$ annonces regroupées en
	351	+ \code{\_\_other\_\_}.
	352	+\end{itemize}
	353	+
	354	+%======================================================================
	355	+\section{Fichiers produits}
	356	+
	357	+\begin{longtable}{@{}p{5.4cm}p{9.2cm}@{}}
	358	+\toprule
	359	+\textbf{Fichier} & \textbf{Contenu}\\
	360	+\midrule
	361	+\endhead
	362	+\code{hedonic.py}, \code{hedonic\_v4.py} & pipelines reproductibles (nettoyage $+$
	363	+modèles)\\
	364	+\code{extract\_specs.py}, \code{facade\_vision.py} & extracteurs LLM (texte /
	365	+vision), parallèles et résumables\\
	366	+\code{specs.jsonl}, \code{facade\_specs.jsonl} & sorties LLM brutes (clé \code{uid})\\
	367	+\code{maisons\_hedonic*.csv} & dataset final $+$ prix prédit $+$ \code{resid\_pct}\\
	368	+\code{hedonic\_*\_summary.txt} & sorties statistiques complètes\\
	369	+\bottomrule
	370	+\end{longtable}
	371	+
	372	+\end{document}

added paper/.latexmkrc +5 −0

		`@@ -0,0 +1,5 @@`
	1	+$pdf_mode = 1;
	2	+$pdflatex = 'pdflatex -interaction=nonstopmode -halt-on-error -synctex=1 %O %S';
	3	+$bibtex_use = 2;
	4	+$clean_ext = 'synctex.gz run.xml bbl bcf fdb_latexmk fls log aux out toc lof lot blg';
	5	+@default_files = ('main.tex');

added paper/Makefile +11 −0

		`@@ -0,0 +1,11 @@`
	1	+.PHONY: all clean distclean
	2	+
	3	+all:
	4	+ latexmk
	5	+
	6	+clean:
	7	+ latexmk -c
	8	+
	9	+distclean:
	10	+ latexmk -C
	11	+ rm -f .synctex.gz .run.xml .bbl .bcf

added paper/appendix/appendix.tex +92 −0

		`@@ -0,0 +1,92 @@`
	1	+% ============================================================================
	2	+% Appendix
	3	+% ============================================================================
	4	+\newpage
	5	+\begin{appendices}
	6	+
	7	+\renewcommand{\thetable}{A\arabic{table}}
	8	+\renewcommand{\thefigure}{A\arabic{figure}}
	9	+\setcounter{table}{0}
	10	+\setcounter{figure}{0}
	11	+
	12	+\section{Machine-Generated Feature Definitions}
	13	+\label{app:features}
	14	+
	15	+Table~\ref{tab:app_features} lists the attributes extracted by the language and
	16	+vision models under their respective strict JSON schemas. All fields use closed
	17	+enumerations or booleans; absent information is coded as \texttt{false},
	18	+\texttt{unknown}, or an empty list.
	19	+
	20	+\begin{table}[htbp]
	21	+ \centering
	22	+ \begin{threeparttable}
	23	+ \caption{Attributes Extracted by the AI Models}
	24	+ \label{tab:app_features}
	25	+ \small
	26	+ \begin{tabular}{@{}p{2.9cm}p{4.4cm}p{6.3cm}@{}}
	27	+ \toprule
	28	+ Layer & Field & Domain \\
	29	+ \midrule
	30	+ \multicolumn{3}{@{}l}{\textit{Panel A: Text (LLM, \texttt{gpt-4o-mini})}} \\[3pt]
	31	+ Condition & \texttt{condition} & to renovate / original / good / renovated / like-new / new \\
	32	+ Turnkey & \texttt{turnkey} & boolean \\
	33	+ Renovated kitchen & \texttt{renovated\_kitchen} & boolean \\
	34	+ Renovated bathroom & \texttt{renovated\_bathroom} & boolean \\
	35	+ Flooring & \texttt{flooring} & hardwood / ceramic / laminate / carpet / vinyl / concrete \\
	36	+ Countertop & \texttt{countertop} & granite / quartz / marble / laminate / wood \\
	37	+ Heating & \texttt{heating} & electric / heat pump / wood / gas / oil / radiant / geothermal \\
	38	+ Water view & \texttt{view} & water / mountain / city / none \\
	39	+ Income property & \texttt{income\_property} & boolean \\
	40	+ Large lot; quiet area & \texttt{large\_lot}, \texttt{quiet\_area} & boolean \\[6pt]
	41	+ \multicolumn{3}{@{}l}{\textit{Panel B: Vision (\texttt{claude-haiku-4-5}, façade)}} \\[3pt]
	42	+ Building type (visual) & \texttt{building\_type\_visual} & detached / semi / row / condo / cottage / land \\
	43	+ Condition score & \texttt{condition\_score} & integer 1--10 \\
	44	+ Curb appeal & \texttt{curb\_appeal} & integer 1--10 \\
	45	+ Renovation level & \texttt{renovation\_level} & to renovate / dated / standard / renovated / high-end \\
	46	+ Exterior materials & \texttt{exterior\_materials} & brick / stone / vinyl / wood / aluminum / stucco / concrete \\
	47	+ Landscaping & \texttt{landscaping} & none / minimal / maintained / manicured \\
	48	+ Setting & \texttt{setting} & dense urban / suburban / rural / waterfront / wooded \\
	49	+ Luxury indicator & \texttt{curb\_luxury} & boolean \\
	50	+ Photo quality & \texttt{photo\_quality} & integer 1--10 \\
	51	+ \bottomrule
	52	+ \end{tabular}
	53	+ \begin{tablenotes}[flushleft]
	54	+ \footnotesize
	55	+ \item \textit{Notes:} Both extractions use schema-constrained decoding with
	56	+ closed enumerations, required fields, and \texttt{additionalProperties:false};
	57	+ decoding temperature is zero. The text layer reads the free-text description;
	58	+ the vision layer analyzes one façade photograph per listing. Fields are mapped
	59	+ to numeric regressors (booleans to $0/1$, ordinal scales to integers, categories
	60	+ to indicators), each with an availability flag.
	61	+ \end{tablenotes}
	62	+ \end{threeparttable}
	63	+\end{table}
	64	+
	65	+\section{Sample Construction Flow}
	66	+\label{app:flow}
	67	+
	68	+\begin{table}[htbp]
	69	+ \centering
	70	+ \begin{threeparttable}
	71	+ \caption{From Raw Records to Estimation Sample}
	72	+ \label{tab:app_flow}
	73	+ \begin{tabular}{@{}lr@{}}
	74	+ \toprule
	75	+ Step & Records \\
	76	+ \midrule
	77	+ Realtor.ca raw & 27{,}302 \\
	78	+ DuProprio raw & 7{,}693 \\
	79	+ Houses (single-family + townhouse) with price & 19{,}273 \\
	80	+ After outlier cleaning & 19{,}123 \\
	81	+ \quad of which with measured living area (Model~A) & 9{,}446 \\
	82	+ \bottomrule
	83	+ \end{tabular}
	84	+ \begin{tablenotes}[flushleft]
	85	+ \footnotesize
	86	+ \item \textit{Notes:} See Section~\ref{sec:data}. ``Measured living area'' means
	87	+ reported area or area reconstructed from room-level dimensions.
	88	+ \end{tablenotes}
	89	+ \end{threeparttable}
	90	+\end{table}
	91	+
	92	+\end{appendices}

added paper/figures/fig_ablation.pdf +0 −0

Binary file not shown.

added paper/figures/fig_calibration.pdf +0 −0

Binary file not shown.

added paper/figures/fig_importance.pdf +0 −0

Binary file not shown.

added paper/figures/fig_premia.pdf +0 −0

Binary file not shown.

added paper/figures/fig_validation.pdf +0 −0

Binary file not shown.

added paper/figures/nn_diagram.tex +80 −0

		`@@ -0,0 +1,80 @@`
	1	+% ============================================================================
	2	+% TikZ: Robust tabular neural network architecture
	3	+% ============================================================================
	4	+\begin{tikzpicture}[
	5	+ font=\small,
	6	+ >=Latex,
	7	+ node distance=6mm and 6mm,
	8	+ box/.style={rounded corners=2pt, draw=black!55, align=center,
	9	+ minimum height=8mm, inner sep=3pt, text width=24mm},
	10	+ cat/.style={box, fill=NavyBlue!8},
	11	+ num/.style={box, fill=OliveGreen!10, text width=30mm},
	12	+ emb/.style={box, fill=BurntOrange!14, text width=22mm},
	13	+ op/.style={box, fill=black!5},
	14	+ spine/.style={box, fill=black!6, text width=52mm},
	15	+ res/.style={box, fill=Plum!10, text width=52mm, minimum height=11mm},
	16	+ outb/.style={box, fill=NavyBlue!16, text width=34mm, font=\bfseries},
	17	+ detail/.style={rounded corners=2pt, draw=black!40, dashed, fill=Plum!5,
	18	+ align=left, text width=40mm, inner sep=4pt, font=\footnotesize},
	19	+ ar/.style={->, thick, black!65}
	20	+]
	21	+
	22	+% ---- inputs ----
	23	+\node[cat] (city) {City index\\{\footnotesize (195 cats)}};
	24	+\node[cat, right=of city] (type) {Type index};
	25	+\node[cat, right=of type] (src) {Source index};
	26	+\node[num, right=10mm of src] (numin) {Numeric features\\{\footnotesize $\approx$45: area, lot, beds,\\baths, age, amenities,\\rooms, GPT specs}};
	27	+
	28	+% ---- embeddings / bn ----
	29	+\node[emb, below=of city] (ecity) {Embedding\\{\footnotesize dim 29}};
	30	+\node[emb, below=of type] (etype) {Embedding\\{\footnotesize dim 4}};
	31	+\node[emb, below=of src] (esrc) {Embedding\\{\footnotesize dim 2}};
	32	+\node[op, below=of numin, text width=30mm] (bn) {BatchNorm\\{\footnotesize + missingness flags}};
	33	+
	34	+\draw[ar] (city)--(ecity); \draw[ar] (type)--(etype);
	35	+\draw[ar] (src)--(esrc); \draw[ar] (numin)--(bn);
	36	+
	37	+% ---- concat ----
	38	+\node[spine, below=12mm of etype.south, xshift=8mm] (concat)
	39	+ {Concatenate \;$\rightarrow$\; $\mathbb{R}^{\,\text{emb}+\text{num}}$};
	40	+\draw[ar] (ecity) \|- (concat.west);
	41	+\draw[ar] (etype) -- (concat);
	42	+\draw[ar] (esrc) -- (concat);
	43	+\draw[ar] (bn) -\| (concat.east);
	44	+
	45	+% ---- projection ----
	46	+\node[spine, below=of concat] (proj) {Linear $\rightarrow$ 256\;\; (trunk)};
	47	+\draw[ar] (concat)--(proj);
	48	+
	49	+% ---- residual blocks ----
	50	+\node[res, below=of proj] (resb)
	51	+ {Residual block \;$\times\,3$\\{\footnotesize BN $\to$ GELU $\to$ Lin$_{512}$ $\to$ Dropout $\to$ Lin$_{256}$ \;$\oplus$\; skip}};
	52	+\draw[ar] (proj)--(resb);
	53	+
	54	+% ---- head ----
	55	+\node[spine, below=of resb] (head)
	56	+ {Head:\; BatchNorm $\to$ GELU $\to$ Linear$_{1}$};
	57	+\draw[ar] (resb)--(head);
	58	+
	59	+% ---- output ----
	60	+\node[outb, below=of head] (out) {$\widehat{\ln(\text{Price})}$};
	61	+\draw[ar] (head)--(out);
	62	+
	63	+% ---- residual detail callout ----
	64	+\node[detail, right=10mm of resb] (dtl)
	65	+ {\textbf{Res block:} $x \mapsto x + f(x)$\\
	66	+ $f=$ Lin$_{256}\!\circ$Drop$\circ$Lin$_{512}\!\circ$GELU$\circ$BN};
	67	+\draw[->, dashed, black!45] (resb.east)--(dtl.west);
	68	+
	69	+% ---- training / ensemble annotation ----
	70	+\node[detail, right=10mm of head, text width=42mm, fill=OliveGreen!6, draw=black!40] (train)
	71	+ {\textbf{Training (robust):}\\
	72	+ Huber loss $\cdot$ AdamW $\cdot$ cosine LR\\
	73	+ grad-clip 5 $\cdot$ dropout $\cdot$ early stop};
	74	+
	75	+% ---- ensemble + blend ----
	76	+\node[op, below=8mm of out, text width=64mm, fill=NavyBlue!7] (ens)
	77	+ {Ensemble of 3 seeds (average) \;$\Rightarrow$\; blend 50/50 with gradient boosting};
	78	+\draw[ar] (out)--(ens);
	79	+
	80	+\end{tikzpicture}

added paper/main.pdf +0 −0

Binary file not shown.

added paper/main.tex +150 −0

		`@@ -0,0 +1,150 @@`
	1	+% ============================================================================
	2	+% UQO Working Paper — Hedonic House Pricing with LLM & Vision Features
	3	+% ============================================================================
	4	+\documentclass[12pt,letterpaper]{article}
	5	+
	6	+% --- Encoding & Language ---
	7	+\usepackage[utf8]{inputenc}
	8	+\usepackage[T1]{fontenc}
	9	+\usepackage[english]{babel}
	10	+
	11	+% --- Page Layout ---
	12	+\usepackage[letterpaper,top=1in,bottom=1in,left=1in,right=1in]{geometry}
	13	+
	14	+% --- Typography ---
	15	+\usepackage{mathptmx}
	16	+\usepackage{microtype}
	17	+\usepackage{setspace}
	18	+\onehalfspacing
	19	+
	20	+% --- Math ---
	21	+\usepackage{amsmath,amssymb,amsthm}
	22	+\usepackage{bm}
	23	+
	24	+% --- Tables ---
	25	+\usepackage{booktabs}
	26	+\usepackage{tabularx}
	27	+\usepackage{multirow}
	28	+\usepackage{threeparttable}
	29	+
	30	+% --- Figures ---
	31	+\usepackage{graphicx}
	32	+\usepackage{subcaption}
	33	+\usepackage{float}
	34	+\usepackage{tikz}
	35	+\usetikzlibrary{arrows.meta,positioning,fit,backgrounds,calc,shapes.geometric}
	36	+
	37	+% --- Colors & Links ---
	38	+\usepackage[dvipsnames]{xcolor}
	39	+\usepackage[colorlinks=true,linkcolor=NavyBlue,citecolor=NavyBlue,urlcolor=NavyBlue,breaklinks=true,pdftitle={Hedonic House Pricing with Language-Model and Vision Features: Evidence from Quebec Real-Estate Listings},pdfauthor={Simon-Pierre Boucher}]{hyperref}
	40	+
	41	+% --- Bibliography ---
	42	+\usepackage[authoryear,round,semicolon]{natbib}
	43	+\bibliographystyle{aer}
	44	+
	45	+% --- Captions ---
	46	+\usepackage[font=small,labelfont=bf,labelsep=period,justification=justified,singlelinecheck=false]{caption}
	47	+
	48	+% --- Headers & Footers ---
	49	+\usepackage{fancyhdr}
	50	+\pagestyle{fancy}
	51	+\fancyhf{}
	52	+\renewcommand{\headrulewidth}{0pt}
	53	+\fancyfoot[C]{\thepage}
	54	+
	55	+% --- Section Formatting ---
	56	+\usepackage{titlesec}
	57	+\titleformat{\section}{\large\bfseries}{\thesection.}{0.5em}{}
	58	+\titleformat{\subsection}{\normalsize\bfseries}{\thesubsection.}{0.5em}{}
	59	+\titleformat{\subsubsection}{\normalsize\itshape}{\thesubsubsection.}{0.5em}{}
	60	+
	61	+% --- Appendix Support ---
	62	+\usepackage[toc,page]{appendix}
	63	+
	64	+% --- Misc ---
	65	+\usepackage{enumitem}
	66	+\setlist{nosep,leftmargin=*}
	67	+
	68	+% ============================================================================
	69	+% METADATA
	70	+% ============================================================================
	71	+\newcommand{\WPnumber}{14}
	72	+\newcommand{\WPtitle}{Hedonic House Pricing with Language-Model and Vision Features}
	73	+\newcommand{\WPsubtitle}{Evidence from Quebec Real-Estate Listings}
	74	+\newcommand{\WPdate}{July 2026}
	75	+\newcommand{\WPversion}{2.0}
	76	+\newcommand{\WPabstract}{%
	77	+ Hedonic house-price models are bounded by the characteristics an analyst can
	78	+ measure, yet much of a home's value is recorded only in the free-text description
	79	+ and photographs of its listing. We treat general-purpose AI models as a
	80	+ feature-engineering technology for hedonic analysis: a large language model
	81	+ converts each description into structured property specifications, and a
	82	+ vision--language model converts the façade photograph into objective visual
	83	+ attributes, both under strict JSON schemas that render the outputs reproducible
	84	+ econometric variables. Using 19{,}123 Quebec house listings, the two AI layers
	85	+ raise the $R^2$ of an already rich log--linear hedonic specification from 0.717 to
	86	+ 0.755, and the vision layer delivers the largest out-of-sample gain of any feature
	87	+ block for every estimator, from OLS to gradient boosting and neural networks (best
	88	+ ensemble: out-of-sample $R^2$ of 0.779, median absolute error 14.0\%). The
	89	+ extracted attributes carry economically sensible implicit prices---2.3\% per point
	90	+ of curb appeal, 3--4\% for brick or stone façades---that survive
	91	+ municipality-clustered inference.
	92	+}
	93	+\newcommand{\WPkeywords}{hedonic pricing, house prices, large language models, computer vision, feature engineering, Quebec}
	94	+\newcommand{\WPjel}{R31, C21, C55, R21}
	95	+
	96	+% --- Author ---
	97	+\newcommand{\WPauthor}{Simon-Pierre Boucher}
	98	+\newcommand{\WPaffiliation}{%
	99	+ Département des sciences administratives\\
	100	+ Université du Québec en Outaouais}
	101	+\newcommand{\WPemail}{simon-pierre.boucher@uqo.ca}
	102	+\newcommand{\WPaddress}{%
	103	+ Gatineau -- Pavillon Alexandre-Taché\\
	104	+ 283, boulevard Alexandre-Taché\\
	105	+ Gatineau, Québec, Canada J9A 1L8}
	106	+
	107	+% ============================================================================
	108	+\begin{document}
	109	+
	110	+\input{sections/titlepage}
	111	+
	112	+\input{sections/introduction}
	113	+\input{sections/literature}
	114	+\input{sections/data}
	115	+\input{sections/methodology}
	116	+\input{sections/results}
	117	+\input{sections/benchmark}
	118	+\input{sections/robustness}
	119	+\input{sections/conclusion}
	120	+
	121	+% ============================================================================
	122	+% Declarations (Elsevier requirements)
	123	+% ============================================================================
	124	+\section*{Declarations}
	125	+
	126	+\paragraph{Funding.} This research did not receive any specific grant from funding
	127	+agencies in the public, commercial, or not-for-profit sectors.
	128	+
	129	+\paragraph{Declaration of competing interest.} The author declares that he has no
	130	+known competing financial interests or personal relationships that could have
	131	+appeared to influence the work reported in this paper.
	132	+
	133	+\paragraph{Data availability.} The listing data were collected from publicly
	134	+accessible web platforms (Realtor.ca and DuProprio) on 3~July~2026. The full
	135	+data-construction code, the extraction schemas, and the derived analysis dataset
	136	+(excluding copyrighted listing text and images) are available from the author upon
	137	+reasonable request.
	138	+
	139	+\paragraph{Use of AI models.} Large language and vision--language models
	140	+(\texttt{gpt-4o-mini} and \texttt{claude-haiku-4-5}) are the object of study of this
	141	+paper and were used exclusively as documented in Section~\ref{sec:methodology} to
	142	+extract structured features from listing text and photographs. They were not used to
	143	+generate the analysis or the conclusions of the paper.
	144	+
	145	+\newpage
	146	+\bibliography{references}
	147	+
	148	+\input{appendix/appendix}
	149	+
	150	+\end{document}

added paper/references.bib +316 −0

		`@@ -0,0 +1,316 @@`
	1	+@article{rosen1974hedonic,
	2	+ title={Hedonic prices and implicit markets: product differentiation in pure competition},
	3	+ author={Rosen, Sherwin},
	4	+ journal={Journal of Political Economy},
	5	+ volume={82},
	6	+ number={1},
	7	+ pages={34--55},
	8	+ year={1974}
	9	+}
	10	+
	11	+@article{lancaster1966new,
	12	+ title={A new approach to consumer theory},
	13	+ author={Lancaster, Kelvin J.},
	14	+ journal={Journal of Political Economy},
	15	+ volume={74},
	16	+ number={2},
	17	+ pages={132--157},
	18	+ year={1966}
	19	+}
	20	+
	21	+@article{sirmans2005composition,
	22	+ title={The composition of hedonic pricing models},
	23	+ author={Sirmans, Stacy G. and Macpherson, David A. and Zietz, Emily N.},
	24	+ journal={Journal of Real Estate Literature},
	25	+ volume={13},
	26	+ number={1},
	27	+ pages={1--44},
	28	+ year={2005}
	29	+}
	30	+
	31	+@incollection{malpezzi2003hedonic,
	32	+ title={Hedonic pricing models: a selective and applied review},
	33	+ author={Malpezzi, Stephen},
	34	+ booktitle={Housing Economics and Public Policy},
	35	+ editor={O'Sullivan, Tony and Gibb, Kenneth},
	36	+ pages={67--89},
	37	+ publisher={Blackwell Science},
	38	+ year={2003}
	39	+}
	40	+
	41	+@article{duan1983smearing,
	42	+ title={Smearing estimate: a nonparametric retransformation method},
	43	+ author={Duan, Naihua},
	44	+ journal={Journal of the American Statistical Association},
	45	+ volume={78},
	46	+ number={383},
	47	+ pages={605--610},
	48	+ year={1983}
	49	+}
	50	+
	51	+@article{white1980heteroskedasticity,
	52	+ title={A heteroskedasticity-consistent covariance matrix estimator and a direct test for heteroskedasticity},
	53	+ author={White, Halbert},
	54	+ journal={Econometrica},
	55	+ volume={48},
	56	+ number={4},
	57	+ pages={817--838},
	58	+ year={1980}
	59	+}
	60	+
	61	+@article{mackinnon1985some,
	62	+ title={Some heteroskedasticity-consistent covariance matrix estimators with improved finite sample properties},
	63	+ author={MacKinnon, James G. and White, Halbert},
	64	+ journal={Journal of Econometrics},
	65	+ volume={29},
	66	+ number={3},
	67	+ pages={305--325},
	68	+ year={1985}
	69	+}
	70	+
	71	+@article{can1992spatial,
	72	+ title={Specification and estimation of hedonic housing price models},
	73	+ author={Can, Ayse},
	74	+ journal={Regional Science and Urban Economics},
	75	+ volume={22},
	76	+ number={3},
	77	+ pages={453--474},
	78	+ year={1992}
	79	+}
	80	+
	81	+@book{anselin1988spatial,
	82	+ title={Spatial Econometrics: Methods and Models},
	83	+ author={Anselin, Luc},
	84	+ publisher={Kluwer Academic Publishers},
	85	+ address={Dordrecht},
	86	+ year={1988}
	87	+}
	88	+
	89	+@article{gentzkow2019text,
	90	+ title={Text as data},
	91	+ author={Gentzkow, Matthew and Kelly, Bryan and Taddy, Matt},
	92	+ journal={Journal of Economic Literature},
	93	+ volume={57},
	94	+ number={3},
	95	+ pages={535--574},
	96	+ year={2019}
	97	+}
	98	+
	99	+@article{poursaeed2018vision,
	100	+ title={Vision-based real estate price estimation},
	101	+ author={Poursaeed, Omid and Matera, Tomas and Belongie, Serge},
	102	+ journal={Machine Vision and Applications},
	103	+ volume={29},
	104	+ number={4},
	105	+ pages={667--676},
	106	+ year={2018}
	107	+}
	108	+
	109	+@inproceedings{ahmed2016house,
	110	+ title={House price estimation from visual and textual features},
	111	+ author={Ahmed, Eman H. and Moustafa, Mohamed},
	112	+ booktitle={Proceedings of the 8th International Joint Conference on Computational Intelligence (NCTA)},
	113	+ pages={62--68},
	114	+ year={2016}
	115	+}
	116	+
	117	+@article{bin2004prediction,
	118	+ title={A prediction comparison of housing sales prices by parametric versus semi-parametric regressions},
	119	+ author={Bin, Okmyung},
	120	+ journal={Journal of Housing Economics},
	121	+ volume={13},
	122	+ number={1},
	123	+ pages={68--84},
	124	+ year={2004}
	125	+}
	126	+
	127	+@article{mullainathan2017machine,
	128	+ title={Machine learning: an applied econometric approach},
	129	+ author={Mullainathan, Sendhil and Spiess, Jann},
	130	+ journal={Journal of Economic Perspectives},
	131	+ volume={31},
	132	+ number={2},
	133	+ pages={87--106},
	134	+ year={2017}
	135	+}
	136	+
	137	+@article{halvorsen1980interpretation,
	138	+ title={The interpretation of dummy variables in semilogarithmic equations},
	139	+ author={Halvorsen, Robert and Palmquist, Raymond},
	140	+ journal={American Economic Review},
	141	+ volume={70},
	142	+ number={3},
	143	+ pages={474--475},
	144	+ year={1980}
	145	+}
	146	+
	147	+@article{breiman2001random,
	148	+ title={Random forests},
	149	+ author={Breiman, Leo},
	150	+ journal={Machine Learning},
	151	+ volume={45},
	152	+ number={1},
	153	+ pages={5--32},
	154	+ year={2001}
	155	+}
	156	+
	157	+@article{friedman2001greedy,
	158	+ title={Greedy function approximation: a gradient boosting machine},
	159	+ author={Friedman, Jerome H.},
	160	+ journal={Annals of Statistics},
	161	+ volume={29},
	162	+ number={5},
	163	+ pages={1189--1232},
	164	+ year={2001}
	165	+}
	166	+
	167	+@article{guo2016entity,
	168	+ title={Entity embeddings of categorical variables},
	169	+ author={Guo, Cheng and Berkhahn, Felix},
	170	+ journal={arXiv preprint arXiv:1604.06737},
	171	+ year={2016}
	172	+}
	173	+
	174	+@inproceedings{loshchilov2019decoupled,
	175	+ title={Decoupled weight decay regularization},
	176	+ author={Loshchilov, Ilya and Hutter, Frank},
	177	+ booktitle={International Conference on Learning Representations (ICLR)},
	178	+ year={2019}
	179	+}
	180	+
	181	+@inproceedings{gorishniy2021revisiting,
	182	+ title={Revisiting deep learning models for tabular data},
	183	+ author={Gorishniy, Yury and Rubachev, Ivan and Khrulkov, Valentin and Babenko, Artem},
	184	+ booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
	185	+ year={2021}
	186	+}
	187	+
	188	+@inproceedings{grinsztajn2022why,
	189	+ title={Why do tree-based models still outperform deep learning on tabular data?},
	190	+ author={Grinsztajn, L{\'e}o and Oyallon, Edouard and Varoquaux, Ga{\"e}l},
	191	+ booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
	192	+ year={2022}
	193	+}
	194	+
	195	+@inproceedings{he2016deep,
	196	+ title={Deep residual learning for image recognition},
	197	+ author={He, Kaiming and Zhang, Xiangyu and Ren, Shaoqing and Sun, Jian},
	198	+ booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
	199	+ pages={770--778},
	200	+ year={2016}
	201	+}
	202	+
	203	+@article{nowak2017textual,
	204	+ title={Textual analysis in real estate},
	205	+ author={Nowak, Adam and Smith, Patrick S.},
	206	+ journal={Journal of Applied Econometrics},
	207	+ volume={32},
	208	+ number={4},
	209	+ pages={896--918},
	210	+ year={2017}
	211	+}
	212	+
	213	+@article{shen2021information,
	214	+ title={Information value of property description: A machine learning approach},
	215	+ author={Shen, Lily and Ross, Stephen L.},
	216	+ journal={Journal of Urban Economics},
	217	+ volume={121},
	218	+ pages={103299},
	219	+ year={2021}
	220	+}
	221	+
	222	+@article{johnson2020curb,
	223	+ title={Valuing curb appeal},
	224	+ author={Johnson, Erik B. and Tidwell, Alan and Villupuram, Sriram V.},
	225	+ journal={Journal of Real Estate Finance and Economics},
	226	+ volume={60},
	227	+ number={1-2},
	228	+ pages={111--133},
	229	+ year={2020}
	230	+}
	231	+
	232	+@article{lindenthal2021architecture,
	233	+ title={Machine learning, architectural styles and property values},
	234	+ author={Lindenthal, Thies and Johnson, Erik B.},
	235	+ journal={Journal of Real Estate Finance and Economics},
	236	+ year={2021}
	237	+}
	238	+
	239	+@techreport{glaeser2018vision,
	240	+ title={Computer vision and real estate: Do looks matter and do incentives determine looks},
	241	+ author={Glaeser, Edward L. and Kincaid, Michael Scott and Naik, Nikhil},
	242	+ institution={National Bureau of Economic Research},
	243	+ type={Working Paper},
	244	+ number={25174},
	245	+ year={2018}
	246	+}
	247	+
	248	+@article{kok2017bigdata,
	249	+ title={Big data in real estate? {F}rom manual appraisal to automated valuation},
	250	+ author={Kok, Nils and Koponen, Eija-Leena and Mart{\'i}nez-Barbosa, Carmen Adriana},
	251	+ journal={Journal of Portfolio Management},
	252	+ volume={43},
	253	+ number={6},
	254	+ pages={202--211},
	255	+ year={2017}
	256	+}
	257	+
	258	+@article{hendel2009mls,
	259	+ title={The relative performance of real estate marketing platforms: {MLS} versus {FSBOMadison.com}},
	260	+ author={Hendel, Igal and Nevo, Aviv and Ortalo-Magn{\'e}, Fran{\c{c}}ois},
	261	+ journal={American Economic Review},
	262	+ volume={99},
	263	+ number={5},
	264	+ pages={1878--1898},
	265	+ year={2009}
	266	+}
	267	+
	268	+@article{horowitz1992list,
	269	+ title={The role of the list price in housing markets: Theory and an econometric model},
	270	+ author={Horowitz, Joel L.},
	271	+ journal={Journal of Applied Econometrics},
	272	+ volume={7},
	273	+ number={2},
	274	+ pages={115--129},
	275	+ year={1992}
	276	+}
	277	+
	278	+@article{knight2002listing,
	279	+ title={Listing price, time on market, and ultimate selling price: Causes and effects of listing price changes},
	280	+ author={Knight, John R.},
	281	+ journal={Real Estate Economics},
	282	+ volume={30},
	283	+ number={2},
	284	+ pages={213--237},
	285	+ year={2002}
	286	+}
	287	+
	288	+@incollection{han2015microstructure,
	289	+ title={The microstructure of housing markets: Search, bargaining, and brokerage},
	290	+ author={Han, Lu and Strange, William C.},
	291	+ booktitle={Handbook of Regional and Urban Economics},
	292	+ volume={5},
	293	+ pages={813--886},
	294	+ publisher={Elsevier},
	295	+ year={2015}
	296	+}
	297	+
	298	+@article{hill2013hedonic,
	299	+ title={Hedonic price indexes for residential housing: A survey, evaluation and taxonomy},
	300	+ author={Hill, Robert J.},
	301	+ journal={Journal of Economic Surveys},
	302	+ volume={27},
	303	+ number={5},
	304	+ pages={879--914},
	305	+ year={2013}
	306	+}
	307	+
	308	+@article{cameron2015cluster,
	309	+ title={A practitioner's guide to cluster-robust inference},
	310	+ author={Cameron, A. Colin and Miller, Douglas L.},
	311	+ journal={Journal of Human Resources},
	312	+ volume={50},
	313	+ number={2},
	314	+ pages={317--372},
	315	+ year={2015}
	316	+}

added paper/sections/benchmark.tex +260 −0

		`@@ -0,0 +1,260 @@`
	1	+% ============================================================================
	2	+\section{Predictive Benchmark: From OLS to Deep Learning}
	3	+\label{sec:benchmark}
	4	+% ============================================================================
	5	+
	6	+The hedonic OLS of Section~\ref{sec:results} is designed for interpretation: its
	7	+coefficients are implicit prices. It is not, however, the most accurate predictor of
	8	+price. To map the accuracy frontier---and to gauge how much curvature and interaction
	9	+the linear model leaves on the table---we benchmark a ladder of estimators of
	10	+increasing flexibility on the \emph{identical} feature set, under a common
	11	+5-fold cross-validation protocol. All rows of Table~\ref{tab:models} share the same
	12	+folds and random seed, so differences reflect the estimator, not the split.
	13	+
	14	+\subsection{The Estimator Ladder}
	15	+\label{subsec:ladder}
	16	+
	17	+\paragraph{(1) Ordinary least squares.} The linear hedonic of
	18	+Equation~\eqref{eq:hedonic}, with one-hot city, property-type, and source indicators
	19	+and within-property-type median imputation of missing continuous attributes. It is
	20	+the interpretable baseline and, mechanically, the least flexible: it imposes constant
	21	+marginal effects and additive separability.
	22	+
	23	+\paragraph{(2) Random forest.} A bagged ensemble of regression trees
	24	+\citep{mullainathan2017machine}. Trees capture non-linearities and low-order
	25	+interactions automatically and handle missing values natively, but averaging
	26	+deep decorrelated trees tends to over-smooth sharp price thresholds.
	27	+
	28	+\paragraph{(3) Gradient boosting.} A histogram-based gradient-boosted tree ensemble
	29	+(600 boosting iterations, learning rate 0.05, up to 63 leaves, $L_2$ regularization).
	30	+Boosting fits trees sequentially to residuals, which lets it model both smooth
	31	+curvature (age, area) and abrupt thresholds; like the forest it ingests missing
	32	+values directly. It is the strongest single model in our comparison.
	33	+
	34	+\paragraph{(4) Neural networks.} A plain multilayer perceptron and a robust
	35	+residual network with entity embeddings, described in
	36	+Section~\ref{subsec:nn}. Neural networks excel when high-cardinality categoricals
	37	+(here, 195 municipalities) can be represented as dense learned vectors rather than
	38	+sparse indicators.
	39	+
	40	+\paragraph{(5) Blend.} A simple 50/50 average of the robust neural network and the
	41	+gradient-boosting predictions. Because the two models make partly decorrelated
	42	+errors---the network smooths through embeddings while the boosted trees cut sharp
	43	+thresholds---their average cancels idiosyncratic noise, a standard stacking result.
	44	+
	45	+\subsection{Formal Definitions}
	46	+\label{subsec:formal}
	47	+
	48	+Let $y_i=\ln P_i$ and let $\mathbf{x}_i$ denote the feature vector of
	49	+Equation~\eqref{eq:blocks}. Each estimator approximates the conditional mean
	50	+$f(\mathbf{x})=\mathbb{E}[y\mid\mathbf{x}]$.
	51	+
	52	+\paragraph{Linear (OLS).} $f(\mathbf{x})=\mathbf{x}^{\!\top}\bm{\theta}$ with the
	53	+closed form of Section~\ref{subsec:econometrics}; categoricals enter as one-hot
	54	+indicators.
	55	+
	56	+\paragraph{Random forest \citep{breiman2001random}.} An average of $B$ regression
	57	+trees, each grown on a bootstrap resample with a random feature subset at every split:
	58	+%
	59	+\begin{equation}
	60	+ \hat f_{\text{RF}}(\mathbf{x})=\frac1B\sum_{b=1}^{B} T_b(\mathbf{x};\Theta_b),
	61	+ \label{eq:rf}
	62	+\end{equation}
	63	+%
	64	+where $\Theta_b$ denotes the randomness (resample and split subsets) of tree~$b$.
	65	+
	66	+\paragraph{Gradient boosting \citep{friedman2001greedy}.} A forward stagewise additive
	67	+expansion built by fitting each new tree to the negative gradient of the loss:
	68	+%
	69	+\begin{equation}
	70	+ \resizebox{.94\textwidth}{!}{$\displaystyle
	71	+ F_M(\mathbf{x}) = F_0 + \nu\sum_{m=1}^{M} f_m(\mathbf{x}),
	72	+ \quad
	73	+ f_m = \arg\min_{f}\sum_{i=1}^{n}
	74	+ \big(\!-g_{mi}-f(\mathbf{x}_i)\big)^2,
	75	+ \quad
	76	+ g_{mi}=\frac{\partial L\big(y_i,F_{m-1}(\mathbf{x}_i)\big)}{\partial F_{m-1}(\mathbf{x}_i)}
	77	+ $}
	78	+ \label{eq:gbm}
	79	+\end{equation}
	80	+%
	81	+with learning rate $\nu$ and $L_2$-regularized histogram trees $f_m$; missing values
	82	+are routed to a learned default direction at each split.
	83	+
	84	+\paragraph{Neural network (robust).} High-cardinality categoricals are mapped to
	85	+learned embeddings \citep{guo2016entity}, concatenated with the batch-normalized
	86	+numeric block and projected to a $d$-dimensional trunk:
	87	+%
	88	+\begin{equation}
	89	+ \mathbf{h}_0 = \mathbf{W}_0
	90	+ \Big[\, \textstyle\bigoplus_{c\in\{\text{city},\text{type},\text{src}\}}
	91	+ \mathbf{E}_c(\text{cat}_{ci})
	92	+ \;\Vert\; \mathrm{BN}(\mathbf{x}^{\text{num}}_i)\,\Big].
	93	+ \label{eq:nn_in}
	94	+\end{equation}
	95	+%
	96	+The trunk passes through $L=3$ pre-activation residual blocks
	97	+\citep{he2016deep,gorishniy2021revisiting},
	98	+%
	99	+\begin{equation}
	100	+ \mathbf{h}_{\ell}
	101	+ = \mathbf{h}_{\ell-1}
	102	+ + \mathrm{Drop}\!\Big(\mathbf{W}_2^{(\ell)}\,
	103	+ \mathrm{Drop}\!\big(\phi(\mathbf{W}_1^{(\ell)}\,
	104	+ \mathrm{BN}(\mathbf{h}_{\ell-1}))\big)\Big),
	105	+ \qquad \phi=\text{GELU},
	106	+ \label{eq:resblock}
	107	+\end{equation}
	108	+%
	109	+and a normalization head yields $\hat y_i = \mathbf{w}^{\!\top}\phi(\mathrm{BN}(\mathbf{h}_L))+b$.
	110	+Parameters minimize the Huber (smooth-$L_1$) loss, robust to price outliers,
	111	+%
	112	+\begin{equation}
	113	+ L_\delta(r)=
	114	+ \begin{cases}
	115	+ \tfrac12 r^2, & \|r\|\le \delta,\\[2pt]
	116	+ \delta\big(\|r\|-\tfrac12\delta\big), & \|r\|>\delta,
	117	+ \end{cases}
	118	+ \qquad r_i=\hat y_i-y_i,\;\delta=1,
	119	+ \label{eq:huber}
	120	+\end{equation}
	121	+%
	122	+optimized with AdamW (decoupled weight decay) \citep{loshchilov2019decoupled} under a
	123	+cosine-annealed learning rate $\eta_t=\tfrac12\eta_0\big(1+\cos(\pi t/T)\big)$,
	124	+gradient-norm clipping, and early stopping. Three seeds are averaged per fold.
	125	+
	126	+\paragraph{Blend.} A convex combination of the two strongest models in log space,
	127	+%
	128	+\begin{equation}
	129	+ \widehat{\ln P}^{\,\text{blend}}
	130	+ = w\,\widehat{\ln P}^{\,\text{GBM}}
	131	+ + (1-w)\,\widehat{\ln P}^{\,\text{NN}},
	132	+ \qquad w=0.5.
	133	+ \label{eq:blend}
	134	+\end{equation}
	135	+
	136	+\paragraph{Evaluation.} With $\hat y_{-k(i)}$ the prediction for $i$ from the model
	137	+trained on the folds excluding $i$'s fold, we report out-of-sample fit and level
	138	+accuracy,
	139	+%
	140	+\begin{equation}
	141	+ R^2_{\text{OOS}} = 1-\frac{\sum_i\big(y_i-\hat y_{-k(i)}\big)^2}
	142	+ {\sum_i\big(y_i-\bar y\big)^2},
	143	+ \qquad
	144	+ \text{MdAPE}=\operatorname*{median}_i\Big\|e^{\,\hat y_{-k(i)}-y_i}-1\Big\|.
	145	+ \label{eq:cvmetrics}
	146	+\end{equation}
	147	+
	148	+\subsection{Results}
	149	+\label{subsec:bench_results}
	150	+
	151	+\input{tables/models}
	152	+
	153	+Table~\ref{tab:models} reports the ladder in two panels. In Panel~A (structural,
	154	+amenity, and text features) out-of-sample $R^2$ rises monotonically with flexibility:
	155	+from 0.628 for OLS, to 0.677 for the random forest, 0.711 for the plain network, and a
	156	+near-tie at the top between the robust network (0.741) and gradient boosting (0.743);
	157	+their blend attains $R^2_{\text{OOS}}=0.760$. Panel~B adds the twelve façade vision
	158	+attributes. Vision lifts \emph{every} estimator---OLS gains 5.8 points
	159	+(0.628 $\to$ 0.686), the random forest 2.7 (0.677 $\to$ 0.704), and gradient boosting
	160	+2.7 (0.743 $\to$ 0.770), the single largest increment from any one feature layer at
	161	+the frontier---and the \textbf{text\,+\,vision blend attains
	162	+$R^2_{\text{OOS}}=0.779$ with a median absolute error of 14.0\%}, the best model in
	163	+the study.
	164	+
	165	+Three lessons follow. First, on the full feature set roughly nine points of
	166	+out-of-sample $R^2$ separate the interpretable linear model (0.686) from the
	167	+predictive frontier (0.779)---the price of interpretability, and a measure of the
	168	+curvature and interactions the hedonic form omits. Notably, the vision layer buys the
	169	+linear model more accuracy (+5.8 points) than switching from OLS to a random forest
	170	+does (+4.9), so better measurement rivals better estimation. Second, on this tabular
	171	+problem gradient boosting and a carefully regularized neural network are
	172	+statistically comparable, consistent with the broader evidence that tree ensembles
	173	+remain strong baselines on tabular data \citep{grinsztajn2022why}; the gains come
	174	+from \emph{combining} them. Third, the façade-vision layer is not redundant with the
	175	+text and structural fields: it moves the frontier more than any other block we add,
	176	+confirming that image-derived quality is genuinely orthogonal information.
	177	+
	178	+\begin{figure}[htbp]
	179	+ \centering
	180	+ \includegraphics[width=.72\textwidth]{figures/fig_ablation.pdf}
	181	+ \caption{Out-of-sample $R^2$ as feature layers are added (nested, left to right).
	182	+ Each layer improves both the linear and boosted models; the façade-vision layer
	183	+ delivers the largest single jump---about $+5.8$ points for OLS and $+2.7$ for
	184	+ gradient boosting.}
	185	+ \label{fig:ablation}
	186	+\end{figure}
	187	+
	188	+\subsection{Feature Ablation and Importance}
	189	+\label{subsec:ablation}
	190	+
	191	+Table~\ref{tab:ablation} and Figure~\ref{fig:ablation} decompose the accuracy gains
	192	+layer by layer. Reading down the table, room detail and listing signals add the first
	193	+large increment, the LLM text layer a smaller one, and the vision layer the single
	194	+largest jump for every estimator---raising OLS from 0.628 to 0.686 and gradient
	195	+boosting from 0.743 to 0.770 out of sample. The pattern is monotone: no layer degrades
	196	+fit, and the two AI layers together account for a substantial share of the distance
	197	+between the structural baseline and the frontier.
	198	+
	199	+\input{tables/ablation}
	200	+
	201	+Figure~\ref{fig:importance} aggregates the gradient-boosting permutation importances
	202	+into feature blocks. Structural attributes and location fixed effects dominate, as
	203	+expected, but the \textbf{vision block alone accounts for 13\% of total importance}---
	204	+more than the rule-based amenities and the LLM text block combined---reinforcing that
	205	+the façade photograph is a first-order, not marginal, source of price information.
	206	+Figure~\ref{fig:calibration} shows the blend is well calibrated across the price range,
	207	+with an approximately symmetric, tight error distribution centered near zero.
	208	+
	209	+\begin{figure}[htbp]
	210	+ \centering
	211	+ \includegraphics[width=.72\textwidth]{figures/fig_importance.pdf}
	212	+ \caption{Share of gradient-boosting permutation importance by feature block
	213	+ (text\,+\,vision model). The vision block accounts for 13\% of total importance,
	214	+ exceeding the rule-based amenity and LLM-text blocks.}
	215	+ \label{fig:importance}
	216	+\end{figure}
	217	+
	218	+\begin{figure}[htbp]
	219	+ \centering
	220	+ \includegraphics[width=\textwidth]{figures/fig_calibration.pdf}
	221	+ \caption{Out-of-sample calibration of the text\,+\,vision blend. Left: predicted
	222	+ versus actual price (log--log; dashed line is the 45$^\circ$ identity). Right:
	223	+ distribution of percentage prediction errors, with median absolute error of 14.0\%.}
	224	+ \label{fig:calibration}
	225	+\end{figure}
	226	+
	227	+\subsection{A Robust Tabular Neural Network}
	228	+\label{subsec:nn}
	229	+
	230	+Figure~\ref{fig:nn} details the robust network. High-cardinality categoricals enter
	231	+as learned \textbf{entity embeddings}: each municipality, property type, and source is
	232	+mapped to a dense vector (dimension $\approx 1.6\,C^{0.56}$ in cardinality $C$, so
	233	+about 29 dimensions for the 195 cities). The embeddings are concatenated with the
	234	+batch-normalized numeric block and projected to a 256-dimensional trunk, which passes
	235	+through three \textbf{residual blocks}
	236	+(BatchNorm $\to$ GELU $\to$ Linear$_{512}$ $\to$ Dropout $\to$ Linear$_{256}$, with a
	237	+skip connection), followed by a normalization head that outputs $\ln(\text{Price})$.
	238	+
	239	+Robustness comes from the training recipe rather than raw capacity:
	240	+a \textbf{Huber (smooth-$L_1$) loss} that damps the influence of price outliers;
	241	+\textbf{AdamW} with decoupled weight decay and \textbf{cosine-annealed} learning rate;
	242	+\textbf{gradient-norm clipping} at 5; \textbf{dropout} and batch normalization
	243	+throughout; \textbf{learned missingness indicators} appended to the numeric block so
	244	+the network can exploit the non-random pattern of missing fields; and
	245	+\textbf{early stopping} on a held-out 10\% of each training fold. Finally, three
	246	+networks with different random seeds are trained per fold and \textbf{averaged}, which
	247	+reduces prediction variance. The result (0.741) is competitive with gradient boosting
	248	+and, once blended with it, produces the best model in the study.
	249	+
	250	+\begin{figure}[htbp]
	251	+ \centering
	252	+ \input{figures/nn_diagram}
	253	+ \caption{Architecture of the robust tabular neural network. High-cardinality
	254	+ categoricals enter as learned embeddings; numeric features are batch-normalized and
	255	+ augmented with missingness indicators. The concatenated representation is projected
	256	+ to a 256-dimensional trunk, refined by three residual blocks, and mapped to
	257	+ $\ln(\text{Price})$. Three seeds are ensembled per fold; the final model blends this
	258	+ network 50/50 with gradient boosting.}
	259	+ \label{fig:nn}
	260	+\end{figure}

added paper/sections/conclusion.tex +41 −0

		`@@ -0,0 +1,41 @@`
	1	+% ============================================================================
	2	+\section{Conclusion}
	3	+\label{sec:conclusion}
	4	+% ============================================================================
	5	+
	6	+We have estimated a hedonic price model for Quebec houses that augments the standard
	7	+set of structural, locational, and amenity attributes with two machine-generated
	8	+feature layers: structured specifications extracted from free-text descriptions by a
	9	+large language model, and objective visual attributes extracted from façade
	10	+photographs by a vision--language model. Schema-constrained decoding renders these
	11	+outputs as ordinary, reproducible econometric variables, obtained at a marginal cost
	12	+of a fraction of a cent per listing. The central message is methodological:
	13	+general-purpose AI models, disciplined by strict output schemas, function as a
	14	+low-cost feature-engineering technology that measurably expands the characteristics
	15	+set available to hedonic analysis.
	16	+
	17	+The gains are economically meaningful. The two AI layers add nearly four points of
	18	+in-sample $R^2$ to an already rich log--linear specification (0.717 to 0.755 with
	19	+living area observed), and they raise the out-of-sample accuracy of every estimator
	20	+in our ladder, from OLS to gradient boosting and an embedding-based residual neural
	21	+network; the best blend attains an out-of-sample $R^2$ of 0.779 with a median
	22	+absolute prediction error of 14.0\%. The façade-vision layer alone contributes the
	23	+largest increment of any feature block (2.7 points for gradient boosting, 5.8 for
	24	+OLS), confirming that image-derived quality is genuinely orthogonal to text and
	25	+structured fields. Crucially for hedonic practice, the extracted attributes carry
	26	+interpretable implicit prices---curb appeal, renovation level, luxury cues, and
	27	+exterior materials all price sensibly and survive municipality-clustered
	28	+inference---so the technology augments the classical regression rather than
	29	+replacing it with a black box.
	30	+
	31	+Three extensions are natural. First, the same schema-constrained approach scales
	32	+from a single façade image to a small panel of interior photographs (kitchen,
	33	+bathroom, living space), which our design already accommodates. Second, the
	34	+DuProprio coordinates permit continuous spatial features---distance to the urban
	35	+core, to water, or to amenities---that would refine the geography beyond discrete
	36	+city effects. Third, and most important for external validity, linking listings to
	37	+registry transaction prices would establish how much of the AI-measured quality
	38	+premium survives bargaining. For appraisal, automated valuation, and property-tax
	39	+assessment---applications that consume hedonic models at scale---the practical
	40	+implication is immediate: the unstructured majority of a listing's information
	41	+content is no longer out of reach.

added paper/sections/data.tex +74 −0

		`@@ -0,0 +1,74 @@`
	1	+% ============================================================================
	2	+\section{Data}
	3	+\label{sec:data}
	4	+% ============================================================================
	5	+
	6	+\subsection{Data Sources}
	7	+\label{subsec:data_sources}
	8	+
	9	+We assemble a cross-section of residential listings from two complementary Quebec
	10	+platforms, both retrieved on 3~July~2026 via web-scraping actors. The first,
	11	+\textbf{Realtor.ca}, is the national Multiple Listing Service (MLS) portal and
	12	+provides broker-listed inventory (27{,}302 raw records). The second,
	13	+\textbf{DuProprio}, is the dominant for-sale-by-owner platform in Quebec (7{,}693
	14	+raw records). Pooling the two sources broadens coverage across price segments and
	15	+sale channels---brokered and FSBO listings differ systematically in seller
	16	+composition and marketing \citep{hendel2009mls}---and a source indicator absorbs any
	17	+systematic level difference between them. The price we observe is the \emph{asking}
	18	+price; Section~\ref{subsec:limits} discusses the interpretation of hedonic
	19	+coefficients estimated on list rather than transaction prices
	20	+\citep{horowitz1992list,knight2002listing}.
	21	+
	22	+Each record is a structured JSON object containing a price, an address, a
	23	+property-attribute block, a free-text description, a list of features, and an array
	24	+of image URLs. The two sources differ in their auxiliary fields: Realtor.ca records
	25	+include room-by-room dimensions but no geographic coordinates, whereas DuProprio
	26	+records include coordinates and captioned images but no room block. We construct a
	27	+unique key by concatenating the source and the listing identifier. Because the data
	28	+form a single-date cross-section, no time dimension is modeled.
	29	+
	30	+\subsection{Sample Construction}
	31	+\label{subsec:sample}
	32	+
	33	+We restrict attention to \emph{houses}, defined as single-family dwellings and
	34	+townhouses. On Realtor.ca this corresponds to building types \textit{House} and
	35	+\textit{Row/Townhouse}; on DuProprio it corresponds to the property types
	36	+\textit{Home}, \textit{Bungalow}, \textit{Semi Detached}, \textit{Intergenerational},
	37	+\textit{Country Home}, \textit{Cottage}, \textit{Villa}, and \textit{Townhouse}.
	38	+Condominiums, income properties (duplex through quintuplex), vacant land, farms,
	39	+commercial properties, and mobile homes are excluded. After requiring a non-missing,
	40	+positive price, the pooled sample contains 19{,}273 houses.
	41	+
	42	+\subsection{Outlier Cleaning}
	43	+\label{subsec:cleaning}
	44	+
	45	+Scraped prices contain parsing artifacts spanning several orders of magnitude (the
	46	+raw minimum and maximum are \$899 and \$368~billion). We apply conservative,
	47	+transparent bounds rather than model-based trimming: prices are retained in the
	48	+\$50{,}000--\$15{,}000{,}000 range; living area in 300--20{,}000~ft$^2$; lot area in
	49	+200--5{,}000{,}000~ft$^2$; bedrooms in 0--15 and bathrooms in 0--12. When a single
	50	+attribute violates its bound it is set to missing (and later imputed) rather than
	51	+deleting the observation, preserving sample size. Finally, we trim the
	52	+price-per-square-foot distribution at its 0.5th and 99.5th percentiles
	53	+(approximately \$73--\$1{,}276/ft$^2$) among listings with a measured area, removing
	54	+residual price- or area-parsing errors. The cleaned sample contains
	55	+\textbf{19{,}123 houses}, of which 80.3\% originate from Realtor.ca.
	56	+
	57	+\subsection{Summary Statistics}
	58	+\label{subsec:summary_stats}
	59	+
	60	+Table~\ref{tab:summary_stats} reports descriptive statistics for the key variables.
	61	+The median list price is \$599{,}900. The distribution of living area is right-skewed,
	62	+motivating the logarithmic transformation adopted in Section~\ref{sec:methodology}.
	63	+Living area is reported for only a minority of Realtor.ca listings; where it is
	64	+missing we reconstruct it from room-level dimensions (Section~\ref{subsec:structural}),
	65	+and where neither is available we rely on the area-free specification.
	66	+
	67	+\input{tables/summary_stats}
	68	+
	69	+Table~\ref{tab:ai_summary} previews the machine-generated attributes described in
	70	+Section~\ref{sec:methodology}: roughly a quarter of listings are described as recently
	71	+renovated, the mean visual condition and curb-appeal scores are 7.4 and 7.2 on a
	72	+10-point scale, and 43\% of façades display a visual luxury cue.
	73	+
	74	+\input{tables/ai_summary}

added paper/sections/introduction.tex +87 −0

		`@@ -0,0 +1,87 @@`
	1	+% ============================================================================
	2	+\section{Introduction}
	3	+\label{sec:introduction}
	4	+% ============================================================================
	5	+
	6	+The hedonic price framework decomposes the market value of a heterogeneous good
	7	+into the implicit prices of its characteristics \citep{lancaster1966new,
	8	+rosen1974hedonic}. In housing markets it remains the workhorse tool for appraisal,
	9	+index construction, property-tax assessment, and policy analysis. Yet the empirical
	10	+performance of hedonic models is bounded by the richness of the characteristics an
	11	+analyst can measure. A large share of what determines a home's value---its state of
	12	+repair, the quality of its finishes, its curb appeal, the character of its
	13	+street---is recorded not in structured fields but in the free-text description and
	14	+the photographs that accompany each listing. These modalities have historically been
	15	+discarded because they were costly to encode at scale, leaving hedonic estimates
	16	+exposed to omitted-quality bias and hedonic predictions short of their potential
	17	+accuracy.
	18	+
	19	+This paper treats general-purpose artificial-intelligence models as a
	20	+\emph{feature-engineering technology} for hedonic analysis, and quantifies the
	21	+explanatory power that technology adds. We assemble a cross-section of 19{,}123
	22	+single-family houses and townhouses listed for sale in Quebec, pooling the
	23	+brokerage-listed MLS inventory from Realtor.ca with the for-sale-by-owner inventory
	24	+from DuProprio. On top of standard structural, locational, and amenity attributes,
	25	+we build two machine-generated feature layers. First, a large language model (LLM)
	26	+reads each free-text description and returns a fixed set of property
	27	+specifications---state of repair, renovation of the kitchen and bathroom, flooring
	28	+and countertop materials, heating system, view, and income potential---under a
	29	+strict JSON schema that guarantees valid, reproducible output. Second, a
	30	+vision--language model inspects the façade photograph of each property and returns
	31	+objective visual attributes: overall condition, curb appeal, renovation level,
	32	+exterior materials, landscaping, setting, and a luxury indicator. Because both
	33	+extractions are schema-constrained, deterministic, and blind to the listing price,
	34	+their outputs behave as ordinary tabulated regressors rather than as noisy text.
	35	+
	36	+Our contribution is threefold. First, we show that AI-based feature engineering
	37	+delivers economically meaningful gains in explanatory power over an already rich
	38	+hedonic specification: the two machine-generated layers raise in-sample $R^2$ from
	39	+0.717 to 0.755 (nearly four points) in our preferred specification, and raise the
	40	+out-of-sample $R^2$ of every estimator we consider, with the façade-vision layer
	41	+alone contributing the largest single increment of any feature block (2.7 points for
	42	+gradient boosting, 5.8 points for OLS). Second, we show that the extracted
	43	+attributes carry interpretable implicit prices---a one-point increase in visual curb
	44	+appeal is associated with a 2.3\% higher asking price, a visible luxury cue with
	45	+5.2\%, brick or stone façades with premia of 3--4\% relative to vinyl siding, and a
	46	+renovated kitchen (read from the text) with 2.6\%---so the approach augments rather
	47	+than replaces the classical hedonic regression. Third, we document a complete,
	48	+reproducible pipeline from raw scraped listings to an analysis-ready dataset,
	49	+including rules-based outlier cleaning, unit harmonization that reconstructs living
	50	+area from room-level dimensions, and the schema-constrained extraction protocol,
	51	+with a total AI cost of about \$60 for the full sample---roughly a third of a cent
	52	+per listing.
	53	+
	54	+Our main findings can be summarized as follows. A log--linear specification with 195
	55	+municipality fixed effects, property-type fixed effects, and
	56	+municipality-clustered standard errors explains 75.5\% of the variation in log
	57	+asking prices when living area is observed, and 72.5\% on the full sample when it is
	58	+not. The estimated implicit prices are economically sensible: the living-area
	59	+elasticity is 0.36, each additional bathroom is associated with a 7.6\% premium, and
	60	+the age profile traces the familiar depreciation-then-heritage pattern. The
	61	+municipality effects span a wide range---from roughly $+231$\% in Westmount to about
	62	+$-30$\% in remote municipalities---and order the province's markets as expected.
	63	+Benchmarking a ladder of estimators on the identical feature set, an ensemble blend
	64	+of a robust embedding-based neural network and gradient boosting attains an
	65	+out-of-sample $R^2$ of 0.779 with a median absolute prediction error of 14.0\%. We
	66	+interpret all estimates as \emph{predictive} associations rather than causal
	67	+effects, since listing content is chosen by sellers and correlates with unobserved
	68	+determinants of value, and we discuss the asking-price nature of the dependent
	69	+variable at length in Section~\ref{sec:robustness}.
	70	+
	71	+Beyond the specific Quebec application, the exercise speaks to a general measurement
	72	+problem in housing economics. Text- and image-derived quality measures have until
	73	+now required bespoke supervised models trained on labeled data
	74	+\citep{nowak2017textual,shen2021information,johnson2020curb}. General-purpose
	75	+foundation models with schema-constrained decoding lower the fixed cost of that
	76	+measurement step to nearly zero, produce attributes that are human-readable rather
	77	+than latent, and can be audited field by field. Our results suggest that this
	78	+technology moves the effective characteristics frontier of hedonic analysis by
	79	+enough to matter for appraisal, automated valuation, and assessment applications.
	80	+
	81	+The remainder of the paper is organized as follows. Section~\ref{sec:literature}
	82	+reviews the related literature. Section~\ref{sec:data} describes the data sources
	83	+and cleaning. Section~\ref{sec:methodology} presents the feature-engineering
	84	+pipeline and the econometric specification. Section~\ref{sec:results} reports the
	85	+main results. Section~\ref{sec:benchmark} maps the predictive frontier across
	86	+estimators. Section~\ref{sec:robustness} discusses robustness and limitations.
	87	+Section~\ref{sec:conclusion} concludes.

added paper/sections/literature.tex +74 −0

		`@@ -0,0 +1,74 @@`
	1	+% ============================================================================
	2	+\section{Related Literature}
	3	+\label{sec:literature}
	4	+% ============================================================================
	5	+
	6	+This paper connects four strands of research.
	7	+
	8	+The first is the classical \textbf{hedonic pricing} literature. Building on
	9	+\citet{lancaster1966new}, \citet{rosen1974hedonic} formalized the market for
	10	+differentiated goods as a market for their underlying characteristics, providing the
	11	+theoretical foundation for interpreting regression coefficients as implicit prices.
	12	+The empirical housing literature that followed is vast; \citet{sirmans2005composition}
	13	+survey the characteristics most commonly included in hedonic house-price models and
	14	+document substantial heterogeneity in specification, \citet{malpezzi2003hedonic}
	15	+provides a critical review of functional-form and identification choices, and
	16	+\citet{hill2013hedonic} surveys the use of hedonic methods for house-price
	17	+measurement. A recurring theme in this literature is omitted-variable bias from
	18	+unmeasured quality: the state of repair, the standard of finishes, and the visual
	19	+quality of a dwelling are priced by the market but rarely recorded in the structured
	20	+fields available to the econometrician. Our contribution speaks directly to this
	21	+measurement problem. Our log--linear specification, with logged continuous attributes
	22	+and a dependent variable in logs, follows the mainstream of this literature.
	23	+
	24	+The second strand concerns \textbf{spatial structure, functional form, and
	25	+prediction}. \citet{can1992spatial} and \citet{anselin1988spatial} emphasize that
	26	+unobserved locational amenities induce spatial dependence that, if ignored, biases
	27	+coefficient estimates; we absorb the first-order component of this problem through a
	28	+rich set of municipality fixed effects and cluster inference at the same level.
	29	+On the prediction side, \citet{bin2004prediction} compares parametric and
	30	+semi-parametric hedonic predictors, \citet{mullainathan2017machine} frame
	31	+machine-learning methods as estimators of the conditional expectation
	32	+$\mathbb{E}[P\mid\mathbf{z}]$ well suited to prediction tasks, and
	33	+\citet{kok2017bigdata} document the rise of automated valuation models in industry.
	34	+This literature generally finds that flexible learners beat linear hedonics on
	35	+accuracy at some cost in interpretability; our estimator ladder
	36	+(Section~\ref{sec:benchmark}) reproduces that pattern on a common feature set, and
	37	+our focus is on how much of the gap is closed not by a better \emph{estimator} but by
	38	+better \emph{measurement}.
	39	+
	40	+The third strand is the use of \textbf{unstructured listing content}---text and
	41	+images---in real-estate valuation. \citet{gentzkow2019text} survey text-as-data
	42	+methods in economics. In real estate specifically, \citet{nowak2017textual} show that
	43	+words in MLS remarks predict transaction prices and reduce hedonic pricing errors,
	44	+and \citet{shen2021information} extract quality signals from property descriptions
	45	+with supervised machine learning, finding that the text carries information about
	46	+unobserved quality that is capitalized into prices. On the image side,
	47	+\citet{ahmed2016house} combine visual and textual features for price estimation;
	48	+\citet{poursaeed2018vision} show that learned measures of interior and exterior
	49	+quality carry independent predictive power; \citet{glaeser2018vision} use street-level
	50	+imagery of Boston houses to measure exterior quality and study its interaction with
	51	+incentives; \citet{johnson2020curb} construct a curb-appeal index from Google Street
	52	+View images and estimate its capitalization into sale prices; and
	53	+\citet{lindenthal2021architecture} classify architectural styles from photographs and
	54	+document style premia. Relative to this work, our approach differs in the
	55	+\emph{technology of measurement}: rather than training bespoke supervised models on
	56	+labeled data---which requires labels, task-specific engineering, and yields opaque
	57	+scores---we use general-purpose large language and vision--language models with
	58	+schema-constrained decoding to produce a fixed set of interpretable, human-readable
	59	+attributes (e.g., ``renovated kitchen,'' ``brick façade,'' a 1--10 curb-appeal score)
	60	+at a marginal cost of a fraction of a cent per listing, and we embed those attributes
	61	+in a transparent hedonic regression whose coefficients can be read as implicit
	62	+prices. The models never observe the price, so the extracted attributes cannot
	63	+mechanically encode the outcome.
	64	+
	65	+The fourth strand is the \textbf{microstructure of listing markets}, which frames two
	66	+institutional features of our data. Our sample pools a brokered MLS platform
	67	+(Realtor.ca) with a for-sale-by-owner platform (DuProprio); \citet{hendel2009mls}
	68	+compare exactly these two market forms and document differences in seller
	69	+composition, marketing, and outcomes, motivating our platform indicator and
	70	+platform-specific robustness checks. And because we observe asking rather than
	71	+transaction prices, the interpretation of our dependent variable rests on the
	72	+list-price literature: the asking price is a strategic choice that anchors search and
	73	+bargaining \citep{horowitz1992list,knight2002listing,han2015microstructure}. We
	74	+return to this limitation in Section~\ref{sec:robustness}.

added paper/sections/methodology.tex +241 −0

		`@@ -0,0 +1,241 @@`
	1	+% ============================================================================
	2	+\section{Methodology}
	3	+\label{sec:methodology}
	4	+% ============================================================================
	5	+
	6	+Our empirical strategy has two parts: a feature-engineering pipeline that turns raw
	7	+listings into an analysis-ready design matrix (Sections~\ref{subsec:structural}--%
	8	+\ref{subsec:vision}), and a hedonic regression that maps those features to price
	9	+(Section~\ref{subsec:econometrics}).
	10	+
	11	+\subsection{Theoretical Framework}
	12	+\label{subsec:theory}
	13	+
	14	+Following \citet{rosen1974hedonic}, a house is a bundle of $K$ characteristics
	15	+$\mathbf{z}=(z_1,\dots,z_K)$, and the market clears through a hedonic price function
	16	+$P(\mathbf{z})$ that equates the marginal bid of buyers to the marginal offer of
	17	+sellers. A buyer with income $y$ and preferences $U(\mathbf{x},\mathbf{z})$ over a
	18	+numéraire good $\mathbf{x}$ and housing attributes chooses $\mathbf{z}$ to solve
	19	+%
	20	+\begin{equation}
	21	+ \max_{\mathbf{x},\,\mathbf{z}}\; U(\mathbf{x},\mathbf{z})
	22	+ \quad\text{s.t.}\quad y = \mathbf{x} + P(\mathbf{z}),
	23	+ \label{eq:consumer}
	24	+\end{equation}
	25	+%
	26	+which yields the first-order condition that, at the optimum, the implicit
	27	+(\emph{marginal}) price of each attribute equals the buyer's marginal willingness to
	28	+pay:
	29	+%
	30	+\begin{equation}
	31	+ p_k(\mathbf{z}) \;\equiv\; \frac{\partial P(\mathbf{z})}{\partial z_k}
	32	+ \;=\; \frac{\partial U/\partial z_k}{\partial U/\partial x},
	33	+ \qquad k = 1,\dots,K.
	34	+ \label{eq:implicit}
	35	+\end{equation}
	36	+%
	37	+Equation~\eqref{eq:implicit} is the object of interest: the slope of the hedonic
	38	+surface in the direction of attribute~$z_k$. The empirical task is to recover
	39	+$P(\mathbf{z})$---and hence its gradient---from observed prices and characteristics.
	40	+The linear model of Section~\ref{subsec:econometrics} imposes a parametric form on
	41	+$P(\cdot)$ and reads the implicit prices off its coefficients; the machine-learning
	42	+estimators of Section~\ref{sec:benchmark} approximate $P(\cdot)$ nonparametrically,
	43	+trading the closed-form gradient for predictive flexibility.
	44	+
	45	+\subsection{Structural and Locational Attributes}
	46	+\label{subsec:structural}
	47	+
	48	+Structural attributes are extracted by deterministic rules. Living area and lot area
	49	+are reported in heterogeneous units (\texttt{"1535 sqft"}, \texttt{"1,333.2 ft²"},
	50	+\texttt{"11 561.52 ft²"}, \texttt{"961.6 m2"}); we parse each string and convert all
	51	+areas to square feet. For Realtor.ca listings that report room-by-room dimensions
	52	+(e.g.\ \texttt{"23 ft ,4 in x 13 ft ,8 in"}) but no aggregate living area, we sum the
	53	+parsed room areas and calibrate the total against listings where both are observed;
	54	+the median ratio of reported area to summed-room area is~0.98, so the reconstructed
	55	+value is used directly to fill missing living area. Age is computed as
	56	+$2026-\text{year built}$, entered together with its square to allow a non-monotone
	57	+depreciation profile. Bedrooms, bathrooms, parking spaces, and storeys are taken
	58	+from the attribute block. Missing values are imputed by the within-property-type
	59	+median, and each imputed field is accompanied by a ``\texttt{has\_}'' indicator so
	60	+that the regression can distinguish observed from imputed values.
	61	+
	62	+Location is captured by \textbf{municipality fixed effects}. City names are
	63	+normalized (Realtor.ca appends neighborhood qualifiers in parentheses, which we
	64	+strip), and municipalities with fewer than 20 listings are pooled into a residual
	65	+category, leaving 195 city effects. Property-type fixed effects and a source
	66	+indicator complete the locational and platform controls.
	67	+
	68	+\subsection{Rule-Based Amenities and Room Detail}
	69	+\label{subsec:amenities}
	70	+
	71	+Seventeen amenity indicators are detected by bilingual (French/English) keyword
	72	+matching over the concatenated feature list, description, and title: swimming pool,
	73	+waterfront, garage, fireplace, central air conditioning or heat pump, finished
	74	+basement, cul-de-sac, proximity to a golf course, ski hill, park, school, highway,
	75	+and public transit, among others. From the Realtor.ca room block we additionally
	76	+construct room-detail indicators---presence of a primary bedroom, a basement, an
	77	+office, a family room, a workshop, a wine cellar, a sunroom, a mezzanine, a library,
	78	+and a laundry room---together with the room and storey counts. Two listing-quality
	79	+signals, the number of photographs (capped at~60) and the presence of a virtual
	80	+tour, proxy for the marketing tier of the listing.
	81	+
	82	+\subsection{Language-Model Specifications from Text}
	83	+\label{subsec:llm}
	84	+
	85	+Free-text descriptions contain valuation-relevant information that resists keyword
	86	+extraction: whether the home is turnkey or a fixer-upper, whether the kitchen or
	87	+bathroom has been renovated, the flooring and countertop materials, the heating
	88	+system, the presence of a water view, and income potential. We extract these as a
	89	+fixed set of structured specifications using a large language model
	90	+(\texttt{gpt-4o-mini}) with \emph{schema-constrained decoding}: the request specifies
	91	+a strict JSON schema with closed enumerations and required fields, so the API
	92	+validates the output at generation time and the model returns exactly the expected
	93	+fields in the expected value domains. Decoding temperature is set to zero for
	94	+determinism, and the system instruction constrains the model to report only what the
	95	+text supports, coding absent information as \texttt{false}/\texttt{unknown}/empty.
	96	+Each call consumes roughly 600 input and 110 output tokens; the extraction is run in
	97	+parallel, is resumable through an on-disk cache keyed by the listing identifier, and
	98	+retries transient errors with exponential backoff, for a total cost of about \$3
	99	+across the sample.
	100	+
	101	+\subsection{Vision Attributes from the Façade Photograph}
	102	+\label{subsec:vision}
	103	+
	104	+To capture the visual quality that neither structured fields nor text convey, we
	105	+analyze one photograph per listing---the façade---with a vision--language model
	106	+(\texttt{claude-haiku-4-5}). For Realtor.ca the first image is the exterior shot; for
	107	+DuProprio, whose images are captioned, we select the image whose caption is the
	108	+address rather than a room. The model returns twelve objective attributes under a
	109	+strict JSON schema: an overall condition score (1--10), a curb-appeal score (1--10),
	110	+a renovation level, exterior materials, landscaping quality, visible garage, storeys,
	111	+setting (urban/suburban/rural/waterfront/wooded), a luxury indicator, and photo
	112	+quality. The system instruction directs the model to ground each score in concrete,
	113	+reproducible criteria (surface condition, materials, landscaping) rather than
	114	+subjective taste, and to flag non-exterior images. Images are passed by URL without
	115	+local download; each call consumes roughly 2{,}470 input and 100 output tokens, for a
	116	+total cost of about \$57 across the sample. As with the text layer, the pass is
	117	+parallelized, resumable, and retried.
	118	+
	119	+\paragraph{Why schema-constrained decoding matters.}
	120	+Imposing a typed schema with enumerations, required fields, and
	121	+\texttt{additionalProperties:false} moves validation into the API layer. The models
	122	+are forced to emit exactly the expected fields within the expected value domains, so
	123	+their outputs behave as ordinary tabulated variables---reproducible and free of
	124	+free-text post-processing. The outputs are then mapped to numeric regressors:
	125	+booleans to $0/1$, the ordinal condition scale to $0$--$5$, materials to indicator
	126	+variables, and so on, each with an availability flag.
	127	+
	128	+\subsection{Econometric Specification}
	129	+\label{subsec:econometrics}
	130	+
	131	+We adopt a semi-logarithmic functional form, the standard workhorse in the hedonic
	132	+literature \citep{malpezzi2003hedonic}. Collect the attributes of listing~$i$ into
	133	+disjoint blocks,
	134	+%
	135	+\begin{equation}
	136	+ \resizebox{.94\textwidth}{!}{$\displaystyle
	137	+ \mathbf{X}_i =
	138	+ \big(\underbrace{\ln\text{Area}_i,\ \ln\text{Lot}_i}_{\text{size (logged)}},\
	139	+ \underbrace{\mathbf{s}_i}_{\text{structural}},\
	140	+ \underbrace{\mathbf{a}_i}_{\text{amenities}},\
	141	+ \underbrace{\mathbf{r}_i}_{\text{rooms}},\
	142	+ \underbrace{\mathbf{m}_i}_{\text{listing}},\
	143	+ \underbrace{\mathbf{g}_i}_{\text{LLM text}},\
	144	+ \underbrace{\mathbf{v}_i}_{\text{vision}}\big)
	145	+ $}
	146	+ \label{eq:blocks}
	147	+\end{equation}
	148	+%
	149	+and estimate, by ordinary least squares,
	150	+%
	151	+\begin{equation}
	152	+ \ln P_i \;=\; \alpha
	153	+ \;+\; \beta\,\ln\text{Area}_i
	154	+ \;+\; \gamma\,\ln\text{Lot}_i
	155	+ \;+\; \mathbf{X}_i^{\!\top}\bm{\delta}
	156	+ \;+\; \underbrace{\mu_{c(i)}}_{\text{city FE}}
	157	+ \;+\; \underbrace{\tau_{t(i)}}_{\text{type FE}}
	158	+ \;+\; \varepsilon_i ,
	159	+ \label{eq:hedonic}
	160	+\end{equation}
	161	+%
	162	+where $c(i)$ and $t(i)$ index the municipality and property type of listing~$i$, and
	163	+$\mu_c$, $\tau_t$ are the corresponding fixed effects that sweep out
	164	+location- and type-specific price levels. In matrix form
	165	+$\mathbf{y}=\mathbf{Z}\bm{\theta}+\bm{\varepsilon}$ with
	166	+$\widehat{\bm{\theta}}=(\mathbf{Z}^{\!\top}\mathbf{Z})^{-1}\mathbf{Z}^{\!\top}\mathbf{y}$.
	167	+
	168	+\paragraph{Interpretation of the coefficients.}
	169	+Because the size variables enter in logs, $\beta$ and $\gamma$ are \emph{elasticities}:
	170	+$\partial \ln P/\partial \ln\text{Area}=\beta$. For an indicator regressor $d$ with
	171	+coefficient $b$, the exact percentage price effect is not $b$ but
	172	+%
	173	+\begin{equation}
	174	+ \%\Delta P \;=\; \big(e^{\,b}-1\big)\times 100,
	175	+ \label{eq:halvorsen}
	176	+\end{equation}
	177	+%
	178	+the \citet{halvorsen1980interpretation} correction we apply throughout (for small
	179	+$b$, $e^{b}-1\approx b$). Under Equation~\eqref{eq:hedonic} the implicit price of a
	180	+continuous attribute $z_k$ from \eqref{eq:implicit} is $p_k=\delta_k\,P/z_k$, so the
	181	+semi-log form lets the \emph{dollar} marginal price scale with the price level while
	182	+keeping the \emph{percentage} effect constant.
	183	+
	184	+\paragraph{Two-part coverage design.}
	185	+Because living area is reported for only about 38\% of Realtor.ca listings, we
	186	+estimate two nested versions of \eqref{eq:hedonic} and combine their predictions:
	187	+%
	188	+\begin{equation}
	189	+ \widehat{\ln P}_i =
	190	+ \begin{cases}
	191	+ \widehat{\ln P}_i^{\,A} & \text{if area is measured (Model A, rich),}\\[2pt]
	192	+ \widehat{\ln P}_i^{\,B} & \text{otherwise (Model B, broad; }\beta\equiv 0).
	193	+ \end{cases}
	194	+ \label{eq:twopart}
	195	+\end{equation}
	196	+%
	197	+Model~A includes $\ln\text{Area}$ and is fit on the measured-area subsample; Model~B
	198	+drops it and is fit on the full cleaned sample, so every listing can be priced.
	199	+
	200	+\paragraph{Inference.}
	201	+House-price residuals are strongly heteroskedastic, and listings within a
	202	+municipality share unobserved local shocks that induce within-cluster error
	203	+correlation. Our preferred standard errors are therefore clustered at the
	204	+municipality level (195 clusters), the same level as the fixed effects
	205	+\citep{cameron2015cluster}:
	206	+%
	207	+\begin{equation}
	208	+ \widehat{\mathrm{Var}}_{\text{CL}}(\widehat{\bm{\theta}})
	209	+ = (\mathbf{Z}^{\!\top}\mathbf{Z})^{-1}
	210	+ \Bigg(\sum_{c=1}^{C}
	211	+ \mathbf{Z}_c^{\!\top}\hat{\bm{\varepsilon}}_c
	212	+ \hat{\bm{\varepsilon}}_c^{\!\top}\mathbf{Z}_c\Bigg)
	213	+ (\mathbf{Z}^{\!\top}\mathbf{Z})^{-1},
	214	+ \label{eq:cluster}
	215	+\end{equation}
	216	+%
	217	+where $\mathbf{Z}_c$ and $\hat{\bm{\varepsilon}}_c$ stack the regressors and
	218	+residuals of cluster $c$. As a complement we also compute
	219	+heteroskedasticity-consistent (HC3) standard errors
	220	+\citep{white1980heteroskedasticity,mackinnon1985some}, which inflate the
	221	+contribution of high-leverage points; Section~\ref{subsec:robust} shows that the two
	222	+choices agree on the significance of the property-level AI attributes and differ
	223	+only for keyword amenities with a strong spatial component.
	224	+
	225	+\paragraph{Retransformation.}
	226	+Predicting the \emph{level} of price requires undoing the log. Because
	227	+$\mathbb{E}[P\mid\mathbf{Z}]=\exp(\mathbf{Z}\bm{\theta})\,\mathbb{E}[e^{\varepsilon}]
	228	+\neq \exp(\mathbf{Z}\bm{\theta})$, naive exponentiation is biased downward. Under
	229	+$\varepsilon\sim\mathcal N(0,\sigma^2)$ we use the closed-form correction, and more
	230	+generally Duan's nonparametric smearing estimate \citep{duan1983smearing}:
	231	+%
	232	+\begin{equation}
	233	+ \widehat{P}_i
	234	+ = \exp\!\big(\widehat{\ln P}_i\big)\cdot
	235	+ \underbrace{\exp\!\big(\tfrac12\hat\sigma^{2}\big)}_{\text{normal}}
	236	+ \qquad\text{or}\qquad
	237	+ \widehat{P}_i
	238	+ = \exp\!\big(\widehat{\ln P}_i\big)\cdot
	239	+ \underbrace{\frac1n\sum_{j=1}^{n} e^{\hat\varepsilon_j}}_{\text{smearing }\hat\phi}.
	240	+ \label{eq:smearing}
	241	+\end{equation}

added paper/sections/results.tex +105 −0

		`@@ -0,0 +1,105 @@`
	1	+% ============================================================================
	2	+\section{Results}
	3	+\label{sec:results}
	4	+% ============================================================================
	5	+
	6	+\subsection{Baseline Estimates}
	7	+\label{subsec:baseline}
	8	+
	9	+Table~\ref{tab:regression} reports the two hedonic specifications estimated on the
	10	+full feature set (structural, amenity, room-detail, listing-signal, LLM text-derived,
	11	+and vision-derived attributes). Model~A, which includes living area and is estimated
	12	+on the 9{,}446 listings with a measured area, explains 75.5\% of the variation in log
	13	+prices ($R^2=0.755$, adjusted $0.748$). Model~B, the area-free specification estimated
	14	+on all 19{,}123 listings, explains 72.5\% ($R^2=0.725$, adjusted $0.721$). Adding the
	15	+machine-generated feature layers raises Model~A's $R^2$ from 0.717 (structural,
	16	+amenity, room, and listing attributes only) to 0.725 with the text layer and 0.755
	17	+with the vision layer---a cumulative gain of nearly four points from the two
	18	+AI-generated feature sets, over and above an already rich hedonic specification.
	19	+
	20	+The estimated implicit prices are economically sensible and, with few exceptions,
	21	+statistically significant with standard errors clustered by municipality. In
	22	+Model~A, the living-area elasticity is 0.36 (a 10\% larger home commands roughly a
	23	+3.6\% higher price) and the lot-area elasticity is small but precisely estimated.
	24	+Each additional \textbf{bathroom} is associated with a 7.6\% premium and each parking
	25	+space with about 1\%. \textbf{Central air conditioning} carries a 5.5\% premium and
	26	+\textbf{public-transit} proximity 9.4\%, although the transit coefficient---like
	27	+other keyword amenities with a strong spatial component---is estimated much less
	28	+precisely once errors are allowed to correlate within municipalities.
	29	+\textbf{Waterfront} access is associated with a 6.0--6.4\% premium, significant at
	30	+conventional levels in the full-sample model. Age enters with a small negative linear
	31	+term and a positive quadratic, tracing the familiar depreciation-then-heritage
	32	+profile.
	33	+
	34	+\input{tables/regression}
	35	+
	36	+\subsection{Text- and Vision-Derived Attributes}
	37	+\label{subsec:ai_coeffs}
	38	+
	39	+The two AI feature layers contribute distinct, largely orthogonal signal.
	40	+
	41	+\paragraph{Text (LLM).} Beyond the keyword amenities, the structured specifications
	42	+carry independent information about finish and condition: a \textbf{renovated
	43	+kitchen} is associated with a 2.6\% premium, \textbf{hardwood flooring} with 2.4\%,
	44	+and the ordinal \textbf{condition} scale is positive and significant in both models
	45	+(1.6\% per step). Listings the model reads as \textbf{income properties} sell about
	46	+1.9\% below otherwise-comparable houses. The \textbf{new-construction} indicator,
	47	+by contrast, is indistinguishable from zero once age, the visual renovation level,
	48	+and the condition scales are controlled for---the premium a naive specification
	49	+would attribute to the label ``new'' is fully absorbed by the measured quality
	50	+attributes.
	51	+
	52	+\paragraph{Vision (façade).} The image-derived attributes are jointly significant and
	53	+add the largest single increment to fit. Holding all structural, amenity, and text
	54	+attributes fixed, a one-point increase in \textbf{curb appeal} (1--10) is associated
	55	+with a 2.3\% higher price, and a one-step increase in the visual \textbf{renovation
	56	+level} with 2.4\%. A visible \textbf{luxury} cue adds 5.2\%, a \textbf{visible garage}
	57	+4.7\%, and each step of \textbf{landscaping} quality 1.5\%. Exterior materials price
	58	+as expected: \textbf{brick} ($+3.4$\%) and \textbf{stone} ($+3.7$\%) command premia,
	59	+while \textbf{vinyl} siding carries a 4.4\% discount. The visual \textbf{rural}
	60	+setting is associated with an 11.5\% discount relative to the omitted category. That
	61	+these visual measures remain significant after conditioning on the full text and
	62	+structural set---and with inference clustered at the municipality level---confirms
	63	+that the façade photograph carries price-relevant information available in no other
	64	+field.
	65	+
	66	+\subsection{Validation of Machine-Generated Features}
	67	+\label{subsec:validation}
	68	+
	69	+Before relying on the AI-generated attributes, we check that they order prices in the
	70	+expected direction. Figure~\ref{fig:validation} plots the median list price against the
	71	+vision model's curb-appeal and condition scores among listings with a detected
	72	+exterior shot. Both relationships are strongly monotone: median price rises steadily
	73	+from the lowest to the highest visual scores, with no reliance on the regression to
	74	+impose the pattern. Because these scores are produced by a model that never sees the
	75	+price, the monotonicity is independent evidence that the vision layer extracts a real,
	76	+price-relevant quality signal rather than noise. Table~\ref{tab:ai_summary} reports the
	77	+prevalence of the main text- and vision-derived attributes.
	78	+
	79	+\begin{figure}[htbp]
	80	+ \centering
	81	+ \includegraphics[width=\textwidth]{figures/fig_validation.pdf}
	82	+ \caption{Validation of the vision layer. Median list price rises monotonically with
	83	+ the model's curb-appeal (left) and condition (right) scores, among listings with a
	84	+ detected façade shot. The scores are generated without access to price.}
	85	+ \label{fig:validation}
	86	+\end{figure}
	87	+
	88	+\subsection{Location Fixed Effects}
	89	+\label{subsec:location}
	90	+
	91	+The 195 municipality fixed effects absorb the dominant locational component of price.
	92	+Relative to the reference city, the largest positive premia are estimated for
	93	+\textbf{Westmount} ($\approx+231$\%), \textbf{Hampstead} ($+173$\%), and
	94	+\textbf{Mont-Royal} ($+169$\%), followed by outlying municipalities at the low end
	95	+carrying discounts of 23--30\%. Figure~\ref{fig:premia} plots the extreme effects. The
	96	+magnitude and ordering align with well-known patterns in the Quebec housing market and
	97	+lend face validity to the specification.
	98	+
	99	+\begin{figure}[htbp]
	100	+ \centering
	101	+ \includegraphics[width=.74\textwidth]{figures/fig_premia.pdf}
	102	+ \caption{Estimated municipality fixed effects (Model~A), the twelve largest and eight
	103	+ smallest, expressed as percentage premia relative to the reference city.}
	104	+ \label{fig:premia}
	105	+\end{figure}

added paper/sections/robustness.tex +99 −0

		`@@ -0,0 +1,99 @@`
	1	+% ============================================================================
	2	+\section{Robustness and Limitations}
	3	+\label{sec:robustness}
	4	+% ============================================================================
	5	+
	6	+\subsection{Robustness}
	7	+\label{subsec:robust}
	8	+
	9	+Several features of the design guard against spurious results. The dual
	10	+specification (Section~\ref{subsec:econometrics}) ensures that estimates are not
	11	+driven by the non-random availability of living area: coefficients on attributes
	12	+common to both models retain their sign and rough magnitude whether or not area is
	13	+conditioned on. The ``\texttt{has\_}'' indicators accompanying each imputed field let
	14	+the regression separate observed from imputed information, so median imputation does
	15	+not silently attenuate the size coefficients. The price-per-square-foot trim removes
	16	+the residual parsing errors that would otherwise dominate the tails, and the
	17	+municipality fixed effects absorb the first-order spatial confound emphasized by
	18	+\citet{can1992spatial} and \citet{anselin1988spatial}.
	19	+
	20	+\paragraph{Inference.} Because listings within a municipality share unobserved
	21	+local shocks, our preferred standard errors are clustered at the municipality level
	22	+(195 clusters; \citealp{cameron2015cluster}), the same level as the fixed effects.
	23	+Clustering matters exactly where theory predicts: the text- and vision-derived
	24	+quality attributes---which vary house by house within a municipality---are
	25	+essentially unaffected (curb appeal, renovation level, luxury, garage, materials,
	26	+landscaping, and the rural-setting discount all remain significant at the 5\% level
	27	+or better in Model~A, as do the LLM condition scale, renovated kitchen, and hardwood
	28	+indicators), whereas keyword amenities with a strong spatial component lose
	29	+precision (public-transit proximity falls to the 5\% boundary in Model~A and is
	30	+insignificant in Model~B; central air conditioning is insignificant in Model~B; the
	31	+curb-appeal score, precisely estimated in Model~A, loses significance in the broader
	32	+Model~B).
	33	+Heteroskedasticity-robust (HC3) standard errors
	34	+\citep{white1980heteroskedasticity,mackinnon1985some} yield the same qualitative
	35	+pattern with uniformly smaller intervals. This asymmetry is reassuring for our main
	36	+claim: the AI-generated attributes capture within-market, property-level quality
	37	+variation rather than repackaged location effects.
	38	+
	39	+Table~\ref{tab:robustness} probes the stability of the predictive blend across
	40	+subsamples. Three patterns stand out. First, accuracy is comparable whether living
	41	+area is observed (median error 13.5\%) or imputed (14.4\%), confirming that the
	42	+area-free model prices the incomplete listings almost as well as the rich model.
	43	+Second, the model fits the broker-listed Realtor.ca inventory better
	44	+($R^2_{\text{OOS}}=0.80$) than the smaller, noisier for-sale-by-owner DuProprio
	45	+segment (0.58), though median errors are identical at 14.0\%; the gap is consistent
	46	+with the compositional and marketing differences between brokered and FSBO
	47	+platforms documented by \citet{hendel2009mls}. Third, the median error is stable
	48	+across price terciles (11--16\%), lowest in the mid-market and modestly higher in
	49	+the tails, where idiosyncratic and luxury attributes are harder to capture. (Within
	50	+a single tercile $R^2$ is uninformative because the price range is restricted; the
	51	+median error is the comparable metric.)
	52	+
	53	+\input{tables/robustness}
	54	+
	55	+\subsection{Limitations}
	56	+\label{subsec:limits}
	57	+
	58	+Five limitations qualify the interpretation of our estimates.
	59	+
	60	+\emph{Asking prices.} The dependent variable is the asking price of properties
	61	+listed for sale, not the transaction price. The list price is a strategic choice
	62	+that anchors buyer search and bargaining
	63	+\citep{horowitz1992list,knight2002listing,han2015microstructure}: it is set with
	64	+sale in mind and is highly correlated with the eventual transaction price, but
	65	+list-to-sale ratios vary systematically with market conditions and overpricing is
	66	+more common in thin segments. Two implications follow. First, our implicit prices
	67	+measure how sellers and their agents \emph{price} attributes into the ask, which
	68	+converges to buyers' valuation only insofar as the market disciplines the ask; the
	69	+economically sensible magnitudes and the monotone validation of
	70	+Section~\ref{subsec:validation} suggest the discipline is substantial. Second, any
	71	+attribute that shifts the list-to-sale ratio itself (e.g., attributes of hard-to-sell
	72	+properties) will carry a coefficient that mixes valuation and listing strategy.
	73	+Replicating the analysis on registry transaction prices is the natural next step.
	74	+
	75	+\emph{Non-causality and content endogeneity.} Descriptions and photographs are
	76	+marketing instruments chosen by sellers and agents. Richer descriptions, more
	77	+photographs, and higher visual quality correlate with higher-end listing agents and
	78	+more valuable properties \citep{nowak2017textual}. The coefficients are therefore
	79	+\emph{predictive} associations, not causal effects; they should not be read as the
	80	+price change from physically adding a given attribute. We control for two observable
	81	+dimensions of marketing effort (photo count and virtual tours), but cannot rule out
	82	+that residual marketing quality loads on the AI-derived attributes.
	83	+
	84	+\emph{Machine-generated features.} LLM and vision outputs can contain errors.
	85	+Zero-temperature decoding, strict schemas, and closed enumerations bound this risk
	86	+and make it reproducible, and classical measurement error in the extracted scores
	87	+would bias their coefficients toward zero---making our estimates conservative---but
	88	+we do not perform systematic human validation, and the visual layer judges each
	89	+property from a single façade photograph, ignoring interiors.
	90	+
	91	+\emph{Selection.} Listed properties are not a random sample of the housing stock,
	92	+and listings with missing area or room detail are not missing at random (they tend
	93	+to be less complete listings). The ``\texttt{has\_}'' indicators and the two-part
	94	+design mitigate, but do not eliminate, the second concern.
	95	+
	96	+\emph{Cross-section.} The data are a single-date snapshot, so the model captures no
	97	+market dynamics and the location premia are fixed as of 3~July~2026. Municipalities
	98	+with fewer than 20 listings are pooled, reducing spatial resolution where data are
	99	+sparse.

added paper/sections/titlepage.tex +77 −0

		`@@ -0,0 +1,77 @@`
	1	+% ============================================================================
	2	+% Title Page
	3	+% ============================================================================
	4	+\thispagestyle{empty}
	5	+
	6	+\begin{center}
	7	+
	8	+% --- Logo ---
	9	+\includegraphics[width=3.5cm]{uq_logo.jpg}
	10	+
	11	+\vspace{0.6cm}
	12	+
	13	+{\footnotesize\textsc{Université du Québec en Outaouais}}\\[0.15cm]
	14	+{\footnotesize\textsc{Département des sciences administratives}}
	15	+
	16	+\vspace{0.8cm}
	17	+
	18	+{\footnotesize\textsc{Working Paper No.~\WPnumber}}
	19	+
	20	+\vspace{1.2cm}
	21	+
	22	+% --- Title ---
	23	+{\LARGE\bfseries \WPtitle\par}
	24	+
	25	+\ifx\WPsubtitle\empty\else
	26	+ \vspace{0.3cm}
	27	+ {\large\itshape \WPsubtitle\par}
	28	+\fi
	29	+
	30	+\vspace{1.2cm}
	31	+
	32	+% --- Author ---
	33	+{\large \WPauthor}\\[0.3cm]
	34	+{\normalsize \WPaffiliation}\\[0.15cm]
	35	+{\normalsize \href{mailto:\WPemail}{\WPemail}}\\[0.15cm]
	36	+{\small \WPaddress}
	37	+
	38	+\vspace{0.8cm}
	39	+
	40	+% --- Date & Version ---
	41	+{\normalsize \WPdate}\\[0.1cm]
	42	+{\small Version~\WPversion}
	43	+
	44	+\end{center}
	45	+
	46	+\vfill
	47	+
	48	+\newpage
	49	+
	50	+% ============================================================================
	51	+% Abstract Page
	52	+% ============================================================================
	53	+\thispagestyle{empty}
	54	+
	55	+\vspace*{1cm}
	56	+
	57	+\noindent\rule{\textwidth}{0.4pt}
	58	+\vspace{0.3cm}
	59	+
	60	+\noindent\textbf{Abstract}
	61	+
	62	+\vspace{0.15cm}
	63	+
	64	+\noindent\WPabstract
	65	+
	66	+\vspace{0.4cm}
	67	+
	68	+\noindent\textbf{Keywords:} \WPkeywords
	69	+
	70	+\vspace{0.15cm}
	71	+
	72	+\noindent\textbf{JEL Classification:} \WPjel
	73	+
	74	+\vspace{0.3cm}
	75	+\noindent\rule{\textwidth}{0.4pt}
	76	+
	77	+\newpage

added paper/tables/ablation.tex +27 −0

		`@@ -0,0 +1,27 @@`
	1	+% Auto-generated (revision: recomputed on identical folds)
	2	+\begin{table}[htbp]
	3	+ \centering
	4	+ \begin{threeparttable}
	5	+ \caption{Feature Ablation: Out-of-Sample $R^2$ by Feature Layer}
	6	+ \label{tab:ablation}
	7	+ \begin{tabular}{@{}lcccc@{}}
	8	+ \toprule
	9	+ Feature set & \# feats & OLS & GBM & Blend \\
	10	+ \midrule
	11	+ Structural + amenity & 24 & 0.521 & 0.692 & --- \\
	12	+ \quad + Rooms + listing & 38 & 0.619 & 0.729 & --- \\
	13	+ \quad + Text (LLM) & 56 & 0.628 & 0.743 & 0.760 \\
	14	+ \quad + Vision (façade) & 77 & 0.686 & 0.770 & 0.779 \\
	15	+ \bottomrule
	16	+ \end{tabular}
	17	+ \begin{tablenotes}[flushleft]
	18	+ \footnotesize
	19	+ \item \textit{Notes:} Each row adds a feature layer to the row above (nested).
	20	+ Entries are out-of-sample $R^2$ on $\ln(\text{Price})$ from 5-fold cross-validation
	21	+ with a common seed. ``\# feats'' counts numeric/indicator features before the city,
	22	+ type, and source fixed effects (added in every specification). The blend averages
	23	+ the robust neural network and gradient boosting and is reported where both were
	24	+ trained. GBM handles missing living area natively; OLS uses median imputation.
	25	+ \end{tablenotes}
	26	+ \end{threeparttable}
	27	+\end{table}

added paper/tables/ai_summary.tex +28 −0

		`@@ -0,0 +1,28 @@`
	1	+% Auto-generated
	2	+\begin{table}[htbp]
	3	+ \centering
	4	+ \begin{threeparttable}
	5	+ \caption{Machine-Generated Feature Summary}
	6	+ \label{tab:ai_summary}
	7	+ \begin{tabular}{@{}lc@{\hskip 1.2cm}lc@{}}
	8	+ \toprule
	9	+ Text (LLM) & Share & Vision (façade) & Mean / share \\
	10	+ \midrule
	11	+ Recently renovated & 26\% & Condition score (1--10) & 7.4 \\
	12	+ Renovated kitchen & 17\% & Curb appeal (1--10) & 7.2 \\
	13	+ Hardwood floor & 16\% & Luxury cue & 43\% \\
	14	+ Quartz/granite counter & 7\% & Garage visible & 46\% \\
	15	+ New construction & 14\% & Brick façade & 47\% \\
	16	+ Income property & 12\% & Waterfront setting & 7\% \\
	17	+ Water view & 24\% & Exterior shot detected & 96\% \\
	18	+ \bottomrule
	19	+ \end{tabular}
	20	+ \begin{tablenotes}[flushleft]
	21	+ \footnotesize
	22	+ \item \textit{Notes:} Prevalence of selected machine-generated attributes across
	23	+ the 19{,}123 cleaned listings. Text attributes are extracted from the
	24	+ description by \texttt{gpt-4o-mini}; vision attributes from the façade photograph by
	25	+ \texttt{claude-haiku-4-5}. Scores are 1--10 means; the remainder are shares.
	26	+ \end{tablenotes}
	27	+ \end{threeparttable}
	28	+\end{table}

added paper/tables/models.tex +39 −0

		`@@ -0,0 +1,39 @@`
	1	+% Auto-generated (revision: full ladder in both panels)
	2	+\begin{table}[htbp]
	3	+ \centering
	4	+ \begin{threeparttable}
	5	+ \caption{Predictive Performance: From OLS to Deep Learning}
	6	+ \label{tab:models}
	7	+ \begin{tabular}{@{}lcccc@{}}
	8	+ \toprule
	9	+ Estimator & $R^2_{\text{OOS}}$ & Median AE & Within $\pm$15\% & Within $\pm$25\% \\
	10	+ \midrule
	11	+ \multicolumn{5}{@{}l}{\textit{Panel A: Structural + amenity + LLM-text features}} \\[3pt]
	12	+ OLS (linear hedonic) & 0.628 & 17.9\% & 43\% & 64\% \\
	13	+ Random Forest & 0.677 & 17.9\% & 43\% & 64\% \\
	14	+ Neural network (plain MLP) & 0.711 & 16.1\% & 47\% & 69\% \\
	15	+ Neural network (robust ResNet) & 0.741 & 15.2\% & 49\% & 70\% \\
	16	+ Gradient boosting (HistGBM) & 0.743 & 15.4\% & 49\% & 70\% \\
	17	+ Blend (NN + GBM) & 0.760 & 14.7\% & 51\% & 72\% \\[4pt]
	18	+ \multicolumn{5}{@{}l}{\textit{Panel B: \ + Vision (façade) features}} \\[3pt]
	19	+ OLS (linear hedonic) & 0.686 & 16.9\% & 45\% & 67\% \\
	20	+ Random Forest & 0.704 & 17.2\% & 45\% & 66\% \\
	21	+ Neural network (robust ResNet) & 0.747 & 14.8\% & 51\% & 71\% \\
	22	+ Gradient boosting (HistGBM) & 0.770 & 14.5\% & 51\% & 71\% \\
	23	+ \textbf{Blend (NN + GBM)} & \textbf{0.779} & \textbf{14.0\%} & \textbf{53\%} & \textbf{73\%} \\
	24	+ \bottomrule
	25	+ \end{tabular}
	26	+ \begin{tablenotes}[flushleft]
	27	+ \footnotesize
	28	+ \item \textit{Notes:} Out-of-sample performance from 5-fold cross-validation with a
	29	+ common seed (identical folds across rows). $R^2_{\text{OOS}}$ is on
	30	+ $\ln(\text{Price})$; ``Median AE'' is the median absolute percentage error in
	31	+ levels (Equation~\ref{eq:cvmetrics}). Panel~A uses structural, amenity, room-detail,
	32	+ listing-signal, and LLM text-derived features; Panel~B adds the twelve façade
	33	+ vision attributes. City, type, and source enter as one-hot indicators for the
	34	+ linear and tree models and as learned embeddings for the neural networks. The tree
	35	+ and neural models handle missing living area natively; OLS uses within-type median
	36	+ imputation. The best model is the text\,+\,vision blend.
	37	+ \end{tablenotes}
	38	+ \end{threeparttable}
	39	+\end{table}

added paper/tables/regression.tex +82 −0

		`@@ -0,0 +1,82 @@`
	1	+% Auto-generated (revision: clustered SEs)
	2	+\begin{table}[htbp]
	3	+ \centering
	4	+ \begin{threeparttable}
	5	+ \caption{Hedonic Regression Results (Text + Vision Features)}
	6	+ \label{tab:regression}
	7	+ \begin{tabular}{@{}lcc@{}}
	8	+ \toprule
	9	+ & (1) Model~A & (2) Model~B \\
	10	+ & \textit{rich} & \textit{broad} \\
	11	+ \midrule
	12	+ $\ln(\text{Living area})$ & 0.358$^{***}$ & \\
	13	+ & (0.012) & \\[3pt]
	14	+ $\ln(\text{Lot area})$ & 0.027$^{*}$ & 0.040$^{*}$ \\
	15	+ & (0.006) & (0.005) \\[3pt]
	16	+ Bedrooms & 0.014$^{*}$ & 0.032$^{*}$ \\
	17	+ & (0.005) & (0.003) \\[3pt]
	18	+ Bathrooms & 0.073$^{*}$ & 0.113$^{*}$ \\
	19	+ & (0.007) & (0.007) \\[3pt]
	20	+ Parking & 0.011$^{*}$ & 0.015$^{*}$ \\
	21	+ & (0.001) & (0.002) \\[3pt]
	22	+ Age & $-$0.004$^{*}$ & $-$0.005$^{*}$ \\
	23	+ & (0.001) & (0.001) \\[3pt]
	24	+ \multicolumn{3}{@{}l}{\textit{Amenities \& text (LLM)}} \\[2pt]
	25	+ Central A/C & 0.053$^{***}$ & 0.020 \\
	26	+ & (0.018) & (0.018) \\[3pt]
	27	+ Public transit & 0.090$^{**}$ & 0.058 \\
	28	+ & (0.046) & (0.039) \\[3pt]
	29	+ Waterfront & 0.058$^{}$ & 0.062$^{**}$ \\
	30	+ & (0.033) & (0.019) \\[3pt]
	31	+ Condition (LLM, ord.) & 0.016$^{*}$ & 0.016$^{*}$ \\
	32	+ & (0.006) & (0.005) \\[3pt]
	33	+ Reno.\ kitchen (LLM) & 0.026$^{*}$ & 0.027$^{*}$ \\
	34	+ & (0.007) & (0.010) \\[3pt]
	35	+ Hardwood floor (LLM) & 0.024$^{}$ & 0.030$^{*}$ \\
	36	+ & (0.009) & (0.011) \\[3pt]
	37	+ New construction (LLM) & $-$0.000 & $-$0.016 \\
	38	+ & (0.013) & (0.014) \\[3pt]
	39	+ Income property (LLM) & $-$0.019$^{*}$ & $-$0.013$^{}$ \\
	40	+ & (0.009) & (0.008) \\[3pt]
	41	+ \multicolumn{3}{@{}l}{\textit{Vision (façade)}} \\[2pt]
	42	+ Curb appeal (1--10) & 0.023$^{***}$ & 0.016 \\
	43	+ & (0.009) & (0.013) \\[3pt]
	44	+ Renovation level (ord.) & 0.024$^{}$ & 0.020$^{}$ \\
	45	+ & (0.010) & (0.009) \\[3pt]
	46	+ Luxury (visual) & 0.051$^{*}$ & 0.062$^{*}$ \\
	47	+ & (0.012) & (0.009) \\[3pt]
	48	+ Garage visible & 0.046$^{*}$ & 0.047$^{*}$ \\
	49	+ & (0.006) & (0.007) \\[3pt]
	50	+ Brick façade & 0.033$^{*}$ & 0.044$^{*}$ \\
	51	+ & (0.009) & (0.011) \\[3pt]
	52	+ Stone façade & 0.036$^{*}$ & 0.047$^{*}$ \\
	53	+ & (0.008) & (0.005) \\[3pt]
	54	+ Vinyl façade & $-$0.045$^{*}$ & $-$0.062$^{*}$ \\
	55	+ & (0.015) & (0.012) \\[3pt]
	56	+ Rural setting & $-$0.122$^{*}$ & $-$0.076$^{*}$ \\
	57	+ & (0.032) & (0.029) \\[3pt]
	58	+ \midrule
	59	+ City FE / Type FE & Yes & Yes \\
	60	+ Living area & Yes & No \\
	61	+ Observations & 9{,}446 & 19{,}123 \\
	62	+ $R^2$ & 0.755 & 0.725 \\
	63	+ Adjusted $R^2$ & 0.748 & 0.721 \\
	64	+ \bottomrule
	65	+ \end{tabular}
	66	+ \begin{tablenotes}[flushleft]
	67	+ \footnotesize
	68	+ \item \textit{Notes:} OLS estimates of Equation~\eqref{eq:hedonic} with the full
	69	+ feature set. Model~A includes $\ln(\text{Living area})$ (measured-area subsample);
	70	+ Model~B omits it (full sample). ``LLM'' variables are extracted from the
	71	+ description text; ``Vision'' variables from the façade photograph. Curb appeal and
	72	+ renovation level enter as scores; luxury, garage, and material/setting are
	73	+ indicators. All specifications include city and property-type fixed effects and the
	74	+ full amenity, room-detail, and listing-signal controls (only selected coefficients
	75	+ shown). Standard errors clustered by municipality (195 clusters) in parentheses.
	76	+ Heteroskedasticity-robust (HC3) standard errors are uniformly smaller and yield
	77	+ the same significance pattern for the text- and vision-derived attributes, except
	78	+ that curb appeal in Model~B is significant under HC3 but not under clustering.
	79	+ $^{*}$, $^{}$, $^{*}$: significance at 1\%, 5\%, 10\%.
	80	+ \end{tablenotes}
	81	+ \end{threeparttable}
	82	+\end{table}

added paper/tables/robustness.tex +34 −0

		`@@ -0,0 +1,34 @@`
	1	+% Auto-generated
	2	+\begin{table}[htbp]
	3	+ \centering
	4	+ \begin{threeparttable}
	5	+ \caption{Robustness: Blend Performance Across Subsamples}
	6	+ \label{tab:robustness}
	7	+ \begin{tabular}{@{}lcccc@{}}
	8	+ \toprule
	9	+ Subsample & $N$ & $R^2_{\text{OOS}}$ & Median AE & Within $\pm$15\% \\
	10	+ \midrule
	11	+ \textit{Full sample} & 19{,}123 & 0.779 & 14.0\% & 53\% \\
	12	+ \quad Price tercile 1 (low) & 6{,}375 & n/a$^{\dagger}$ & 15.6\% & 49\% \\
	13	+ \quad Price tercile 2 (mid) & 6{,}412 & n/a$^{\dagger}$ & 11.0\% & 64\% \\
	14	+ \quad Price tercile 3 (high) & 6{,}336 & n/a$^{\dagger}$ & 16.4\% & 47\% \\
	15	+ \quad Realtor.ca & 15{,}356 & 0.802 & 14.0\% & 53\% \\
	16	+ \quad DuProprio & 3{,}767 & 0.579 & 14.0\% & 54\% \\
	17	+ \quad Measured area & 9{,}446 & 0.773 & 13.5\% & 55\% \\
	18	+ \quad Imputed area & 9{,}677 & 0.783 & 14.4\% & 52\% \\
	19	+ \bottomrule
	20	+ \end{tabular}
	21	+ \begin{tablenotes}[flushleft]
	22	+ \footnotesize
	23	+ \item \textit{Notes:} Out-of-sample performance of the text\,+\,vision blend
	24	+ (Table~\ref{tab:models}, Panel~B) computed on subsamples of the same 5-fold
	25	+ cross-validated predictions. Price terciles are defined on the actual list price.
	26	+ ``Measured area'' are listings with a reported or room-reconstructed living area
	27	+ (priced by Model~A); ``Imputed area'' listings are priced by the area-free model.
	28	+ $^{\dagger}$Within a single price tercile the range of $\ln P$ is heavily
	29	+ restricted, so $R^2$---which normalizes by within-group variance---is
	30	+ uninformative; the median absolute error is the comparable accuracy metric across
	31	+ terciles.
	32	+ \end{tablenotes}
	33	+ \end{threeparttable}
	34	+\end{table}

added paper/tables/summary_stats.tex +28 −0

		`@@ -0,0 +1,28 @@`
	1	+% Auto-generated
	2	+\begin{table}[htbp]
	3	+ \centering
	4	+ \begin{threeparttable}
	5	+ \caption{Summary Statistics}
	6	+ \label{tab:summary_stats}
	7	+ \begin{tabular}{@{}lcccccc@{}}
	8	+ \toprule
	9	+ Variable & $N$ & Mean & Std.\ Dev. & $p_{25}$ & Median & $p_{75}$ \\
	10	+ \midrule
	11	+ Price (CAD) & 19{,}123 & 764{{,}}708 & 689{{,}}118 & 439{{,}}000 & 599{{,}}900 & 849{{,}}000 \\
	12	+ Living area (ft$^2$) & 9{,}446 & 1{{,}}730 & 990 & 1{{,}}128 & 1{{,}}470 & 2{{,}}026 \\
	13	+ Lot area (ft$^2$) & 19{,}123 & 71{{,}}429 & 309{{,}}209 & 6{{,}}000 & 10{{,}}686 & 34{{,}}978 \\
	14	+ Bedrooms & 19{,}123 & 3.5 & 1.2 & 3.0 & 3.0 & 4.0 \\
	15	+ Bathrooms & 19{,}123 & 1.8 & 0.8 & 1.0 & 2.0 & 2.0 \\
	16	+ Parking spaces & 19{,}123 & 4.0 & 4.4 & 0.0 & 3.0 & 6.0 \\
	17	+ Storeys & 19{,}123 & 1.6 & 1.6 & 1.0 & 2.0 & 2.0 \\
	18	+ Age (years) & 19{,}123 & 42.5 & 28.0 & 24.5 & 40.0 & 53.0 \\
	19	+ Rooms (Realtor) & 19{,}123 & 10.1 & 6.5 & 7.0 & 11.0 & 14.0 \\
	20	+ Photos & 19{,}123 & 42 & 18 & 30 & 40 & 51 \\
	21	+ \bottomrule
	22	+ \end{tabular}
	23	+ \begin{tablenotes}[flushleft]
	24	+ \footnotesize
	25	+ \item \textit{Notes:} Descriptive statistics for the cleaned sample of 19{,}123 single-family and townhouse listings (Realtor.ca and DuProprio), scraped on 3~July~2026. 80\% of listings originate from Realtor.ca. Living area is the reported area, or reconstructed from room dimensions where missing. Continuous variables are bounded by the outlier rules described in Section~\ref{sec:data}; price-per-square-foot is trimmed at the 0.5th and 99.5th percentiles.
	26	+ \end{tablenotes}
	27	+ \end{threeparttable}
	28	+\end{table}

added paper/uq_logo.jpg +0 −0

Binary file not shown.

added requirements.txt +6 −0

		`@@ -0,0 +1,6 @@`
	1	+numpy>=1.26
	2	+pandas>=2.0
	3	+statsmodels>=0.14
	4	+scikit-learn>=1.4
	5	+matplotlib>=3.8
	6	+torch>=2.2

added results/ablation_extras.json +62 −0

		`@@ -0,0 +1,62 @@`
	1	+{
	2	+ "L1": {
	3	+ "nfeat": 24,
	4	+ "ols": {
	5	+ "r2": 0.5208807490583831,
	6	+ "mdae": 19.311838552673066,
	7	+ "w15": 40.46959159127752,
	8	+ "w25": 60.90048632536735
	9	+ },
	10	+ "gbm": {
	11	+ "r2": 0.6922092084886344,
	12	+ "mdae": 17.160081786433278,
	13	+ "w15": 44.70532866182084,
	14	+ "w25": 65.6591538984469
	15	+ }
	16	+ },
	17	+ "L2": {
	18	+ "nfeat": 38,
	19	+ "ols": {
	20	+ "r2": 0.6193226824225797,
	21	+ "mdae": 18.059509413013565,
	22	+ "w15": 42.81754954766512,
	23	+ "w25": 63.755686869215076
	24	+ },
	25	+ "gbm": {
	26	+ "r2": 0.7288688475138769,
	27	+ "mdae": 15.95094002848213,
	28	+ "w15": 47.46640171521205,
	29	+ "w25": 68.38885112168592
	30	+ }
	31	+ },
	32	+ "L3": {
	33	+ "nfeat": 56,
	34	+ "ols": {
	35	+ "r2": 0.6277548901386698,
	36	+ "mdae": 17.877027384670708,
	37	+ "w15": 42.92213564817236,
	38	+ "w25": 64.46687235266432
	39	+ },
	40	+ "gbm": {
	41	+ "r2": 0.7428875872402715,
	42	+ "mdae": 15.421354827745482,
	43	+ "w15": 48.84170893688229,
	44	+ "w25": 69.91580818909166
	45	+ }
	46	+ },
	47	+ "L4": {
	48	+ "nfeat": 77,
	49	+ "ols": {
	50	+ "r2": 0.6862361284610743,
	51	+ "mdae": 16.886250547842508,
	52	+ "w15": 45.25440568948387,
	53	+ "w25": 66.72070281859541
	54	+ },
	55	+ "gbm": {
	56	+ "r2": 0.7695283346486282,
	57	+ "mdae": 14.51330744579874,
	58	+ "w15": 51.383151179208284,
	59	+ "w25": 71.29634471578727
	60	+ }
	61	+ }
	62	+}
		\ No newline at end of file

added results/hedonic_summary.txt +536 −0

		`@@ -0,0 +1,536 @@`
	1	+=== MODELE A (avec superficie) ===
	2	+ OLS Regression Results
	3	+==============================================================================
	4	+Dep. Variable: ln_price R-squared: 0.717
	5	+Model: OLS Adj. R-squared: 0.709
	6	+Method: Least Squares F-statistic: 93.16
	7	+Date: Fri, 03 Jul 2026 Prob (F-statistic): 0.00
	8	+Time: 07:41:43 Log-Likelihood: -1866.7
	9	+No. Observations: 9444 AIC: 4215.
	10	+Df Residuals: 9203 BIC: 5939.
	11	+Df Model: 240
	12	+Covariance Type: HC3
	13	+========================================================================================================================
	14	+ coef std err z P>\|z\| [0.025 0.975]
	15	+------------------------------------------------------------------------------------------------------------------------
	16	+Intercept 9.7330 0.094 103.249 0.000 9.548 9.918
	17	+C(source)[T.realtor] -0.2189 0.022 -10.109 0.000 -0.261 -0.176
	18	+C(ptype_fe)[T.Country Home] 0.0045 0.051 0.087 0.930 -0.096 0.105
	19	+C(ptype_fe)[T.Home] -0.0071 0.013 -0.525 0.599 -0.034 0.019
	20	+C(ptype_fe)[T.House] -0.0307 0.012 -2.558 0.011 -0.054 -0.007
	21	+C(ptype_fe)[T.Intergenerational] -0.0200 0.036 -0.558 0.577 -0.090 0.050
	22	+C(ptype_fe)[T.Other] 0.2178 0.120 1.822 0.068 -0.016 0.452
	23	+C(ptype_fe)[T.Row / Townhouse] -0.1882 0.016 -12.125 0.000 -0.219 -0.158
	24	+C(ptype_fe)[T.Semi Detached] -0.0601 0.015 -3.893 0.000 -0.090 -0.030
	25	+C(ptype_fe)[T.Townhouse] -0.0038 0.025 -0.151 0.880 -0.053 0.045
	26	+C(city_fe)[T.acton vale] -0.0615 0.058 -1.065 0.287 -0.175 0.052
	27	+C(city_fe)[T.alma] -0.0170 0.054 -0.317 0.751 -0.122 0.088
	28	+C(city_fe)[T.amherst] -0.0746 0.076 -0.980 0.327 -0.224 0.075
	29	+C(city_fe)[T.amos] -0.0308 0.063 -0.490 0.624 -0.154 0.092
	30	+C(city_fe)[T.austin] 0.2786 0.091 3.061 0.002 0.100 0.457
	31	+C(city_fe)[T.baie-comeau] -0.3656 0.083 -4.411 0.000 -0.528 -0.203
	32	+C(city_fe)[T.baie-saint-paul] -0.0934 0.053 -1.765 0.078 -0.197 0.010
	33	+C(city_fe)[T.beaconsfield] 0.6128 0.052 11.867 0.000 0.512 0.714
	34	+C(city_fe)[T.beauport] 0.0424 0.037 1.149 0.250 -0.030 0.115
	35	+C(city_fe)[T.beloeil] 0.2255 0.052 4.354 0.000 0.124 0.327
	36	+C(city_fe)[T.blainville] 0.3095 0.030 10.218 0.000 0.250 0.369
	37	+C(city_fe)[T.blue sea] -0.0272 0.060 -0.450 0.652 -0.146 0.091
	38	+C(city_fe)[T.bois-des-filion] 0.1842 0.057 3.221 0.001 0.072 0.296
	39	+C(city_fe)[T.boisbriand] 0.2661 0.039 6.854 0.000 0.190 0.342
	40	+C(city_fe)[T.boischatel] 0.1368 0.082 1.664 0.096 -0.024 0.298
	41	+C(city_fe)[T.boucherville] 0.3473 0.040 8.688 0.000 0.269 0.426
	42	+C(city_fe)[T.bromont] 0.3122 0.037 8.391 0.000 0.239 0.385
	43	+C(city_fe)[T.brossard] 0.4100 0.030 13.746 0.000 0.352 0.468
	44	+C(city_fe)[T.bécancour] -0.0918 0.039 -2.323 0.020 -0.169 -0.014
	45	+C(city_fe)[T.cantley] 0.1223 0.033 3.758 0.000 0.059 0.186
	46	+C(city_fe)[T.carignan] 0.2488 0.086 2.881 0.004 0.080 0.418
	47	+C(city_fe)[T.chandler] -0.2905 0.141 -2.068 0.039 -0.566 -0.015
	48	+C(city_fe)[T.chateauguay] 0.2290 0.036 6.421 0.000 0.159 0.299
	49	+C(city_fe)[T.chelsea] 0.2572 0.038 6.779 0.000 0.183 0.332
	50	+C(city_fe)[T.chertsey] -0.0464 0.064 -0.730 0.466 -0.171 0.078
	51	+C(city_fe)[T.chicoutimi] -0.1007 0.046 -2.203 0.028 -0.190 -0.011
	52	+C(city_fe)[T.chomedey] 0.3195 0.050 6.430 0.000 0.222 0.417
	53	+C(city_fe)[T.châteauguay] 0.2060 0.038 5.390 0.000 0.131 0.281
	54	+C(city_fe)[T.contrecoeur] 0.0915 0.057 1.607 0.108 -0.020 0.203
	55	+C(city_fe)[T.cowansville] 0.1177 0.050 2.376 0.017 0.021 0.215
	56	+C(city_fe)[T.crabtree] 0.0241 0.096 0.251 0.802 -0.164 0.213
	57	+C(city_fe)[T.côte-saint-luc] 0.6071 0.050 12.165 0.000 0.509 0.705
	58	+C(city_fe)[T.dolbeau-mistassini] -0.2645 0.091 -2.905 0.004 -0.443 -0.086
	59	+C(city_fe)[T.dollard-des-ormeaux] 0.4023 0.038 10.535 0.000 0.327 0.477
	60	+C(city_fe)[T.donnacona] 0.2465 0.039 6.358 0.000 0.171 0.322
	61	+C(city_fe)[T.dorval] 0.5515 0.055 10.101 0.000 0.445 0.659
	62	+C(city_fe)[T.drummondville] -0.0196 0.022 -0.895 0.371 -0.062 0.023
	63	+C(city_fe)[T.duhamel] 0.1373 0.076 1.815 0.069 -0.011 0.286
	64	+C(city_fe)[T.dunham] 0.2726 0.064 4.237 0.000 0.147 0.399
	65	+C(city_fe)[T.eastman] 0.3041 0.051 5.919 0.000 0.203 0.405
	66	+C(city_fe)[T.entrelacs] 0.2617 0.135 1.945 0.052 -0.002 0.525
	67	+C(city_fe)[T.farnham] 0.0545 0.062 0.885 0.376 -0.066 0.175
	68	+C(city_fe)[T.fossambault-sur-le-lac] 0.2664 0.079 3.365 0.001 0.111 0.422
	69	+C(city_fe)[T.gaspé] -0.2214 0.069 -3.228 0.001 -0.356 -0.087
	70	+C(city_fe)[T.gatineau] 0.0901 0.012 7.233 0.000 0.066 0.114
	71	+C(city_fe)[T.gore] 0.0786 0.104 0.753 0.451 -0.126 0.283
	72	+C(city_fe)[T.granby] 0.1117 0.024 4.632 0.000 0.064 0.159
	73	+C(city_fe)[T.grenville-sur-la-rouge] 0.1287 0.078 1.656 0.098 -0.024 0.281
	74	+C(city_fe)[T.hampstead] 1.0843 0.057 19.163 0.000 0.973 1.195
	75	+C(city_fe)[T.harrington] 0.1429 0.152 0.940 0.347 -0.155 0.441
	76	+C(city_fe)[T.hudson] 0.2555 0.199 1.282 0.200 -0.135 0.646
	77	+C(city_fe)[T.joliette] -0.0275 0.071 -0.387 0.699 -0.167 0.112
	78	+C(city_fe)[T.jonquière] -0.1380 0.045 -3.096 0.002 -0.225 -0.051
	79	+C(city_fe)[T.kirkland] 0.3686 0.037 9.967 0.000 0.296 0.441
	80	+C(city_fe)[T.l'ange-gardien] 0.0142 0.158 0.090 0.928 -0.295 0.323
	81	+C(city_fe)[T.l'assomption] 0.1479 0.060 2.460 0.014 0.030 0.266
	82	+C(city_fe)[T.la conception] 0.3351 0.053 6.306 0.000 0.231 0.439
	83	+C(city_fe)[T.la malbaie] -0.1347 0.063 -2.142 0.032 -0.258 -0.011
	84	+C(city_fe)[T.la minerve] 0.0449 0.069 0.646 0.518 -0.091 0.181
	85	+C(city_fe)[T.la prairie] 0.3741 0.035 10.542 0.000 0.305 0.444
	86	+C(city_fe)[T.la pêche] 0.0354 0.037 0.968 0.333 -0.036 0.107
	87	+C(city_fe)[T.la tuque] -0.3119 0.067 -4.627 0.000 -0.444 -0.180
	88	+C(city_fe)[T.labelle] 0.1111 0.113 0.980 0.327 -0.111 0.333
	89	+C(city_fe)[T.lac-beauport] 0.3609 0.094 3.852 0.000 0.177 0.545
	90	+C(city_fe)[T.lac-brome] 0.5002 0.062 8.091 0.000 0.379 0.621
	91	+C(city_fe)[T.lac-des-écorces] -0.0964 0.159 -0.607 0.544 -0.408 0.215
	92	+C(city_fe)[T.lac-etchemin] -0.1890 0.135 -1.396 0.163 -0.454 0.076
	93	+C(city_fe)[T.lachute] 0.1234 0.075 1.655 0.098 -0.023 0.269
	94	+C(city_fe)[T.lanoraie] -0.0824 0.051 -1.623 0.105 -0.182 0.017
	95	+C(city_fe)[T.laval] 0.3952 0.020 19.308 0.000 0.355 0.435
	96	+C(city_fe)[T.lavaltrie] 0.1283 0.065 1.990 0.047 0.002 0.255
	97	+C(city_fe)[T.les éboulements] 0.2007 0.080 2.516 0.012 0.044 0.357
	98	+C(city_fe)[T.longueuil] 0.3160 0.020 16.113 0.000 0.278 0.354
	99	+C(city_fe)[T.lorraine] 0.1128 0.113 0.995 0.320 -0.109 0.335
	100	+C(city_fe)[T.lévis] 0.0663 0.027 2.486 0.013 0.014 0.119
	101	+C(city_fe)[T.magog] 0.3570 0.034 10.628 0.000 0.291 0.423
	102	+C(city_fe)[T.mandeville] -0.1781 0.080 -2.221 0.026 -0.335 -0.021
	103	+C(city_fe)[T.mansfield-et-pontefract] -0.3813 0.079 -4.847 0.000 -0.536 -0.227
	104	+C(city_fe)[T.marieville] 0.1717 0.105 1.634 0.102 -0.034 0.377
	105	+C(city_fe)[T.mascouche] 0.1768 0.042 4.225 0.000 0.095 0.259
	106	+C(city_fe)[T.matane] -0.2714 0.042 -6.488 0.000 -0.353 -0.189
	107	+C(city_fe)[T.mercier / hochelaga / maisonneuve] 0.3175 0.034 9.297 0.000 0.251 0.384
	108	+C(city_fe)[T.messines] -0.1638 0.061 -2.701 0.007 -0.283 -0.045
	109	+C(city_fe)[T.mille-isles] 0.2388 0.055 4.380 0.000 0.132 0.346
	110	+C(city_fe)[T.mirabel] 0.2407 0.027 9.076 0.000 0.189 0.293
	111	+C(city_fe)[T.mont-joli] -0.2476 0.177 -1.397 0.162 -0.595 0.100
	112	+C(city_fe)[T.mont-laurier] -0.0494 0.038 -1.316 0.188 -0.123 0.024
	113	+C(city_fe)[T.mont-royal] 1.0590 0.060 17.641 0.000 0.941 1.177
	114	+C(city_fe)[T.mont-saint-hilaire] 0.3721 0.109 3.421 0.001 0.159 0.585
	115	+C(city_fe)[T.mont-tremblant] 0.3943 0.040 9.855 0.000 0.316 0.473
	116	+C(city_fe)[T.montréal] 0.6721 0.021 31.548 0.000 0.630 0.714
	117	+C(city_fe)[T.morin-heights] 0.2972 0.057 5.204 0.000 0.185 0.409
	118	+C(city_fe)[T.nominingue] 0.1336 0.105 1.276 0.202 -0.072 0.339
	119	+C(city_fe)[T.notre-dame-de-l'île-perrot] 0.2087 0.061 3.406 0.001 0.089 0.329
	120	+C(city_fe)[T.notre-dame-de-pontmain] 0.0094 0.061 0.154 0.878 -0.111 0.130
	121	+C(city_fe)[T.notre-dame-des-bois] -0.2375 0.074 -3.230 0.001 -0.382 -0.093
	122	+C(city_fe)[T.notre-dame-des-prairies] 0.0822 0.118 0.697 0.486 -0.149 0.313
	123	+C(city_fe)[T.notre-dame-du-laus] 0.1009 0.053 1.894 0.058 -0.003 0.205
	124	+C(city_fe)[T.orford] 0.3230 0.063 5.122 0.000 0.199 0.447
	125	+C(city_fe)[T.papineauville] -0.0886 0.124 -0.717 0.473 -0.331 0.153
	126	+C(city_fe)[T.petite-rivière-saint-françois] -0.0343 0.084 -0.407 0.684 -0.200 0.131
	127	+C(city_fe)[T.piedmont] 0.4643 0.067 6.959 0.000 0.334 0.595
	128	+C(city_fe)[T.pincourt] 0.1811 0.055 3.304 0.001 0.074 0.289
	129	+C(city_fe)[T.pointe-claire] 0.4251 0.038 11.133 0.000 0.350 0.500
	130	+C(city_fe)[T.pont-rouge] 0.0628 0.056 1.115 0.265 -0.048 0.173
	131	+C(city_fe)[T.prévost] 0.1183 0.037 3.231 0.001 0.047 0.190
	132	+C(city_fe)[T.québec] 0.1687 0.020 8.570 0.000 0.130 0.207
	133	+C(city_fe)[T.rawdon] 0.0579 0.047 1.244 0.213 -0.033 0.149
	134	+C(city_fe)[T.repentigny] 0.1908 0.032 5.948 0.000 0.128 0.254
	135	+C(city_fe)[T.rigaud] 0.2981 0.064 4.632 0.000 0.172 0.424
	136	+C(city_fe)[T.rimouski] -0.0613 0.034 -1.824 0.068 -0.127 0.005
	137	+C(city_fe)[T.rivière-du-loup] -0.1841 0.094 -1.952 0.051 -0.369 0.001
	138	+C(city_fe)[T.rivière-rouge] -0.0045 0.071 -0.063 0.950 -0.143 0.135
	139	+C(city_fe)[T.roberval] -0.2600 0.081 -3.206 0.001 -0.419 -0.101
	140	+C(city_fe)[T.rosemère] 0.3854 0.089 4.317 0.000 0.210 0.560
	141	+C(city_fe)[T.saguenay] -0.0722 0.044 -1.628 0.103 -0.159 0.015
	142	+C(city_fe)[T.saint-adolphe-d'howard] 0.3369 0.054 6.269 0.000 0.232 0.442
	143	+C(city_fe)[T.saint-amable] 0.3185 0.071 4.461 0.000 0.179 0.458
	144	+C(city_fe)[T.saint-anicet] 0.1207 0.072 1.684 0.092 -0.020 0.261
	145	+C(city_fe)[T.saint-basile-le-grand] 0.3421 0.093 3.686 0.000 0.160 0.524
	146	+C(city_fe)[T.saint-bruno-de-montarville] 0.6408 0.073 8.723 0.000 0.497 0.785
	147	+C(city_fe)[T.saint-calixte] 0.1561 0.090 1.730 0.084 -0.021 0.333
	148	+C(city_fe)[T.saint-charles-borromée] 0.1371 0.163 0.839 0.401 -0.183 0.457
	149	+C(city_fe)[T.saint-colomban] 0.2111 0.070 3.016 0.003 0.074 0.348
	150	+C(city_fe)[T.saint-constant] 0.2154 0.052 4.115 0.000 0.113 0.318
	151	+C(city_fe)[T.saint-donat] 0.2569 0.067 3.862 0.000 0.127 0.387
	152	+C(city_fe)[T.saint-eustache] 0.4482 0.071 6.307 0.000 0.309 0.587
	153	+C(city_fe)[T.saint-félicien] -0.2423 0.336 -0.721 0.471 -0.901 0.416
	154	+C(city_fe)[T.saint-gabriel-de-brandon] -0.1880 0.080 -2.346 0.019 -0.345 -0.031
	155	+C(city_fe)[T.saint-georges] -0.1023 0.063 -1.626 0.104 -0.226 0.021
	156	+C(city_fe)[T.saint-hippolyte] 0.3575 0.082 4.378 0.000 0.197 0.517
	157	+C(city_fe)[T.saint-honoré] -0.1282 0.072 -1.788 0.074 -0.269 0.012
	158	+C(city_fe)[T.saint-hyacinthe] 0.1600 0.061 2.616 0.009 0.040 0.280
	159	+C(city_fe)[T.saint-jean-de-matha] 0.1081 0.147 0.734 0.463 -0.181 0.397
	160	+C(city_fe)[T.saint-jean-sur-richelieu] 0.1894 0.029 6.448 0.000 0.132 0.247
	161	+C(city_fe)[T.saint-jérôme] 0.1558 0.039 4.037 0.000 0.080 0.231
	162	+C(city_fe)[T.saint-lambert] 0.6472 0.071 9.070 0.000 0.507 0.787
	163	+C(city_fe)[T.saint-lambert-de-lauzon] 0.0527 0.081 0.654 0.513 -0.105 0.210
	164	+C(city_fe)[T.saint-lin/laurentides] 0.0940 0.120 0.783 0.433 -0.141 0.329
	165	+C(city_fe)[T.saint-marc-des-carrières] -0.2484 0.047 -5.271 0.000 -0.341 -0.156
	166	+C(city_fe)[T.saint-michel-des-saints] -0.2325 0.080 -2.906 0.004 -0.389 -0.076
	167	+C(city_fe)[T.saint-paul-de-l'île-aux-noix] -0.1222 0.249 -0.491 0.623 -0.610 0.365
	168	+C(city_fe)[T.saint-philippe] 0.3232 0.040 8.134 0.000 0.245 0.401
	169	+C(city_fe)[T.saint-raymond] 0.1551 0.088 1.768 0.077 -0.017 0.327
	170	+C(city_fe)[T.saint-sauveur] 0.2986 0.058 5.155 0.000 0.185 0.412
	171	+C(city_fe)[T.saint-zotique] 0.4566 0.074 6.189 0.000 0.312 0.601
	172	+C(city_fe)[T.saint-zénon] -0.3094 86.426 -0.004 0.997 -169.702 169.083
	173	+C(city_fe)[T.sainte-adèle] 0.3608 0.078 4.614 0.000 0.208 0.514
	174	+C(city_fe)[T.sainte-agathe-des-monts] 0.1222 0.092 1.334 0.182 -0.057 0.302
	175	+C(city_fe)[T.sainte-anne-des-plaines] 0.8940 0.773 1.156 0.248 -0.622 2.409
	176	+C(city_fe)[T.sainte-barbe] 0.1909 0.291 0.655 0.512 -0.380 0.762
	177	+C(city_fe)[T.sainte-brigitte-de-laval] 0.1374 0.051 2.680 0.007 0.037 0.238
	178	+C(city_fe)[T.sainte-béatrix] 0.0338 0.044 0.764 0.445 -0.053 0.121
	179	+C(city_fe)[T.sainte-catherine] 0.3921 0.073 5.382 0.000 0.249 0.535
	180	+C(city_fe)[T.sainte-catherine-de-la-jacques-cartier] 0.2108 0.068 3.084 0.002 0.077 0.345
	181	+C(city_fe)[T.sainte-julie] 0.3731 0.082 4.540 0.000 0.212 0.534
	182	+C(city_fe)[T.sainte-julienne] 0.0598 0.076 0.787 0.431 -0.089 0.209
	183	+C(city_fe)[T.sainte-lucie-des-laurentides] 0.3382 0.073 4.613 0.000 0.195 0.482
	184	+C(city_fe)[T.sainte-marguerite-du-lac-masson] 0.1056 0.051 2.065 0.039 0.005 0.206
	185	+C(city_fe)[T.sainte-marthe-sur-le-lac] 0.2464 0.035 7.007 0.000 0.177 0.315
	186	+C(city_fe)[T.sainte-sophie] 0.0938 0.043 2.179 0.029 0.009 0.178
	187	+C(city_fe)[T.sainte-thérèse] 0.2976 0.087 3.428 0.001 0.127 0.468
	188	+C(city_fe)[T.sainte-émélie-de-l'énergie] -0.1963 0.064 -3.071 0.002 -0.322 -0.071
	189	+C(city_fe)[T.salaberry-de-valleyfield] 0.0587 0.049 1.205 0.228 -0.037 0.154
	190	+C(city_fe)[T.sept-îles] -0.1510 0.085 -1.774 0.076 -0.318 0.016
	191	+C(city_fe)[T.shawinigan] -0.1856 0.053 -3.475 0.001 -0.290 -0.081
	192	+C(city_fe)[T.shefford] 0.1276 0.089 1.437 0.151 -0.046 0.302
	193	+C(city_fe)[T.sherbrooke] 0.0976 0.018 5.321 0.000 0.062 0.133
	194	+C(city_fe)[T.sorel-tracy] -0.0858 0.046 -1.883 0.060 -0.175 0.003
	195	+C(city_fe)[T.st-colomban] 0.2324 0.059 3.962 0.000 0.117 0.347
	196	+C(city_fe)[T.st-hippolyte] 0.1484 0.057 2.614 0.009 0.037 0.260
	197	+C(city_fe)[T.st-jean-sur-richelieu] 0.1674 0.033 5.005 0.000 0.102 0.233
	198	+C(city_fe)[T.st-jérôme] 0.1660 0.032 5.221 0.000 0.104 0.228
	199	+C(city_fe)[T.st-sauveur] 0.2742 0.053 5.170 0.000 0.170 0.378
	200	+C(city_fe)[T.ste-adèle] 0.1521 0.071 2.155 0.031 0.014 0.290
	201	+C(city_fe)[T.ste-julienne] 0.0571 0.043 1.321 0.187 -0.028 0.142
	202	+C(city_fe)[T.sutton] 0.3865 0.068 5.668 0.000 0.253 0.520
	203	+C(city_fe)[T.terrebonne] 0.2097 0.029 7.341 0.000 0.154 0.266
	204	+C(city_fe)[T.thetford mines] -0.3094 0.118 -2.617 0.009 -0.541 -0.078
	205	+C(city_fe)[T.trois-rivières] -0.0763 0.023 -3.372 0.001 -0.121 -0.032
	206	+C(city_fe)[T.très-saint-rédempteur] 0.4880 0.032 15.400 0.000 0.426 0.550
	207	+C(city_fe)[T.témiscouata-sur-le-lac] -0.2810 0.139 -2.018 0.044 -0.554 -0.008
	208	+C(city_fe)[T.val-d'or] 0.0673 0.031 2.182 0.029 0.007 0.128
	209	+C(city_fe)[T.val-david] 0.0033 0.100 0.033 0.973 -0.193 0.199
	210	+C(city_fe)[T.val-des-monts] 0.1145 0.025 4.660 0.000 0.066 0.163
	211	+C(city_fe)[T.val-morin] -0.1213 0.092 -1.311 0.190 -0.302 0.060
	212	+C(city_fe)[T.vaudreuil-dorion] 0.0676 0.039 1.731 0.083 -0.009 0.144
	213	+C(city_fe)[T.venise-en-québec] 0.0774 0.073 1.065 0.287 -0.065 0.220
	214	+C(city_fe)[T.victoriaville] -0.0961 0.031 -3.150 0.002 -0.156 -0.036
	215	+C(city_fe)[T.waterloo] 0.1009 0.051 1.969 0.049 0.000 0.201
	216	+C(city_fe)[T.weedon] -0.1440 0.081 -1.780 0.075 -0.303 0.015
	217	+C(city_fe)[T.wentworth-nord] 0.2091 0.088 2.378 0.017 0.037 0.381
	218	+C(city_fe)[T.westmount] 1.3088 0.042 31.188 0.000 1.227 1.391
	219	+ln_sqft 0.4171 0.013 31.566 0.000 0.391 0.443
	220	+ln_lot 0.0257 0.004 6.727 0.000 0.018 0.033
	221	+has_lot -0.0242 0.014 -1.683 0.092 -0.052 0.004
	222	+beds 0.0022 0.004 0.513 0.608 -0.006 0.011
	223	+baths 0.0897 0.007 13.613 0.000 0.077 0.103
	224	+parking 0.0109 0.002 7.110 0.000 0.008 0.014
	225	+storeys -0.0023 0.007 -0.312 0.755 -0.017 0.012
	226	+age -0.0060 0.000 -12.824 0.000 -0.007 -0.005
	227	+age2 2.97e-05 3.88e-06 7.650 0.000 2.21e-05 3.73e-05
	228	+has_age -0.1804 0.021 -8.685 0.000 -0.221 -0.140
	229	+pool 0.0341 0.008 4.444 0.000 0.019 0.049
	230	+waterfront 0.0623 0.021 2.965 0.003 0.021 0.104
	231	+garage 0.0279 0.007 3.953 0.000 0.014 0.042
	232	+fireplace 0.0172 0.008 2.130 0.033 0.001 0.033
	233	+central_air 0.1016 0.011 9.291 0.000 0.080 0.123
	234	+finished_bsmt -0.0115 0.010 -1.191 0.234 -0.030 0.007
	235	+near_water 0.0972 0.023 4.149 0.000 0.051 0.143
	236	+cul_de_sac 0.0279 0.012 2.409 0.016 0.005 0.051
	237	+golf 0.0225 0.007 3.138 0.002 0.008 0.037
	238	+ski 0.0218 0.013 1.649 0.099 -0.004 0.048
	239	+near_park 0.0037 0.011 0.328 0.743 -0.019 0.026
	240	+near_school -0.0695 0.011 -6.157 0.000 -0.092 -0.047
	241	+near_highway 0.0514 0.008 6.362 0.000 0.036 0.067
	242	+transit 0.1011 0.009 11.372 0.000 0.084 0.119
	243	+no_rear_neighbor -0.0377 0.010 -3.802 0.000 -0.057 -0.018
	244	+n_images_c 0.0045 0.000 14.507 0.000 0.004 0.005
	245	+virtual_tour 0.0444 0.022 1.998 0.046 0.001 0.088
	246	+n_rooms 0.0172 0.002 9.260 0.000 0.014 0.021
	247	+n_levels_r -0.0248 0.009 -2.825 0.005 -0.042 -0.008
	248	+has_primary_bed 0.0775 0.016 4.958 0.000 0.047 0.108
	249	+has_basement -0.0251 0.015 -1.714 0.086 -0.054 0.004
	250	+r_office 0.0300 0.011 2.631 0.009 0.008 0.052
	251	+r_family 0.0422 0.011 3.973 0.000 0.021 0.063
	252	+r_workshop -0.0401 0.014 -2.897 0.004 -0.067 -0.013
	253	+r_winecellar 0.1086 0.036 3.034 0.002 0.038 0.179
	254	+r_sunroom -0.0042 0.021 -0.201 0.840 -0.046 0.037
	255	+r_mezzanine 0.0136 0.026 0.527 0.598 -0.037 0.064
	256	+r_library 0.0945 0.084 1.123 0.262 -0.070 0.260
	257	+r_laundry 0.0202 0.009 2.222 0.026 0.002 0.038
	258	+==============================================================================
	259	+Omnibus: 403.871 Durbin-Watson: 1.849
	260	+Prob(Omnibus): 0.000 Jarque-Bera (JB): 1350.435
	261	+Skew: 0.024 Prob(JB): 5.71e-294
	262	+Kurtosis: 4.852 Cond. No. 5.36e+16
	263	+==============================================================================
	264	+
	265	+Notes:
	266	+[1] Standard Errors are heteroscedasticity robust (HC3)
	267	+[2] The smallest eigenvalue is 5.8e-23. This might indicate that there are
	268	+strong multicollinearity problems or that the design matrix is singular.
	269	+
	270	+=== MODELE B (sans superficie) ===
	271	+ OLS Regression Results
	272	+==============================================================================
	273	+Dep. Variable: ln_price R-squared: 0.677
	274	+Model: OLS Adj. R-squared: 0.673
	275	+Method: Least Squares F-statistic: 149.6
	276	+Date: Fri, 03 Jul 2026 Prob (F-statistic): 0.00
	277	+Time: 07:41:43 Log-Likelihood: -6293.3
	278	+No. Observations: 19115 AIC: 1.307e+04
	279	+Df Residuals: 18875 BIC: 1.495e+04
	280	+Df Model: 239
	281	+Covariance Type: HC3
	282	+========================================================================================================================
	283	+ coef std err z P>\|z\| [0.025 0.975]
	284	+------------------------------------------------------------------------------------------------------------------------
	285	+Intercept 12.4073 0.115 108.355 0.000 12.183 12.632
	286	+C(source)[T.realtor] -0.3680 0.021 -17.300 0.000 -0.410 -0.326
	287	+C(ptype_fe)[T.Country Home] 0.0513 0.067 0.765 0.444 -0.080 0.183
	288	+C(ptype_fe)[T.Home] 0.1075 0.083 1.297 0.195 -0.055 0.270
	289	+C(ptype_fe)[T.House] -0.0659 0.014 -4.694 0.000 -0.093 -0.038
	290	+C(ptype_fe)[T.Intergenerational] 0.1281 0.062 2.051 0.040 0.006 0.250
	291	+C(ptype_fe)[T.Other] 0.2713 0.121 2.243 0.025 0.034 0.508
	292	+C(ptype_fe)[T.Row / Townhouse] -0.3021 0.014 -21.062 0.000 -0.330 -0.274
	293	+C(ptype_fe)[T.Semi Detached] -0.0065 0.057 -0.115 0.909 -0.118 0.105
	294	+C(ptype_fe)[T.Townhouse] 0.0963 0.100 0.965 0.335 -0.099 0.292
	295	+C(city_fe)[T.acton vale] -0.0167 0.033 -0.499 0.618 -0.082 0.049
	296	+C(city_fe)[T.alma] -0.0480 0.060 -0.796 0.426 -0.166 0.070
	297	+C(city_fe)[T.amherst] -0.0109 0.050 -0.217 0.828 -0.110 0.088
	298	+C(city_fe)[T.amos] -0.1497 0.046 -3.239 0.001 -0.240 -0.059
	299	+C(city_fe)[T.austin] 0.4089 0.071 5.776 0.000 0.270 0.548
	300	+C(city_fe)[T.baie-comeau] -0.3409 0.071 -4.806 0.000 -0.480 -0.202
	301	+C(city_fe)[T.baie-saint-paul] 0.0980 0.074 1.327 0.184 -0.047 0.243
	302	+C(city_fe)[T.beaconsfield] 0.7818 0.039 20.221 0.000 0.706 0.858
	303	+C(city_fe)[T.beauport] 0.0722 0.041 1.775 0.076 -0.008 0.152
	304	+C(city_fe)[T.beloeil] 0.3670 0.049 7.506 0.000 0.271 0.463
	305	+C(city_fe)[T.blainville] 0.4382 0.025 17.427 0.000 0.389 0.487
	306	+C(city_fe)[T.blue sea] -0.1010 0.064 -1.585 0.113 -0.226 0.024
	307	+C(city_fe)[T.bois-des-filion] 0.2889 0.048 5.999 0.000 0.194 0.383
	308	+C(city_fe)[T.boisbriand] 0.3584 0.038 9.505 0.000 0.284 0.432
	309	+C(city_fe)[T.boischatel] 0.4149 0.076 5.483 0.000 0.267 0.563
	310	+C(city_fe)[T.boucherville] 0.4653 0.037 12.616 0.000 0.393 0.538
	311	+C(city_fe)[T.bromont] 0.4068 0.034 11.856 0.000 0.340 0.474
	312	+C(city_fe)[T.brossard] 0.5145 0.020 25.986 0.000 0.476 0.553
	313	+C(city_fe)[T.bécancour] -0.0914 0.033 -2.788 0.005 -0.156 -0.027
	314	+C(city_fe)[T.cantley] 0.1594 0.048 3.356 0.001 0.066 0.253
	315	+C(city_fe)[T.carignan] 0.3871 0.041 9.472 0.000 0.307 0.467
	316	+C(city_fe)[T.chandler] -0.3044 0.081 -3.746 0.000 -0.464 -0.145
	317	+C(city_fe)[T.chateauguay] 0.2448 0.040 6.056 0.000 0.166 0.324
	318	+C(city_fe)[T.chelsea] 0.2986 0.040 7.393 0.000 0.219 0.378
	319	+C(city_fe)[T.chertsey] -0.0194 0.038 -0.508 0.612 -0.094 0.056
	320	+C(city_fe)[T.chicoutimi] -0.0770 0.041 -1.880 0.060 -0.157 0.003
	321	+C(city_fe)[T.chomedey] 0.4087 0.056 7.260 0.000 0.298 0.519
	322	+C(city_fe)[T.châteauguay] 0.2797 0.023 12.337 0.000 0.235 0.324
	323	+C(city_fe)[T.contrecoeur] 0.1257 0.034 3.670 0.000 0.059 0.193
	324	+C(city_fe)[T.cowansville] 0.1495 0.036 4.133 0.000 0.079 0.220
	325	+C(city_fe)[T.crabtree] 0.0639 0.046 1.394 0.163 -0.026 0.154
	326	+C(city_fe)[T.côte-saint-luc] 0.7419 0.052 14.316 0.000 0.640 0.843
	327	+C(city_fe)[T.dolbeau-mistassini] -0.1920 0.066 -2.895 0.004 -0.322 -0.062
	328	+C(city_fe)[T.dollard-des-ormeaux] 0.5263 0.029 18.095 0.000 0.469 0.583
	329	+C(city_fe)[T.donnacona] 0.1454 0.035 4.112 0.000 0.076 0.215
	330	+C(city_fe)[T.dorval] 0.6752 0.050 13.510 0.000 0.577 0.773
	331	+C(city_fe)[T.drummondville] 0.0561 0.019 2.952 0.003 0.019 0.093
	332	+C(city_fe)[T.duhamel] 0.0808 0.077 1.049 0.294 -0.070 0.232
	333	+C(city_fe)[T.dunham] 0.3488 0.063 5.542 0.000 0.225 0.472
	334	+C(city_fe)[T.eastman] 0.3943 0.055 7.157 0.000 0.286 0.502
	335	+C(city_fe)[T.entrelacs] 0.2226 0.086 2.601 0.009 0.055 0.390
	336	+C(city_fe)[T.farnham] 0.0560 0.038 1.463 0.143 -0.019 0.131
	337	+C(city_fe)[T.fossambault-sur-le-lac] 0.3607 0.067 5.395 0.000 0.230 0.492
	338	+C(city_fe)[T.gaspé] -0.3110 0.062 -4.989 0.000 -0.433 -0.189
	339	+C(city_fe)[T.gatineau] 0.1163 0.013 8.687 0.000 0.090 0.143
	340	+C(city_fe)[T.gore] 0.0808 0.058 1.403 0.161 -0.032 0.194
	341	+C(city_fe)[T.granby] 0.1488 0.019 7.962 0.000 0.112 0.185
	342	+C(city_fe)[T.grenville-sur-la-rouge] 0.1281 0.046 2.801 0.005 0.038 0.218
	343	+C(city_fe)[T.hampstead] 1.2858 0.051 25.203 0.000 1.186 1.386
	344	+C(city_fe)[T.harrington] -0.0622 0.130 -0.478 0.633 -0.318 0.193
	345	+C(city_fe)[T.hudson] 0.4820 0.071 6.816 0.000 0.343 0.621
	346	+C(city_fe)[T.joliette] 0.1230 0.046 2.684 0.007 0.033 0.213
	347	+C(city_fe)[T.jonquière] -0.1590 0.046 -3.470 0.001 -0.249 -0.069
	348	+C(city_fe)[T.kirkland] 0.5347 0.035 15.112 0.000 0.465 0.604
	349	+C(city_fe)[T.l'ange-gardien] 0.2734 0.073 3.767 0.000 0.131 0.416
	350	+C(city_fe)[T.l'assomption] 0.1985 0.027 7.277 0.000 0.145 0.252
	351	+C(city_fe)[T.la conception] 0.2903 0.052 5.600 0.000 0.189 0.392
	352	+C(city_fe)[T.la malbaie] -0.1641 0.055 -2.997 0.003 -0.271 -0.057
	353	+C(city_fe)[T.la minerve] 0.0741 0.054 1.378 0.168 -0.031 0.180
	354	+C(city_fe)[T.la prairie] 0.4728 0.030 15.749 0.000 0.414 0.532
	355	+C(city_fe)[T.la pêche] 0.0037 0.039 0.094 0.925 -0.073 0.081
	356	+C(city_fe)[T.la tuque] -0.3610 0.049 -7.307 0.000 -0.458 -0.264
	357	+C(city_fe)[T.labelle] 0.0955 0.057 1.687 0.092 -0.015 0.206
	358	+C(city_fe)[T.lac-beauport] 0.4166 0.089 4.659 0.000 0.241 0.592
	359	+C(city_fe)[T.lac-brome] 0.6311 0.057 11.018 0.000 0.519 0.743
	360	+C(city_fe)[T.lac-des-écorces] 0.0133 0.068 0.196 0.845 -0.120 0.146
	361	+C(city_fe)[T.lac-etchemin] -0.2591 0.057 -4.521 0.000 -0.371 -0.147
	362	+C(city_fe)[T.lachute] 0.1603 0.038 4.268 0.000 0.087 0.234
	363	+C(city_fe)[T.lanoraie] 0.1578 0.038 4.195 0.000 0.084 0.232
	364	+C(city_fe)[T.laval] 0.4618 0.014 33.004 0.000 0.434 0.489
	365	+C(city_fe)[T.lavaltrie] 0.1654 0.039 4.251 0.000 0.089 0.242
	366	+C(city_fe)[T.les éboulements] 0.1792 0.064 2.798 0.005 0.054 0.305
	367	+C(city_fe)[T.longueuil] 0.3736 0.016 23.052 0.000 0.342 0.405
	368	+C(city_fe)[T.lorraine] 0.4812 0.056 8.517 0.000 0.370 0.592
	369	+C(city_fe)[T.lévis] 0.1144 0.022 5.189 0.000 0.071 0.158
	370	+C(city_fe)[T.magog] 0.3435 0.032 10.704 0.000 0.281 0.406
	371	+C(city_fe)[T.mandeville] -0.0847 0.041 -2.084 0.037 -0.164 -0.005
	372	+C(city_fe)[T.mansfield-et-pontefract] -0.5178 0.067 -7.737 0.000 -0.649 -0.387
	373	+C(city_fe)[T.marieville] 0.2018 0.034 5.959 0.000 0.135 0.268
	374	+C(city_fe)[T.mascouche] 0.2837 0.021 13.524 0.000 0.243 0.325
	375	+C(city_fe)[T.matane] -0.1801 0.047 -3.836 0.000 -0.272 -0.088
	376	+C(city_fe)[T.mercier / hochelaga / maisonneuve] 0.3725 0.047 7.859 0.000 0.280 0.465
	377	+C(city_fe)[T.messines] -0.2559 0.053 -4.805 0.000 -0.360 -0.152
	378	+C(city_fe)[T.mille-isles] 0.1934 0.052 3.750 0.000 0.092 0.294
	379	+C(city_fe)[T.mirabel] 0.2915 0.020 14.237 0.000 0.251 0.332
	380	+C(city_fe)[T.mont-joli] -0.2792 0.099 -2.832 0.005 -0.472 -0.086
	381	+C(city_fe)[T.mont-laurier] -0.0419 0.040 -1.042 0.298 -0.121 0.037
	382	+C(city_fe)[T.mont-royal] 1.2687 0.054 23.406 0.000 1.162 1.375
	383	+C(city_fe)[T.mont-saint-hilaire] 0.5053 0.041 12.197 0.000 0.424 0.586
	384	+C(city_fe)[T.mont-tremblant] 0.3895 0.036 10.890 0.000 0.319 0.460
	385	+C(city_fe)[T.montréal] 0.7168 0.017 42.732 0.000 0.684 0.750
	386	+C(city_fe)[T.morin-heights] 0.2959 0.038 7.784 0.000 0.221 0.370
	387	+C(city_fe)[T.nominingue] 0.0840 0.085 0.990 0.322 -0.082 0.250
	388	+C(city_fe)[T.notre-dame-de-l'île-perrot] 0.4132 0.043 9.698 0.000 0.330 0.497
	389	+C(city_fe)[T.notre-dame-de-pontmain] -0.1086 0.060 -1.801 0.072 -0.227 0.010
	390	+C(city_fe)[T.notre-dame-des-bois] -0.3513 0.066 -5.303 0.000 -0.481 -0.221
	391	+C(city_fe)[T.notre-dame-des-prairies] 0.0815 0.039 2.105 0.035 0.006 0.157
	392	+C(city_fe)[T.notre-dame-du-laus] 0.0340 0.057 0.597 0.550 -0.078 0.146
	393	+C(city_fe)[T.orford] 0.3442 0.065 5.321 0.000 0.217 0.471
	394	+C(city_fe)[T.papineauville] -0.0271 0.113 -0.241 0.810 -0.248 0.194
	395	+C(city_fe)[T.petite-rivière-saint-françois] -0.0295 0.054 -0.544 0.587 -0.136 0.077
	396	+C(city_fe)[T.piedmont] 0.3505 0.055 6.383 0.000 0.243 0.458
	397	+C(city_fe)[T.pincourt] 0.3096 0.038 8.225 0.000 0.236 0.383
	398	+C(city_fe)[T.pointe-claire] 0.6052 0.042 14.568 0.000 0.524 0.687
	399	+C(city_fe)[T.pont-rouge] -0.0087 0.045 -0.196 0.845 -0.096 0.079
	400	+C(city_fe)[T.prévost] 0.1693 0.027 6.376 0.000 0.117 0.221
	401	+C(city_fe)[T.québec] 0.1918 0.018 10.706 0.000 0.157 0.227
	402	+C(city_fe)[T.rawdon] 0.0432 0.031 1.373 0.170 -0.018 0.105
	403	+C(city_fe)[T.repentigny] 0.2854 0.019 14.912 0.000 0.248 0.323
	404	+C(city_fe)[T.rigaud] 0.1455 0.046 3.193 0.001 0.056 0.235
	405	+C(city_fe)[T.rimouski] -0.0705 0.033 -2.115 0.034 -0.136 -0.005
	406	+C(city_fe)[T.rivière-du-loup] -0.0883 0.056 -1.585 0.113 -0.197 0.021
	407	+C(city_fe)[T.rivière-rouge] -0.1303 0.058 -2.240 0.025 -0.244 -0.016
	408	+C(city_fe)[T.roberval] -0.1680 0.067 -2.506 0.012 -0.299 -0.037
	409	+C(city_fe)[T.rosemère] 0.4821 0.039 12.318 0.000 0.405 0.559
	410	+C(city_fe)[T.saguenay] -0.0384 0.023 -1.664 0.096 -0.084 0.007
	411	+C(city_fe)[T.saint-adolphe-d'howard] 0.1647 0.045 3.665 0.000 0.077 0.253
	412	+C(city_fe)[T.saint-amable] 0.2714 0.033 8.280 0.000 0.207 0.336
	413	+C(city_fe)[T.saint-anicet] 0.2760 0.070 3.950 0.000 0.139 0.413
	414	+C(city_fe)[T.saint-basile-le-grand] 0.3655 0.044 8.222 0.000 0.278 0.453
	415	+C(city_fe)[T.saint-bruno-de-montarville] 0.6120 0.039 15.504 0.000 0.535 0.689
	416	+C(city_fe)[T.saint-calixte] 0.0123 0.033 0.367 0.714 -0.053 0.078
	417	+C(city_fe)[T.saint-charles-borromée] 0.1447 0.050 2.874 0.004 0.046 0.243
	418	+C(city_fe)[T.saint-colomban] 0.1238 0.042 2.966 0.003 0.042 0.206
	419	+C(city_fe)[T.saint-constant] 0.3327 0.030 11.057 0.000 0.274 0.392
	420	+C(city_fe)[T.saint-donat] 0.3373 0.047 7.155 0.000 0.245 0.430
	421	+C(city_fe)[T.saint-eustache] 0.3390 0.036 9.470 0.000 0.269 0.409
	422	+C(city_fe)[T.saint-félicien] -0.0495 0.059 -0.832 0.405 -0.166 0.067
	423	+C(city_fe)[T.saint-gabriel-de-brandon] 0.0228 0.061 0.371 0.711 -0.098 0.143
	424	+C(city_fe)[T.saint-georges] -0.1857 0.038 -4.869 0.000 -0.260 -0.111
	425	+C(city_fe)[T.saint-hippolyte] 0.2228 0.041 5.377 0.000 0.142 0.304
	426	+C(city_fe)[T.saint-honoré] -0.0445 0.061 -0.728 0.466 -0.164 0.075
	427	+C(city_fe)[T.saint-hyacinthe] 0.1802 0.034 5.373 0.000 0.114 0.246
	428	+C(city_fe)[T.saint-jean-de-matha] 0.0361 0.068 0.530 0.596 -0.097 0.170
	429	+C(city_fe)[T.saint-jean-sur-richelieu] 0.2700 0.020 13.718 0.000 0.231 0.309
	430	+C(city_fe)[T.saint-jérôme] 0.1966 0.020 9.627 0.000 0.157 0.237
	431	+C(city_fe)[T.saint-lambert] 0.8612 0.048 17.881 0.000 0.767 0.956
	432	+C(city_fe)[T.saint-lambert-de-lauzon] 0.0811 0.049 1.641 0.101 -0.016 0.178
	433	+C(city_fe)[T.saint-lin/laurentides] 0.0843 0.030 2.770 0.006 0.025 0.144
	434	+C(city_fe)[T.saint-marc-des-carrières] -0.1463 0.030 -4.804 0.000 -0.206 -0.087
	435	+C(city_fe)[T.saint-michel-des-saints] -0.0296 0.067 -0.444 0.657 -0.160 0.101
	436	+C(city_fe)[T.saint-paul-de-l'île-aux-noix] 0.0415 0.102 0.407 0.684 -0.159 0.242
	437	+C(city_fe)[T.saint-philippe] 0.4171 0.027 15.486 0.000 0.364 0.470
	438	+C(city_fe)[T.saint-raymond] 0.0777 0.068 1.140 0.254 -0.056 0.211
	439	+C(city_fe)[T.saint-sauveur] 0.2484 0.034 7.368 0.000 0.182 0.314
	440	+C(city_fe)[T.saint-zotique] 0.2794 0.039 7.099 0.000 0.202 0.357
	441	+C(city_fe)[T.saint-zénon] -0.0054 0.107 -0.050 0.960 -0.215 0.204
	442	+C(city_fe)[T.sainte-adèle] 0.2888 0.049 5.952 0.000 0.194 0.384
	443	+C(city_fe)[T.sainte-agathe-des-monts] 0.1307 0.052 2.523 0.012 0.029 0.232
	444	+C(city_fe)[T.sainte-anne-des-plaines] 0.2630 0.059 4.451 0.000 0.147 0.379
	445	+C(city_fe)[T.sainte-barbe] 0.4472 0.064 6.992 0.000 0.322 0.573
	446	+C(city_fe)[T.sainte-brigitte-de-laval] 0.1144 0.040 2.844 0.004 0.036 0.193
	447	+C(city_fe)[T.sainte-béatrix] -0.0542 0.044 -1.231 0.218 -0.141 0.032
	448	+C(city_fe)[T.sainte-catherine] 0.3326 0.033 10.090 0.000 0.268 0.397
	449	+C(city_fe)[T.sainte-catherine-de-la-jacques-cartier] 0.1289 0.061 2.125 0.034 0.010 0.248
	450	+C(city_fe)[T.sainte-julie] 0.4505 0.054 8.316 0.000 0.344 0.557
	451	+C(city_fe)[T.sainte-julienne] 0.0339 0.035 0.981 0.327 -0.034 0.102
	452	+C(city_fe)[T.sainte-lucie-des-laurentides] 0.1581 0.039 4.057 0.000 0.082 0.234
	453	+C(city_fe)[T.sainte-marguerite-du-lac-masson] 0.0159 0.053 0.298 0.766 -0.089 0.121
	454	+C(city_fe)[T.sainte-marthe-sur-le-lac] 0.3085 0.023 13.594 0.000 0.264 0.353
	455	+C(city_fe)[T.sainte-sophie] 0.0743 0.038 1.969 0.049 0.000 0.148
	456	+C(city_fe)[T.sainte-thérèse] 0.3650 0.041 8.871 0.000 0.284 0.446
	457	+C(city_fe)[T.sainte-émélie-de-l'énergie] -0.2224 0.060 -3.680 0.000 -0.341 -0.104
	458	+C(city_fe)[T.salaberry-de-valleyfield] 0.0989 0.045 2.196 0.028 0.011 0.187
	459	+C(city_fe)[T.sept-îles] -0.1031 0.037 -2.808 0.005 -0.175 -0.031
	460	+C(city_fe)[T.shawinigan] -0.1030 0.026 -4.024 0.000 -0.153 -0.053
	461	+C(city_fe)[T.shefford] 0.2575 0.059 4.350 0.000 0.141 0.373
	462	+C(city_fe)[T.sherbrooke] 0.0970 0.017 5.655 0.000 0.063 0.131
	463	+C(city_fe)[T.sorel-tracy] 0.0076 0.027 0.280 0.780 -0.045 0.061
	464	+C(city_fe)[T.st-colomban] 0.2945 0.071 4.138 0.000 0.155 0.434
	465	+C(city_fe)[T.st-hippolyte] 0.1163 0.079 1.468 0.142 -0.039 0.272
	466	+C(city_fe)[T.st-jean-sur-richelieu] 0.2196 0.035 6.245 0.000 0.151 0.289
	467	+C(city_fe)[T.st-jérôme] 0.2480 0.039 6.409 0.000 0.172 0.324
	468	+C(city_fe)[T.st-sauveur] 0.4105 0.091 4.505 0.000 0.232 0.589
	469	+C(city_fe)[T.ste-adèle] 0.2046 0.066 3.088 0.002 0.075 0.335
	470	+C(city_fe)[T.ste-julienne] 0.0905 0.048 1.903 0.057 -0.003 0.184
	471	+C(city_fe)[T.sutton] 0.4707 0.053 8.949 0.000 0.368 0.574
	472	+C(city_fe)[T.terrebonne] 0.2919 0.018 16.339 0.000 0.257 0.327
	473	+C(city_fe)[T.thetford mines] -0.2933 0.058 -5.069 0.000 -0.407 -0.180
	474	+C(city_fe)[T.trois-rivières] 0.0072 0.020 0.357 0.721 -0.032 0.047
	475	+C(city_fe)[T.très-saint-rédempteur] 0.5979 0.049 12.126 0.000 0.501 0.695
	476	+C(city_fe)[T.témiscouata-sur-le-lac] -0.1648 0.074 -2.223 0.026 -0.310 -0.020
	477	+C(city_fe)[T.val-d'or] 0.0067 0.030 0.226 0.821 -0.051 0.065
	478	+C(city_fe)[T.val-david] 0.0280 0.051 0.551 0.582 -0.072 0.127
	479	+C(city_fe)[T.val-des-monts] 0.0953 0.027 3.504 0.000 0.042 0.149
	480	+C(city_fe)[T.val-morin] 0.0441 0.060 0.734 0.463 -0.074 0.162
	481	+C(city_fe)[T.vaudreuil-dorion] 0.2448 0.026 9.301 0.000 0.193 0.296
	482	+C(city_fe)[T.venise-en-québec] 0.1940 0.041 4.773 0.000 0.114 0.274
	483	+C(city_fe)[T.victoriaville] -0.0391 0.025 -1.593 0.111 -0.087 0.009
	484	+C(city_fe)[T.waterloo] 0.1392 0.045 3.091 0.002 0.051 0.227
	485	+C(city_fe)[T.weedon] -0.1707 0.060 -2.823 0.005 -0.289 -0.052
	486	+C(city_fe)[T.wentworth-nord] 0.0653 0.073 0.892 0.372 -0.078 0.209
	487	+C(city_fe)[T.westmount] 1.6044 0.038 41.845 0.000 1.529 1.680
	488	+ln_lot 0.0424 0.003 13.299 0.000 0.036 0.049
	489	+has_lot -0.0854 0.014 -5.897 0.000 -0.114 -0.057
	490	+beds 0.0237 0.004 5.676 0.000 0.016 0.032
	491	+baths 0.1352 0.005 25.581 0.000 0.125 0.146
	492	+parking 0.0154 0.002 8.288 0.000 0.012 0.019
	493	+storeys 0.0020 0.097 0.021 0.983 -0.188 0.192
	494	+age -0.0071 0.000 -18.661 0.000 -0.008 -0.006
	495	+age2 3.559e-05 2.82e-06 12.637 0.000 3.01e-05 4.11e-05
	496	+has_age -0.1532 0.018 -8.628 0.000 -0.188 -0.118
	497	+pool 0.0548 0.007 8.093 0.000 0.042 0.068
	498	+waterfront 0.1091 0.018 6.132 0.000 0.074 0.144
	499	+garage 0.0533 0.006 9.016 0.000 0.042 0.065
	500	+fireplace 0.0116 0.007 1.768 0.077 -0.001 0.024
	501	+central_air 0.0651 0.009 7.208 0.000 0.047 0.083
	502	+finished_bsmt -0.0169 0.007 -2.438 0.015 -0.031 -0.003
	503	+near_water 0.0769 0.021 3.748 0.000 0.037 0.117
	504	+cul_de_sac 0.0316 0.009 3.509 0.000 0.014 0.049
	505	+golf 0.0197 0.006 3.395 0.001 0.008 0.031
	506	+ski 0.0468 0.010 4.760 0.000 0.028 0.066
	507	+near_park 0.0043 0.008 0.544 0.587 -0.011 0.020
	508	+near_school -0.0571 0.009 -6.578 0.000 -0.074 -0.040
	509	+near_highway 0.0499 0.006 7.721 0.000 0.037 0.063
	510	+transit 0.0667 0.008 8.828 0.000 0.052 0.082
	511	+no_rear_neighbor -0.0478 0.009 -5.132 0.000 -0.066 -0.030
	512	+n_images_c 0.0059 0.000 19.406 0.000 0.005 0.006
	513	+virtual_tour 0.0277 0.016 1.751 0.080 -0.003 0.059
	514	+n_rooms 0.0262 0.001 17.890 0.000 0.023 0.029
	515	+n_levels_r 0.0432 0.048 0.897 0.370 -0.051 0.138
	516	+has_primary_bed 0.0696 0.010 7.033 0.000 0.050 0.089
	517	+has_basement -0.1168 0.044 -2.657 0.008 -0.203 -0.031
	518	+r_office 0.0348 0.008 4.562 0.000 0.020 0.050
	519	+r_family 0.0372 0.009 4.122 0.000 0.020 0.055
	520	+r_workshop -0.0346 0.009 -3.928 0.000 -0.052 -0.017
	521	+r_winecellar 0.1413 0.026 5.384 0.000 0.090 0.193
	522	+r_sunroom 0.0166 0.014 1.224 0.221 -0.010 0.043
	523	+r_mezzanine 0.0202 0.018 1.114 0.265 -0.015 0.056
	524	+r_library 0.1466 0.070 2.082 0.037 0.009 0.285
	525	+r_laundry 0.0382 0.006 6.301 0.000 0.026 0.050
	526	+==============================================================================
	527	+Omnibus: 2418.650 Durbin-Watson: 1.758
	528	+Prob(Omnibus): 0.000 Jarque-Bera (JB): 15665.854
	529	+Skew: 0.429 Prob(JB): 0.00
	530	+Kurtosis: 7.351 Cond. No. 2.81e+16
	531	+==============================================================================
	532	+
	533	+Notes:
	534	+[1] Standard Errors are heteroscedasticity robust (HC3)
	535	+[2] The smallest eigenvalue is 5.65e-22. This might indicate that there are
	536	+strong multicollinearity problems or that the design matrix is singular.
		\ No newline at end of file

added results/hedonic_v4_summary.txt +576 −0

		`@@ -0,0 +1,576 @@`
	1	+=== A+GPT ===
	2	+ OLS Regression Results
	3	+==============================================================================
	4	+Dep. Variable: ln_price R-squared: 0.725
	5	+Model: OLS Adj. R-squared: 0.717
	6	+Method: Least Squares F-statistic: 6.503e+04
	7	+Date: Fri, 03 Jul 2026 Prob (F-statistic): 0.00
	8	+Time: 08:11:28 Log-Likelihood: -1724.8
	9	+No. Observations: 9446 AIC: 3970.
	10	+Df Residuals: 9186 BIC: 5829.
	11	+Df Model: 259
	12	+Covariance Type: HC3
	13	+========================================================================================================================
	14	+ coef std err z P>\|z\| [0.025 0.975]
	15	+------------------------------------------------------------------------------------------------------------------------
	16	+Intercept 4.8574 0.047 102.393 0.000 4.764 4.950
	17	+C(source)[T.realtor] -0.2485 0.023 -10.988 0.000 -0.293 -0.204
	18	+C(ptype_fe)[T.Country Home] 0.0088 0.051 0.174 0.862 -0.090 0.108
	19	+C(ptype_fe)[T.Home] -0.0125 0.014 -0.905 0.365 -0.040 0.015
	20	+C(ptype_fe)[T.House] -0.0462 0.012 -3.711 0.000 -0.071 -0.022
	21	+C(ptype_fe)[T.Intergenerational] 0.0010 0.038 0.026 0.979 -0.073 0.075
	22	+C(ptype_fe)[T.Other] 0.2105 0.120 1.749 0.080 -0.025 0.446
	23	+C(ptype_fe)[T.Row / Townhouse] -0.2023 0.016 -12.633 0.000 -0.234 -0.171
	24	+C(ptype_fe)[T.Semi Detached] -0.0788 0.016 -5.067 0.000 -0.109 -0.048
	25	+C(ptype_fe)[T.Townhouse] -0.0203 0.026 -0.770 0.442 -0.072 0.031
	26	+C(city_fe)[T.acton vale] -0.0447 0.060 -0.747 0.455 -0.162 0.072
	27	+C(city_fe)[T.alma] -0.0355 0.050 -0.704 0.481 -0.134 0.063
	28	+C(city_fe)[T.amherst] -0.0742 0.075 -0.983 0.325 -0.222 0.074
	29	+C(city_fe)[T.amos] -0.0204 0.062 -0.327 0.744 -0.143 0.102
	30	+C(city_fe)[T.austin] 0.2867 0.088 3.250 0.001 0.114 0.460
	31	+C(city_fe)[T.baie-comeau] -0.3689 0.079 -4.688 0.000 -0.523 -0.215
	32	+C(city_fe)[T.baie-saint-paul] -0.0794 0.062 -1.280 0.200 -0.201 0.042
	33	+C(city_fe)[T.beaconsfield] 0.6183 0.050 12.411 0.000 0.521 0.716
	34	+C(city_fe)[T.beauport] 0.0450 0.035 1.289 0.198 -0.023 0.113
	35	+C(city_fe)[T.beloeil] 0.2256 0.052 4.347 0.000 0.124 0.327
	36	+C(city_fe)[T.blainville] 0.3035 0.031 9.859 0.000 0.243 0.364
	37	+C(city_fe)[T.blue sea] -0.0483 0.058 -0.837 0.403 -0.161 0.065
	38	+C(city_fe)[T.bois-des-filion] 0.1607 0.035 4.566 0.000 0.092 0.230
	39	+C(city_fe)[T.boisbriand] 0.2550 0.043 5.892 0.000 0.170 0.340
	40	+C(city_fe)[T.boischatel] 0.1434 0.074 1.929 0.054 -0.002 0.289
	41	+C(city_fe)[T.boucherville] 0.3369 0.039 8.695 0.000 0.261 0.413
	42	+C(city_fe)[T.bromont] 0.3139 0.035 8.946 0.000 0.245 0.383
	43	+C(city_fe)[T.brossard] 0.4180 0.030 13.774 0.000 0.359 0.477
	44	+C(city_fe)[T.bécancour] -0.1107 0.039 -2.846 0.004 -0.187 -0.034
	45	+C(city_fe)[T.cantley] 0.1315 0.033 4.040 0.000 0.068 0.195
	46	+C(city_fe)[T.carignan] 0.2473 0.091 2.724 0.006 0.069 0.425
	47	+C(city_fe)[T.chandler] -0.2870 0.136 -2.106 0.035 -0.554 -0.020
	48	+C(city_fe)[T.chateauguay] 0.2320 0.037 6.320 0.000 0.160 0.304
	49	+C(city_fe)[T.chelsea] 0.2610 0.038 6.931 0.000 0.187 0.335
	50	+C(city_fe)[T.chertsey] -0.0870 0.060 -1.458 0.145 -0.204 0.030
	51	+C(city_fe)[T.chicoutimi] -0.1131 0.042 -2.703 0.007 -0.195 -0.031
	52	+C(city_fe)[T.chomedey] 0.3214 0.048 6.746 0.000 0.228 0.415
	53	+C(city_fe)[T.châteauguay] 0.2100 0.040 5.244 0.000 0.132 0.289
	54	+C(city_fe)[T.contrecoeur] 0.0760 0.059 1.284 0.199 -0.040 0.192
	55	+C(city_fe)[T.cowansville] 0.1233 0.043 2.845 0.004 0.038 0.208
	56	+C(city_fe)[T.crabtree] 0.0300 0.090 0.334 0.738 -0.146 0.206
	57	+C(city_fe)[T.côte-saint-luc] 0.6067 0.051 11.909 0.000 0.507 0.707
	58	+C(city_fe)[T.dolbeau-mistassini] -0.2804 0.086 -3.253 0.001 -0.449 -0.111
	59	+C(city_fe)[T.dollard-des-ormeaux] 0.4198 0.039 10.848 0.000 0.344 0.496
	60	+C(city_fe)[T.donnacona] 0.2326 0.039 5.989 0.000 0.156 0.309
	61	+C(city_fe)[T.dorval] 0.5505 0.048 11.354 0.000 0.455 0.645
	62	+C(city_fe)[T.drummondville] -0.0231 0.021 -1.108 0.268 -0.064 0.018
	63	+C(city_fe)[T.duhamel] 0.1073 0.071 1.511 0.131 -0.032 0.246
	64	+C(city_fe)[T.dunham] 0.2692 0.068 3.981 0.000 0.137 0.402
	65	+C(city_fe)[T.eastman] 0.2749 0.055 5.019 0.000 0.168 0.382
	66	+C(city_fe)[T.entrelacs] 0.2622 0.126 2.080 0.037 0.015 0.509
	67	+C(city_fe)[T.farnham] 0.0785 0.058 1.343 0.179 -0.036 0.193
	68	+C(city_fe)[T.fossambault-sur-le-lac] 0.2222 0.078 2.844 0.004 0.069 0.375
	69	+C(city_fe)[T.gaspé] -0.2403 0.071 -3.372 0.001 -0.380 -0.101
	70	+C(city_fe)[T.gatineau] 0.1031 0.013 8.192 0.000 0.078 0.128
	71	+C(city_fe)[T.gore] 0.0422 0.099 0.425 0.671 -0.152 0.237
	72	+C(city_fe)[T.granby] 0.1272 0.024 5.254 0.000 0.080 0.175
	73	+C(city_fe)[T.grenville-sur-la-rouge] 0.1171 0.082 1.433 0.152 -0.043 0.277
	74	+C(city_fe)[T.hampstead] 1.0869 0.057 19.051 0.000 0.975 1.199
	75	+C(city_fe)[T.harrington] 0.1032 0.145 0.713 0.476 -0.180 0.387
	76	+C(city_fe)[T.hudson] 0.2472 0.179 1.383 0.167 -0.103 0.598
	77	+C(city_fe)[T.joliette] -0.0242 0.070 -0.346 0.729 -0.161 0.113
	78	+C(city_fe)[T.jonquière] -0.1325 0.044 -3.001 0.003 -0.219 -0.046
	79	+C(city_fe)[T.kirkland] 0.3831 0.038 10.128 0.000 0.309 0.457
	80	+C(city_fe)[T.l'ange-gardien] 0.0259 0.163 0.159 0.874 -0.294 0.346
	81	+C(city_fe)[T.l'assomption] 0.1521 0.063 2.426 0.015 0.029 0.275
	82	+C(city_fe)[T.l'île-perrot] 0.4022 0.061 6.630 0.000 0.283 0.521
	83	+C(city_fe)[T.la conception] 0.3810 0.055 6.871 0.000 0.272 0.490
	84	+C(city_fe)[T.la malbaie] -0.1254 0.063 -1.980 0.048 -0.250 -0.001
	85	+C(city_fe)[T.la minerve] 0.0403 0.069 0.585 0.559 -0.095 0.175
	86	+C(city_fe)[T.la prairie] 0.3665 0.036 10.091 0.000 0.295 0.438
	87	+C(city_fe)[T.la pêche] 0.0226 0.036 0.621 0.535 -0.049 0.094
	88	+C(city_fe)[T.la tuque] -0.3102 0.065 -4.784 0.000 -0.437 -0.183
	89	+C(city_fe)[T.labelle] 0.0982 0.113 0.868 0.386 -0.124 0.320
	90	+C(city_fe)[T.lac-beauport] 0.3579 0.088 4.045 0.000 0.185 0.531
	91	+C(city_fe)[T.lac-brome] 0.5121 0.063 8.090 0.000 0.388 0.636
	92	+C(city_fe)[T.lac-des-écorces] -0.0499 0.154 -0.324 0.746 -0.352 0.252
	93	+C(city_fe)[T.lac-etchemin] -0.2092 0.144 -1.457 0.145 -0.491 0.072
	94	+C(city_fe)[T.lachute] 0.1368 0.073 1.878 0.060 -0.006 0.280
	95	+C(city_fe)[T.lanoraie] -0.1106 0.057 -1.941 0.052 -0.222 0.001
	96	+C(city_fe)[T.laval] 0.4064 0.020 19.826 0.000 0.366 0.447
	97	+C(city_fe)[T.lavaltrie] 0.1330 0.058 2.301 0.021 0.020 0.246
	98	+C(city_fe)[T.les éboulements] 0.1814 0.088 2.065 0.039 0.009 0.354
	99	+C(city_fe)[T.longueuil] 0.3166 0.019 16.430 0.000 0.279 0.354
	100	+C(city_fe)[T.lorraine] 0.1245 0.106 1.172 0.241 -0.084 0.333
	101	+C(city_fe)[T.lévis] 0.0630 0.025 2.508 0.012 0.014 0.112
	102	+C(city_fe)[T.magog] 0.3398 0.033 10.422 0.000 0.276 0.404
	103	+C(city_fe)[T.mandeville] -0.1594 0.080 -1.980 0.048 -0.317 -0.002
	104	+C(city_fe)[T.mansfield-et-pontefract] -0.4148 0.082 -5.049 0.000 -0.576 -0.254
	105	+C(city_fe)[T.marieville] 0.1990 0.092 2.158 0.031 0.018 0.380
	106	+C(city_fe)[T.mascouche] 0.1885 0.042 4.469 0.000 0.106 0.271
	107	+C(city_fe)[T.matane] -0.2595 0.037 -6.937 0.000 -0.333 -0.186
	108	+C(city_fe)[T.mercier / hochelaga / maisonneuve] 0.3424 0.036 9.489 0.000 0.272 0.413
	109	+C(city_fe)[T.messines] -0.1621 0.058 -2.784 0.005 -0.276 -0.048
	110	+C(city_fe)[T.mille-isles] 0.2411 0.055 4.390 0.000 0.133 0.349
	111	+C(city_fe)[T.mirabel] 0.2446 0.027 8.899 0.000 0.191 0.298
	112	+C(city_fe)[T.mont-joli] -0.2369 0.183 -1.292 0.196 -0.596 0.123
	113	+C(city_fe)[T.mont-laurier] -0.0547 0.037 -1.490 0.136 -0.127 0.017
	114	+C(city_fe)[T.mont-royal] 1.0680 0.061 17.561 0.000 0.949 1.187
	115	+C(city_fe)[T.mont-saint-hilaire] 0.3694 0.102 3.633 0.000 0.170 0.569
	116	+C(city_fe)[T.mont-tremblant] 0.4024 0.040 10.104 0.000 0.324 0.480
	117	+C(city_fe)[T.montréal] 0.6861 0.021 32.059 0.000 0.644 0.728
	118	+C(city_fe)[T.morin-heights] 0.3081 0.058 5.336 0.000 0.195 0.421
	119	+C(city_fe)[T.nominingue] 0.1255 0.095 1.322 0.186 -0.061 0.311
	120	+C(city_fe)[T.notre-dame-de-l'île-perrot] 0.2077 0.066 3.151 0.002 0.079 0.337
	121	+C(city_fe)[T.notre-dame-de-pontmain] 0.0126 0.061 0.206 0.837 -0.108 0.133
	122	+C(city_fe)[T.notre-dame-des-bois] -0.1992 0.075 -2.650 0.008 -0.347 -0.052
	123	+C(city_fe)[T.notre-dame-des-prairies] 0.0993 0.116 0.853 0.394 -0.129 0.327
	124	+C(city_fe)[T.notre-dame-du-laus] 0.0915 0.053 1.742 0.081 -0.011 0.194
	125	+C(city_fe)[T.orford] 0.3190 0.062 5.126 0.000 0.197 0.441
	126	+C(city_fe)[T.papineauville] -0.0746 0.124 -0.602 0.547 -0.317 0.168
	127	+C(city_fe)[T.petite-rivière-saint-françois] -0.0072 0.088 -0.083 0.934 -0.179 0.165
	128	+C(city_fe)[T.piedmont] 0.4510 0.061 7.336 0.000 0.330 0.571
	129	+C(city_fe)[T.pincourt] 0.2002 0.052 3.826 0.000 0.098 0.303
	130	+C(city_fe)[T.pointe-claire] 0.4417 0.037 11.947 0.000 0.369 0.514
	131	+C(city_fe)[T.pont-rouge] 0.1126 0.058 1.925 0.054 -0.002 0.227
	132	+C(city_fe)[T.prévost] 0.1344 0.038 3.571 0.000 0.061 0.208
	133	+C(city_fe)[T.québec] 0.1808 0.020 9.213 0.000 0.142 0.219
	134	+C(city_fe)[T.rawdon] 0.0837 0.045 1.854 0.064 -0.005 0.172
	135	+C(city_fe)[T.repentigny] 0.1814 0.033 5.527 0.000 0.117 0.246
	136	+C(city_fe)[T.rigaud] 0.2917 0.064 4.548 0.000 0.166 0.417
	137	+C(city_fe)[T.rimouski] -0.0524 0.036 -1.452 0.147 -0.123 0.018
	138	+C(city_fe)[T.rivière-du-loup] -0.1793 0.089 -2.009 0.044 -0.354 -0.004
	139	+C(city_fe)[T.rivière-rouge] -0.0220 0.071 -0.309 0.758 -0.161 0.118
	140	+C(city_fe)[T.roberval] -0.2385 0.076 -3.129 0.002 -0.388 -0.089
	141	+C(city_fe)[T.rosemère] 0.4051 0.088 4.628 0.000 0.234 0.577
	142	+C(city_fe)[T.saguenay] -0.0684 0.044 -1.544 0.123 -0.155 0.018
	143	+C(city_fe)[T.saint-adolphe-d'howard] 0.2906 0.054 5.420 0.000 0.186 0.396
	144	+C(city_fe)[T.saint-amable] 0.2992 0.084 3.543 0.000 0.134 0.465
	145	+C(city_fe)[T.saint-anicet] 0.0324 0.081 0.400 0.689 -0.126 0.191
	146	+C(city_fe)[T.saint-basile-le-grand] 0.3769 0.091 4.138 0.000 0.198 0.555
	147	+C(city_fe)[T.saint-bruno-de-montarville] 0.6511 0.071 9.114 0.000 0.511 0.791
	148	+C(city_fe)[T.saint-calixte] 0.1676 0.088 1.904 0.057 -0.005 0.340
	149	+C(city_fe)[T.saint-charles-borromée] 0.1796 0.174 1.034 0.301 -0.161 0.520
	150	+C(city_fe)[T.saint-colomban] 0.2115 0.068 3.114 0.002 0.078 0.345
	151	+C(city_fe)[T.saint-constant] 0.2084 0.056 3.745 0.000 0.099 0.317
	152	+C(city_fe)[T.saint-donat] 0.2478 0.066 3.738 0.000 0.118 0.378
	153	+C(city_fe)[T.saint-eustache] 0.4254 0.068 6.255 0.000 0.292 0.559
	154	+C(city_fe)[T.saint-félicien] -0.1964 0.320 -0.613 0.540 -0.824 0.431
	155	+C(city_fe)[T.saint-gabriel-de-brandon] -0.2016 0.100 -2.014 0.044 -0.398 -0.005
	156	+C(city_fe)[T.saint-georges] -0.0958 0.067 -1.426 0.154 -0.228 0.036
	157	+C(city_fe)[T.saint-hippolyte] 0.3319 0.079 4.193 0.000 0.177 0.487
	158	+C(city_fe)[T.saint-honoré] -0.1112 0.075 -1.478 0.139 -0.259 0.036
	159	+C(city_fe)[T.saint-hyacinthe] 0.1609 0.065 2.481 0.013 0.034 0.288
	160	+C(city_fe)[T.saint-jean-de-matha] 0.0935 0.157 0.597 0.550 -0.213 0.400
	161	+C(city_fe)[T.saint-jean-sur-richelieu] 0.1990 0.031 6.451 0.000 0.139 0.259
	162	+C(city_fe)[T.saint-jérôme] 0.1716 0.037 4.677 0.000 0.100 0.243
	163	+C(city_fe)[T.saint-lambert] 0.6578 0.069 9.598 0.000 0.523 0.792
	164	+C(city_fe)[T.saint-lambert-de-lauzon] 0.0260 0.087 0.298 0.766 -0.145 0.197
	165	+C(city_fe)[T.saint-lin/laurentides] 0.1117 0.124 0.898 0.369 -0.132 0.356
	166	+C(city_fe)[T.saint-marc-des-carrières] -0.2794 0.048 -5.829 0.000 -0.373 -0.185
	167	+C(city_fe)[T.saint-michel-des-saints] -0.2187 0.085 -2.572 0.010 -0.385 -0.052
	168	+C(city_fe)[T.saint-paul-de-l'île-aux-noix] -0.1315 0.223 -0.591 0.555 -0.568 0.305
	169	+C(city_fe)[T.saint-philippe] 0.3302 0.040 8.276 0.000 0.252 0.408
	170	+C(city_fe)[T.saint-raymond] 0.1480 0.085 1.750 0.080 -0.018 0.314
	171	+C(city_fe)[T.saint-sauveur] 0.3199 0.059 5.434 0.000 0.205 0.435
	172	+C(city_fe)[T.saint-zotique] 0.4330 0.073 5.923 0.000 0.290 0.576
	173	+C(city_fe)[T.saint-zénon] -0.2586 76.240 -0.003 0.997 -149.686 149.169
	174	+C(city_fe)[T.sainte-adèle] 0.3672 0.077 4.781 0.000 0.217 0.518
	175	+C(city_fe)[T.sainte-agathe-des-monts] 0.1184 0.092 1.291 0.197 -0.061 0.298
	176	+C(city_fe)[T.sainte-anne-des-plaines] 0.9383 0.778 1.206 0.228 -0.586 2.463
	177	+C(city_fe)[T.sainte-barbe] 0.1802 0.339 0.532 0.595 -0.483 0.844
	178	+C(city_fe)[T.sainte-brigitte-de-laval] 0.1292 0.052 2.507 0.012 0.028 0.230
	179	+C(city_fe)[T.sainte-béatrix] 0.0314 0.045 0.703 0.482 -0.056 0.119
	180	+C(city_fe)[T.sainte-catherine] 0.4067 0.072 5.638 0.000 0.265 0.548
	181	+C(city_fe)[T.sainte-catherine-de-la-jacques-cartier] 0.1900 0.066 2.877 0.004 0.061 0.319
	182	+C(city_fe)[T.sainte-julie] 0.4254 0.085 4.980 0.000 0.258 0.593
	183	+C(city_fe)[T.sainte-julienne] 0.0580 0.073 0.798 0.425 -0.084 0.200
	184	+C(city_fe)[T.sainte-lucie-des-laurentides] 0.3427 0.064 5.320 0.000 0.216 0.469
	185	+C(city_fe)[T.sainte-marguerite-du-lac-masson] 0.0785 0.052 1.501 0.133 -0.024 0.181
	186	+C(city_fe)[T.sainte-marthe-sur-le-lac] 0.2436 0.036 6.705 0.000 0.172 0.315
	187	+C(city_fe)[T.sainte-sophie] 0.1004 0.064 1.561 0.119 -0.026 0.227
	188	+C(city_fe)[T.sainte-thérèse] 0.3162 0.085 3.721 0.000 0.150 0.483
	189	+C(city_fe)[T.sainte-émélie-de-l'énergie] -0.2048 0.075 -2.730 0.006 -0.352 -0.058
	190	+C(city_fe)[T.salaberry-de-valleyfield] 0.0382 0.046 0.826 0.409 -0.052 0.129
	191	+C(city_fe)[T.sept-îles] -0.1273 0.079 -1.615 0.106 -0.282 0.027
	192	+C(city_fe)[T.shawinigan] -0.1806 0.051 -3.520 0.000 -0.281 -0.080
	193	+C(city_fe)[T.shefford] 0.1555 0.089 1.741 0.082 -0.020 0.331
	194	+C(city_fe)[T.sherbrooke] 0.0978 0.018 5.524 0.000 0.063 0.133
	195	+C(city_fe)[T.sorel-tracy] -0.1019 0.044 -2.339 0.019 -0.187 -0.017
	196	+C(city_fe)[T.st-colomban] 0.2146 0.057 3.780 0.000 0.103 0.326
	197	+C(city_fe)[T.st-hippolyte] 0.1331 0.055 2.435 0.015 0.026 0.240
	198	+C(city_fe)[T.st-jean-sur-richelieu] 0.1735 0.032 5.503 0.000 0.112 0.235
	199	+C(city_fe)[T.st-jérôme] 0.1551 0.031 5.014 0.000 0.094 0.216
	200	+C(city_fe)[T.st-sauveur] 0.2612 0.049 5.298 0.000 0.165 0.358
	201	+C(city_fe)[T.ste-adèle] 0.1396 0.069 2.012 0.044 0.004 0.276
	202	+C(city_fe)[T.ste-julienne] 0.0377 0.035 1.083 0.279 -0.031 0.106
	203	+C(city_fe)[T.sutton] 0.4078 0.066 6.196 0.000 0.279 0.537
	204	+C(city_fe)[T.terrebonne] 0.1965 0.028 6.923 0.000 0.141 0.252
	205	+C(city_fe)[T.thetford mines] -0.3029 0.117 -2.596 0.009 -0.532 -0.074
	206	+C(city_fe)[T.trois-rivières] -0.0810 0.022 -3.706 0.000 -0.124 -0.038
	207	+C(city_fe)[T.très-saint-rédempteur] 0.4839 0.032 15.354 0.000 0.422 0.546
	208	+C(city_fe)[T.témiscouata-sur-le-lac] -0.3066 0.124 -2.473 0.013 -0.550 -0.064
	209	+C(city_fe)[T.val-d'or] 0.0788 0.031 2.543 0.011 0.018 0.139
	210	+C(city_fe)[T.val-david] 0.0115 0.117 0.099 0.921 -0.217 0.240
	211	+C(city_fe)[T.val-des-monts] 0.1031 0.025 4.142 0.000 0.054 0.152
	212	+C(city_fe)[T.val-morin] -0.1267 0.093 -1.357 0.175 -0.310 0.056
	213	+C(city_fe)[T.vaudreuil-dorion] 0.0802 0.044 1.807 0.071 -0.007 0.167
	214	+C(city_fe)[T.venise-en-québec] 0.0614 0.082 0.744 0.457 -0.100 0.223
	215	+C(city_fe)[T.victoriaville] -0.0752 0.029 -2.593 0.010 -0.132 -0.018
	216	+C(city_fe)[T.waterloo] 0.0905 0.056 1.608 0.108 -0.020 0.201
	217	+C(city_fe)[T.weedon] -0.1348 0.080 -1.675 0.094 -0.293 0.023
	218	+C(city_fe)[T.wentworth-nord] 0.1764 0.090 1.951 0.051 -0.001 0.354
	219	+C(city_fe)[T.westmount] 1.3152 0.041 31.987 0.000 1.235 1.396
	220	+ln_sqft 0.4106 0.013 30.648 0.000 0.384 0.437
	221	+ln_lot 0.0253 0.004 6.506 0.000 0.018 0.033
	222	+has_lot -0.0201 0.015 -1.369 0.171 -0.049 0.009
	223	+beds 0.0057 0.004 1.361 0.173 -0.003 0.014
	224	+baths 0.0867 0.007 13.252 0.000 0.074 0.099
	225	+parking 0.0112 0.002 7.312 0.000 0.008 0.014
	226	+storeys -0.0018 0.009 -0.199 0.843 -0.019 0.016
	227	+age -0.0053 0.000 -11.284 0.000 -0.006 -0.004
	228	+age2 2.612e-05 3.72e-06 7.024 0.000 1.88e-05 3.34e-05
	229	+has_age -0.1623 0.022 -7.477 0.000 -0.205 -0.120
	230	+pool 0.0059 0.011 0.526 0.599 -0.016 0.028
	231	+waterfront 0.0166 0.022 0.751 0.452 -0.027 0.060
	232	+garage 0.0025 0.021 0.119 0.905 -0.038 0.043
	233	+fireplace -0.0611 0.016 -3.776 0.000 -0.093 -0.029
	234	+central_air 0.0785 0.012 6.669 0.000 0.055 0.102
	235	+finished_bsmt -0.0316 0.011 -2.757 0.006 -0.054 -0.009
	236	+near_water 0.0910 0.024 3.842 0.000 0.045 0.137
	237	+cul_de_sac 0.0268 0.012 2.317 0.020 0.004 0.050
	238	+golf 0.0232 0.007 3.268 0.001 0.009 0.037
	239	+ski 0.0139 0.013 1.066 0.286 -0.012 0.040
	240	+near_park 0.0039 0.011 0.340 0.734 -0.018 0.026
	241	+near_school -0.0637 0.011 -5.687 0.000 -0.086 -0.042
	242	+near_highway 0.0475 0.008 5.929 0.000 0.032 0.063
	243	+transit 0.1036 0.009 11.762 0.000 0.086 0.121
	244	+no_rear_neighbor -0.0396 0.010 -4.038 0.000 -0.059 -0.020
	245	+n_images_c 0.0044 0.000 14.018 0.000 0.004 0.005
	246	+virtual_tour 0.0324 0.022 1.460 0.144 -0.011 0.076
	247	+n_rooms 0.0170 0.002 9.265 0.000 0.013 0.021
	248	+n_levels_r -0.0226 0.009 -2.526 0.012 -0.040 -0.005
	249	+has_primary_bed 0.0753 0.016 4.839 0.000 0.045 0.106
	250	+has_basement -0.0169 0.015 -1.148 0.251 -0.046 0.012
	251	+r_office 0.0308 0.011 2.735 0.006 0.009 0.053
	252	+r_family 0.0359 0.011 3.375 0.001 0.015 0.057
	253	+r_workshop -0.0386 0.014 -2.828 0.005 -0.065 -0.012
	254	+r_winecellar 0.1108 0.036 3.073 0.002 0.040 0.181
	255	+r_sunroom 0.0007 0.021 0.033 0.974 -0.040 0.042
	256	+r_mezzanine 0.0175 0.025 0.690 0.490 -0.032 0.067
	257	+r_library 0.0971 0.083 1.171 0.242 -0.065 0.260
	258	+r_laundry 0.0178 0.009 1.980 0.048 0.000 0.035
	259	+g_cond 0.0115 0.004 2.567 0.010 0.003 0.020
	260	+g_hardwood 0.0256 0.008 3.211 0.001 0.010 0.041
	261	+g_quartz_granite 0.0186 0.011 1.771 0.077 -0.002 0.039
	262	+g_view_water 0.0132 0.016 0.845 0.398 -0.017 0.044
	263	+has_gpt 4.8574 0.047 102.393 0.000 4.764 4.950
	264	+g_turnkey 0.0198 0.008 2.562 0.010 0.005 0.035
	265	+g_recently_renovated 0.0050 0.010 0.513 0.608 -0.014 0.024
	266	+g_garage 0.0235 0.021 1.131 0.258 -0.017 0.064
	267	+g_pool 0.0492 0.012 4.204 0.000 0.026 0.072
	268	+g_waterfront 0.0744 0.018 4.135 0.000 0.039 0.110
	269	+g_finished_basement 0.0189 0.009 2.203 0.028 0.002 0.036
	270	+g_fireplace 0.0907 0.016 5.618 0.000 0.059 0.122
	271	+g_central_ac 0.0140 0.010 1.406 0.160 -0.006 0.034
	272	+g_new_construction 0.0680 0.014 4.909 0.000 0.041 0.095
	273	+g_income_property -0.0331 0.011 -3.083 0.002 -0.054 -0.012
	274	+g_renovated_kitchen 0.0298 0.010 2.906 0.004 0.010 0.050
	275	+g_renovated_bathroom -0.0060 0.010 -0.614 0.539 -0.025 0.013
	276	+g_large_lot -0.0033 0.009 -0.370 0.711 -0.021 0.014
	277	+g_quiet_area -0.0028 0.008 -0.369 0.712 -0.018 0.012
	278	+==============================================================================
	279	+Omnibus: 426.401 Durbin-Watson: 1.855
	280	+Prob(Omnibus): 0.000 Jarque-Bera (JB): 1443.982
	281	+Skew: 0.076 Prob(JB): 0.00
	282	+Kurtosis: 4.909 Cond. No. 2.08e+16
	283	+==============================================================================
	284	+
	285	+Notes:
	286	+[1] Standard Errors are heteroscedasticity robust (HC3)
	287	+[2] The smallest eigenvalue is 3.84e-22. This might indicate that there are
	288	+strong multicollinearity problems or that the design matrix is singular.
	289	+
	290	+=== B+GPT ===
	291	+ OLS Regression Results
	292	+==============================================================================
	293	+Dep. Variable: ln_price R-squared: 0.684
	294	+Model: OLS Adj. R-squared: 0.680
	295	+Method: Least Squares F-statistic: 1.873e+05
	296	+Date: Fri, 03 Jul 2026 Prob (F-statistic): 0.00
	297	+Time: 08:11:28 Log-Likelihood: -6089.2
	298	+No. Observations: 19123 AIC: 1.270e+04
	299	+Df Residuals: 18864 BIC: 1.473e+04
	300	+Df Model: 258
	301	+Covariance Type: HC3
	302	+========================================================================================================================
	303	+ coef std err z P>\|z\| [0.025 0.975]
	304	+------------------------------------------------------------------------------------------------------------------------
	305	+Intercept 6.1815 0.057 108.357 0.000 6.070 6.293
	306	+C(source)[T.realtor] -0.3800 0.021 -17.682 0.000 -0.422 -0.338
	307	+C(ptype_fe)[T.Country Home] 0.0525 0.067 0.788 0.431 -0.078 0.183
	308	+C(ptype_fe)[T.Home] 0.0936 0.081 1.154 0.249 -0.065 0.253
	309	+C(ptype_fe)[T.House] -0.0724 0.014 -5.164 0.000 -0.100 -0.045
	310	+C(ptype_fe)[T.Intergenerational] 0.1435 0.063 2.291 0.022 0.021 0.266
	311	+C(ptype_fe)[T.Other] 0.2526 0.121 2.086 0.037 0.015 0.490
	312	+C(ptype_fe)[T.Row / Townhouse] -0.3076 0.014 -21.318 0.000 -0.336 -0.279
	313	+C(ptype_fe)[T.Semi Detached] -0.0284 0.057 -0.501 0.616 -0.140 0.083
	314	+C(ptype_fe)[T.Townhouse] 0.0700 0.098 0.713 0.476 -0.122 0.262
	315	+C(city_fe)[T.acton vale] -0.0098 0.034 -0.284 0.776 -0.077 0.058
	316	+C(city_fe)[T.alma] -0.0584 0.054 -1.079 0.281 -0.165 0.048
	317	+C(city_fe)[T.amherst] -0.0225 0.049 -0.456 0.648 -0.119 0.074
	318	+C(city_fe)[T.amos] -0.1343 0.044 -3.020 0.003 -0.222 -0.047
	319	+C(city_fe)[T.austin] 0.4205 0.070 5.972 0.000 0.282 0.559
	320	+C(city_fe)[T.baie-comeau] -0.3294 0.069 -4.787 0.000 -0.464 -0.195
	321	+C(city_fe)[T.baie-saint-paul] 0.1285 0.077 1.672 0.094 -0.022 0.279
	322	+C(city_fe)[T.beaconsfield] 0.7858 0.037 21.069 0.000 0.713 0.859
	323	+C(city_fe)[T.beauport] 0.0776 0.037 2.076 0.038 0.004 0.151
	324	+C(city_fe)[T.beloeil] 0.3812 0.049 7.840 0.000 0.286 0.476
	325	+C(city_fe)[T.blainville] 0.4461 0.025 17.924 0.000 0.397 0.495
	326	+C(city_fe)[T.blue sea] -0.1151 0.062 -1.863 0.062 -0.236 0.006
	327	+C(city_fe)[T.bois-des-filion] 0.2942 0.048 6.129 0.000 0.200 0.388
	328	+C(city_fe)[T.boisbriand] 0.3654 0.038 9.665 0.000 0.291 0.439
	329	+C(city_fe)[T.boischatel] 0.4011 0.071 5.616 0.000 0.261 0.541
	330	+C(city_fe)[T.boucherville] 0.4671 0.037 12.493 0.000 0.394 0.540
	331	+C(city_fe)[T.bromont] 0.4244 0.034 12.535 0.000 0.358 0.491
	332	+C(city_fe)[T.brossard] 0.5263 0.020 26.802 0.000 0.488 0.565
	333	+C(city_fe)[T.bécancour] -0.0973 0.032 -3.045 0.002 -0.160 -0.035
	334	+C(city_fe)[T.cantley] 0.1731 0.046 3.757 0.000 0.083 0.263
	335	+C(city_fe)[T.carignan] 0.3790 0.042 8.990 0.000 0.296 0.462
	336	+C(city_fe)[T.chandler] -0.2950 0.079 -3.757 0.000 -0.449 -0.141
	337	+C(city_fe)[T.chateauguay] 0.2492 0.040 6.176 0.000 0.170 0.328
	338	+C(city_fe)[T.chelsea] 0.3027 0.040 7.590 0.000 0.225 0.381
	339	+C(city_fe)[T.chertsey] -0.0418 0.038 -1.107 0.268 -0.116 0.032
	340	+C(city_fe)[T.chicoutimi] -0.0792 0.040 -1.976 0.048 -0.158 -0.001
	341	+C(city_fe)[T.chomedey] 0.4118 0.055 7.515 0.000 0.304 0.519
	342	+C(city_fe)[T.châteauguay] 0.2811 0.024 11.697 0.000 0.234 0.328
	343	+C(city_fe)[T.contrecoeur] 0.1282 0.035 3.614 0.000 0.059 0.198
	344	+C(city_fe)[T.cowansville] 0.1731 0.034 5.103 0.000 0.107 0.240
	345	+C(city_fe)[T.crabtree] 0.0682 0.048 1.421 0.155 -0.026 0.162
	346	+C(city_fe)[T.côte-saint-luc] 0.7496 0.053 14.252 0.000 0.647 0.853
	347	+C(city_fe)[T.dolbeau-mistassini] -0.1900 0.063 -3.033 0.002 -0.313 -0.067
	348	+C(city_fe)[T.dollard-des-ormeaux] 0.5396 0.029 18.879 0.000 0.484 0.596
	349	+C(city_fe)[T.donnacona] 0.1291 0.035 3.685 0.000 0.060 0.198
	350	+C(city_fe)[T.dorval] 0.6824 0.048 14.112 0.000 0.588 0.777
	351	+C(city_fe)[T.drummondville] 0.0599 0.018 3.286 0.001 0.024 0.096
	352	+C(city_fe)[T.duhamel] 0.0493 0.071 0.697 0.485 -0.089 0.188
	353	+C(city_fe)[T.dunham] 0.3393 0.064 5.282 0.000 0.213 0.465
	354	+C(city_fe)[T.eastman] 0.3752 0.056 6.645 0.000 0.265 0.486
	355	+C(city_fe)[T.entrelacs] 0.2115 0.084 2.512 0.012 0.046 0.377
	356	+C(city_fe)[T.farnham] 0.0658 0.035 1.873 0.061 -0.003 0.135
	357	+C(city_fe)[T.fossambault-sur-le-lac] 0.3149 0.065 4.826 0.000 0.187 0.443
	358	+C(city_fe)[T.gaspé] -0.3176 0.065 -4.918 0.000 -0.444 -0.191
	359	+C(city_fe)[T.gatineau] 0.1320 0.013 10.162 0.000 0.107 0.158
	360	+C(city_fe)[T.gore] 0.0680 0.055 1.227 0.220 -0.041 0.177
	361	+C(city_fe)[T.granby] 0.1692 0.019 9.137 0.000 0.133 0.206
	362	+C(city_fe)[T.grenville-sur-la-rouge] 0.1187 0.046 2.558 0.011 0.028 0.210
	363	+C(city_fe)[T.hampstead] 1.2817 0.052 24.725 0.000 1.180 1.383
	364	+C(city_fe)[T.harrington] -0.0772 0.125 -0.618 0.537 -0.322 0.168
	365	+C(city_fe)[T.hudson] 0.4854 0.066 7.303 0.000 0.355 0.616
	366	+C(city_fe)[T.joliette] 0.1347 0.045 3.023 0.003 0.047 0.222
	367	+C(city_fe)[T.jonquière] -0.1481 0.044 -3.388 0.001 -0.234 -0.062
	368	+C(city_fe)[T.kirkland] 0.5562 0.035 15.874 0.000 0.488 0.625
	369	+C(city_fe)[T.l'ange-gardien] 0.2606 0.073 3.560 0.000 0.117 0.404
	370	+C(city_fe)[T.l'assomption] 0.2026 0.028 7.206 0.000 0.147 0.258
	371	+C(city_fe)[T.l'île-perrot] 0.4466 0.049 9.037 0.000 0.350 0.543
	372	+C(city_fe)[T.la conception] 0.3260 0.052 6.254 0.000 0.224 0.428
	373	+C(city_fe)[T.la malbaie] -0.1569 0.054 -2.898 0.004 -0.263 -0.051
	374	+C(city_fe)[T.la minerve] 0.0588 0.052 1.128 0.260 -0.043 0.161
	375	+C(city_fe)[T.la prairie] 0.4831 0.030 16.099 0.000 0.424 0.542
	376	+C(city_fe)[T.la pêche] -0.0050 0.039 -0.129 0.897 -0.081 0.071
	377	+C(city_fe)[T.la tuque] -0.3458 0.048 -7.231 0.000 -0.439 -0.252
	378	+C(city_fe)[T.labelle] 0.0865 0.054 1.599 0.110 -0.020 0.193
	379	+C(city_fe)[T.lac-beauport] 0.4153 0.085 4.883 0.000 0.249 0.582
	380	+C(city_fe)[T.lac-brome] 0.6355 0.056 11.357 0.000 0.526 0.745
	381	+C(city_fe)[T.lac-des-écorces] 0.0014 0.067 0.021 0.983 -0.130 0.133
	382	+C(city_fe)[T.lac-etchemin] -0.2994 0.057 -5.288 0.000 -0.410 -0.188
	383	+C(city_fe)[T.lachute] 0.1773 0.037 4.791 0.000 0.105 0.250
	384	+C(city_fe)[T.lanoraie] 0.1419 0.037 3.844 0.000 0.070 0.214
	385	+C(city_fe)[T.laval] 0.4740 0.014 34.317 0.000 0.447 0.501
	386	+C(city_fe)[T.lavaltrie] 0.1688 0.037 4.538 0.000 0.096 0.242
	387	+C(city_fe)[T.les éboulements] 0.1753 0.069 2.547 0.011 0.040 0.310
	388	+C(city_fe)[T.longueuil] 0.3851 0.016 24.101 0.000 0.354 0.416
	389	+C(city_fe)[T.lorraine] 0.4987 0.056 8.885 0.000 0.389 0.609
	390	+C(city_fe)[T.lévis] 0.1158 0.021 5.394 0.000 0.074 0.158
	391	+C(city_fe)[T.magog] 0.3283 0.032 10.286 0.000 0.266 0.391
	392	+C(city_fe)[T.mandeville] -0.0735 0.040 -1.826 0.068 -0.152 0.005
	393	+C(city_fe)[T.mansfield-et-pontefract] -0.5322 0.067 -7.928 0.000 -0.664 -0.401
	394	+C(city_fe)[T.marieville] 0.2115 0.031 6.931 0.000 0.152 0.271
	395	+C(city_fe)[T.mascouche] 0.3035 0.022 13.981 0.000 0.261 0.346
	396	+C(city_fe)[T.matane] -0.1678 0.047 -3.568 0.000 -0.260 -0.076
	397	+C(city_fe)[T.mercier / hochelaga / maisonneuve] 0.4115 0.048 8.594 0.000 0.318 0.505
	398	+C(city_fe)[T.messines] -0.2594 0.051 -5.045 0.000 -0.360 -0.159
	399	+C(city_fe)[T.mille-isles] 0.2025 0.051 3.999 0.000 0.103 0.302
	400	+C(city_fe)[T.mirabel] 0.3088 0.021 14.650 0.000 0.268 0.350
	401	+C(city_fe)[T.mont-joli] -0.2580 0.097 -2.649 0.008 -0.449 -0.067
	402	+C(city_fe)[T.mont-laurier] -0.0438 0.041 -1.076 0.282 -0.124 0.036
	403	+C(city_fe)[T.mont-royal] 1.2714 0.054 23.624 0.000 1.166 1.377
	404	+C(city_fe)[T.mont-saint-hilaire] 0.5104 0.040 12.618 0.000 0.431 0.590
	405	+C(city_fe)[T.mont-tremblant] 0.3946 0.035 11.119 0.000 0.325 0.464
	406	+C(city_fe)[T.montréal] 0.7272 0.017 43.952 0.000 0.695 0.760
	407	+C(city_fe)[T.morin-heights] 0.3128 0.039 7.963 0.000 0.236 0.390
	408	+C(city_fe)[T.nominingue] 0.0819 0.081 1.012 0.311 -0.077 0.241
	409	+C(city_fe)[T.notre-dame-de-l'île-perrot] 0.3960 0.042 9.361 0.000 0.313 0.479
	410	+C(city_fe)[T.notre-dame-de-pontmain] -0.1176 0.060 -1.973 0.049 -0.234 -0.001
	411	+C(city_fe)[T.notre-dame-des-bois] -0.3186 0.067 -4.775 0.000 -0.449 -0.188
	412	+C(city_fe)[T.notre-dame-des-prairies] 0.0959 0.042 2.279 0.023 0.013 0.178
	413	+C(city_fe)[T.notre-dame-du-laus] 0.0251 0.056 0.448 0.654 -0.085 0.135
	414	+C(city_fe)[T.orford] 0.3476 0.061 5.701 0.000 0.228 0.467
	415	+C(city_fe)[T.papineauville] -0.0121 0.115 -0.105 0.916 -0.238 0.213
	416	+C(city_fe)[T.petite-rivière-saint-françois] -0.0216 0.055 -0.391 0.696 -0.130 0.087
	417	+C(city_fe)[T.piedmont] 0.3718 0.051 7.301 0.000 0.272 0.472
	418	+C(city_fe)[T.pincourt] 0.3208 0.036 8.976 0.000 0.251 0.391
	419	+C(city_fe)[T.pointe-claire] 0.6096 0.041 14.960 0.000 0.530 0.689
	420	+C(city_fe)[T.pont-rouge] 0.0167 0.044 0.378 0.705 -0.070 0.103
	421	+C(city_fe)[T.prévost] 0.1910 0.027 7.046 0.000 0.138 0.244
	422	+C(city_fe)[T.québec] 0.2113 0.018 11.997 0.000 0.177 0.246
	423	+C(city_fe)[T.rawdon] 0.0548 0.032 1.711 0.087 -0.008 0.118
	424	+C(city_fe)[T.repentigny] 0.2910 0.019 15.040 0.000 0.253 0.329
	425	+C(city_fe)[T.rigaud] 0.1464 0.045 3.257 0.001 0.058 0.235
	426	+C(city_fe)[T.rimouski] -0.0604 0.034 -1.779 0.075 -0.127 0.006
	427	+C(city_fe)[T.rivière-du-loup] -0.0744 0.054 -1.380 0.168 -0.180 0.031
	428	+C(city_fe)[T.rivière-rouge] -0.1228 0.059 -2.083 0.037 -0.238 -0.007
	429	+C(city_fe)[T.roberval] -0.1505 0.064 -2.335 0.020 -0.277 -0.024
	430	+C(city_fe)[T.rosemère] 0.4912 0.038 12.813 0.000 0.416 0.566
	431	+C(city_fe)[T.saguenay] -0.0362 0.023 -1.577 0.115 -0.081 0.009
	432	+C(city_fe)[T.saint-adolphe-d'howard] 0.1426 0.045 3.142 0.002 0.054 0.231
	433	+C(city_fe)[T.saint-amable] 0.2897 0.035 8.261 0.000 0.221 0.358
	434	+C(city_fe)[T.saint-anicet] 0.2175 0.068 3.178 0.001 0.083 0.352
	435	+C(city_fe)[T.saint-basile-le-grand] 0.3978 0.045 8.829 0.000 0.310 0.486
	436	+C(city_fe)[T.saint-bruno-de-montarville] 0.6239 0.040 15.791 0.000 0.546 0.701
	437	+C(city_fe)[T.saint-calixte] 0.0061 0.034 0.177 0.859 -0.061 0.073
	438	+C(city_fe)[T.saint-charles-borromée] 0.1596 0.053 3.015 0.003 0.056 0.263
	439	+C(city_fe)[T.saint-colomban] 0.1400 0.041 3.385 0.001 0.059 0.221
	440	+C(city_fe)[T.saint-constant] 0.3374 0.031 10.768 0.000 0.276 0.399
	441	+C(city_fe)[T.saint-donat] 0.3168 0.047 6.726 0.000 0.224 0.409
	442	+C(city_fe)[T.saint-eustache] 0.3452 0.036 9.568 0.000 0.274 0.416
	443	+C(city_fe)[T.saint-félicien] -0.0434 0.054 -0.807 0.420 -0.149 0.062
	444	+C(city_fe)[T.saint-gabriel-de-brandon] -0.0090 0.063 -0.143 0.887 -0.133 0.115
	445	+C(city_fe)[T.saint-georges] -0.1772 0.038 -4.679 0.000 -0.251 -0.103
	446	+C(city_fe)[T.saint-hippolyte] 0.2118 0.041 5.225 0.000 0.132 0.291
	447	+C(city_fe)[T.saint-honoré] -0.0330 0.062 -0.536 0.592 -0.154 0.088
	448	+C(city_fe)[T.saint-hyacinthe] 0.1948 0.033 5.974 0.000 0.131 0.259
	449	+C(city_fe)[T.saint-jean-de-matha] 0.0276 0.065 0.422 0.673 -0.101 0.156
	450	+C(city_fe)[T.saint-jean-sur-richelieu] 0.2785 0.019 14.288 0.000 0.240 0.317
	451	+C(city_fe)[T.saint-jérôme] 0.2141 0.020 10.479 0.000 0.174 0.254
	452	+C(city_fe)[T.saint-lambert] 0.8773 0.047 18.632 0.000 0.785 0.970
	453	+C(city_fe)[T.saint-lambert-de-lauzon] 0.0684 0.053 1.298 0.194 -0.035 0.172
	454	+C(city_fe)[T.saint-lin/laurentides] 0.1066 0.032 3.313 0.001 0.044 0.170
	455	+C(city_fe)[T.saint-marc-des-carrières] -0.1398 0.032 -4.307 0.000 -0.203 -0.076
	456	+C(city_fe)[T.saint-michel-des-saints] -0.0394 0.067 -0.589 0.556 -0.171 0.092
	457	+C(city_fe)[T.saint-paul-de-l'île-aux-noix] 0.0212 0.099 0.214 0.831 -0.173 0.215
	458	+C(city_fe)[T.saint-philippe] 0.4244 0.029 14.752 0.000 0.368 0.481
	459	+C(city_fe)[T.saint-raymond] 0.0769 0.067 1.154 0.249 -0.054 0.207
	460	+C(city_fe)[T.saint-sauveur] 0.2746 0.033 8.211 0.000 0.209 0.340
	461	+C(city_fe)[T.saint-zotique] 0.2721 0.037 7.329 0.000 0.199 0.345
	462	+C(city_fe)[T.saint-zénon] -0.0122 0.106 -0.115 0.908 -0.221 0.196
	463	+C(city_fe)[T.sainte-adèle] 0.3087 0.048 6.408 0.000 0.214 0.403
	464	+C(city_fe)[T.sainte-agathe-des-monts] 0.1241 0.050 2.457 0.014 0.025 0.223
	465	+C(city_fe)[T.sainte-anne-des-plaines] 0.2834 0.062 4.584 0.000 0.162 0.405
	466	+C(city_fe)[T.sainte-barbe] 0.4826 0.071 6.781 0.000 0.343 0.622
	467	+C(city_fe)[T.sainte-brigitte-de-laval] 0.1199 0.041 2.899 0.004 0.039 0.201
	468	+C(city_fe)[T.sainte-béatrix] -0.0687 0.043 -1.582 0.114 -0.154 0.016
	469	+C(city_fe)[T.sainte-catherine] 0.3381 0.034 9.976 0.000 0.272 0.405
	470	+C(city_fe)[T.sainte-catherine-de-la-jacques-cartier] 0.1270 0.059 2.152 0.031 0.011 0.243
	471	+C(city_fe)[T.sainte-julie] 0.4826 0.055 8.787 0.000 0.375 0.590
	472	+C(city_fe)[T.sainte-julienne] 0.0375 0.035 1.079 0.281 -0.031 0.106
	473	+C(city_fe)[T.sainte-lucie-des-laurentides] 0.1493 0.037 3.986 0.000 0.076 0.223
	474	+C(city_fe)[T.sainte-marguerite-du-lac-masson] 0.0068 0.053 0.129 0.897 -0.097 0.110
	475	+C(city_fe)[T.sainte-marthe-sur-le-lac] 0.3182 0.023 13.870 0.000 0.273 0.363
	476	+C(city_fe)[T.sainte-sophie] 0.0978 0.038 2.566 0.010 0.023 0.173
	477	+C(city_fe)[T.sainte-thérèse] 0.3765 0.041 9.286 0.000 0.297 0.456
	478	+C(city_fe)[T.sainte-émélie-de-l'énergie] -0.2157 0.060 -3.580 0.000 -0.334 -0.098
	479	+C(city_fe)[T.salaberry-de-valleyfield] 0.0841 0.043 1.978 0.048 0.001 0.168
	480	+C(city_fe)[T.sept-îles] -0.0848 0.036 -2.364 0.018 -0.155 -0.014
	481	+C(city_fe)[T.shawinigan] -0.1015 0.025 -4.056 0.000 -0.151 -0.052
	482	+C(city_fe)[T.shefford] 0.2935 0.059 4.934 0.000 0.177 0.410
	483	+C(city_fe)[T.sherbrooke] 0.1051 0.017 6.332 0.000 0.073 0.138
	484	+C(city_fe)[T.sorel-tracy] 0.0059 0.027 0.220 0.826 -0.047 0.059
	485	+C(city_fe)[T.st-colomban] 0.2744 0.069 3.978 0.000 0.139 0.410
	486	+C(city_fe)[T.st-hippolyte] 0.1024 0.077 1.324 0.185 -0.049 0.254
	487	+C(city_fe)[T.st-jean-sur-richelieu] 0.2296 0.034 6.754 0.000 0.163 0.296
	488	+C(city_fe)[T.st-jérôme] 0.2389 0.036 6.713 0.000 0.169 0.309
	489	+C(city_fe)[T.st-sauveur] 0.4023 0.088 4.549 0.000 0.229 0.576
	490	+C(city_fe)[T.ste-adèle] 0.1944 0.064 3.044 0.002 0.069 0.320
	491	+C(city_fe)[T.ste-julienne] 0.0665 0.038 1.740 0.082 -0.008 0.141
	492	+C(city_fe)[T.sutton] 0.4900 0.051 9.559 0.000 0.390 0.590
	493	+C(city_fe)[T.terrebonne] 0.2995 0.018 16.761 0.000 0.264 0.335
	494	+C(city_fe)[T.thetford mines] -0.2784 0.058 -4.762 0.000 -0.393 -0.164
	495	+C(city_fe)[T.trois-rivières] 0.0129 0.020 0.634 0.526 -0.027 0.053
	496	+C(city_fe)[T.très-saint-rédempteur] 0.5938 0.052 11.367 0.000 0.491 0.696
	497	+C(city_fe)[T.témiscouata-sur-le-lac] -0.1940 0.074 -2.618 0.009 -0.339 -0.049
	498	+C(city_fe)[T.val-d'or] 0.0212 0.029 0.742 0.458 -0.035 0.077
	499	+C(city_fe)[T.val-david] 0.0449 0.054 0.828 0.408 -0.061 0.151
	500	+C(city_fe)[T.val-des-monts] 0.0856 0.027 3.163 0.002 0.033 0.139
	501	+C(city_fe)[T.val-morin] 0.0522 0.057 0.916 0.360 -0.060 0.164
	502	+C(city_fe)[T.vaudreuil-dorion] 0.2565 0.029 8.949 0.000 0.200 0.313
	503	+C(city_fe)[T.venise-en-québec] 0.1799 0.044 4.085 0.000 0.094 0.266
	504	+C(city_fe)[T.victoriaville] -0.0166 0.024 -0.691 0.489 -0.064 0.031
	505	+C(city_fe)[T.waterloo] 0.1307 0.049 2.687 0.007 0.035 0.226
	506	+C(city_fe)[T.weedon] -0.1577 0.062 -2.550 0.011 -0.279 -0.036
	507	+C(city_fe)[T.wentworth-nord] 0.0592 0.072 0.828 0.408 -0.081 0.199
	508	+C(city_fe)[T.westmount] 1.6031 0.038 42.458 0.000 1.529 1.677
	509	+ln_lot 0.0400 0.003 12.549 0.000 0.034 0.046
	510	+has_lot -0.0842 0.015 -5.753 0.000 -0.113 -0.056
	511	+beds 0.0281 0.004 6.804 0.000 0.020 0.036
	512	+baths 0.1331 0.005 25.473 0.000 0.123 0.143
	513	+parking 0.0153 0.002 8.276 0.000 0.012 0.019
	514	+storeys 0.0023 0.096 0.024 0.981 -0.186 0.190
	515	+age -0.0066 0.000 -16.853 0.000 -0.007 -0.006
	516	+age2 3.287e-05 2.79e-06 11.795 0.000 2.74e-05 3.83e-05
	517	+has_age -0.1356 0.019 -7.296 0.000 -0.172 -0.099
	518	+pool 0.0217 0.011 1.945 0.052 -0.000 0.044
	519	+waterfront 0.0497 0.019 2.636 0.008 0.013 0.087
	520	+garage 0.0322 0.018 1.748 0.080 -0.004 0.068
	521	+fireplace -0.0667 0.016 -4.270 0.000 -0.097 -0.036
	522	+central_air 0.0445 0.010 4.577 0.000 0.025 0.064
	523	+finished_bsmt -0.0063 0.008 -0.773 0.439 -0.022 0.010
	524	+near_water 0.0709 0.020 3.459 0.001 0.031 0.111
	525	+cul_de_sac 0.0285 0.009 3.179 0.001 0.011 0.046
	526	+golf 0.0180 0.006 3.122 0.002 0.007 0.029
	527	+ski 0.0382 0.010 3.899 0.000 0.019 0.057
	528	+near_park 0.0074 0.008 0.929 0.353 -0.008 0.023
	529	+near_school -0.0520 0.009 -6.027 0.000 -0.069 -0.035
	530	+near_highway 0.0475 0.006 7.384 0.000 0.035 0.060
	531	+transit 0.0684 0.007 9.180 0.000 0.054 0.083
	532	+no_rear_neighbor -0.0453 0.009 -4.890 0.000 -0.063 -0.027
	533	+n_images_c 0.0057 0.000 18.731 0.000 0.005 0.006
	534	+virtual_tour 0.0207 0.016 1.327 0.184 -0.010 0.051
	535	+n_rooms 0.0254 0.001 17.680 0.000 0.023 0.028
	536	+n_levels_r 0.0413 0.047 0.871 0.384 -0.052 0.134
	537	+has_primary_bed 0.0664 0.010 6.743 0.000 0.047 0.086
	538	+has_basement -0.1032 0.043 -2.392 0.017 -0.188 -0.019
	539	+r_office 0.0359 0.008 4.728 0.000 0.021 0.051
	540	+r_family 0.0359 0.009 4.050 0.000 0.019 0.053
	541	+r_workshop -0.0297 0.009 -3.377 0.001 -0.047 -0.012
	542	+r_winecellar 0.1393 0.026 5.335 0.000 0.088 0.190
	543	+r_sunroom 0.0193 0.013 1.438 0.150 -0.007 0.046
	544	+r_mezzanine 0.0208 0.018 1.150 0.250 -0.015 0.056
	545	+r_library 0.1475 0.069 2.125 0.034 0.011 0.284
	546	+r_laundry 0.0363 0.006 6.025 0.000 0.025 0.048
	547	+g_cond 0.0119 0.004 3.005 0.003 0.004 0.020
	548	+g_hardwood 0.0315 0.008 3.705 0.000 0.015 0.048
	549	+g_quartz_granite 0.0177 0.008 2.091 0.037 0.001 0.034
	550	+g_view_water 0.0309 0.014 2.142 0.032 0.003 0.059
	551	+has_gpt 6.1815 0.057 108.357 0.000 6.070 6.293
	552	+g_turnkey 0.0044 0.006 0.692 0.489 -0.008 0.017
	553	+g_recently_renovated 0.0067 0.008 0.868 0.385 -0.008 0.022
	554	+g_garage 0.0193 0.018 1.076 0.282 -0.016 0.054
	555	+g_pool 0.0484 0.012 4.089 0.000 0.025 0.072
	556	+g_waterfront 0.0703 0.016 4.356 0.000 0.039 0.102
	557	+g_finished_basement -0.0274 0.007 -4.097 0.000 -0.040 -0.014
	558	+g_fireplace 0.0852 0.016 5.392 0.000 0.054 0.116
	559	+g_central_ac 0.0312 0.010 3.279 0.001 0.013 0.050
	560	+g_new_construction 0.0628 0.013 4.949 0.000 0.038 0.088
	561	+g_income_property -0.0340 0.009 -3.966 0.000 -0.051 -0.017
	562	+g_renovated_kitchen 0.0338 0.008 4.022 0.000 0.017 0.050
	563	+g_renovated_bathroom -0.0109 0.008 -1.389 0.165 -0.026 0.004
	564	+g_large_lot 0.0117 0.007 1.781 0.075 -0.001 0.025
	565	+g_quiet_area -0.0209 0.006 -3.661 0.000 -0.032 -0.010
	566	+==============================================================================
	567	+Omnibus: 2524.251 Durbin-Watson: 1.763
	568	+Prob(Omnibus): 0.000 Jarque-Bera (JB): 16663.299
	569	+Skew: 0.454 Prob(JB): 0.00
	570	+Kurtosis: 7.482 Cond. No. 9.65e+16
	571	+==============================================================================
	572	+
	573	+Notes:
	574	+[1] Standard Errors are heteroscedasticity robust (HC3)
	575	+[2] The smallest eigenvalue is 4.82e-23. This might indicate that there are
	576	+strong multicollinearity problems or that the design matrix is singular.
		\ No newline at end of file

added results/revision_extras.json +486 −0

		`@@ -0,0 +1,486 @@`
	1	+{
	2	+ "modelA": {
	3	+ "r2": 0.7552103110293896,
	4	+ "adj": 0.7477317389713678,
	5	+ "n": 9446,
	6	+ "coef": {
	7	+ "ln_sqft": {
	8	+ "b": 0.3583611466783186,
	9	+ "se_hc3": 0.013021312719264802,
	10	+ "p_hc3": 9.811436146831853e-167,
	11	+ "se_cl": 0.011546608518322881,
	12	+ "p_cl": 1.7598083257322608e-211
	13	+ },
	14	+ "ln_lot": {
	15	+ "b": 0.027321695727434744,
	16	+ "se_hc3": 0.003968643673932088,
	17	+ "p_hc3": 5.803517449373336e-12,
	18	+ "se_cl": 0.0057217322173256846,
	19	+ "p_cl": 1.7964103147193876e-06
	20	+ },
	21	+ "beds": {
	22	+ "b": 0.013623293126953095,
	23	+ "se_hc3": 0.0040602600210150885,
	24	+ "p_hc3": 0.0007928582892261605,
	25	+ "se_cl": 0.00457074886728697,
	26	+ "p_cl": 0.002877420342379003
	27	+ },
	28	+ "baths": {
	29	+ "b": 0.07348769487973825,
	30	+ "se_hc3": 0.006121072722376004,
	31	+ "p_hc3": 3.3169032296614814e-33,
	32	+ "se_cl": 0.0069532806823969125,
	33	+ "p_cl": 4.1586858986169373e-26
	34	+ },
	35	+ "parking": {
	36	+ "b": 0.01149153979340771,
	37	+ "se_hc3": 0.0015211101627679384,
	38	+ "p_hc3": 4.198076313292844e-14,
	39	+ "se_cl": 0.0014144294006288624,
	40	+ "p_cl": 4.491890458706105e-16
	41	+ },
	42	+ "age": {
	43	+ "b": -0.003778162499817968,
	44	+ "se_hc3": 0.0004166167508553082,
	45	+ "p_hc3": 1.2047149502398342e-19,
	46	+ "se_cl": 0.0010761773989556892,
	47	+ "p_cl": 0.00044688706454431995
	48	+ },
	49	+ "age2": {
	50	+ "b": 2.0412213402030788e-05,
	51	+ "se_hc3": 3.1634803715459156e-06,
	52	+ "p_hc3": 1.1005287533895908e-10,
	53	+ "se_cl": 4.403903629319602e-06,
	54	+ "p_cl": 3.568904338805579e-06
	55	+ },
	56	+ "central_air": {
	57	+ "b": 0.053384334694057345,
	58	+ "se_hc3": 0.011164185372283303,
	59	+ "p_hc3": 1.737763866667577e-06,
	60	+ "se_cl": 0.017763303901713427,
	61	+ "p_cl": 0.0026530539182792814
	62	+ },
	63	+ "transit": {
	64	+ "b": 0.09000078993798896,
	65	+ "se_hc3": 0.00842353892882497,
	66	+ "p_hc3": 1.2037429378492829e-26,
	67	+ "se_cl": 0.04560367671561142,
	68	+ "p_cl": 0.04843379796570755
	69	+ },
	70	+ "waterfront": {
	71	+ "b": 0.05808657079776136,
	72	+ "se_hc3": 0.04276278415929502,
	73	+ "p_hc3": 0.17435455833645552,
	74	+ "se_cl": 0.033496164012264705,
	75	+ "p_cl": 0.08289571923560418
	76	+ },
	77	+ "g_cond": {
	78	+ "b": 0.016168128372704954,
	79	+ "se_hc3": 0.004284326158407061,
	80	+ "p_hc3": 0.0001607889544697841,
	81	+ "se_cl": 0.0058314343073510595,
	82	+ "p_cl": 0.005561354758076747
	83	+ },
	84	+ "g_renovated_kitchen": {
	85	+ "b": 0.026089835765711222,
	86	+ "se_hc3": 0.009731092534231904,
	87	+ "p_hc3": 0.007338501052577358,
	88	+ "se_cl": 0.006899975364881851,
	89	+ "p_cl": 0.00015610612945056893
	90	+ },
	91	+ "g_hardwood": {
	92	+ "b": 0.023654684502188772,
	93	+ "se_hc3": 0.007568398763791719,
	94	+ "p_hc3": 0.0017753064896647113,
	95	+ "se_cl": 0.009325076436933125,
	96	+ "p_cl": 0.011191092284423674
	97	+ },
	98	+ "g_new_construction": {
	99	+ "b": -0.00030428071460375516,
	100	+ "se_hc3": 0.013498897144460655,
	101	+ "p_hc3": 0.9820162844667388,
	102	+ "se_cl": 0.013315085313141847,
	103	+ "p_cl": 0.9817680661387342
	104	+ },
	105	+ "g_income_property": {
	106	+ "b": -0.01916827298970958,
	107	+ "se_hc3": 0.010125989321693481,
	108	+ "p_hc3": 0.05836082179559654,
	109	+ "se_cl": 0.009252803603413885,
	110	+ "p_cl": 0.03830109511355022
	111	+ },
	112	+ "v_curb": {
	113	+ "b": 0.023017903057193298,
	114	+ "se_hc3": 0.007507554436150879,
	115	+ "p_hc3": 0.0021696833221349896,
	116	+ "se_cl": 0.00865392944593514,
	117	+ "p_cl": 0.007818219565218354
	118	+ },
	119	+ "v_reno": {
	120	+ "b": 0.023593099837405124,
	121	+ "se_hc3": 0.007459829833637167,
	122	+ "p_hc3": 0.0015632080662522378,
	123	+ "se_cl": 0.009529503385437559,
	124	+ "p_cl": 0.013293976823511309
	125	+ },
	126	+ "v_luxury": {
	127	+ "b": 0.05100475597004259,
	128	+ "se_hc3": 0.009917011195885339,
	129	+ "p_hc3": 2.7015815598254944e-07,
	130	+ "se_cl": 0.011932756209389966,
	131	+ "p_cl": 1.916971879355069e-05
	132	+ },
	133	+ "v_garage_vis": {
	134	+ "b": 0.04626239493908085,
	135	+ "se_hc3": 0.007095142592775905,
	136	+ "p_hc3": 7.017108386426899e-11,
	137	+ "se_cl": 0.005814520452429543,
	138	+ "p_cl": 1.7718006523061964e-15
	139	+ },
	140	+ "v_mat_brique": {
	141	+ "b": 0.033343794442173225,
	142	+ "se_hc3": 0.00827594130032085,
	143	+ "p_hc3": 5.601384271322806e-05,
	144	+ "se_cl": 0.009143815895444862,
	145	+ "p_cl": 0.0002657381339945666
	146	+ },
	147	+ "v_mat_pierre": {
	148	+ "b": 0.03615198096489164,
	149	+ "se_hc3": 0.007737924779006411,
	150	+ "p_hc3": 2.982065560443421e-06,
	151	+ "se_cl": 0.008011891510590393,
	152	+ "p_cl": 6.413130561843427e-06
	153	+ },
	154	+ "v_mat_vinyle": {
	155	+ "b": -0.04493374989989597,
	156	+ "se_hc3": 0.007597159274294025,
	157	+ "p_hc3": 3.3279171858711188e-09,
	158	+ "se_cl": 0.015100349535056819,
	159	+ "p_cl": 0.0029234340683073927
	160	+ },
	161	+ "v_set_rural": {
	162	+ "b": -0.12230705352273197,
	163	+ "se_hc3": 0.04529347936059,
	164	+ "p_hc3": 0.006927205074830596,
	165	+ "se_cl": 0.032184971001596396,
	166	+ "p_cl": 0.00014462056902901985
	167	+ },
	168	+ "v_land": {
	169	+ "b": 0.015079214851207945,
	170	+ "se_hc3": 0.005163307230556136,
	171	+ "p_hc3": 0.003495188385196164,
	172	+ "se_cl": 0.0062394992136362536,
	173	+ "p_cl": 0.015660422546723243
	174	+ },
	175	+ "v_cond": {
	176	+ "b": 0.011435724114234713,
	177	+ "se_hc3": 0.009935939848613642,
	178	+ "p_hc3": 0.24975470480811712,
	179	+ "se_cl": 0.00793900430747386,
	180	+ "p_cl": 0.1497406477456507
	181	+ }
	182	+ }
	183	+ },
	184	+ "modelB": {
	185	+ "r2": 0.7250019979121325,
	186	+ "adj": 0.7209302236414477,
	187	+ "n": 19123,
	188	+ "coef": {
	189	+ "ln_lot": {
	190	+ "b": 0.040421819255842366,
	191	+ "se_hc3": 0.0030793415163729103,
	192	+ "p_hc3": 2.3130287315281885e-39,
	193	+ "se_cl": 0.005130684552116609,
	194	+ "p_cl": 3.314786914788146e-15
	195	+ },
	196	+ "beds": {
	197	+ "b": 0.03225644920405396,
	198	+ "se_hc3": 0.0037380829672411845,
	199	+ "p_hc3": 6.181384242360143e-18,
	200	+ "se_cl": 0.0026461656800326736,
	201	+ "p_cl": 3.5192764358646113e-34
	202	+ },
	203	+ "baths": {
	204	+ "b": 0.11334162478693238,
	205	+ "se_hc3": 0.004820025879176098,
	206	+ "p_hc3": 2.882946420348713e-122,
	207	+ "se_cl": 0.0065680707475647485,
	208	+ "p_cl": 1.0006690034057575e-66
	209	+ },
	210	+ "parking": {
	211	+ "b": 0.014887132710857906,
	212	+ "se_hc3": 0.001722223609697608,
	213	+ "p_hc3": 5.4214611052329344e-18,
	214	+ "se_cl": 0.0019170158855884335,
	215	+ "p_cl": 8.114125760065666e-15
	216	+ },
	217	+ "age": {
	218	+ "b": -0.00454728435076453,
	219	+ "se_hc3": 0.00031568867613853834,
	220	+ "p_hc3": 4.859972485857048e-47,
	221	+ "se_cl": 0.0005421521451234677,
	222	+ "p_cl": 4.9671978979020763e-17
	223	+ },
	224	+ "age2": {
	225	+ "b": 2.4738741307502545e-05,
	226	+ "se_hc3": 2.2589789920551392e-06,
	227	+ "p_hc3": 6.55056354106643e-28,
	228	+ "se_cl": 1.6280767754927894e-06,
	229	+ "p_cl": 3.8125346654408024e-52
	230	+ },
	231	+ "central_air": {
	232	+ "b": 0.020145753639079278,
	233	+ "se_hc3": 0.009178809775081053,
	234	+ "p_hc3": 0.028177150453293415,
	235	+ "se_cl": 0.017821940423307536,
	236	+ "p_cl": 0.25831168008287975
	237	+ },
	238	+ "transit": {
	239	+ "b": 0.058387955324457536,
	240	+ "se_hc3": 0.006823518187251239,
	241	+ "p_hc3": 1.1597359094737698e-17,
	242	+ "se_cl": 0.039429380632120564,
	243	+ "p_cl": 0.1386535998125059
	244	+ },
	245	+ "waterfront": {
	246	+ "b": 0.061910921726998136,
	247	+ "se_hc3": 0.023452608193077887,
	248	+ "p_hc3": 0.008294738330249088,
	249	+ "se_cl": 0.019005260291841415,
	250	+ "p_cl": 0.0011237145722294154
	251	+ },
	252	+ "g_cond": {
	253	+ "b": 0.016317524818277362,
	254	+ "se_hc3": 0.003676322607356966,
	255	+ "p_hc3": 9.056887962882854e-06,
	256	+ "se_cl": 0.0051984051513270825,
	257	+ "p_cl": 0.0016955538138371947
	258	+ },
	259	+ "g_renovated_kitchen": {
	260	+ "b": 0.02694519585550771,
	261	+ "se_hc3": 0.00778196222799486,
	262	+ "p_hc3": 0.0005351431624444984,
	263	+ "se_cl": 0.00963562440448091,
	264	+ "p_cl": 0.005167314672277988
	265	+ },
	266	+ "g_hardwood": {
	267	+ "b": 0.029508607751202176,
	268	+ "se_hc3": 0.007732639804181234,
	269	+ "p_hc3": 0.000135571823960277,
	270	+ "se_cl": 0.010501088270130441,
	271	+ "p_cl": 0.004953344261501498
	272	+ },
	273	+ "g_new_construction": {
	274	+ "b": -0.015913022424995732,
	275	+ "se_hc3": 0.012366399085437902,
	276	+ "p_hc3": 0.19816569619711843,
	277	+ "se_cl": 0.013844198802557594,
	278	+ "p_cl": 0.2503761852618782
	279	+ },
	280	+ "g_income_property": {
	281	+ "b": -0.012880740908525923,
	282	+ "se_hc3": 0.008090151720423983,
	283	+ "p_hc3": 0.11135084238149032,
	284	+ "se_cl": 0.007646640747981612,
	285	+ "p_cl": 0.09208572167107143
	286	+ },
	287	+ "v_curb": {
	288	+ "b": 0.01595794997392595,
	289	+ "se_hc3": 0.006208945541452705,
	290	+ "p_hc3": 0.010165313745874592,
	291	+ "se_cl": 0.013160921694546399,
	292	+ "p_cl": 0.22531138001532158
	293	+ },
	294	+ "v_reno": {
	295	+ "b": 0.01998000405352155,
	296	+ "se_hc3": 0.005938481115424498,
	297	+ "p_hc3": 0.0007668321703145232,
	298	+ "se_cl": 0.00862597465890367,
	299	+ "p_cl": 0.020544051694669916
	300	+ },
	301	+ "v_luxury": {
	302	+ "b": 0.061660294830383405,
	303	+ "se_hc3": 0.009917957983108048,
	304	+ "p_hc3": 5.066358372641718e-10,
	305	+ "se_cl": 0.008697717532597092,
	306	+ "p_cl": 1.3484065713602217e-12
	307	+ },
	308	+ "v_garage_vis": {
	309	+ "b": 0.04651267305326928,
	310	+ "se_hc3": 0.005473235777759556,
	311	+ "p_hc3": 1.9254685227194248e-17,
	312	+ "se_cl": 0.0067205744806021995,
	313	+ "p_cl": 4.4866578070246505e-12
	314	+ },
	315	+ "v_mat_brique": {
	316	+ "b": 0.04403606617642199,
	317	+ "se_hc3": 0.006434026368289839,
	318	+ "p_hc3": 7.687913299168285e-12,
	319	+ "se_cl": 0.011001756372510255,
	320	+ "p_cl": 6.263967386269365e-05
	321	+ },
	322	+ "v_mat_pierre": {
	323	+ "b": 0.04658529674361125,
	324	+ "se_hc3": 0.006389001415391058,
	325	+ "p_hc3": 3.0656103177722586e-13,
	326	+ "se_cl": 0.004986368566238395,
	327	+ "p_cl": 9.405939222542233e-21
	328	+ },
	329	+ "v_mat_vinyle": {
	330	+ "b": -0.06180816681575538,
	331	+ "se_hc3": 0.0059550197354287316,
	332	+ "p_hc3": 3.0844255288655324e-25,
	333	+ "se_cl": 0.012339743471244337,
	334	+ "p_cl": 5.475063124724954e-07
	335	+ },
	336	+ "v_set_rural": {
	337	+ "b": -0.07600759882521751,
	338	+ "se_hc3": 0.03798172502754968,
	339	+ "p_hc3": 0.04537489488193291,
	340	+ "se_cl": 0.02898506811408183,
	341	+ "p_cl": 0.008733802916670029
	342	+ },
	343	+ "v_land": {
	344	+ "b": 0.025364605559571263,
	345	+ "se_hc3": 0.004436697037848911,
	346	+ "p_hc3": 1.0841988159731501e-08,
	347	+ "se_cl": 0.004475099360773844,
	348	+ "p_cl": 1.4452250155571873e-08
	349	+ },
	350	+ "v_cond": {
	351	+ "b": 0.03275330527845849,
	352	+ "se_hc3": 0.008042409776296826,
	353	+ "p_hc3": 4.649651177320347e-05,
	354	+ "se_cl": 0.00741253911271998,
	355	+ "p_cl": 9.932619742557207e-06
	356	+ }
	357	+ }
	358	+ },
	359	+ "ablationA_insample": {
	360	+ "base": 0.7168982211740981,
	361	+ "plus_text": 0.7252251732782795,
	362	+ "plus_text_vision": 0.7552103110293896
	363	+ },
	364	+ "city_fe_top": [
	365	+ [
	366	+ "westmount",
	367	+ 230.7041938252226
	368	+ ],
	369	+ [
	370	+ "hampstead",
	371	+ 173.35777258613155
	372	+ ],
	373	+ [
	374	+ "mont-royal",
	375	+ 168.59619659817113
	376	+ ],
	377	+ [
	378	+ "sainte-anne-des-plaines",
	379	+ 142.9419234222625
	380	+ ],
	381	+ [
	382	+ "montr\u00e9al",
	383	+ 83.5002446059172
	384	+ ],
	385	+ [
	386	+ "saint-lambert",
	387	+ 81.58672431832005
	388	+ ],
	389	+ [
	390	+ "beaconsfield",
	391	+ 80.76456609139153
	392	+ ],
	393	+ [
	394	+ "c\u00f4te-saint-luc",
	395	+ 76.87439413605533
	396	+ ],
	397	+ [
	398	+ "saint-bruno-de-montarville",
	399	+ 70.85936417552094
	400	+ ],
	401	+ [
	402	+ "dorval",
	403	+ 68.65665951427297
	404	+ ],
	405	+ [
	406	+ "lac-brome",
	407	+ 62.14364632969884
	408	+ ],
	409	+ [
	410	+ "pointe-claire",
	411	+ 55.677246677163005
	412	+ ]
	413	+ ],
	414	+ "city_fe_bot": [
	415	+ [
	416	+ "mansfield-et-pontefract",
	417	+ -30.06708220027833
	418	+ ],
	419	+ [
	420	+ "baie-comeau",
	421	+ -29.663061228503672
	422	+ ],
	423	+ [
	424	+ "t\u00e9miscouata-sur-le-lac",
	425	+ -28.11245530496782
	426	+ ],
	427	+ [
	428	+ "la tuque",
	429	+ -25.71095774442922
	430	+ ],
	431	+ [
	432	+ "saint-marc-des-carri\u00e8res",
	433	+ -25.50454270323863
	434	+ ],
	435	+ [
	436	+ "thetford mines",
	437	+ -24.723126467416044
	438	+ ],
	439	+ [
	440	+ "chandler",
	441	+ -22.75037961497599
	442	+ ],
	443	+ [
	444	+ "dolbeau-mistassini",
	445	+ -22.659205442967977
	446	+ ]
	447	+ ],
	448	+ "cv": {
	449	+ "ols_v4": {
	450	+ "r2": 0.6277548901390929,
	451	+ "mdae": 17.87702738532425,
	452	+ "w15": 42.92213564817236,
	453	+ "w25": 64.46687235266432
	454	+ },
	455	+ "ols_v5": {
	456	+ "r2": 0.6862361284588729,
	457	+ "mdae": 16.88625054752524,
	458	+ "w15": 45.25440568948387,
	459	+ "w25": 66.72070281859541
	460	+ },
	461	+ "gbm_v4": {
	462	+ "r2": 0.7417951731584451,
	463	+ "mdae": 15.490609148904856,
	464	+ "w15": 48.826021021806206,
	465	+ "w25": 69.8687444438634
	466	+ },
	467	+ "gbm_v5": {
	468	+ "r2": 0.7688886936527248,
	469	+ "mdae": 14.68537838434596,
	470	+ "w15": 50.75040527113946,
	471	+ "w25": 71.32249124091409
	472	+ },
	473	+ "rf_v4": {
	474	+ "r2": 0.6765341495717297,
	475	+ "mdae": 17.899891187986572,
	476	+ "w15": 42.84369607279193,
	477	+ "w25": 64.13219683104116
	478	+ },
	479	+ "rf_v5": {
	480	+ "r2": 0.7039491915758698,
	481	+ "mdae": 17.201324709196374,
	482	+ "w15": 44.77330962715055,
	483	+ "w25": 65.63823667834545
	484	+ }
	485	+ }
	486	+}
		\ No newline at end of file

added scripts/00a_test_openai_extraction.py +71 −0

		`@@ -0,0 +1,71 @@`
	1	+#!/usr/bin/env python3
	2	+# -- coding: utf-8 --
	3	+"""Test : extraction de specs via GPT-4o-mini (API REST, sans dépendance)."""
	4	+import json, os, urllib.request, urllib.error
	5	+
	6	+# charge la clé depuis .env
	7	+for line in open(".env", encoding="utf-8"):
	8	+ if line.startswith("OPENAI_API_KEY="):
	9	+ os.environ["OPENAI_API_KEY"] = line.split("=", 1)[1].strip()
	10	+KEY = os.environ["OPENAI_API_KEY"]
	11	+
	12	+SCHEMA = {
	13	+ "type": "object",
	14	+ "additionalProperties": False,
	15	+ "properties": {
	16	+ "condition": {"type": "string", "enum": ["a_renover","original","bon","renove","refait_neuf","neuf","inconnu"]},
	17	+ "turnkey": {"type": "boolean"},
	18	+ "recently_renovated": {"type": "boolean"},
	19	+ "flooring": {"type": "array", "items": {"type": "string", "enum": ["bois_franc","ceramique","plancher_flottant","tapis","vinyle","beton","autre"]}},
	20	+ "countertop": {"type": "string", "enum": ["granit","quartz","marbre","stratifie","bois","autre","inconnu"]},
	21	+ "heating": {"type": "array", "items": {"type": "string", "enum": ["electrique","plinthes","thermopompe","bois","gaz","mazout","radiant","geothermie","autre"]}},
	22	+ "garage": {"type": "boolean"},
	23	+ "pool": {"type": "boolean"},
	24	+ "waterfront": {"type": "boolean"},
	25	+ "view": {"type": "string", "enum": ["eau","montagne","ville","aucune","inconnu"]},
	26	+ "finished_basement":{"type": "boolean"},
	27	+ "fireplace": {"type": "boolean"},
	28	+ "central_ac": {"type": "boolean"},
	29	+ "new_construction":{"type": "boolean"},
	30	+ "income_property": {"type": "boolean"},
	31	+ "highlights": {"type": "array", "items": {"type": "string"}, "description": "3-6 mots-clés de vente saillants"}
	32	+ },
	33	+ "required": ["condition","turnkey","recently_renovated","flooring","countertop",
	34	+ "heating","garage","pool","waterfront","view","finished_basement",
	35	+ "fireplace","central_ac","new_construction","income_property","highlights"]
	36	+}
	37	+
	38	+SYS = ("Tu es un extracteur d'attributs immobiliers. À partir de la description "
	39	+ "d'une propriété (français ou anglais), renvoie les specs au format JSON "
	40	+ "strict. Base-toi UNIQUEMENT sur le texte ; si une info est absente, mets "
	41	+ "false / 'inconnu' / [] selon le champ. highlights = 3 à 6 mots-clés.")
	42	+
	43	+def extract(desc):
	44	+ body = {
	45	+ "model": "gpt-4o-mini",
	46	+ "temperature": 0,
	47	+ "messages": [{"role": "system", "content": SYS},
	48	+ {"role": "user", "content": desc[:4000]}],
	49	+ "response_format": {"type": "json_schema", "json_schema":
	50	+ {"name": "specs", "strict": True, "schema": SCHEMA}},
	51	+ }
	52	+ req = urllib.request.Request(
	53	+ "https://api.openai.com/v1/chat/completions",
	54	+ data=json.dumps(body).encode(),
	55	+ headers={"Authorization": f"Bearer {KEY}",
	56	+ "Content-Type": "application/json"})
	57	+ with urllib.request.urlopen(req, timeout=60) as r:
	58	+ out = json.load(r)
	59	+ return out["choices"][0]["message"]["content"], out.get("usage")
	60	+
	61	+# 3 descriptions du dataset
	62	+data = json.load(open("data/processed/maisons_avec_prix.json", encoding="utf-8"))
	63	+descs = [d["description"] for d in data if d.get("description")][:3]
	64	+for i, dsc in enumerate(descs):
	65	+ try:
	66	+ content, usage = extract(dsc)
	67	+ print(f"\n=== Exemple {i+1} (tokens: {usage}) ===")
	68	+ print("DESC:", dsc[:160].replace("\n", " "), "...")
	69	+ print("SPECS:", json.dumps(json.loads(content), ensure_ascii=False))
	70	+ except urllib.error.HTTPError as e:
	71	+ print("HTTP", e.code, e.read().decode()[:400]); break

added scripts/00b_test_vision_claude.py +65 −0

		`@@ -0,0 +1,65 @@`
	1	+#!/usr/bin/env python3
	2	+# -- coding: utf-8 --
	3	+"""Test : analyse d'images immobilières via Claude Haiku 4.5 (REST, sans SDK)."""
	4	+import json, os, urllib.request, urllib.error
	5	+
	6	+for line in open(".env", encoding="utf-8"):
	7	+ if line.startswith("ANTHROPIC_API_KEY="):
	8	+ KEY = line.split("=", 1)[1].strip()
	9	+
	10	+SCHEMA = {
	11	+ "type": "object", "additionalProperties": False,
	12	+ "properties": {
	13	+ "shot_type": {"type": "string", "enum": ["exterieur_facade","cuisine","salle_bain","salon","chambre","sous_sol","cour_terrain","vue","autre_interieur","autre"]},
	14	+ "condition_score": {"type": "integer", "description": "1=très délabré, 10=neuf/luxueux"},
	15	+ "renovation_level": {"type": "string", "enum": ["a_renover","date","standard","renove","haut_de_gamme","inconnu"]},
	16	+ "natural_light": {"type": "string", "enum": ["sombre","moyen","lumineux","tres_lumineux"]},
	17	+ "clutter": {"type": "string", "enum": ["vide","epure","normal","encombre"]},
	18	+ "curb_appeal": {"type": "integer", "description": "attrait extérieur 1-10; 0 si non applicable"},
	19	+ "materials": {"type": "array", "items": {"type": "string", "enum": ["bois_franc","ceramique","granit_quartz","stainless","brique","pierre","vinyle","beton","tapis"]}},
	20	+ "view_quality": {"type": "string", "enum": ["aucune","urbaine","verdure","eau","montagne","inconnu"]},
	21	+ "staged": {"type": "boolean", "description": "mise en scène pro / home staging"},
	22	+ "photo_quality": {"type": "integer", "description": "qualité de la photo elle-même 1-10"}
	23	+ },
	24	+ "required": ["shot_type","condition_score","renovation_level","natural_light","clutter","curb_appeal","materials","view_quality","staged","photo_quality"]
	25	+}
	26	+SYS = ("Tu es un évaluateur immobilier objectif. Analyse la photo et renvoie des "
	27	+ "attributs FACTUELS et reproductibles au format JSON. Sois objectif : base "
	28	+ "les scores sur des critères concrets (état des surfaces, finitions, "
	29	+ "luminosité), pas sur un goût personnel.")
	30	+
	31	+def analyze(url):
	32	+ body = {"model": "claude-haiku-4-5", "max_tokens": 400,
	33	+ "system": SYS,
	34	+ "messages": [{"role": "user", "content": [
	35	+ {"type": "image", "source": {"type": "url", "url": url}},
	36	+ {"type": "text", "text": "Analyse cette photo de propriété."}]}],
	37	+ "output_config": {"format": {"type": "json_schema", "schema": SCHEMA}}}
	38	+ req = urllib.request.Request(
	39	+ "https://api.anthropic.com/v1/messages", data=json.dumps(body).encode(),
	40	+ headers={"x-api-key": KEY, "anthropic-version": "2023-06-01",
	41	+ "content-type": "application/json"})
	42	+ with urllib.request.urlopen(req, timeout=90) as r:
	43	+ out = json.load(r)
	44	+ txt = next(b["text"] for b in out["content"] if b["type"] == "text")
	45	+ return json.loads(txt), out.get("usage")
	46	+
	47	+data_r = json.load(open("data/raw/dataset_realtor-property-crawler_2026-07-03_10-36-26-001.json", encoding="utf-8"))
	48	+data_d = json.load(open("data/raw/dataset_duproprio-property-crawler_2026-07-03_00-37-49-449.json", encoding="utf-8"))
	49	+# realtor: 1re image = façade ; duproprio: image dont caption ~ Kitchen
	50	+tests = []
	51	+r0 = next(d for d in data_r if d.get("images"))
	52	+tests.append(("REALTOR 1re photo (façade attendue)", r0["images"][0]["url"]))
	53	+d0 = next(d for d in data_d if d.get("images"))
	54	+kit = next((im for im in d0["images"] if "kitchen" in (im.get("caption") or "").lower()), d0["images"][0])
	55	+tests.append(("DUPROPRIO (caption=%s)" % (kit.get("caption","?")[:30]), kit["url"]))
	56	+
	57	+for label, url in tests:
	58	+ try:
	59	+ specs, usage = analyze(url)
	60	+ print(f"\n=== {label} ===")
	61	+ print("URL:", url[:90])
	62	+ print("tokens:", usage)
	63	+ print(json.dumps(specs, ensure_ascii=False, indent=1))
	64	+ except urllib.error.HTTPError as e:
	65	+ print("HTTP", e.code, e.read().decode()[:500])

added scripts/01_extract_specs_llm.py +135 −0

		`@@ -0,0 +1,135 @@`
	1	+#!/usr/bin/env python3
	2	+# -- coding: utf-8 --
	3	+"""
	4	+Extraction des specs de propriété via GPT-4o-mini, à partir des descriptions.
	5	+Parallèle, résumable (cache specs.jsonl), avec retry/backoff. Aucune dépendance.
	6	+"""
	7	+import json, os, time, threading, urllib.request, urllib.error
	8	+from concurrent.futures import ThreadPoolExecutor, as_completed
	9	+
	10	+SRC = "data/processed/maisons_avec_prix.json"
	11	+OUT = "data/processed/specs.jsonl"
	12	+WORKERS = int(os.environ.get("WORKERS", "24"))
	13	+NSHARDS = int(os.environ.get("NSHARDS", "1"))
	14	+SHARD = int(os.environ.get("SHARD", "0"))
	15	+MODEL = "gpt-4o-mini"
	16	+
	17	+# clé
	18	+for line in open(".env", encoding="utf-8"):
	19	+ if line.startswith("OPENAI_API_KEY="):
	20	+ os.environ["OPENAI_API_KEY"] = line.split("=", 1)[1].strip()
	21	+KEY = os.environ["OPENAI_API_KEY"]
	22	+
	23	+SCHEMA = {
	24	+ "type": "object", "additionalProperties": False,
	25	+ "properties": {
	26	+ "condition": {"type": "string", "enum": ["a_renover","original","bon","renove","refait_neuf","neuf","inconnu"]},
	27	+ "turnkey": {"type": "boolean"},
	28	+ "recently_renovated": {"type": "boolean"},
	29	+ "flooring": {"type": "array", "items": {"type": "string", "enum": ["bois_franc","ceramique","plancher_flottant","tapis","vinyle","beton","autre"]}},
	30	+ "countertop": {"type": "string", "enum": ["granit","quartz","marbre","stratifie","bois","autre","inconnu"]},
	31	+ "heating": {"type": "array", "items": {"type": "string", "enum": ["electrique","plinthes","thermopompe","bois","gaz","mazout","radiant","geothermie","autre"]}},
	32	+ "garage": {"type": "boolean"},
	33	+ "pool": {"type": "boolean"},
	34	+ "waterfront": {"type": "boolean"},
	35	+ "view": {"type": "string", "enum": ["eau","montagne","ville","aucune","inconnu"]},
	36	+ "finished_basement": {"type": "boolean"},
	37	+ "fireplace": {"type": "boolean"},
	38	+ "central_ac": {"type": "boolean"},
	39	+ "new_construction": {"type": "boolean"},
	40	+ "income_property": {"type": "boolean"},
	41	+ "renovated_kitchen": {"type": "boolean"},
	42	+ "renovated_bathroom": {"type": "boolean"},
	43	+ "large_lot": {"type": "boolean"},
	44	+ "quiet_area": {"type": "boolean"},
	45	+ "highlights": {"type": "array", "items": {"type": "string"}}
	46	+ },
	47	+ "required": ["condition","turnkey","recently_renovated","flooring","countertop",
	48	+ "heating","garage","pool","waterfront","view","finished_basement",
	49	+ "fireplace","central_ac","new_construction","income_property",
	50	+ "renovated_kitchen","renovated_bathroom","large_lot","quiet_area",
	51	+ "highlights"]
	52	+}
	53	+SYS = ("Tu es un extracteur d'attributs immobiliers. À partir de la description "
	54	+ "d'une propriété (français ou anglais), renvoie les specs au format JSON "
	55	+ "strict. Base-toi UNIQUEMENT sur le texte ; si une info est absente, mets "
	56	+ "false / 'inconnu' / [] selon le champ. highlights = 3 à 6 mots-clés de vente.")
	57	+
	58	+def call(desc, tries=5):
	59	+ body = {"model": MODEL, "temperature": 0,
	60	+ "messages": [{"role": "system", "content": SYS},
	61	+ {"role": "user", "content": desc[:4000]}],
	62	+ "response_format": {"type": "json_schema", "json_schema":
	63	+ {"name": "specs", "strict": True, "schema": SCHEMA}}}
	64	+ data = json.dumps(body).encode()
	65	+ for t in range(tries):
	66	+ try:
	67	+ req = urllib.request.Request(
	68	+ "https://api.openai.com/v1/chat/completions", data=data,
	69	+ headers={"Authorization": f"Bearer {KEY}",
	70	+ "Content-Type": "application/json"})
	71	+ with urllib.request.urlopen(req, timeout=90) as r:
	72	+ out = json.load(r)
	73	+ return json.loads(out["choices"][0]["message"]["content"])
	74	+ except urllib.error.HTTPError as e:
	75	+ if e.code in (429, 500, 502, 503, 504) and t < tries - 1:
	76	+ time.sleep(2 ** t + 0.5); continue
	77	+ raise
	78	+ except (urllib.error.URLError, TimeoutError):
	79	+ if t < tries - 1:
	80	+ time.sleep(2 ** t + 0.5); continue
	81	+ raise
	82	+ return None
	83	+
	84	+# charge données + cache
	85	+data = json.load(open(SRC, encoding="utf-8"))
	86	+done = set()
	87	+if os.path.exists(OUT):
	88	+ for line in open(OUT, encoding="utf-8"):
	89	+ try:
	90	+ done.add(json.loads(line)["uid"])
	91	+ except Exception:
	92	+ pass
	93	+
	94	+todo = []
	95	+for d in data:
	96	+ uid = f"{d.get('source')}:{d.get('listingId')}"
	97	+ desc = d.get("description")
	98	+ if uid in done or not desc:
	99	+ continue
	100	+ todo.append((uid, desc))
	101	+
	102	+# sharding : partition disjointe et complète par index (déterministe)
	103	+if NSHARDS > 1:
	104	+ todo = [t for i, t in enumerate(todo) if i % NSHARDS == SHARD]
	105	+
	106	+print(f"[shard {SHARD}/{NSHARDS}, workers={WORKERS}] Total={len(data)} "
	107	+ f"déjà fait={len(done)} à traiter (ce shard)={len(todo)}", flush=True)
	108	+
	109	+lock = threading.Lock()
	110	+counter = {"n": 0, "err": 0}
	111	+fh = open(OUT, "a", encoding="utf-8")
	112	+
	113	+def work(item):
	114	+ uid, desc = item
	115	+ try:
	116	+ specs = call(desc)
	117	+ rec = {"uid": uid, "specs": specs}
	118	+ except Exception as e:
	119	+ rec = {"uid": uid, "error": str(e)[:200]}
	120	+ with lock:
	121	+ fh.write(json.dumps(rec, ensure_ascii=False) + "\n")
	122	+ fh.flush()
	123	+ counter["n"] += 1
	124	+ if "error" in rec:
	125	+ counter["err"] += 1
	126	+ if counter["n"] % 250 == 0:
	127	+ print(f" {counter['n']}/{len(todo)} (err={counter['err']})",
	128	+ flush=True)
	129	+
	130	+t0 = time.time()
	131	+with ThreadPoolExecutor(max_workers=WORKERS) as ex:
	132	+ list(ex.map(work, todo))
	133	+fh.close()
	134	+print(f"Terminé : {counter['n']} traités, {counter['err']} erreurs, "
	135	+ f"{time.time()-t0:.0f}s", flush=True)

added scripts/02_facade_vision.py +135 −0

		`@@ -0,0 +1,135 @@`
	1	+#!/usr/bin/env python3
	2	+# -- coding: utf-8 --
	3	+"""
	4	+Analyse de la photo de façade (1/annonce) via Claude Haiku 4.5.
	5	+Attributs visuels OBJECTIFS. Parallèle, résumable (facade_specs.jsonl), retry.
	6	+"""
	7	+import json, os, time, threading, urllib.request, urllib.error
	8	+from concurrent.futures import ThreadPoolExecutor
	9	+
	10	+SRC = "data/processed/maisons_avec_prix.json"
	11	+OUT = "data/processed/facade_specs.jsonl"
	12	+WORKERS = int(os.environ.get("WORKERS", "16"))
	13	+NSHARDS = int(os.environ.get("NSHARDS", "1"))
	14	+SHARD = int(os.environ.get("SHARD", "0"))
	15	+MODEL = "claude-haiku-4-5"
	16	+
	17	+for line in open(".env", encoding="utf-8"):
	18	+ if line.startswith("ANTHROPIC_API_KEY="):
	19	+ KEY = line.split("=", 1)[1].strip()
	20	+
	21	+SCHEMA = {
	22	+ "type": "object", "additionalProperties": False,
	23	+ "properties": {
	24	+ "is_exterior": {"type": "boolean", "description": "la photo montre-t-elle l'extérieur/façade"},
	25	+ "building_type_visual": {"type": "string", "enum": ["maison_detachee","jumele","rangee","condo_immeuble","chalet","terrain","autre"]},
	26	+ "condition_score": {"type": "integer", "description": "état visuel 1=délabré à 10=neuf/luxueux"},
	27	+ "curb_appeal": {"type": "integer", "description": "attrait extérieur 1-10"},
	28	+ "renovation_level": {"type": "string", "enum": ["a_renover","date","standard","renove","haut_de_gamme","inconnu"]},
	29	+ "exterior_materials": {"type": "array", "items": {"type": "string", "enum": ["brique","pierre","vinyle","bois","aluminium","stucco","crepi","beton","autre"]}},
	30	+ "landscaping": {"type": "string", "enum": ["aucun","minimal","entretenu","amenage_soigne"]},
	31	+ "has_garage_visible": {"type": "boolean"},
	32	+ "storeys_visual": {"type": "string", "enum": ["1","1.5","2","3+","inconnu"]},
	33	+ "setting": {"type": "string", "enum": ["urbain_dense","banlieue","rural","bord_eau","boise","inconnu"]},
	34	+ "photo_quality": {"type": "integer", "description": "qualité de la photo 1-10"},
	35	+ "curb_luxury": {"type": "boolean", "description": "signes de luxe (grande taille, matériaux nobles, prestige)"}
	36	+ },
	37	+ "required": ["is_exterior","building_type_visual","condition_score","curb_appeal",
	38	+ "renovation_level","exterior_materials","landscaping","has_garage_visible",
	39	+ "storeys_visual","setting","photo_quality","curb_luxury"]
	40	+}
	41	+SYS = ("Tu es un évaluateur immobilier objectif. Analyse la photo (censée être la "
	42	+ "façade d'une propriété) et renvoie des attributs FACTUELS et reproductibles "
	43	+ "au format JSON. Fonde les scores sur des critères concrets (état des "
	44	+ "surfaces, matériaux, aménagement), pas sur un goût personnel. Si la photo "
	45	+ "n'est pas un extérieur, mets is_exterior=false et remplis au mieux.")
	46	+
	47	+ROOM_KW = ("kitchen","cuisine","bathroom","salle de bain","bedroom","chambre",
	48	+ "living","salon","basement","sous-sol","dining","salle à manger",
	49	+ "garage","office","bureau","laundry","closet","stair")
	50	+
	51	+def pick_facade(d):
	52	+ imgs = d.get("images") or []
	53	+ if not imgs:
	54	+ return None
	55	+ if d.get("source") == "realtor":
	56	+ return imgs[0].get("url") # 1re photo = extérieur
	57	+ # duproprio : caption = adresse (sans mot-clé de pièce) = photo extérieure
	58	+ for im in imgs:
	59	+ cap = (im.get("caption") or "").lower()
	60	+ if cap and not any(k in cap for k in ROOM_KW):
	61	+ return im.get("url")
	62	+ return imgs[0].get("url")
	63	+
	64	+def call(url, tries=5):
	65	+ body = {"model": MODEL, "max_tokens": 400, "system": SYS,
	66	+ "messages": [{"role": "user", "content": [
	67	+ {"type": "image", "source": {"type": "url", "url": url}},
	68	+ {"type": "text", "text": "Analyse cette photo de propriété."}]}],
	69	+ "output_config": {"format": {"type": "json_schema", "schema": SCHEMA}}}
	70	+ data = json.dumps(body).encode()
	71	+ for t in range(tries):
	72	+ try:
	73	+ req = urllib.request.Request(
	74	+ "https://api.anthropic.com/v1/messages", data=data,
	75	+ headers={"x-api-key": KEY, "anthropic-version": "2023-06-01",
	76	+ "content-type": "application/json"})
	77	+ with urllib.request.urlopen(req, timeout=120) as r:
	78	+ out = json.load(r)
	79	+ txt = next(b["text"] for b in out["content"] if b["type"] == "text")
	80	+ return json.loads(txt)
	81	+ except urllib.error.HTTPError as e:
	82	+ # 400 inclus : fetch d'URL image transitoire côté Anthropic sous charge
	83	+ if e.code in (400, 429, 500, 502, 503, 504, 529) and t < tries - 1:
	84	+ time.sleep(2 ** t + 1); continue
	85	+ raise
	86	+ except (urllib.error.URLError, TimeoutError, ValueError):
	87	+ if t < tries - 1:
	88	+ time.sleep(2 ** t + 1); continue
	89	+ raise
	90	+ return None
	91	+
	92	+data = json.load(open(SRC, encoding="utf-8"))
	93	+done = set()
	94	+if os.path.exists(OUT):
	95	+ for line in open(OUT, encoding="utf-8"):
	96	+ try: done.add(json.loads(line)["uid"])
	97	+ except Exception: pass
	98	+
	99	+todo = []
	100	+for d in data:
	101	+ uid = f"{d.get('source')}:{d.get('listingId')}"
	102	+ if uid in done:
	103	+ continue
	104	+ url = pick_facade(d)
	105	+ if url:
	106	+ todo.append((uid, url))
	107	+
	108	+if NSHARDS > 1:
	109	+ todo = [t for i, t in enumerate(todo) if i % NSHARDS == SHARD]
	110	+
	111	+print(f"[shard {SHARD}/{NSHARDS}, workers={WORKERS}] Total={len(data)} "
	112	+ f"déjà fait={len(done)} à traiter (ce shard)={len(todo)}", flush=True)
	113	+
	114	+lock = threading.Lock()
	115	+c = {"n": 0, "err": 0}
	116	+fh = open(OUT, "a", encoding="utf-8")
	117	+
	118	+def work(item):
	119	+ uid, url = item
	120	+ try:
	121	+ rec = {"uid": uid, "facade": call(url)}
	122	+ except Exception as e:
	123	+ rec = {"uid": uid, "error": str(e)[:200]}
	124	+ with lock:
	125	+ fh.write(json.dumps(rec, ensure_ascii=False) + "\n"); fh.flush()
	126	+ c["n"] += 1
	127	+ if "error" in rec: c["err"] += 1
	128	+ if c["n"] % 250 == 0:
	129	+ print(f" {c['n']}/{len(todo)} (err={c['err']})", flush=True)
	130	+
	131	+t0 = time.time()
	132	+with ThreadPoolExecutor(max_workers=WORKERS) as ex:
	133	+ list(ex.map(work, todo))
	134	+fh.close()
	135	+print(f"Terminé : {c['n']} traités, {c['err']} erreurs, {time.time()-t0:.0f}s", flush=True)

added scripts/03_merge_llm_features.py +84 −0

		`@@ -0,0 +1,84 @@`
	1	+#!/usr/bin/env python3
	2	+# -- coding: utf-8 --
	3	+"""Fusionne les features GPT (specs.jsonl) aux données maison (maisons_avec_prix.json)."""
	4	+import json, csv
	5	+
	6	+# 1) charger specs GPT
	7	+specs = {}
	8	+for line in open("data/processed/specs.jsonl", encoding="utf-8"):
	9	+ r = json.loads(line)
	10	+ if r.get("specs"):
	11	+ specs[r["uid"]] = r["specs"]
	12	+print(f"Specs GPT chargées : {len(specs)}")
	13	+
	14	+# 2) charger maisons
	15	+data = json.load(open("data/processed/maisons_avec_prix.json", encoding="utf-8"))
	16	+
	17	+# 3) JSON fusionné : record complet + gptSpecs imbriqué
	18	+merged = []
	19	+n_match = 0
	20	+for d in data:
	21	+ uid = f"{d.get('source')}:{d.get('listingId')}"
	22	+ d2 = dict(d)
	23	+ d2["uid"] = uid
	24	+ s = specs.get(uid)
	25	+ d2["gptSpecs"] = s # None si absent
	26	+ if s:
	27	+ n_match += 1
	28	+ merged.append(d2)
	29	+json.dump(merged, open("data/processed/maisons_specs.json", "w", encoding="utf-8"),
	30	+ ensure_ascii=False)
	31	+print(f"JSON fusionné -> maisons_specs.json ({len(merged)} records, "
	32	+ f"{n_match} avec specs)")
	33	+
	34	+# 4) CSV aplati pour analyse
	35	+BOOL = ["turnkey","recently_renovated","garage","pool","waterfront",
	36	+ "finished_basement","fireplace","central_ac","new_construction",
	37	+ "income_property","renovated_kitchen","renovated_bathroom",
	38	+ "large_lot","quiet_area"]
	39	+FLOOR = ["bois_franc","ceramique","plancher_flottant","tapis","vinyle","beton","autre"]
	40	+HEAT = ["electrique","plinthes","thermopompe","bois","gaz","mazout","radiant","geothermie","autre"]
	41	+
	42	+cols = (["uid","source","listingId","price","currency","city","province",
	43	+ "ptype","sqft_raw","lot_raw","beds","baths","yearBuilt","parking",
	44	+ "has_gpt","g_condition","g_countertop","g_view"]
	45	+ + ["g_"+b for b in BOOL]
	46	+ + ["fl_"+f for f in FLOOR]
	47	+ + ["heat_"+h for h in HEAT]
	48	+ + ["g_highlights"])
	49	+
	50	+with open("data/processed/maisons_specs.csv", "w", newline="", encoding="utf-8") as fh:
	51	+ w = csv.writer(fh)
	52	+ w.writerow(cols)
	53	+ for d in data:
	54	+ uid = f"{d.get('source')}:{d.get('listingId')}"
	55	+ p = d.get("property") or {}
	56	+ ad = d.get("address") or {}
	57	+ s = specs.get(uid)
	58	+ row = [uid, d.get("source"), d.get("listingId"), d.get("priceNumber"),
	59	+ d.get("currency"), ad.get("city"), ad.get("province"),
	60	+ (p.get("buildingType") or p.get("type")),
	61	+ p.get("squareFeet"), p.get("lotSize"),
	62	+ p.get("bedrooms"), p.get("bathrooms"), p.get("yearBuilt"),
	63	+ p.get("parkingSpaces"),
	64	+ 1 if s else 0,
	65	+ (s or {}).get("condition",""),
	66	+ (s or {}).get("countertop",""),
	67	+ (s or {}).get("view","")]
	68	+ row += [int(bool((s or {}).get(b))) if s else "" for b in BOOL]
	69	+ fl = set((s or {}).get("flooring") or [])
	70	+ row += [int(f in fl) if s else "" for f in FLOOR]
	71	+ ht = set((s or {}).get("heating") or [])
	72	+ row += [int(h in ht) if s else "" for h in HEAT]
	73	+ row += ["; ".join((s or {}).get("highlights") or [])]
	74	+ w.writerow(row)
	75	+print(f"CSV aplati -> maisons_specs.csv ({len(cols)} colonnes)")
	76	+
	77	+# 5) petit récap de distribution
	78	+import collections
	79	+cond = collections.Counter(s.get("condition") for s in specs.values())
	80	+print("\nRépartition 'condition' :", dict(cond))
	81	+for b in ["turnkey","recently_renovated","central_ac","pool","waterfront",
	82	+ "renovated_kitchen","income_property"]:
	83	+ n = sum(1 for s in specs.values() if s.get(b))
	84	+ print(f" {b:20s}: {n:5d} ({n/len(specs)*100:.1f}%)")

added scripts/04_hedonic_ols_base.py +338 −0

		`@@ -0,0 +1,338 @@`
	1	+#!/usr/bin/env python3
	2	+# -- coding: utf-8 --
	3	+"""
	4	+Nettoyage des aberrations + modèle hédonique de prix des maisons
	5	+(Realtor + DuProprio combinés). Log-linéaire, effets fixes ville, SE robustes.
	6	+"""
	7	+import json, re, warnings
	8	+import numpy as np
	9	+import pandas as pd
	10	+import statsmodels.formula.api as smf
	11	+
	12	+warnings.filterwarnings("ignore")
	13	+SRC = "data/processed/maisons_avec_prix.json"
	14	+CUR_YEAR = 2026
	15	+
	16	+# ---------- Chargement ----------
	17	+with open(SRC, encoding="utf-8") as f:
	18	+ data = json.load(f)
	19	+print(f"Chargé : {len(data)} enregistrements")
	20	+
	21	+# ---------- Parsers ----------
	22	+def parse_area(s):
	23	+ """Retourne une superficie en pieds carrés (ft²)."""
	24	+ if not s or not isinstance(s, str):
	25	+ return np.nan
	26	+ txt = s.lower()
	27	+ is_m2 = ("m2" in txt) or ("m²" in txt)
	28	+ # garder chiffres, points, virgules, espaces ; retirer unités
	29	+ num = re.sub(r"[^0-9.,\s]", "", txt).strip()
	30	+ num = num.replace(" ", "").replace(",", "") # séparateurs de milliers
	31	+ try:
	32	+ v = float(num)
	33	+ except ValueError:
	34	+ return np.nan
	35	+ if v <= 0:
	36	+ return np.nan
	37	+ return v * 10.7639 if is_m2 else v
	38	+
	39	+def clean_city(city, source):
	40	+ if not city or not isinstance(city, str):
	41	+ return None
	42	+ c = city.split("(")[0].strip() # retire quartiers realtor
	43	+ c = re.sub(r"\s+", " ", c).lower().strip()
	44	+ return c or None
	45	+
	46	+def to_int(x):
	47	+ try:
	48	+ return int(x)
	49	+ except (TypeError, ValueError):
	50	+ return np.nan
	51	+
	52	+# amenity keywords -> recherche dans features + description + title (bilingue)
	53	+AMENITIES = {
	54	+ "pool": ["swimming pool", "piscine", "inground pool", "pool"],
	55	+ "waterfront": ["waterfront", "bord de l'eau", "navigable", "water access",
	56	+ "lakefront", "riverfront", "au bord de", "waterfront"],
	57	+ "garage": ["garage"],
	58	+ "fireplace": ["fireplace", "foyer", "wood stove", "poêle"],
	59	+ "central_air": ["central air", "air conditionné", "climatisation", "a/c",
	60	+ "heat-pump", "thermopompe", "air exchanger"],
	61	+ "finished_bsmt":["totally finished", "finished basement", "sous-sol fini",
	62	+ "basement"],
	63	+ "near_water": ["near water", "waterfront", "navigable"],
	64	+ "cul_de_sac": ["cul-de-sac", "cul de sac"],
	65	+ "golf": ["golf"],
	66	+ "ski": ["ski"],
	67	+ "near_park": ["park", "parc"],
	68	+ "near_school": ["school", "école", "cegep", "university", "université",
	69	+ "elementary", "high school"],
	70	+ "near_highway":["highway", "autoroute"],
	71	+ "transit": ["public transit", "public transportation", "transport"],
	72	+ "no_rear_neighbor":["no backyard neighbors", "no rear neighbor",
	73	+ "sans voisin arrière"],
	74	+}
	75	+
	76	+# ---------- Parsing des pièces (Realtor) ----------
	77	+def _side_ft(side):
	78	+ s = side.lower()
	79	+ if "ft" in s:
	80	+ ft = re.search(r"([\d.]+)\s*ft", s)
	81	+ inch = re.search(r"([\d.]+)\s*in", s)
	82	+ return (float(ft.group(1)) if ft else 0) + \
	83	+ (float(inch.group(1)) / 12 if inch else 0)
	84	+ m = re.search(r"([\d.,]+)\s*m", s)
	85	+ if m:
	86	+ try:
	87	+ return float(m.group(1).replace(",", ".")) * 3.28084
	88	+ except ValueError:
	89	+ return None
	90	+ return None
	91	+
	92	+def _room_area(dim):
	93	+ if not dim or "x" not in dim:
	94	+ return None
	95	+ a, _, b = dim.partition("x")
	96	+ la, lb = _side_ft(a), _side_ft(b)
	97	+ if la and lb and 1 < la < 200 and 1 < lb < 200:
	98	+ return la * lb
	99	+ return None
	100	+
	101	+ROOM_FLAGS = { # pièces à valeur potentielle
	102	+ "r_office": ["office"],
	103	+ "r_family": ["family room"],
	104	+ "r_workshop": ["workshop"],
	105	+ "r_winecellar": ["wine cellar"],
	106	+ "r_sunroom": ["sunroom", "solarium"],
	107	+ "r_mezzanine": ["mezzanine"],
	108	+ "r_library": ["library"],
	109	+ "r_laundry": ["laundry"],
	110	+}
	111	+
	112	+def parse_rooms(rooms):
	113	+ out = {"room_area": np.nan, "n_rooms": 0, "n_bed_r": 0, "n_bath_r": 0,
	114	+ "has_primary_bed": 0, "has_basement": 0, "n_levels_r": 0}
	115	+ for k in ROOM_FLAGS:
	116	+ out[k] = 0
	117	+ if not rooms:
	118	+ return out
	119	+ area, levels = 0.0, set()
	120	+ for r in rooms:
	121	+ nm = (r.get("name") or "").lower()
	122	+ lv = (r.get("level") or "").lower()
	123	+ out["n_rooms"] += 1
	124	+ if lv:
	125	+ levels.add(lv)
	126	+ if "basement" in lv or "sous-sol" in lv:
	127	+ out["has_basement"] = 1
	128	+ if "primary bedroom" in nm:
	129	+ out["has_primary_bed"] = 1
	130	+ out["n_bed_r"] += 1
	131	+ elif "bedroom" in nm:
	132	+ out["n_bed_r"] += 1
	133	+ if "bathroom" in nm:
	134	+ out["n_bath_r"] += 1
	135	+ for k, kws in ROOM_FLAGS.items():
	136	+ if any(w in nm for w in kws):
	137	+ out[k] = 1
	138	+ a = _room_area(r.get("dimensions"))
	139	+ if a:
	140	+ area += a
	141	+ out["room_area"] = area if area > 0 else np.nan
	142	+ out["n_levels_r"] = len(levels)
	143	+ return out
	144	+
	145	+rows = []
	146	+for d in data:
	147	+ prop = d.get("property") or {}
	148	+ addr = d.get("address") or {}
	149	+ feats = d.get("features") or []
	150	+ blob = (" ".join(feats) + " " + (d.get("description") or "") + " " +
	151	+ (d.get("title") or "")).lower()
	152	+
	153	+ yb = to_int(prop.get("yearBuilt"))
	154	+ age = (CUR_YEAR - yb) if (not pd.isna(yb) and 1700 <= yb <= CUR_YEAR) else np.nan
	155	+
	156	+ ptype = prop.get("buildingType") or prop.get("type") # realtor/duproprio
	157	+ storeys = prop.get("storeys") or prop.get("levels")
	158	+ try:
	159	+ storeys = float(str(storeys).replace(",", "."))
	160	+ except (TypeError, ValueError):
	161	+ storeys = np.nan
	162	+
	163	+ rec = {
	164	+ "price": d.get("priceNumber"),
	165	+ "source": d.get("source"),
	166	+ "sqft": parse_area(prop.get("squareFeet")),
	167	+ "lot_sqft": parse_area(prop.get("lotSize")),
	168	+ "beds": to_int(prop.get("bedrooms")),
	169	+ "baths": to_int(prop.get("bathrooms")),
	170	+ "parking": to_int(prop.get("parkingSpaces")), # realtor seulement
	171	+ "storeys": storeys,
	172	+ "age": age,
	173	+ "ptype": (ptype or "Unknown").strip(),
	174	+ "city": clean_city(addr.get("city"), d.get("source")),
	175	+ "region": (addr.get("region") or "").strip().lower() or None,
	176	+ }
	177	+ for name, kws in AMENITIES.items():
	178	+ rec[name] = int(any(k in blob for k in kws))
	179	+ # signaux "autre chose" : photos + visite virtuelle
	180	+ rec["n_images"] = len(d.get("images") or [])
	181	+ rec["virtual_tour"] = int(bool(d.get("virtualTourUrl")))
	182	+ # détail des pièces
	183	+ rec.update(parse_rooms(d.get("rooms")))
	184	+ rows.append(rec)
	185	+
	186	+df = pd.DataFrame(rows)
	187	+print(f"DataFrame brut : {df.shape}")
	188	+
	189	+# ---------- Nettoyage des aberrations ----------
	190	+n0 = len(df)
	191	+df = df[df["price"].notna()]
	192	+# prix plausibles pour une maison au Québec
	193	+df = df[(df["price"] >= 50_000) & (df["price"] <= 15_000_000)]
	194	+# superficie habitable : borner si présente, garder les NA (imputés + flag)
	195	+df.loc[(df["sqft"] < 300) \| (df["sqft"] > 20_000), "sqft"] = np.nan
	196	+# terrain plausible (garder NA autorisé -> mais on impute plus bas)
	197	+df.loc[(df["lot_sqft"] < 200) \| (df["lot_sqft"] > 5_000_000), "lot_sqft"] = np.nan
	198	+# chambres / sdb plausibles
	199	+df = df[(df["beds"].between(0, 15)) \| (df["beds"].isna())]
	200	+df = df[(df["baths"].between(0, 12)) \| (df["baths"].isna())]
	201	+df.loc[df["beds"] > 15, "beds"] = np.nan
	202	+df.loc[df["baths"] > 12, "baths"] = np.nan
	203	+# prix au pied carré aberrant (sur les lignes qui ont une superficie réelle)
	204	+df["ppsf"] = df["price"] / df["sqft"]
	205	+lo, hi = df["ppsf"].quantile([0.005, 0.995])
	206	+bad = df["ppsf"].notna() & ((df["ppsf"] < lo) \| (df["ppsf"] > hi))
	207	+df = df[~bad]
	208	+print(f"Après nettoyage : {len(df)} (retirés : {n0 - len(df)}) "
	209	+ f"[filtre ppsf {lo:.0f}–{hi:.0f} $/ft² sur superficies réelles]")
	210	+
	211	+# ---------- Superficie : réelle > reconstruite via pièces > imputée ----------
	212	+# calibration : ratio médian (superficie réelle / somme des pièces) sur recouvrement
	213	+ov = df[df["sqft"].notna() & df["room_area"].notna() &
	214	+ (df["room_area"] > 200)]
	215	+ratio = (ov["sqft"] / ov["room_area"]).median()
	216	+print(f"Calibration surface pièces : ratio réel/pièces = {ratio:.3f} "
	217	+ f"(n={len(ov)})")
	218	+room_est = df["room_area"] * ratio
	219	+room_ok = df["room_area"].notna() & (room_est.between(300, 20_000))
	220	+
	221	+df["sqft_final"] = df["sqft"]
	222	+fill = df["sqft_final"].isna() & room_ok
	223	+df.loc[fill, "sqft_final"] = room_est[fill]
	224	+df["sqft_measured"] = df["sqft_final"].notna().astype(int) # réelle ou pièces
	225	+n_room_fill = int(fill.sum())
	226	+print(f"Superficies : réelles={int(df['sqft'].notna().sum())}, "
	227	+ f"complétées via pièces=+{n_room_fill}, "
	228	+ f"mesurées au total={int(df['sqft_measured'].sum())}/{len(df)}")
	229	+
	230	+# imputation du reste = médiane par type
	231	+df["has_sqft"] = df["sqft_measured"]
	232	+df["sqft_imp"] = df.groupby("ptype")["sqft_final"].transform(
	233	+ lambda s: s.fillna(s.median()))
	234	+df["sqft_imp"] = df["sqft_imp"].fillna(df["sqft_final"].median())
	235	+
	236	+df["has_lot"] = df["lot_sqft"].notna().astype(int)
	237	+df["lot_sqft"] = df["lot_sqft"].fillna(df["lot_sqft"].median())
	238	+df["has_age"] = df["age"].notna().astype(int)
	239	+df["age"] = df["age"].fillna(df["age"].median()).clip(lower=0)
	240	+df["beds"] = df["beds"].fillna(df["beds"].median())
	241	+df["baths"] = df["baths"].fillna(df["baths"].median())
	242	+df["parking"] = df["parking"].fillna(0)
	243	+df["storeys"] = df["storeys"].fillna(df["storeys"].median())
	244	+
	245	+df["ln_price"] = np.log(df["price"])
	246	+df["ln_sqft"] = np.log(df["sqft_imp"])
	247	+df["ln_lot"] = np.log(df["lot_sqft"])
	248	+df["age2"] = df["age"] ** 2
	249	+
	250	+# Effets fixes ville : regrouper les villes rares
	251	+vc = df["city"].value_counts()
	252	+keep = set(vc[vc >= 20].index)
	253	+df["city_fe"] = df["city"].where(df["city"].isin(keep), other="__other__")
	254	+df["city_fe"] = df["city_fe"].fillna("__other__")
	255	+print(f"Villes en effets fixes : {df['city_fe'].nunique()} "
	256	+ f"(seuil ≥20 listings ; {len(keep)} villes nommées)")
	257	+
	258	+# Type de propriété harmonisé (regrouper rares)
	259	+tc = df["ptype"].value_counts()
	260	+keept = set(tc[tc >= 30].index)
	261	+df["ptype_fe"] = df["ptype"].where(df["ptype"].isin(keept), other="Other")
	262	+
	263	+# ---------- Modèles hédoniques ----------
	264	+amen = list(AMENITIES.keys())
	265	+def pct(b):
	266	+ return (np.exp(b) - 1) * 100
	267	+
	268	+# variables "détail des pièces" + autres signaux (Realtor surtout)
	269	+df["n_images_c"] = df["n_images"].clip(upper=60) # écrête les extrêmes
	270	+room_vars = ["n_rooms", "n_levels_r", "has_primary_bed", "has_basement",
	271	+ "r_office", "r_family", "r_workshop", "r_winecellar",
	272	+ "r_sunroom", "r_mezzanine", "r_library", "r_laundry"]
	273	+extra = ["n_images_c", "virtual_tour"] + room_vars
	274	+
	275	+base = ("ln_lot + has_lot + beds + baths + parking + storeys "
	276	+ "+ age + age2 + has_age + C(source) + C(ptype_fe) + "
	277	+ + " + ".join(amen) + " + " + " + ".join(extra)
	278	+ + " + C(city_fe)")
	279	+fA = "ln_price ~ ln_sqft + " + base # riche (superficie réelle)
	280	+fB = "ln_price ~ " + base # large (sans superficie)
	281	+
	282	+dfA = df[df["has_sqft"] == 1].copy()
	283	+print(f"\nModèle A (avec superficie réelle) : N={len(dfA)}")
	284	+mA = smf.ols(fA, data=dfA).fit(cov_type="HC3")
	285	+print(f"Modèle B (sans superficie, couverture large) : N={len(df)}")
	286	+mB = smf.ols(fB, data=df).fit(cov_type="HC3")
	287	+
	288	+sdA, sdB = np.std(mA.resid), np.std(mB.resid)
	289	+
	290	+def report(m, sd, name):
	291	+ print("\n" + "=" * 70)
	292	+ print(f"{name} N={int(m.nobs)} R²={m.rsquared:.4f} "
	293	+ f"R²_adj={m.rsquared_adj:.4f} "
	294	+ f"erreur typique ±{(np.exp(sd)-1)*100:.1f}%")
	295	+ print("=" * 70)
	296	+ pr, pv = m.params, m.pvalues
	297	+ show = [p for p in pr.index if not p.startswith("C(city_fe)")]
	298	+ print(f"{'variable':<30}{'coef':>10}{'~%impact':>11}{'p':>9}")
	299	+ for p in show:
	300	+ st = "*" if pv[p]<.01 else "" if pv[p]<.05 else "*" if pv[p]<.1 else ""
	301	+ pc = f"{pct(pr[p]):+.1f}%" if abs(pr[p]) < 3 else ""
	302	+ print(f"{p:<30}{pr[p]:>10.4f}{pc:>11}{pv[p]:>9.3g} {st}")
	303	+
	304	+report(mA, sdA, "MODÈLE A — hédonique riche (superficie incluse)")
	305	+report(mB, sdB, "MODÈLE B — couverture large (sans superficie)")
	306	+
	307	+# ---------- Prédictions combinées ----------
	308	+# A pour les biens à superficie réelle, B sinon
	309	+predA = np.exp(mA.predict(df)) * np.exp(sdA**2 / 2)
	310	+predB = np.exp(mB.predict(df)) * np.exp(sdB**2 / 2)
	311	+df["pred_price"] = np.where(df["has_sqft"] == 1, predA, predB)
	312	+df["model_used"] = np.where(df["has_sqft"] == 1, "A", "B")
	313	+df["resid_pct"] = (df["price"] / df["pred_price"] - 1) * 100
	314	+
	315	+out_cols = (["source", "city", "ptype", "price", "pred_price", "resid_pct",
	316	+ "model_used", "sqft", "room_area", "sqft_final", "sqft_measured",
	317	+ "lot_sqft", "beds", "baths", "parking", "storeys", "age",
	318	+ "n_images", "virtual_tour"] + amen + room_vars)
	319	+df[out_cols].to_csv("data/processed/maisons_hedonic.csv", index=False)
	320	+print(f"\nDataset final + prédictions -> maisons_hedonic.csv ({len(df)} lignes)")
	321	+
	322	+# top villes (effets fixes du modèle A)
	323	+cityfe = {k.split("T.")[1].rstrip("]"): v for k, v in mA.params.items()
	324	+ if k.startswith("C(city_fe)")}
	325	+cf = pd.Series(cityfe).sort_values(ascending=False)
	326	+print("\nTop 12 villes — prime de localisation (Modèle A, vs référence):")
	327	+for c, b in cf.head(12).items():
	328	+ print(f" {c:<30}{pct(b):+.1f}%")
	329	+print("Bottom 6:")
	330	+for c, b in cf.tail(6).items():
	331	+ print(f" {c:<30}{pct(b):+.1f}%")
	332	+
	333	+with open("results/hedonic_summary.txt", "w", encoding="utf-8") as fh:
	334	+ fh.write("=== MODELE A (avec superficie) ===\n")
	335	+ fh.write(str(mA.summary()))
	336	+ fh.write("\n\n=== MODELE B (sans superficie) ===\n")
	337	+ fh.write(str(mB.summary()))
	338	+print("Résumés complets -> hedonic_summary.txt")

added scripts/05_hedonic_ols_v4_text.py +249 −0

		`@@ -0,0 +1,249 @@`
	1	+#!/usr/bin/env python3
	2	+# -- coding: utf-8 --
	3	+"""
	4	+Modèle hédonique v4 : ajoute les specs extraites par GPT-4o-mini (specs.jsonl)
	5	+au modèle v3. Réutilise la même logique de nettoyage/features + effets fixes.
	6	+"""
	7	+import json, re, warnings
	8	+import numpy as np
	9	+import pandas as pd
	10	+import statsmodels.formula.api as smf
	11	+
	12	+warnings.filterwarnings("ignore")
	13	+CUR_YEAR = 2026
	14	+
	15	+# ---------- charge specs GPT ----------
	16	+specs = {}
	17	+for line in open("data/processed/specs.jsonl", encoding="utf-8"):
	18	+ try:
	19	+ r = json.loads(line)
	20	+ except Exception:
	21	+ continue
	22	+ if "specs" in r and r.get("specs"):
	23	+ specs[r["uid"]] = r["specs"]
	24	+print(f"Specs GPT chargées : {len(specs)}")
	25	+
	26	+# specs -> features numériques (dummies + catégorielles réduites)
	27	+COND_SCORE = {"a_renover":0,"original":1,"bon":2,"renove":3,"refait_neuf":4,
	28	+ "neuf":5,"inconnu":2}
	29	+BOOL_SPECS = ["turnkey","recently_renovated","garage","pool","waterfront",
	30	+ "finished_basement","fireplace","central_ac","new_construction",
	31	+ "income_property","renovated_kitchen","renovated_bathroom",
	32	+ "large_lot","quiet_area"]
	33	+
	34	+def spec_feats(uid):
	35	+ s = specs.get(uid)
	36	+ out = {"g_" + b: 0 for b in BOOL_SPECS}
	37	+ out["g_cond"] = 2
	38	+ out["g_hardwood"] = 0; out["g_quartz_granite"] = 0; out["g_view_water"] = 0
	39	+ out["has_gpt"] = 0
	40	+ if not s:
	41	+ return out
	42	+ out["has_gpt"] = 1
	43	+ for b in BOOL_SPECS:
	44	+ out["g_" + b] = int(bool(s.get(b)))
	45	+ out["g_cond"] = COND_SCORE.get(s.get("condition", "inconnu"), 2)
	46	+ fl = s.get("flooring") or []
	47	+ out["g_hardwood"] = int("bois_franc" in fl)
	48	+ out["g_quartz_granite"] = int(s.get("countertop") in ("quartz", "granit", "marbre"))
	49	+ out["g_view_water"] = int(s.get("view") == "eau")
	50	+ return out
	51	+
	52	+# ===================== pipeline v3 (repris) =====================
	53	+SRC = "data/processed/maisons_avec_prix.json"
	54	+data = json.load(open(SRC, encoding="utf-8"))
	55	+
	56	+def parse_area(s):
	57	+ if not s or not isinstance(s, str): return np.nan
	58	+ txt = s.lower(); is_m2 = ("m2" in txt) or ("m²" in txt)
	59	+ num = re.sub(r"[^0-9.,\s]", "", txt).strip().replace(" ", "").replace(",", "")
	60	+ try: v = float(num)
	61	+ except ValueError: return np.nan
	62	+ if v <= 0: return np.nan
	63	+ return v * 10.7639 if is_m2 else v
	64	+
	65	+def clean_city(city, source):
	66	+ if not city or not isinstance(city, str): return None
	67	+ c = city.split("(")[0].strip(); c = re.sub(r"\s+", " ", c).lower().strip()
	68	+ return c or None
	69	+
	70	+def to_int(x):
	71	+ try: return int(x)
	72	+ except (TypeError, ValueError): return np.nan
	73	+
	74	+AMENITIES = {
	75	+ "pool":["swimming pool","piscine","inground pool","pool"],
	76	+ "waterfront":["waterfront","bord de l'eau","navigable","water access","lakefront","riverfront","au bord de"],
	77	+ "garage":["garage"],"fireplace":["fireplace","foyer","wood stove","poêle"],
	78	+ "central_air":["central air","air conditionné","climatisation","a/c","heat-pump","thermopompe","air exchanger"],
	79	+ "finished_bsmt":["totally finished","finished basement","sous-sol fini","basement"],
	80	+ "near_water":["near water","waterfront","navigable"],
	81	+ "cul_de_sac":["cul-de-sac","cul de sac"],"golf":["golf"],"ski":["ski"],
	82	+ "near_park":["park","parc"],
	83	+ "near_school":["school","école","cegep","university","université","elementary","high school"],
	84	+ "near_highway":["highway","autoroute"],
	85	+ "transit":["public transit","public transportation","transport"],
	86	+ "no_rear_neighbor":["no backyard neighbors","no rear neighbor","sans voisin arrière"],
	87	+}
	88	+
	89	+def _side_ft(side):
	90	+ s = side.lower()
	91	+ if "ft" in s:
	92	+ ft = re.search(r"([\d.]+)\sft", s); inch = re.search(r"([\d.]+)\sin", s)
	93	+ return (float(ft.group(1)) if ft else 0)+(float(inch.group(1))/12 if inch else 0)
	94	+ m = re.search(r"([\d.,]+)\s*m", s)
	95	+ if m:
	96	+ try: return float(m.group(1).replace(",", ".")) * 3.28084
	97	+ except ValueError: return None
	98	+ return None
	99	+
	100	+def _room_area(dim):
	101	+ if not dim or "x" not in dim: return None
	102	+ a, _, b = dim.partition("x"); la, lb = _side_ft(a), _side_ft(b)
	103	+ if la and lb and 1 < la < 200 and 1 < lb < 200: return la * lb
	104	+ return None
	105	+
	106	+ROOM_FLAGS = {"r_office":["office"],"r_family":["family room"],"r_workshop":["workshop"],
	107	+ "r_winecellar":["wine cellar"],"r_sunroom":["sunroom","solarium"],
	108	+ "r_mezzanine":["mezzanine"],"r_library":["library"],"r_laundry":["laundry"]}
	109	+
	110	+def parse_rooms(rooms):
	111	+ out = {"room_area":np.nan,"n_rooms":0,"has_primary_bed":0,"has_basement":0,"n_levels_r":0}
	112	+ for k in ROOM_FLAGS: out[k] = 0
	113	+ if not rooms: return out
	114	+ area, levels = 0.0, set()
	115	+ for r in rooms:
	116	+ nm = (r.get("name") or "").lower(); lv = (r.get("level") or "").lower()
	117	+ out["n_rooms"] += 1
	118	+ if lv: levels.add(lv)
	119	+ if "basement" in lv or "sous-sol" in lv: out["has_basement"] = 1
	120	+ if "primary bedroom" in nm: out["has_primary_bed"] = 1
	121	+ for k, kws in ROOM_FLAGS.items():
	122	+ if any(w in nm for w in kws): out[k] = 1
	123	+ a = _room_area(r.get("dimensions"))
	124	+ if a: area += a
	125	+ out["room_area"] = area if area > 0 else np.nan
	126	+ out["n_levels_r"] = len(levels)
	127	+ return out
	128	+
	129	+rows = []
	130	+for d in data:
	131	+ prop = d.get("property") or {}; addr = d.get("address") or {}
	132	+ feats = d.get("features") or []
	133	+ blob = (" ".join(feats)+" "+(d.get("description") or "")+" "+(d.get("title") or "")).lower()
	134	+ yb = to_int(prop.get("yearBuilt"))
	135	+ age = (CUR_YEAR-yb) if (not pd.isna(yb) and 1700 <= yb <= CUR_YEAR) else np.nan
	136	+ ptype = prop.get("buildingType") or prop.get("type")
	137	+ storeys = prop.get("storeys") or prop.get("levels")
	138	+ try: storeys = float(str(storeys).replace(",", "."))
	139	+ except (TypeError, ValueError): storeys = np.nan
	140	+ uid = f"{d.get('source')}:{d.get('listingId')}"
	141	+ rec = {"uid":uid,"price":d.get("priceNumber"),"source":d.get("source"),
	142	+ "sqft":parse_area(prop.get("squareFeet")),"lot_sqft":parse_area(prop.get("lotSize")),
	143	+ "beds":to_int(prop.get("bedrooms")),"baths":to_int(prop.get("bathrooms")),
	144	+ "parking":to_int(prop.get("parkingSpaces")),"storeys":storeys,"age":age,
	145	+ "ptype":(ptype or "Unknown").strip(),"city":clean_city(addr.get("city"),d.get("source"))}
	146	+ for name, kws in AMENITIES.items():
	147	+ rec[name] = int(any(k in blob for k in kws))
	148	+ rec["n_images"] = len(d.get("images") or [])
	149	+ rec["virtual_tour"] = int(bool(d.get("virtualTourUrl")))
	150	+ rec.update(parse_rooms(d.get("rooms")))
	151	+ rec.update(spec_feats(uid))
	152	+ rows.append(rec)
	153	+
	154	+df = pd.DataFrame(rows)
	155	+
	156	+# nettoyage identique v3
	157	+n0 = len(df)
	158	+df = df[df["price"].notna()]
	159	+df = df[(df["price"] >= 50_000) & (df["price"] <= 15_000_000)]
	160	+df.loc[(df["sqft"] < 300) \| (df["sqft"] > 20_000), "sqft"] = np.nan
	161	+df.loc[(df["lot_sqft"] < 200) \| (df["lot_sqft"] > 5_000_000), "lot_sqft"] = np.nan
	162	+df.loc[df["beds"] > 15, "beds"] = np.nan
	163	+df.loc[df["baths"] > 12, "baths"] = np.nan
	164	+df["ppsf"] = df["price"] / df["sqft"]
	165	+lo, hi = df["ppsf"].quantile([0.005, 0.995])
	166	+df = df[~(df["ppsf"].notna() & ((df["ppsf"] < lo) \| (df["ppsf"] > hi)))]
	167	+
	168	+# surface finale
	169	+ov = df[df["sqft"].notna() & df["room_area"].notna() & (df["room_area"] > 200)]
	170	+ratio = (ov["sqft"] / ov["room_area"]).median()
	171	+room_est = df["room_area"] * ratio
	172	+room_ok = df["room_area"].notna() & (room_est.between(300, 20_000))
	173	+df["sqft_final"] = df["sqft"]
	174	+fill = df["sqft_final"].isna() & room_ok
	175	+df.loc[fill, "sqft_final"] = room_est[fill]
	176	+df["has_sqft"] = df["sqft_final"].notna().astype(int)
	177	+df["sqft_imp"] = df.groupby("ptype")["sqft_final"].transform(lambda s: s.fillna(s.median()))
	178	+df["sqft_imp"] = df["sqft_imp"].fillna(df["sqft_final"].median())
	179	+
	180	+df["has_lot"] = df["lot_sqft"].notna().astype(int)
	181	+df["lot_sqft"] = df["lot_sqft"].fillna(df["lot_sqft"].median())
	182	+df["has_age"] = df["age"].notna().astype(int)
	183	+df["age"] = df["age"].fillna(df["age"].median()).clip(lower=0)
	184	+for c in ["beds","baths","storeys"]:
	185	+ df[c] = df[c].fillna(df[c].median())
	186	+df["parking"] = df["parking"].fillna(0)
	187	+df["ln_price"] = np.log(df["price"]); df["ln_sqft"] = np.log(df["sqft_imp"])
	188	+df["ln_lot"] = np.log(df["lot_sqft"]); df["age2"] = df["age"]**2
	189	+df["n_images_c"] = df["n_images"].clip(upper=60)
	190	+
	191	+vc = df["city"].value_counts(); keep = set(vc[vc >= 20].index)
	192	+df["city_fe"] = df["city"].where(df["city"].isin(keep), "__other__").fillna("__other__")
	193	+tc = df["ptype"].value_counts(); keept = set(tc[tc >= 30].index)
	194	+df["ptype_fe"] = df["ptype"].where(df["ptype"].isin(keept), "Other")
	195	+print(f"N nettoyé={len(df)} villes FE={df['city_fe'].nunique()} "
	196	+ f"avec specs GPT={int(df['has_gpt'].sum())}")
	197	+
	198	+# ---------- modèles ----------
	199	+amen = list(AMENITIES.keys())
	200	+room_vars = ["n_rooms","n_levels_r","has_primary_bed","has_basement","r_office",
	201	+ "r_family","r_workshop","r_winecellar","r_sunroom","r_mezzanine",
	202	+ "r_library","r_laundry"]
	203	+gpt_vars = ["g_cond","g_hardwood","g_quartz_granite","g_view_water","has_gpt"] + \
	204	+ ["g_" + b for b in BOOL_SPECS]
	205	+extra = ["n_images_c","virtual_tour"] + room_vars + gpt_vars
	206	+base = ("ln_lot + has_lot + beds + baths + parking + storeys + age + age2 + has_age "
	207	+ "+ C(source) + C(ptype_fe) + " + " + ".join(amen) + " + " + " + ".join(extra)
	208	+ + " + C(city_fe)")
	209	+fA = "ln_price ~ ln_sqft + " + base
	210	+fB = "ln_price ~ " + base
	211	+
	212	+def pct(b): return (np.exp(b)-1)*100
	213	+dfA = df[df["has_sqft"] == 1].copy()
	214	+mA = smf.ols(fA, data=dfA).fit(cov_type="HC3")
	215	+mB = smf.ols(fB, data=df).fit(cov_type="HC3")
	216	+sdA, sdB = np.std(mA.resid), np.std(mB.resid)
	217	+
	218	+def report(m, sd, name):
	219	+ print("\n" + "=" * 70)
	220	+ print(f"{name} N={int(m.nobs)} R²={m.rsquared:.4f} R²adj={m.rsquared_adj:.4f} "
	221	+ f"±{(np.exp(sd)-1)*100:.1f}%")
	222	+ print("=" * 70)
	223	+ pr, pv = m.params, m.pvalues
	224	+ print("-- specs GPT (impact %) --")
	225	+ for p in pr.index:
	226	+ if p.startswith("g_") or p == "has_gpt":
	227	+ st = "*" if pv[p]<.01 else "" if pv[p]<.05 else "*" if pv[p]<.1 else ""
	228	+ print(f" {p:<22}{pr[p]:>9.4f}{pct(pr[p]):>+8.1f}% p={pv[p]:.2g} {st}")
	229	+
	230	+report(mA, sdA, "MODÈLE A+GPT (avec superficie)")
	231	+report(mB, sdB, "MODÈLE B+GPT (couverture large)")
	232	+
	233	+predA = np.exp(mA.predict(df)) * np.exp(sdA**2/2)
	234	+predB = np.exp(mB.predict(df)) * np.exp(sdB**2/2)
	235	+df["pred_price"] = np.where(df["has_sqft"] == 1, predA, predB)
	236	+df["model_used"] = np.where(df["has_sqft"] == 1, "A", "B")
	237	+df["resid_pct"] = (df["price"]/df["pred_price"] - 1) * 100
	238	+print(f"\nErreur abs médiane: {df.resid_pct.abs().median():.1f}% "
	239	+ f"\| ±15%: {(df.resid_pct.abs()<=15).mean()*100:.0f}% "
	240	+ f"±25%: {(df.resid_pct.abs()<=25).mean()*100:.0f}%")
	241	+
	242	+out = (["uid","source","city","ptype","price","pred_price","resid_pct","model_used",
	243	+ "sqft_final","has_sqft","lot_sqft","beds","baths","parking","age","n_images",
	244	+ "virtual_tour"] + amen + room_vars + gpt_vars)
	245	+df[out].to_csv("data/processed/maisons_hedonic_v4.csv", index=False)
	246	+print(f"-> maisons_hedonic_v4.csv ({len(df)} lignes)")
	247	+with open("results/hedonic_v4_summary.txt", "w", encoding="utf-8") as fh:
	248	+ fh.write("=== A+GPT ===\n"+str(mA.summary())+"\n\n=== B+GPT ===\n"+str(mB.summary()))
	249	+print("-> hedonic_v4_summary.txt")

added scripts/06_hedonic_ml_cv.py +171 −0

		`@@ -0,0 +1,171 @@`
	1	+#!/usr/bin/env python3
	2	+# -- coding: utf-8 --
	3	+"""Comparaison OLS vs modèles ML (RF, Gradient Boosting) en validation croisée 5-fold.
	4	+Même jeu de variables que la v4 (structurel + aménités + pièces + listing + GPT + FE)."""
	5	+import json, re, warnings
	6	+import numpy as np, pandas as pd
	7	+from sklearn.linear_model import LinearRegression
	8	+from sklearn.ensemble import RandomForestRegressor, HistGradientBoostingRegressor
	9	+from sklearn.model_selection import KFold, cross_val_predict
	10	+from sklearn.metrics import r2_score
	11	+warnings.filterwarnings("ignore")
	12	+CUR_YEAR = 2026
	13	+
	14	+# ---- specs GPT ----
	15	+specs = {}
	16	+for l in open("data/processed/specs.jsonl", encoding="utf-8"):
	17	+ r = json.loads(l)
	18	+ if r.get("specs"): specs[r["uid"]] = r["specs"]
	19	+COND = {"a_renover":0,"original":1,"bon":2,"renove":3,"refait_neuf":4,"neuf":5,"inconnu":2}
	20	+BOOL = ["turnkey","recently_renovated","garage","pool","waterfront","finished_basement",
	21	+ "fireplace","central_ac","new_construction","income_property","renovated_kitchen",
	22	+ "renovated_bathroom","large_lot","quiet_area"]
	23	+def gfeat(uid):
	24	+ s = specs.get(uid); o = {"g_"+b:0 for b in BOOL}
	25	+ o.update({"g_cond":2,"g_hardwood":0,"g_quartz_granite":0,"g_view_water":0})
	26	+ if not s: return o
	27	+ for b in BOOL: o["g_"+b] = int(bool(s.get(b)))
	28	+ o["g_cond"] = COND.get(s.get("condition","inconnu"),2)
	29	+ fl = s.get("flooring") or []; o["g_hardwood"] = int("bois_franc" in fl)
	30	+ o["g_quartz_granite"] = int(s.get("countertop") in ("quartz","granit","marbre"))
	31	+ o["g_view_water"] = int(s.get("view")=="eau"); return o
	32	+
	33	+# ---- parsing (repris) ----
	34	+def parse_area(s):
	35	+ if not s or not isinstance(s,str): return np.nan
	36	+ t=s.lower(); m=("m2" in t) or ("m²" in t)
	37	+ n=re.sub(r"[^0-9.,\s]","",t).strip().replace(" ","").replace(",","")
	38	+ try: v=float(n)
	39	+ except: return np.nan
	40	+ return (v*10.7639 if m else v) if v>0 else np.nan
	41	+def cc(c):
	42	+ if not c or not isinstance(c,str): return None
	43	+ c=c.split("(")[0].strip(); return re.sub(r"\s+"," ",c).lower().strip() or None
	44	+def ti(x):
	45	+ try: return int(x)
	46	+ except: return np.nan
	47	+AMEN={"pool":["swimming pool","piscine","pool"],"waterfront":["waterfront","bord de l'eau","navigable","lakefront","riverfront","au bord de"],
	48	+"garage":["garage"],"fireplace":["fireplace","foyer","wood stove","poêle"],
	49	+"central_air":["central air","air conditionné","climatisation","a/c","heat-pump","thermopompe","air exchanger"],
	50	+"finished_bsmt":["totally finished","finished basement","sous-sol fini","basement"],
	51	+"near_water":["near water","waterfront","navigable"],"cul_de_sac":["cul-de-sac","cul de sac"],
	52	+"golf":["golf"],"ski":["ski"],"near_park":["park","parc"],
	53	+"near_school":["school","école","cegep","university","université","elementary","high school"],
	54	+"near_highway":["highway","autoroute"],"transit":["public transit","public transportation","transport"],
	55	+"no_rear_neighbor":["no backyard neighbors","sans voisin arrière"]}
	56	+def sd(s):
	57	+ s=s.lower()
	58	+ if "ft" in s:
	59	+ f=re.search(r"([\d.]+)\sft",s);i=re.search(r"([\d.]+)\sin",s)
	60	+ return (float(f.group(1)) if f else 0)+(float(i.group(1))/12 if i else 0)
	61	+ m=re.search(r"([\d.,]+)\s*m",s)
	62	+ if m:
	63	+ try: return float(m.group(1).replace(",","."))*3.28084
	64	+ except: return None
	65	+ return None
	66	+def ra(d):
	67	+ if not d or "x" not in d: return None
	68	+ a,_,b=d.partition("x");la,lb=sd(a),sd(b)
	69	+ return la*lb if (la and lb and 1<la<200 and 1<lb<200) else None
	70	+RF={"r_office":["office"],"r_family":["family room"],"r_workshop":["workshop"],"r_winecellar":["wine cellar"],
	71	+"r_sunroom":["sunroom","solarium"],"r_mezzanine":["mezzanine"],"r_library":["library"],"r_laundry":["laundry"]}
	72	+def pr(rooms):
	73	+ o={"room_area":np.nan,"n_rooms":0,"has_primary_bed":0,"has_basement":0,"n_levels_r":0}
	74	+ for k in RF: o[k]=0
	75	+ if not rooms: return o
	76	+ ar=0.0;lv=set()
	77	+ for r in rooms:
	78	+ nm=(r.get("name") or "").lower();l=(r.get("level") or "").lower();o["n_rooms"]+=1
	79	+ if l: lv.add(l)
	80	+ if "basement" in l or "sous-sol" in l: o["has_basement"]=1
	81	+ if "primary bedroom" in nm: o["has_primary_bed"]=1
	82	+ for k,kw in RF.items():
	83	+ if any(w in nm for w in kw): o[k]=1
	84	+ a=ra(r.get("dimensions"))
	85	+ if a: ar+=a
	86	+ o["room_area"]=ar if ar>0 else np.nan;o["n_levels_r"]=len(lv);return o
	87	+
	88	+data=json.load(open("data/processed/maisons_avec_prix.json",encoding="utf-8"))
	89	+rows=[]
	90	+for d in data:
	91	+ p=d.get("property") or {};ad=d.get("address") or {};ft=d.get("features") or []
	92	+ blob=(" ".join(ft)+" "+(d.get("description") or "")+" "+(d.get("title") or "")).lower()
	93	+ yb=ti(p.get("yearBuilt"));age=(CUR_YEAR-yb) if (not pd.isna(yb) and 1700<=yb<=CUR_YEAR) else np.nan
	94	+ pt=p.get("buildingType") or p.get("type");st=p.get("storeys") or p.get("levels")
	95	+ try: st=float(str(st).replace(",","."))
	96	+ except: st=np.nan
	97	+ uid=f"{d.get('source')}:{d.get('listingId')}"
	98	+ rec={"uid":uid,"price":d.get("priceNumber"),"source":d.get("source"),"sqft":parse_area(p.get("squareFeet")),
	99	+ "lot_sqft":parse_area(p.get("lotSize")),"beds":ti(p.get("bedrooms")),"baths":ti(p.get("bathrooms")),
	100	+ "parking":ti(p.get("parkingSpaces")),"storeys":st,"age":age,"ptype":(pt or "Unknown").strip(),"city":cc(ad.get("city"))}
	101	+ for n,kw in AMEN.items(): rec[n]=int(any(k in blob for k in kw))
	102	+ rec["n_images"]=len(d.get("images") or []);rec["virtual_tour"]=int(bool(d.get("virtualTourUrl")))
	103	+ rec.update(pr(d.get("rooms")));rec.update(gfeat(uid));rows.append(rec)
	104	+df=pd.DataFrame(rows)
	105	+
	106	+# nettoyage (identique)
	107	+df=df[df["price"].notna()];df=df[(df["price"]>=50_000)&(df["price"]<=15_000_000)]
	108	+df.loc[(df["sqft"]<300)\|(df["sqft"]>20_000),"sqft"]=np.nan
	109	+df.loc[(df["lot_sqft"]<200)\|(df["lot_sqft"]>5_000_000),"lot_sqft"]=np.nan
	110	+df.loc[df["beds"]>15,"beds"]=np.nan;df.loc[df["baths"]>12,"baths"]=np.nan
	111	+df["ppsf"]=df["price"]/df["sqft"];lo,hi=df["ppsf"].quantile([0.005,0.995])
	112	+df=df[~(df["ppsf"].notna()&((df["ppsf"]<lo)\|(df["ppsf"]>hi)))]
	113	+ov=df[df["sqft"].notna()&df["room_area"].notna()&(df["room_area"]>200)]
	114	+ratio=(ov["sqft"]/ov["room_area"]).median();re_=df["room_area"]*ratio
	115	+ok=df["room_area"].notna()&(re_.between(300,20000))
	116	+df["sqft_final"]=df["sqft"];fill=df["sqft_final"].isna()&ok;df.loc[fill,"sqft_final"]=re_[fill]
	117	+df["has_sqft"]=df["sqft_final"].notna().astype(int)
	118	+# pour ML : garder NaN (les arbres gèrent) ; pour OLS on impute une copie
	119	+df["ln_price"]=np.log(df["price"]);df["age2"]=df["age"]**2;df["n_images_c"]=df["n_images"].clip(upper=60)
	120	+vc=df["city"].value_counts();keep=set(vc[vc>=20].index)
	121	+df["city_fe"]=df["city"].where(df["city"].isin(keep),"__other__").fillna("__other__")
	122	+tc=df["ptype"].value_counts();df["ptype_fe"]=df["ptype"].where(df["ptype"].isin(set(tc[tc>=30].index)),"Other")
	123	+print(f"N={len(df)}")
	124	+
	125	+amen=list(AMEN.keys());rv=["n_rooms","n_levels_r","has_primary_bed","has_basement","r_office","r_family","r_workshop","r_winecellar","r_sunroom","r_mezzanine","r_library","r_laundry"]
	126	+gv=["g_cond","g_hardwood","g_quartz_granite","g_view_water"]+["g_"+b for b in BOOL]
	127	+num=["sqft_final","has_sqft","lot_sqft","beds","baths","parking","storeys","age","age2","n_images_c","virtual_tour"]+amen+rv+gv
	128	+cat=["city_fe","ptype_fe","source"]
	129	+
	130	+y=df["ln_price"].values
	131	+price=df["price"].values
	132	+Xcat=pd.get_dummies(df[cat],prefix=cat,dtype=float)
	133	+# --- matrice ML (NaN autorisés pour arbres) ---
	134	+Xml=pd.concat([df[num].reset_index(drop=True),Xcat.reset_index(drop=True)],axis=1)
	135	+# --- matrice OLS (imputée médiane par type) ---
	136	+dfi=df.copy()
	137	+for c in ["sqft_final","lot_sqft","beds","baths","storeys","age","age2"]:
	138	+ dfi[c]=dfi.groupby("ptype")[c].transform(lambda s:s.fillna(s.median()))
	139	+ dfi[c]=dfi[c].fillna(dfi[c].median())
	140	+dfi["parking"]=dfi["parking"].fillna(0)
	141	+Xols=pd.concat([dfi[num].reset_index(drop=True),Xcat.reset_index(drop=True)],axis=1)
	142	+
	143	+kf=KFold(n_splits=5,shuffle=True,random_state=42)
	144	+def evalcv(model,X):
	145	+ pred=cross_val_predict(model,X,y,cv=kf,n_jobs=-1)
	146	+ r2=r2_score(y,pred)
	147	+ mape=np.median(np.abs(np.exp(pred-y)-1))*100
	148	+ w15=np.mean(np.abs(np.exp(pred-y)-1)<=.15)*100
	149	+ w25=np.mean(np.abs(np.exp(pred-y)-1)<=.25)*100
	150	+ return r2,mape,w15,w25
	151	+
	152	+print("\nValidation croisée 5-fold (out-of-sample) — cible = ln(prix)\n")
	153	+print(f"{'Modèle':<34}{'R² OOS':>9}{'MAE méd.':>10}{'±15%':>7}{'±25%':>7}")
	154	+res={}
	155	+res["OLS (linéaire, imputé)"]=evalcv(LinearRegression(),Xols)
	156	+res["Random Forest (400)"]=evalcv(RandomForestRegressor(n_estimators=400,max_features=0.5,min_samples_leaf=3,n_jobs=-1,random_state=42),Xml)
	157	+res["HistGradientBoosting"]=evalcv(HistGradientBoostingRegressor(max_iter=600,learning_rate=0.05,max_leaf_nodes=63,l2_regularization=1.0,random_state=42),Xml)
	158	+res["HistGB (tuned depth)"]=evalcv(HistGradientBoostingRegressor(max_iter=900,learning_rate=0.03,max_leaf_nodes=127,min_samples_leaf=20,l2_regularization=2.0,random_state=42),Xml)
	159	+for k,(r2,mape,w15,w25) in res.items():
	160	+ print(f"{k:<34}{r2:>9.4f}{mape:>9.1f}%{w15:>6.0f}%{w25:>6.0f}%")
	161	+
	162	+# importance GBM (permutation rapide via feature_importances_ du HistGB n'existe pas → refit + importances via gain n/a)
	163	+gb=HistGradientBoostingRegressor(max_iter=600,learning_rate=0.05,max_leaf_nodes=63,l2_regularization=1.0,random_state=42).fit(Xml,y)
	164	+from sklearn.inspection import permutation_importance
	165	+import numpy as _np
	166	+idx=_np.random.RandomState(0).choice(len(Xml),4000,replace=False)
	167	+pi=permutation_importance(gb,Xml.iloc[idx],y[idx],n_repeats=3,random_state=0,n_jobs=-1)
	168	+imp=pd.Series(pi.importances_mean,index=Xml.columns).sort_values(ascending=False)
	169	+print("\nTop 20 variables (importance par permutation, GBM) :")
	170	+for name,v in imp.head(20).items(): print(f" {name:<26}{v:.4f}")
	171	+print("\nImportance cumulée des variables GPT (g_*) :","%.4f"%imp[[c for c in Xml.columns if c.startswith('g_')]].sum())

added scripts/07_torch_nn.py +170 −0

		`@@ -0,0 +1,170 @@`
	1	+#!/usr/bin/env python3
	2	+# -- coding: utf-8 --
	3	+"""Réseau de neurones tabulaire (embeddings ville/type + MLP) — 5-fold CV.
	4	+Comparable aux modèles de hedonic_ml.py (même variables, même KFold seed=42)."""
	5	+import json, re, warnings
	6	+import numpy as np, pandas as pd
	7	+import torch, torch.nn as nn
	8	+from sklearn.model_selection import KFold
	9	+from sklearn.metrics import r2_score
	10	+warnings.filterwarnings("ignore")
	11	+torch.manual_seed(42); np.random.seed(42)
	12	+CUR_YEAR=2026
	13	+dev="mps" if torch.backends.mps.is_available() else "cpu"
	14	+print("device:",dev)
	15	+
	16	+# ---- specs GPT ----
	17	+specs={}
	18	+for l in open("data/processed/specs.jsonl",encoding="utf-8"):
	19	+ r=json.loads(l)
	20	+ if r.get("specs"): specs[r["uid"]]=r["specs"]
	21	+COND={"a_renover":0,"original":1,"bon":2,"renove":3,"refait_neuf":4,"neuf":5,"inconnu":2}
	22	+BOOL=["turnkey","recently_renovated","garage","pool","waterfront","finished_basement","fireplace","central_ac","new_construction","income_property","renovated_kitchen","renovated_bathroom","large_lot","quiet_area"]
	23	+def gfeat(uid):
	24	+ s=specs.get(uid);o={"g_"+b:0 for b in BOOL};o.update({"g_cond":2,"g_hardwood":0,"g_quartz_granite":0,"g_view_water":0})
	25	+ if not s: return o
	26	+ for b in BOOL: o["g_"+b]=int(bool(s.get(b)))
	27	+ o["g_cond"]=COND.get(s.get("condition","inconnu"),2)
	28	+ fl=s.get("flooring") or [];o["g_hardwood"]=int("bois_franc" in fl)
	29	+ o["g_quartz_granite"]=int(s.get("countertop") in ("quartz","granit","marbre"));o["g_view_water"]=int(s.get("view")=="eau");return o
	30	+def parse_area(s):
	31	+ if not s or not isinstance(s,str): return np.nan
	32	+ t=s.lower();m=("m2" in t) or ("m²" in t);n=re.sub(r"[^0-9.,\s]","",t).strip().replace(" ","").replace(",","")
	33	+ try: v=float(n)
	34	+ except: return np.nan
	35	+ return (v*10.7639 if m else v) if v>0 else np.nan
	36	+def cc(c):
	37	+ if not c or not isinstance(c,str): return None
	38	+ c=c.split("(")[0].strip();return re.sub(r"\s+"," ",c).lower().strip() or None
	39	+def ti(x):
	40	+ try: return int(x)
	41	+ except: return np.nan
	42	+AMEN={"pool":["swimming pool","piscine","pool"],"waterfront":["waterfront","bord de l'eau","navigable","lakefront","riverfront","au bord de"],"garage":["garage"],"fireplace":["fireplace","foyer","wood stove","poêle"],"central_air":["central air","air conditionné","climatisation","a/c","heat-pump","thermopompe","air exchanger"],"finished_bsmt":["totally finished","finished basement","sous-sol fini","basement"],"near_water":["near water","waterfront","navigable"],"cul_de_sac":["cul-de-sac","cul de sac"],"golf":["golf"],"ski":["ski"],"near_park":["park","parc"],"near_school":["school","école","cegep","university","université","elementary","high school"],"near_highway":["highway","autoroute"],"transit":["public transit","public transportation","transport"],"no_rear_neighbor":["no backyard neighbors","sans voisin arrière"]}
	43	+def sd(s):
	44	+ s=s.lower()
	45	+ if "ft" in s:
	46	+ f=re.search(r"([\d.]+)\sft",s);i=re.search(r"([\d.]+)\sin",s)
	47	+ return (float(f.group(1)) if f else 0)+(float(i.group(1))/12 if i else 0)
	48	+ m=re.search(r"([\d.,]+)\s*m",s)
	49	+ if m:
	50	+ try: return float(m.group(1).replace(",","."))*3.28084
	51	+ except: return None
	52	+ return None
	53	+def ra(d):
	54	+ if not d or "x" not in d: return None
	55	+ a,_,b=d.partition("x");la,lb=sd(a),sd(b);return la*lb if (la and lb and 1<la<200 and 1<lb<200) else None
	56	+RF={"r_office":["office"],"r_family":["family room"],"r_workshop":["workshop"],"r_winecellar":["wine cellar"],"r_sunroom":["sunroom","solarium"],"r_mezzanine":["mezzanine"],"r_library":["library"],"r_laundry":["laundry"]}
	57	+def pr(rooms):
	58	+ o={"room_area":np.nan,"n_rooms":0,"has_primary_bed":0,"has_basement":0,"n_levels_r":0}
	59	+ for k in RF: o[k]=0
	60	+ if not rooms: return o
	61	+ ar=0.0;lv=set()
	62	+ for r in rooms:
	63	+ nm=(r.get("name") or "").lower();l=(r.get("level") or "").lower();o["n_rooms"]+=1
	64	+ if l: lv.add(l)
	65	+ if "basement" in l or "sous-sol" in l: o["has_basement"]=1
	66	+ if "primary bedroom" in nm: o["has_primary_bed"]=1
	67	+ for k,kw in RF.items():
	68	+ if any(w in nm for w in kw): o[k]=1
	69	+ a=ra(r.get("dimensions"))
	70	+ if a: ar+=a
	71	+ o["room_area"]=ar if ar>0 else np.nan;o["n_levels_r"]=len(lv);return o
	72	+data=json.load(open("data/processed/maisons_avec_prix.json",encoding="utf-8"));rows=[]
	73	+for d in data:
	74	+ p=d.get("property") or {};ad=d.get("address") or {};ft=d.get("features") or []
	75	+ blob=(" ".join(ft)+" "+(d.get("description") or "")+" "+(d.get("title") or "")).lower()
	76	+ yb=ti(p.get("yearBuilt"));age=(CUR_YEAR-yb) if (not pd.isna(yb) and 1700<=yb<=CUR_YEAR) else np.nan
	77	+ pt=p.get("buildingType") or p.get("type");st=p.get("storeys") or p.get("levels")
	78	+ try: st=float(str(st).replace(",","."))
	79	+ except: st=np.nan
	80	+ uid=f"{d.get('source')}:{d.get('listingId')}"
	81	+ rec={"uid":uid,"price":d.get("priceNumber"),"source":d.get("source"),"sqft":parse_area(p.get("squareFeet")),"lot_sqft":parse_area(p.get("lotSize")),"beds":ti(p.get("bedrooms")),"baths":ti(p.get("bathrooms")),"parking":ti(p.get("parkingSpaces")),"storeys":st,"age":age,"ptype":(pt or "Unknown").strip(),"city":cc(ad.get("city"))}
	82	+ for n,kw in AMEN.items(): rec[n]=int(any(k in blob for k in kw))
	83	+ rec["n_images"]=len(d.get("images") or []);rec["virtual_tour"]=int(bool(d.get("virtualTourUrl")))
	84	+ rec.update(pr(d.get("rooms")));rec.update(gfeat(uid));rows.append(rec)
	85	+df=pd.DataFrame(rows)
	86	+df=df[df["price"].notna()];df=df[(df["price"]>=50_000)&(df["price"]<=15_000_000)]
	87	+df.loc[(df["sqft"]<300)\|(df["sqft"]>20_000),"sqft"]=np.nan
	88	+df.loc[(df["lot_sqft"]<200)\|(df["lot_sqft"]>5_000_000),"lot_sqft"]=np.nan
	89	+df.loc[df["beds"]>15,"beds"]=np.nan;df.loc[df["baths"]>12,"baths"]=np.nan
	90	+df["ppsf"]=df["price"]/df["sqft"];lo,hi=df["ppsf"].quantile([0.005,0.995])
	91	+df=df[~(df["ppsf"].notna()&((df["ppsf"]<lo)\|(df["ppsf"]>hi)))]
	92	+ov=df[df["sqft"].notna()&df["room_area"].notna()&(df["room_area"]>200)];ratio=(ov["sqft"]/ov["room_area"]).median()
	93	+re_=df["room_area"]*ratio;ok=df["room_area"].notna()&(re_.between(300,20000))
	94	+df["sqft_final"]=df["sqft"];fill=df["sqft_final"].isna()&ok;df.loc[fill,"sqft_final"]=re_[fill]
	95	+df["has_sqft"]=df["sqft_final"].notna().astype(int);df["age2"]=df["age"]**2;df["n_images_c"]=df["n_images"].clip(upper=60)
	96	+vc=df["city"].value_counts();keep=set(vc[vc>=20].index);df["city_fe"]=df["city"].where(df["city"].isin(keep),"__other__").fillna("__other__")
	97	+tc=df["ptype"].value_counts();df["ptype_fe"]=df["ptype"].where(df["ptype"].isin(set(tc[tc>=30].index)),"Other")
	98	+df=df.reset_index(drop=True);print("N=",len(df))
	99	+
	100	+amen=list(AMEN.keys());rv=["n_rooms","n_levels_r","has_primary_bed","has_basement","r_office","r_family","r_workshop","r_winecellar","r_sunroom","r_mezzanine","r_library","r_laundry"]
	101	+gv=["g_cond","g_hardwood","g_quartz_granite","g_view_water"]+["g_"+b for b in BOOL]
	102	+num=["sqft_final","has_sqft","lot_sqft","beds","baths","parking","storeys","age","age2","n_images_c","virtual_tour"]+amen+rv+gv
	103	+y=df["ln_price"]=np.log(df["price"]);y=y.values.astype(np.float32)
	104	+# encodage catégoriel -> indices
	105	+cats=["city_fe","ptype_fe","source"];maps={};card={}
	106	+for c in cats:
	107	+ u=sorted(df[c].astype(str).unique());m={v:i for i,v in enumerate(u)};maps[c]=m;card[c]=len(u)
	108	+ df[c+"_ix"]=df[c].astype(str).map(m).astype(np.int64)
	109	+Xnum=df[num].astype(np.float32).values
	110	+Xcat=df[[c+"_ix" for c in cats]].values.astype(np.int64)
	111	+
	112	+class Net(nn.Module):
	113	+ def __init__(self,ncard,nnum):
	114	+ super().__init__()
	115	+ self.embs=nn.ModuleList()
	116	+ tot=0
	117	+ for c in ncard:
	118	+ d=min(50,(c+1)//2); d=max(d,2); self.embs.append(nn.Embedding(c,d)); tot+=d
	119	+ self.bn=nn.BatchNorm1d(nnum)
	120	+ self.net=nn.Sequential(
	121	+ nn.Linear(tot+nnum,256),nn.ReLU(),nn.BatchNorm1d(256),nn.Dropout(0.3),
	122	+ nn.Linear(256,128),nn.ReLU(),nn.BatchNorm1d(128),nn.Dropout(0.2),
	123	+ nn.Linear(128,64),nn.ReLU(),nn.Dropout(0.1),
	124	+ nn.Linear(64,1))
	125	+ def forward(self,xc,xn):
	126	+ e=[emb(xc[:,i]) for i,emb in enumerate(self.embs)]
	127	+ x=torch.cat(e+[self.bn(xn)],1);return self.net(x).squeeze(1)
	128	+
	129	+def train_predict(tr,te):
	130	+ # imputation médiane (train) + standardisation
	131	+ med=np.nanmedian(Xnum[tr],axis=0);med=np.where(np.isnan(med),0,med)
	132	+ Xn=Xnum.copy()
	133	+ for j in range(Xn.shape[1]):
	134	+ col=Xn[:,j];mask=np.isnan(col);col[mask]=med[j];Xn[:,j]=col
	135	+ mu=Xn[tr].mean(0);sg=Xn[tr].std(0)+1e-6;Xn=(Xn-mu)/sg
	136	+ ym=y[tr].mean();ys=y[tr].std()+1e-6;yz=(y-ym)/ys
	137	+ # split interne train/val pour early stopping
	138	+ ntr=len(tr);perm=np.random.permutation(ntr);vcut=int(ntr*0.9)
	139	+ tri=tr[perm[:vcut]];vai=tr[perm[vcut:]]
	140	+ def T(a,dt): return torch.tensor(a,device=dev,dtype=dt)
	141	+ model=Net([card[c] for c in cats],Xn.shape[1]).to(dev)
	142	+ opt=torch.optim.Adam(model.parameters(),lr=1e-3,weight_decay=1e-5)
	143	+ lossf=nn.MSELoss()
	144	+ Xc_t=T(Xcat,torch.long);Xn_t=T(Xn,torch.float32);y_t=T(yz,torch.float32)
	145	+ best=1e9;bstate=None;patience=0
	146	+ for ep in range(150):
	147	+ model.train();idx=np.random.permutation(tri)
	148	+ for i in range(0,len(idx),512):
	149	+ b=idx[i:i+512]
	150	+ opt.zero_grad();p=model(Xc_t[b],Xn_t[b]);l=lossf(p,y_t[b]);l.backward();opt.step()
	151	+ model.eval()
	152	+ with torch.no_grad():
	153	+ vp=model(Xc_t[vai],Xn_t[vai]);vl=lossf(vp,y_t[vai]).item()
	154	+ if vl<best-1e-4: best=vl;bstate={k:v.clone() for k,v in model.state_dict().items()};patience=0
	155	+ else:
	156	+ patience+=1
	157	+ if patience>=12: break
	158	+ model.load_state_dict(bstate);model.eval()
	159	+ with torch.no_grad():
	160	+ pz=model(Xc_t[te],Xn_t[te]).cpu().numpy()
	161	+ return pz*ys+ym
	162	+
	163	+kf=KFold(n_splits=5,shuffle=True,random_state=42)
	164	+pred=np.zeros_like(y)
	165	+for f,(tr,te) in enumerate(kf.split(Xnum)):
	166	+ pred[te]=train_predict(tr,te);print(f" fold {f+1} fait")
	167	+r2=r2_score(y,pred);err=np.abs(np.exp(pred-y)-1)
	168	+print("\n=== Réseau de neurones (embeddings + MLP) — 5-fold CV OOS ===")
	169	+print(f"R² OOS = {r2:.4f} \| erreur médiane = {np.median(err)100:.1f}% \| ±15% = {np.mean(err<=.15)100:.0f}% ±25% = {np.mean(err<=.25)*100:.0f}%")
	170	+print("\nRappel (mêmes folds) : OLS 0.628 \| RF 0.678 \| HistGBM 0.743")

added scripts/08_torch_nn_robust.py +192 −0

		`@@ -0,0 +1,192 @@`
	1	+#!/usr/bin/env python3
	2	+# -- coding: utf-8 --
	3	+"""Réseau de neurones tabulaire ROBUSTE : embeddings + ResNet-MLP,
	4	+ensemble multi-seed, Huber loss, AdamW + cosine, grad-clip, early stopping.
	5	+5-fold CV out-of-sample, comparable aux autres modèles (seed=42)."""
	6	+import json, re, warnings
	7	+import numpy as np, pandas as pd
	8	+import torch, torch.nn as nn
	9	+from sklearn.model_selection import KFold
	10	+from sklearn.metrics import r2_score
	11	+warnings.filterwarnings("ignore")
	12	+CUR_YEAR=2026
	13	+dev="mps" if torch.backends.mps.is_available() else "cpu"
	14	+print("device:",dev)
	15	+SEEDS=[42,1,7] # ensemble par fold
	16	+D_MAIN=256; N_BLOCKS=3; D_HIDDEN=512
	17	+
	18	+# ======== reconstruction du dataset (identique v4) ========
	19	+specs={}
	20	+for l in open("data/processed/specs.jsonl",encoding="utf-8"):
	21	+ r=json.loads(l)
	22	+ if r.get("specs"): specs[r["uid"]]=r["specs"]
	23	+COND={"a_renover":0,"original":1,"bon":2,"renove":3,"refait_neuf":4,"neuf":5,"inconnu":2}
	24	+BOOL=["turnkey","recently_renovated","garage","pool","waterfront","finished_basement","fireplace","central_ac","new_construction","income_property","renovated_kitchen","renovated_bathroom","large_lot","quiet_area"]
	25	+def gfeat(uid):
	26	+ s=specs.get(uid);o={"g_"+b:0 for b in BOOL};o.update({"g_cond":2,"g_hardwood":0,"g_quartz_granite":0,"g_view_water":0})
	27	+ if not s: return o
	28	+ for b in BOOL: o["g_"+b]=int(bool(s.get(b)))
	29	+ o["g_cond"]=COND.get(s.get("condition","inconnu"),2)
	30	+ fl=s.get("flooring") or [];o["g_hardwood"]=int("bois_franc" in fl)
	31	+ o["g_quartz_granite"]=int(s.get("countertop") in ("quartz","granit","marbre"));o["g_view_water"]=int(s.get("view")=="eau");return o
	32	+def parse_area(s):
	33	+ if not s or not isinstance(s,str): return np.nan
	34	+ t=s.lower();m=("m2" in t) or ("m²" in t);n=re.sub(r"[^0-9.,\s]","",t).strip().replace(" ","").replace(",","")
	35	+ try: v=float(n)
	36	+ except: return np.nan
	37	+ return (v*10.7639 if m else v) if v>0 else np.nan
	38	+def cc(c):
	39	+ if not c or not isinstance(c,str): return None
	40	+ c=c.split("(")[0].strip();return re.sub(r"\s+"," ",c).lower().strip() or None
	41	+def ti(x):
	42	+ try: return int(x)
	43	+ except: return np.nan
	44	+AMEN={"pool":["swimming pool","piscine","pool"],"waterfront":["waterfront","bord de l'eau","navigable","lakefront","riverfront","au bord de"],"garage":["garage"],"fireplace":["fireplace","foyer","wood stove","poêle"],"central_air":["central air","air conditionné","climatisation","a/c","heat-pump","thermopompe","air exchanger"],"finished_bsmt":["totally finished","finished basement","sous-sol fini","basement"],"near_water":["near water","waterfront","navigable"],"cul_de_sac":["cul-de-sac","cul de sac"],"golf":["golf"],"ski":["ski"],"near_park":["park","parc"],"near_school":["school","école","cegep","university","université","elementary","high school"],"near_highway":["highway","autoroute"],"transit":["public transit","public transportation","transport"],"no_rear_neighbor":["no backyard neighbors","sans voisin arrière"]}
	45	+def sd(s):
	46	+ s=s.lower()
	47	+ if "ft" in s:
	48	+ f=re.search(r"([\d.]+)\sft",s);i=re.search(r"([\d.]+)\sin",s)
	49	+ return (float(f.group(1)) if f else 0)+(float(i.group(1))/12 if i else 0)
	50	+ m=re.search(r"([\d.,]+)\s*m",s)
	51	+ if m:
	52	+ try: return float(m.group(1).replace(",","."))*3.28084
	53	+ except: return None
	54	+ return None
	55	+def ra(d):
	56	+ if not d or "x" not in d: return None
	57	+ a,_,b=d.partition("x");la,lb=sd(a),sd(b);return la*lb if (la and lb and 1<la<200 and 1<lb<200) else None
	58	+RF={"r_office":["office"],"r_family":["family room"],"r_workshop":["workshop"],"r_winecellar":["wine cellar"],"r_sunroom":["sunroom","solarium"],"r_mezzanine":["mezzanine"],"r_library":["library"],"r_laundry":["laundry"]}
	59	+def pr(rooms):
	60	+ o={"room_area":np.nan,"n_rooms":0,"has_primary_bed":0,"has_basement":0,"n_levels_r":0}
	61	+ for k in RF: o[k]=0
	62	+ if not rooms: return o
	63	+ ar=0.0;lv=set()
	64	+ for r in rooms:
	65	+ nm=(r.get("name") or "").lower();l=(r.get("level") or "").lower();o["n_rooms"]+=1
	66	+ if l: lv.add(l)
	67	+ if "basement" in l or "sous-sol" in l: o["has_basement"]=1
	68	+ if "primary bedroom" in nm: o["has_primary_bed"]=1
	69	+ for k,kw in RF.items():
	70	+ if any(w in nm for w in kw): o[k]=1
	71	+ a=ra(r.get("dimensions"))
	72	+ if a: ar+=a
	73	+ o["room_area"]=ar if ar>0 else np.nan;o["n_levels_r"]=len(lv);return o
	74	+data=json.load(open("data/processed/maisons_avec_prix.json",encoding="utf-8"));rows=[]
	75	+for d in data:
	76	+ p=d.get("property") or {};ad=d.get("address") or {};ft=d.get("features") or []
	77	+ blob=(" ".join(ft)+" "+(d.get("description") or "")+" "+(d.get("title") or "")).lower()
	78	+ yb=ti(p.get("yearBuilt"));age=(CUR_YEAR-yb) if (not pd.isna(yb) and 1700<=yb<=CUR_YEAR) else np.nan
	79	+ pt=p.get("buildingType") or p.get("type");st=p.get("storeys") or p.get("levels")
	80	+ try: st=float(str(st).replace(",","."))
	81	+ except: st=np.nan
	82	+ uid=f"{d.get('source')}:{d.get('listingId')}"
	83	+ rec={"uid":uid,"price":d.get("priceNumber"),"source":d.get("source"),"sqft":parse_area(p.get("squareFeet")),"lot_sqft":parse_area(p.get("lotSize")),"beds":ti(p.get("bedrooms")),"baths":ti(p.get("bathrooms")),"parking":ti(p.get("parkingSpaces")),"storeys":st,"age":age,"ptype":(pt or "Unknown").strip(),"city":cc(ad.get("city"))}
	84	+ for n,kw in AMEN.items(): rec[n]=int(any(k in blob for k in kw))
	85	+ rec["n_images"]=len(d.get("images") or []);rec["virtual_tour"]=int(bool(d.get("virtualTourUrl")))
	86	+ rec.update(pr(d.get("rooms")));rec.update(gfeat(uid));rows.append(rec)
	87	+df=pd.DataFrame(rows)
	88	+df=df[df["price"].notna()];df=df[(df["price"]>=50_000)&(df["price"]<=15_000_000)]
	89	+df.loc[(df["sqft"]<300)\|(df["sqft"]>20_000),"sqft"]=np.nan
	90	+df.loc[(df["lot_sqft"]<200)\|(df["lot_sqft"]>5_000_000),"lot_sqft"]=np.nan
	91	+df.loc[df["beds"]>15,"beds"]=np.nan;df.loc[df["baths"]>12,"baths"]=np.nan
	92	+df["ppsf"]=df["price"]/df["sqft"];lo,hi=df["ppsf"].quantile([0.005,0.995])
	93	+df=df[~(df["ppsf"].notna()&((df["ppsf"]<lo)\|(df["ppsf"]>hi)))]
	94	+ov=df[df["sqft"].notna()&df["room_area"].notna()&(df["room_area"]>200)];ratio=(ov["sqft"]/ov["room_area"]).median()
	95	+re_=df["room_area"]*ratio;ok=df["room_area"].notna()&(re_.between(300,20000))
	96	+df["sqft_final"]=df["sqft"];fill=df["sqft_final"].isna()&ok;df.loc[fill,"sqft_final"]=re_[fill]
	97	+df["has_sqft"]=df["sqft_final"].notna().astype(int);df["age2"]=df["age"]**2;df["n_images_c"]=df["n_images"].clip(upper=60)
	98	+vc=df["city"].value_counts();keep=set(vc[vc>=20].index);df["city_fe"]=df["city"].where(df["city"].isin(keep),"__other__").fillna("__other__")
	99	+tc=df["ptype"].value_counts();df["ptype_fe"]=df["ptype"].where(df["ptype"].isin(set(tc[tc>=30].index)),"Other")
	100	+df=df.reset_index(drop=True);print("N=",len(df))
	101	+
	102	+amen=list(AMEN.keys());rv=["n_rooms","n_levels_r","has_primary_bed","has_basement","r_office","r_family","r_workshop","r_winecellar","r_sunroom","r_mezzanine","r_library","r_laundry"]
	103	+gv=["g_cond","g_hardwood","g_quartz_granite","g_view_water"]+["g_"+b for b in BOOL]
	104	+num=["sqft_final","has_sqft","lot_sqft","beds","baths","parking","storeys","age","age2","n_images_c","virtual_tour"]+amen+rv+gv
	105	+# indicateurs de manquant (robustesse) pour les colonnes à NaN
	106	+missnum=["sqft_final","lot_sqft","beds","baths","storeys","age","age2"]
	107	+for c in missnum: df["miss_"+c]=df[c].isna().astype(np.float32)
	108	+num=num+["miss_"+c for c in missnum]
	109	+y=np.log(df["price"].values).astype(np.float32)
	110	+cats=["city_fe","ptype_fe","source"];card={}
	111	+for c in cats:
	112	+ u=sorted(df[c].astype(str).unique());m={v:i for i,v in enumerate(u)};card[c]=len(u)
	113	+ df[c+"_ix"]=df[c].astype(str).map(m).astype(np.int64)
	114	+Xnum=df[num].astype(np.float32).values
	115	+Xcat=df[[c+"_ix" for c in cats]].values.astype(np.int64)
	116	+
	117	+class ResBlock(nn.Module):
	118	+ def __init__(self,d,dh,p):
	119	+ super().__init__()
	120	+ self.bn=nn.BatchNorm1d(d);self.l1=nn.Linear(d,dh);self.act=nn.GELU()
	121	+ self.dp1=nn.Dropout(p);self.l2=nn.Linear(dh,d);self.dp2=nn.Dropout(p)
	122	+ def forward(self,x):
	123	+ z=self.bn(x);z=self.act(self.l1(z));z=self.dp1(z);z=self.l2(z);z=self.dp2(z);return x+z
	124	+
	125	+class ResNetTab(nn.Module):
	126	+ def __init__(self,ncard,nnum):
	127	+ super().__init__()
	128	+ self.embs=nn.ModuleList();tot=0
	129	+ for c in ncard:
	130	+ d=max(2,min(64,round(1.6c*0.56)));self.embs.append(nn.Embedding(c,d));tot+=d
	131	+ self.embdp=nn.Dropout(0.1)
	132	+ self.bn0=nn.BatchNorm1d(nnum)
	133	+ self.proj=nn.Linear(tot+nnum,D_MAIN)
	134	+ self.blocks=nn.ModuleList([ResBlock(D_MAIN,D_HIDDEN,0.2) for _ in range(N_BLOCKS)])
	135	+ self.head=nn.Sequential(nn.BatchNorm1d(D_MAIN),nn.GELU(),nn.Linear(D_MAIN,1))
	136	+ def forward(self,xc,xn):
	137	+ e=[emb(xc[:,i]) for i,emb in enumerate(self.embs)]
	138	+ x=torch.cat([self.embdp(torch.cat(e,1)),self.bn0(xn)],1)
	139	+ x=self.proj(x)
	140	+ for b in self.blocks: x=b(x)
	141	+ return self.head(x).squeeze(1)
	142	+
	143	+def train_one(tr,te,seed):
	144	+ torch.manual_seed(seed);np.random.seed(seed)
	145	+ med=np.nanmedian(Xnum[tr],axis=0);med=np.where(np.isnan(med),0,med)
	146	+ Xn=Xnum.copy()
	147	+ for j in range(Xn.shape[1]):
	148	+ col=Xn[:,j];col[np.isnan(col)]=med[j];Xn[:,j]=col
	149	+ mu=Xn[tr].mean(0);sg=Xn[tr].std(0)+1e-6;Xn=(Xn-mu)/sg
	150	+ ym=y[tr].mean();ys=y[tr].std()+1e-6;yz=(y-ym)/ys
	151	+ ntr=len(tr);perm=np.random.permutation(ntr);vc_=int(ntr*0.9)
	152	+ tri=tr[perm[:vc_]];vai=tr[perm[vc_:]]
	153	+ def T(a,dt): return torch.tensor(a,device=dev,dtype=dt)
	154	+ model=ResNetTab([card[c] for c in cats],Xn.shape[1]).to(dev)
	155	+ opt=torch.optim.AdamW(model.parameters(),lr=2e-3,weight_decay=1e-4)
	156	+ EPOCHS=200
	157	+ sched=torch.optim.lr_scheduler.CosineAnnealingLR(opt,T_max=EPOCHS)
	158	+ lossf=nn.SmoothL1Loss(beta=1.0) # Huber : robuste aux queues
	159	+ Xc_t=T(Xcat,torch.long);Xn_t=T(Xn,torch.float32);y_t=T(yz,torch.float32)
	160	+ best=1e9;bstate=None;pat=0
	161	+ for ep in range(EPOCHS):
	162	+ model.train();idx=np.random.permutation(tri)
	163	+ for i in range(0,len(idx),512):
	164	+ b=idx[i:i+512]
	165	+ opt.zero_grad();p=model(Xc_t[b],Xn_t[b]);l=lossf(p,y_t[b])
	166	+ l.backward();torch.nn.utils.clip_grad_norm_(model.parameters(),5.0);opt.step()
	167	+ sched.step()
	168	+ model.eval()
	169	+ with torch.no_grad():
	170	+ vp=model(Xc_t[vai],Xn_t[vai]);vl=((vp-y_t[vai])**2).mean().item()
	171	+ if vl<best-1e-4: best=vl;bstate={k:v.clone() for k,v in model.state_dict().items()};pat=0
	172	+ else:
	173	+ pat+=1
	174	+ if pat>=20: break
	175	+ model.load_state_dict(bstate);model.eval()
	176	+ with torch.no_grad():
	177	+ pz=model(Xc_t[te],Xn_t[te]).cpu().numpy()
	178	+ return pz*ys+ym
	179	+
	180	+kf=KFold(n_splits=5,shuffle=True,random_state=42)
	181	+pred=np.zeros_like(y)
	182	+for f,(tr,te) in enumerate(kf.split(Xnum)):
	183	+ ps=np.zeros(len(te),dtype=np.float64)
	184	+ for sd_ in SEEDS:
	185	+ ps+=train_one(tr,te,sd_)
	186	+ pred[te]=ps/len(SEEDS) # moyenne d'ensemble
	187	+ r2f=r2_score(y[te],pred[te]);print(f" fold {f+1}: R²={r2f:.4f} (ensemble de {len(SEEDS)})")
	188	+r2=r2_score(y,pred);err=np.abs(np.exp(pred-y)-1)
	189	+print("\n=== NN ROBUSTE (ResNet-tab, ensemble multi-seed) — 5-fold CV OOS ===")
	190	+print(f"R² OOS = {r2:.4f} \| erreur médiane = {np.median(err)100:.1f}% \| ±15% = {np.mean(err<=.15)100:.0f}% \| ±25% = {np.mean(err<=.25)*100:.0f}%")
	191	+print("\nRappel (mêmes folds): OLS 0.628 \| RF 0.678 \| NN simple 0.711 \| HistGBM 0.743")
	192	+np.save("data/processed/nn_robust_oos.npy",pred)

added scripts/09_v5_ml_text_vision.py +213 −0

		`@@ -0,0 +1,213 @@`
	1	+#!/usr/bin/env python3
	2	+# -- coding: utf-8 --
	3	+"""v5 : ablation texte (v4) vs texte+vision (v5). GBM + NN robuste + blend, 5-fold CV."""
	4	+import json, re, warnings
	5	+import numpy as np, pandas as pd
	6	+import torch, torch.nn as nn
	7	+from sklearn.ensemble import HistGradientBoostingRegressor
	8	+from sklearn.model_selection import KFold, cross_val_predict
	9	+from sklearn.metrics import r2_score
	10	+warnings.filterwarnings("ignore")
	11	+torch.manual_seed(42); np.random.seed(42)
	12	+CUR_YEAR=2026
	13	+dev="mps" if torch.backends.mps.is_available() else "cpu"; print("device:",dev)
	14	+
	15	+# ---- GPT specs ----
	16	+specs={}
	17	+for l in open("data/processed/specs.jsonl",encoding="utf-8"):
	18	+ r=json.loads(l)
	19	+ if r.get("specs"): specs[r["uid"]]=r["specs"]
	20	+# ---- vision ----
	21	+facade={}
	22	+for l in open("data/processed/facade_specs.jsonl",encoding="utf-8"):
	23	+ r=json.loads(l)
	24	+ if r.get("facade"): facade[r["uid"]]=r["facade"]
	25	+print("specs GPT:",len(specs),"\| façades:",len(facade))
	26	+
	27	+COND={"a_renover":0,"original":1,"bon":2,"renove":3,"refait_neuf":4,"neuf":5,"inconnu":2}
	28	+BOOL=["turnkey","recently_renovated","garage","pool","waterfront","finished_basement","fireplace","central_ac","new_construction","income_property","renovated_kitchen","renovated_bathroom","large_lot","quiet_area"]
	29	+def gfeat(uid):
	30	+ s=specs.get(uid);o={"g_"+b:0 for b in BOOL};o.update({"g_cond":2,"g_hardwood":0,"g_quartz_granite":0,"g_view_water":0})
	31	+ if not s: return o
	32	+ for b in BOOL: o["g_"+b]=int(bool(s.get(b)))
	33	+ o["g_cond"]=COND.get(s.get("condition","inconnu"),2)
	34	+ fl=s.get("flooring") or [];o["g_hardwood"]=int("bois_franc" in fl)
	35	+ o["g_quartz_granite"]=int(s.get("countertop") in ("quartz","granit","marbre"));o["g_view_water"]=int(s.get("view")=="eau");return o
	36	+RENO={"a_renover":0,"date":1,"standard":2,"renove":3,"haut_de_gamme":4,"inconnu":2}
	37	+LAND={"aucun":0,"minimal":1,"entretenu":2,"amenage_soigne":3}
	38	+SETS=["urbain_dense","banlieue","rural","bord_eau","boise"]
	39	+MATS=["brique","pierre","vinyle","bois","aluminium","stucco","crepi","beton"]
	40	+def vfeat(uid):
	41	+ f=facade.get(uid)
	42	+ o={"v_cond":5,"v_curb":5,"v_photoq":5,"v_reno":2,"v_land":1,"v_luxury":0,"v_garage_vis":0,"v_is_ext":0}
	43	+ for s in SETS: o["v_set_"+s]=0
	44	+ for m in MATS: o["v_mat_"+m]=0
	45	+ if not f: return o
	46	+ o["v_cond"]=f.get("condition_score") or 5;o["v_curb"]=f.get("curb_appeal") or 5;o["v_photoq"]=f.get("photo_quality") or 5
	47	+ o["v_reno"]=RENO.get(f.get("renovation_level","inconnu"),2);o["v_land"]=LAND.get(f.get("landscaping","minimal"),1)
	48	+ o["v_luxury"]=int(bool(f.get("curb_luxury")));o["v_garage_vis"]=int(bool(f.get("has_garage_visible")));o["v_is_ext"]=int(bool(f.get("is_exterior")))
	49	+ st=f.get("setting")
	50	+ if st in SETS: o["v_set_"+st]=1
	51	+ for m in (f.get("exterior_materials") or []):
	52	+ if m in MATS: o["v_mat_"+m]=1
	53	+ return o
	54	+# ---- parsing structurel ----
	55	+def pa(s):
	56	+ if not s or not isinstance(s,str): return np.nan
	57	+ t=s.lower();m=("m2" in t) or ("m²" in t);n=re.sub(r"[^0-9.,\s]","",t).strip().replace(" ","").replace(",","")
	58	+ try: v=float(n)
	59	+ except: return np.nan
	60	+ return (v*10.7639 if m else v) if v>0 else np.nan
	61	+def cc(c):
	62	+ if not c or not isinstance(c,str): return None
	63	+ c=c.split("(")[0].strip();return re.sub(r"\s+"," ",c).lower().strip() or None
	64	+def ti(x):
	65	+ try: return int(x)
	66	+ except: return np.nan
	67	+AMEN={"pool":["swimming pool","piscine","pool"],"waterfront":["waterfront","bord de l'eau","navigable","lakefront","riverfront","au bord de"],"garage":["garage"],"fireplace":["fireplace","foyer","wood stove","poêle"],"central_air":["central air","air conditionné","climatisation","a/c","heat-pump","thermopompe","air exchanger"],"finished_bsmt":["totally finished","finished basement","sous-sol fini","basement"],"near_water":["near water","waterfront","navigable"],"cul_de_sac":["cul-de-sac","cul de sac"],"golf":["golf"],"ski":["ski"],"near_park":["park","parc"],"near_school":["school","école","cegep","university","université","elementary","high school"],"near_highway":["highway","autoroute"],"transit":["public transit","public transportation","transport"],"no_rear_neighbor":["no backyard neighbors","sans voisin arrière"]}
	68	+def sd(s):
	69	+ s=s.lower()
	70	+ if "ft" in s:
	71	+ f=re.search(r"([\d.]+)\sft",s);i=re.search(r"([\d.]+)\sin",s)
	72	+ return (float(f.group(1)) if f else 0)+(float(i.group(1))/12 if i else 0)
	73	+ m=re.search(r"([\d.,]+)\s*m",s)
	74	+ if m:
	75	+ try: return float(m.group(1).replace(",","."))*3.28084
	76	+ except: return None
	77	+ return None
	78	+def ra(d):
	79	+ if not d or "x" not in d: return None
	80	+ a,_,b=d.partition("x");la,lb=sd(a),sd(b);return la*lb if (la and lb and 1<la<200 and 1<lb<200) else None
	81	+RFR={"r_office":["office"],"r_family":["family room"],"r_workshop":["workshop"],"r_winecellar":["wine cellar"],"r_sunroom":["sunroom","solarium"],"r_mezzanine":["mezzanine"],"r_library":["library"],"r_laundry":["laundry"]}
	82	+def pr(rooms):
	83	+ o={"room_area":np.nan,"n_rooms":0,"has_primary_bed":0,"has_basement":0,"n_levels_r":0}
	84	+ for k in RFR: o[k]=0
	85	+ if not rooms: return o
	86	+ ar=0.0;lv=set()
	87	+ for r in rooms:
	88	+ nm=(r.get("name") or "").lower();l=(r.get("level") or "").lower();o["n_rooms"]+=1
	89	+ if l: lv.add(l)
	90	+ if "basement" in l or "sous-sol" in l: o["has_basement"]=1
	91	+ if "primary bedroom" in nm: o["has_primary_bed"]=1
	92	+ for k,kw in RFR.items():
	93	+ if any(w in nm for w in kw): o[k]=1
	94	+ a=ra(r.get("dimensions"))
	95	+ if a: ar+=a
	96	+ o["room_area"]=ar if ar>0 else np.nan;o["n_levels_r"]=len(lv);return o
	97	+
	98	+data=json.load(open("data/processed/maisons_avec_prix.json",encoding="utf-8"));rows=[]
	99	+for d in data:
	100	+ p=d.get("property") or {};ad=d.get("address") or {};ft=d.get("features") or []
	101	+ blob=(" ".join(ft)+" "+(d.get("description") or "")+" "+(d.get("title") or "")).lower()
	102	+ yb=ti(p.get("yearBuilt"));age=(CUR_YEAR-yb) if (not pd.isna(yb) and 1700<=yb<=CUR_YEAR) else np.nan
	103	+ pt=p.get("buildingType") or p.get("type");st=p.get("storeys") or p.get("levels")
	104	+ try: st=float(str(st).replace(",","."))
	105	+ except: st=np.nan
	106	+ uid=f"{d.get('source')}:{d.get('listingId')}"
	107	+ rec={"uid":uid,"price":d.get("priceNumber"),"source":d.get("source"),"sqft":pa(p.get("squareFeet")),"lot_sqft":pa(p.get("lotSize")),"beds":ti(p.get("bedrooms")),"baths":ti(p.get("bathrooms")),"parking":ti(p.get("parkingSpaces")),"storeys":st,"age":age,"ptype":(pt or "Unknown").strip(),"city":cc(ad.get("city"))}
	108	+ for n,kw in AMEN.items(): rec[n]=int(any(k in blob for k in kw))
	109	+ rec["n_images"]=len(d.get("images") or []);rec["virtual_tour"]=int(bool(d.get("virtualTourUrl")))
	110	+ rec.update(pr(d.get("rooms")));rec.update(gfeat(uid));rec.update(vfeat(uid));rows.append(rec)
	111	+df=pd.DataFrame(rows)
	112	+df=df[df["price"].notna()];df=df[(df["price"]>=50_000)&(df["price"]<=15_000_000)]
	113	+df.loc[(df["sqft"]<300)\|(df["sqft"]>20_000),"sqft"]=np.nan
	114	+df.loc[(df["lot_sqft"]<200)\|(df["lot_sqft"]>5_000_000),"lot_sqft"]=np.nan
	115	+df.loc[df["beds"]>15,"beds"]=np.nan;df.loc[df["baths"]>12,"baths"]=np.nan
	116	+df["ppsf"]=df["price"]/df["sqft"];lo,hi=df["ppsf"].quantile([0.005,0.995])
	117	+df=df[~(df["ppsf"].notna()&((df["ppsf"]<lo)\|(df["ppsf"]>hi)))]
	118	+ov=df[df["sqft"].notna()&df["room_area"].notna()&(df["room_area"]>200)];ratio=(ov["sqft"]/ov["room_area"]).median()
	119	+re_=df["room_area"]*ratio;ok=df["room_area"].notna()&(re_.between(300,20000))
	120	+df["sqft_final"]=df["sqft"];fill=df["sqft_final"].isna()&ok;df.loc[fill,"sqft_final"]=re_[fill]
	121	+df["has_sqft"]=df["sqft_final"].notna().astype(int);df["age2"]=df["age"]**2;df["n_images_c"]=df["n_images"].clip(upper=60)
	122	+vc=df["city"].value_counts();keep=set(vc[vc>=20].index);df["city_fe"]=df["city"].where(df["city"].isin(keep),"__other__").fillna("__other__")
	123	+tc=df["ptype"].value_counts();df["ptype_fe"]=df["ptype"].where(df["ptype"].isin(set(tc[tc>=30].index)),"Other")
	124	+df=df.reset_index(drop=True);print("N=",len(df))
	125	+y=np.log(df["price"].values).astype(np.float32)
	126	+
	127	+amen=list(AMEN.keys());rv=["n_rooms","n_levels_r","has_primary_bed","has_basement","r_office","r_family","r_workshop","r_winecellar","r_sunroom","r_mezzanine","r_library","r_laundry"]
	128	+gv=["g_cond","g_hardwood","g_quartz_granite","g_view_water"]+["g_"+b for b in BOOL]
	129	+vv=["v_cond","v_curb","v_photoq","v_reno","v_land","v_luxury","v_garage_vis","v_is_ext"]+["v_set_"+s for s in SETS]+["v_mat_"+m for m in MATS]
	130	+base=["sqft_final","has_sqft","lot_sqft","beds","baths","parking","storeys","age","age2","n_images_c","virtual_tour"]+amen+rv
	131	+NUM_v4=base+gv
	132	+NUM_v5=base+gv+vv
	133	+missnum=["sqft_final","lot_sqft","beds","baths","storeys","age","age2"]
	134	+for c in missnum: df["miss_"+c]=df[c].isna().astype(np.float32)
	135	+MISS=["miss_"+c for c in missnum]
	136	+
	137	+cats=["city_fe","ptype_fe","source"];card={}
	138	+for c in cats:
	139	+ u=sorted(df[c].astype(str).unique());m={v:i for i,v in enumerate(u)};card[c]=len(u);df[c+"_ix"]=df[c].astype(str).map(m).astype(np.int64)
	140	+Xcat_ix=df[[c+"_ix" for c in cats]].values.astype(np.int64)
	141	+kf=KFold(n_splits=5,shuffle=True,random_state=42)
	142	+
	143	+def gbm_cv(numcols):
	144	+ X=pd.concat([df[numcols].reset_index(drop=True),pd.get_dummies(df[cats],dtype=float).reset_index(drop=True)],axis=1)
	145	+ m=HistGradientBoostingRegressor(max_iter=600,learning_rate=0.05,max_leaf_nodes=63,l2_regularization=1.0,random_state=42)
	146	+ return cross_val_predict(m,X,y,cv=kf,n_jobs=-1)
	147	+
	148	+# ---- NN robuste ----
	149	+D_MAIN=256;D_HID=512;NB=3;SEEDS=[42,1,7]
	150	+class RB(nn.Module):
	151	+ def __init__(s,d,dh,p):
	152	+ super().__init__();s.bn=nn.BatchNorm1d(d);s.l1=nn.Linear(d,dh);s.act=nn.GELU();s.d1=nn.Dropout(p);s.l2=nn.Linear(dh,d);s.d2=nn.Dropout(p)
	153	+ def forward(s,x): z=s.d1(s.act(s.l1(s.bn(x))));return x+s.d2(s.l2(z))
	154	+class Net(nn.Module):
	155	+ def __init__(s,nc,nn_):
	156	+ super().__init__();s.embs=nn.ModuleList();tot=0
	157	+ for c in nc: d=max(2,min(64,round(1.6c*0.56)));s.embs.append(nn.Embedding(c,d));tot+=d
	158	+ s.ed=nn.Dropout(0.1);s.bn0=nn.BatchNorm1d(nn_);s.proj=nn.Linear(tot+nn_,D_MAIN)
	159	+ s.bl=nn.ModuleList([RB(D_MAIN,D_HID,0.2) for _ in range(NB)])
	160	+ s.head=nn.Sequential(nn.BatchNorm1d(D_MAIN),nn.GELU(),nn.Linear(D_MAIN,1))
	161	+ def forward(s,xc,xn):
	162	+ e=[emb(xc[:,i]) for i,emb in enumerate(s.embs)]
	163	+ x=s.proj(torch.cat([s.ed(torch.cat(e,1)),s.bn0(xn)],1))
	164	+ for b in s.bl: x=b(x)
	165	+ return s.head(x).squeeze(1)
	166	+def nn_fold(Xnum,tr,te,seed):
	167	+ torch.manual_seed(seed);np.random.seed(seed)
	168	+ med=np.nanmedian(Xnum[tr],0);med=np.where(np.isnan(med),0,med);Xn=Xnum.copy()
	169	+ for j in range(Xn.shape[1]): col=Xn[:,j];col[np.isnan(col)]=med[j];Xn[:,j]=col
	170	+ mu=Xn[tr].mean(0);sg=Xn[tr].std(0)+1e-6;Xn=(Xn-mu)/sg
	171	+ ym=y[tr].mean();ys=y[tr].std()+1e-6;yz=(y-ym)/ys
	172	+ perm=np.random.permutation(len(tr));cut=int(len(tr)*0.9);tri=tr[perm[:cut]];vai=tr[perm[cut:]]
	173	+ T=lambda a,dt: torch.tensor(a,device=dev,dtype=dt)
	174	+ model=Net([card[c] for c in cats],Xn.shape[1]).to(dev)
	175	+ opt=torch.optim.AdamW(model.parameters(),lr=2e-3,weight_decay=1e-4)
	176	+ sch=torch.optim.lr_scheduler.CosineAnnealingLR(opt,T_max=200);lf=nn.SmoothL1Loss(beta=1.0)
	177	+ Xc=T(Xcat_ix,torch.long);Xnt=T(Xn,torch.float32);yt=T(yz,torch.float32)
	178	+ best=1e9;bs=None;pat=0
	179	+ for ep in range(200):
	180	+ model.train();idx=np.random.permutation(tri)
	181	+ for i in range(0,len(idx),512):
	182	+ b=idx[i:i+512];opt.zero_grad();l=lf(model(Xc[b],Xnt[b]),yt[b]);l.backward()
	183	+ torch.nn.utils.clip_grad_norm_(model.parameters(),5.0);opt.step()
	184	+ sch.step();model.eval()
	185	+ with torch.no_grad(): vl=((model(Xc[vai],Xnt[vai])-yt[vai])**2).mean().item()
	186	+ if vl<best-1e-4: best=vl;bs={k:v.clone() for k,v in model.state_dict().items()};pat=0
	187	+ else:
	188	+ pat+=1
	189	+ if pat>=20: break
	190	+ model.load_state_dict(bs);model.eval()
	191	+ with torch.no_grad(): return model(Xc[te],Xnt[te]).cpu().numpy()*ys+ym
	192	+def nn_cv(numcols):
	193	+ Xnum=df[numcols+MISS].astype(np.float32).values;pred=np.zeros_like(y)
	194	+ for tr,te in kf.split(Xnum):
	195	+ ps=np.zeros(len(te))
	196	+ for s in SEEDS: ps+=nn_fold(Xnum,tr,te,s)
	197	+ pred[te]=ps/len(SEEDS)
	198	+ return pred
	199	+def stat(p):
	200	+ e=np.abs(np.exp(p-y)-1);return r2_score(y,p),np.median(e)100,np.mean(e<=.15)100,np.mean(e<=.25)*100
	201	+
	202	+print("\n>>> GBM ablation")
	203	+g4=gbm_cv(NUM_v4);g5=gbm_cv(NUM_v5)
	204	+print("\n>>> NN v5 (peut prendre ~2 min)")
	205	+n5=nn_cv(NUM_v5)
	206	+b5=0.5g5+0.5n5
	207	+
	208	+print("\n================ RÉSULTATS v5 (5-fold CV OOS) ================")
	209	+print(f"{'Modèle':<34}{'R²OOS':>8}{'MAEméd':>9}{'±15%':>7}{'±25%':>7}")
	210	+for nm,p in [("GBM v4 (texte)",g4),("GBM v5 (texte+vision)",g5),("NN v5 (texte+vision)",n5),("BLEND v5 (NN+GBM)",b5)]:
	211	+ r2,m,a,b=stat(p);print(f"{nm:<34}{r2:>8.4f}{m:>8.1f}%{a:>6.0f}%{b:>6.0f}%")
	212	+print("\nRappel v4 blend: 0.760 \| GBM 0.743 \| NN 0.741")
	213	+np.save("data/processed/v5_blend_oos.npy",b5)

added scripts/10_final_blend.py +117 −0

		`@@ -0,0 +1,117 @@`
	1	+#!/usr/bin/env python3
	2	+# -- coding: utf-8 --
	3	+"""Blend (stacking) NN robuste + HistGradientBoosting, mêmes folds (seed=42)."""
	4	+import json, re, warnings
	5	+import numpy as np, pandas as pd
	6	+from sklearn.ensemble import HistGradientBoostingRegressor
	7	+from sklearn.model_selection import KFold, cross_val_predict
	8	+from sklearn.metrics import r2_score
	9	+warnings.filterwarnings("ignore")
	10	+CUR_YEAR=2026
	11	+# --- rebuild df exactement comme torch_nn_robust (même ordre) ---
	12	+specs={}
	13	+for l in open("data/processed/specs.jsonl",encoding="utf-8"):
	14	+ r=json.loads(l)
	15	+ if r.get("specs"): specs[r["uid"]]=r["specs"]
	16	+COND={"a_renover":0,"original":1,"bon":2,"renove":3,"refait_neuf":4,"neuf":5,"inconnu":2}
	17	+BOOL=["turnkey","recently_renovated","garage","pool","waterfront","finished_basement","fireplace","central_ac","new_construction","income_property","renovated_kitchen","renovated_bathroom","large_lot","quiet_area"]
	18	+def gfeat(uid):
	19	+ s=specs.get(uid);o={"g_"+b:0 for b in BOOL};o.update({"g_cond":2,"g_hardwood":0,"g_quartz_granite":0,"g_view_water":0})
	20	+ if not s: return o
	21	+ for b in BOOL: o["g_"+b]=int(bool(s.get(b)))
	22	+ o["g_cond"]=COND.get(s.get("condition","inconnu"),2)
	23	+ fl=s.get("flooring") or [];o["g_hardwood"]=int("bois_franc" in fl)
	24	+ o["g_quartz_granite"]=int(s.get("countertop") in ("quartz","granit","marbre"));o["g_view_water"]=int(s.get("view")=="eau");return o
	25	+def parse_area(s):
	26	+ if not s or not isinstance(s,str): return np.nan
	27	+ t=s.lower();m=("m2" in t) or ("m²" in t);n=re.sub(r"[^0-9.,\s]","",t).strip().replace(" ","").replace(",","")
	28	+ try: v=float(n)
	29	+ except: return np.nan
	30	+ return (v*10.7639 if m else v) if v>0 else np.nan
	31	+def cc(c):
	32	+ if not c or not isinstance(c,str): return None
	33	+ c=c.split("(")[0].strip();return re.sub(r"\s+"," ",c).lower().strip() or None
	34	+def ti(x):
	35	+ try: return int(x)
	36	+ except: return np.nan
	37	+AMEN={"pool":["swimming pool","piscine","pool"],"waterfront":["waterfront","bord de l'eau","navigable","lakefront","riverfront","au bord de"],"garage":["garage"],"fireplace":["fireplace","foyer","wood stove","poêle"],"central_air":["central air","air conditionné","climatisation","a/c","heat-pump","thermopompe","air exchanger"],"finished_bsmt":["totally finished","finished basement","sous-sol fini","basement"],"near_water":["near water","waterfront","navigable"],"cul_de_sac":["cul-de-sac","cul de sac"],"golf":["golf"],"ski":["ski"],"near_park":["park","parc"],"near_school":["school","école","cegep","university","université","elementary","high school"],"near_highway":["highway","autoroute"],"transit":["public transit","public transportation","transport"],"no_rear_neighbor":["no backyard neighbors","sans voisin arrière"]}
	38	+def sd(s):
	39	+ s=s.lower()
	40	+ if "ft" in s:
	41	+ f=re.search(r"([\d.]+)\sft",s);i=re.search(r"([\d.]+)\sin",s)
	42	+ return (float(f.group(1)) if f else 0)+(float(i.group(1))/12 if i else 0)
	43	+ m=re.search(r"([\d.,]+)\s*m",s)
	44	+ if m:
	45	+ try: return float(m.group(1).replace(",","."))*3.28084
	46	+ except: return None
	47	+ return None
	48	+def ra(d):
	49	+ if not d or "x" not in d: return None
	50	+ a,_,b=d.partition("x");la,lb=sd(a),sd(b);return la*lb if (la and lb and 1<la<200 and 1<lb<200) else None
	51	+RF={"r_office":["office"],"r_family":["family room"],"r_workshop":["workshop"],"r_winecellar":["wine cellar"],"r_sunroom":["sunroom","solarium"],"r_mezzanine":["mezzanine"],"r_library":["library"],"r_laundry":["laundry"]}
	52	+def pr(rooms):
	53	+ o={"room_area":np.nan,"n_rooms":0,"has_primary_bed":0,"has_basement":0,"n_levels_r":0}
	54	+ for k in RF: o[k]=0
	55	+ if not rooms: return o
	56	+ ar=0.0;lv=set()
	57	+ for r in rooms:
	58	+ nm=(r.get("name") or "").lower();l=(r.get("level") or "").lower();o["n_rooms"]+=1
	59	+ if l: lv.add(l)
	60	+ if "basement" in l or "sous-sol" in l: o["has_basement"]=1
	61	+ if "primary bedroom" in nm: o["has_primary_bed"]=1
	62	+ for k,kw in RF.items():
	63	+ if any(w in nm for w in kw): o[k]=1
	64	+ a=ra(r.get("dimensions"))
	65	+ if a: ar+=a
	66	+ o["room_area"]=ar if ar>0 else np.nan;o["n_levels_r"]=len(lv);return o
	67	+data=json.load(open("data/processed/maisons_avec_prix.json",encoding="utf-8"));rows=[]
	68	+for d in data:
	69	+ p=d.get("property") or {};ad=d.get("address") or {};ft=d.get("features") or []
	70	+ blob=(" ".join(ft)+" "+(d.get("description") or "")+" "+(d.get("title") or "")).lower()
	71	+ yb=ti(p.get("yearBuilt"));age=(CUR_YEAR-yb) if (not pd.isna(yb) and 1700<=yb<=CUR_YEAR) else np.nan
	72	+ pt=p.get("buildingType") or p.get("type");st=p.get("storeys") or p.get("levels")
	73	+ try: st=float(str(st).replace(",","."))
	74	+ except: st=np.nan
	75	+ uid=f"{d.get('source')}:{d.get('listingId')}"
	76	+ rec={"uid":uid,"price":d.get("priceNumber"),"source":d.get("source"),"sqft":parse_area(p.get("squareFeet")),"lot_sqft":parse_area(p.get("lotSize")),"beds":ti(p.get("bedrooms")),"baths":ti(p.get("bathrooms")),"parking":ti(p.get("parkingSpaces")),"storeys":st,"age":age,"ptype":(pt or "Unknown").strip(),"city":cc(ad.get("city"))}
	77	+ for n,kw in AMEN.items(): rec[n]=int(any(k in blob for k in kw))
	78	+ rec["n_images"]=len(d.get("images") or []);rec["virtual_tour"]=int(bool(d.get("virtualTourUrl")))
	79	+ rec.update(pr(d.get("rooms")));rec.update(gfeat(uid));rows.append(rec)
	80	+df=pd.DataFrame(rows)
	81	+df=df[df["price"].notna()];df=df[(df["price"]>=50_000)&(df["price"]<=15_000_000)]
	82	+df.loc[(df["sqft"]<300)\|(df["sqft"]>20_000),"sqft"]=np.nan
	83	+df.loc[(df["lot_sqft"]<200)\|(df["lot_sqft"]>5_000_000),"lot_sqft"]=np.nan
	84	+df.loc[df["beds"]>15,"beds"]=np.nan;df.loc[df["baths"]>12,"baths"]=np.nan
	85	+df["ppsf"]=df["price"]/df["sqft"];lo,hi=df["ppsf"].quantile([0.005,0.995])
	86	+df=df[~(df["ppsf"].notna()&((df["ppsf"]<lo)\|(df["ppsf"]>hi)))]
	87	+ov=df[df["sqft"].notna()&df["room_area"].notna()&(df["room_area"]>200)];ratio=(ov["sqft"]/ov["room_area"]).median()
	88	+re_=df["room_area"]*ratio;ok=df["room_area"].notna()&(re_.between(300,20000))
	89	+df["sqft_final"]=df["sqft"];fill=df["sqft_final"].isna()&ok;df.loc[fill,"sqft_final"]=re_[fill]
	90	+df["has_sqft"]=df["sqft_final"].notna().astype(int);df["age2"]=df["age"]**2;df["n_images_c"]=df["n_images"].clip(upper=60)
	91	+vc=df["city"].value_counts();keep=set(vc[vc>=20].index);df["city_fe"]=df["city"].where(df["city"].isin(keep),"__other__").fillna("__other__")
	92	+tc=df["ptype"].value_counts();df["ptype_fe"]=df["ptype"].where(df["ptype"].isin(set(tc[tc>=30].index)),"Other")
	93	+df=df.reset_index(drop=True)
	94	+y=np.log(df["price"].values)
	95	+amen=list(AMEN.keys());rv=["n_rooms","n_levels_r","has_primary_bed","has_basement","r_office","r_family","r_workshop","r_winecellar","r_sunroom","r_mezzanine","r_library","r_laundry"]
	96	+gv=["g_cond","g_hardwood","g_quartz_granite","g_view_water"]+["g_"+b for b in BOOL]
	97	+num=["sqft_final","has_sqft","lot_sqft","beds","baths","parking","storeys","age","age2","n_images_c","virtual_tour"]+amen+rv+gv
	98	+Xcat=pd.get_dummies(df[["city_fe","ptype_fe","source"]],dtype=float)
	99	+X=pd.concat([df[num].reset_index(drop=True),Xcat.reset_index(drop=True)],axis=1)
	100	+kf=KFold(n_splits=5,shuffle=True,random_state=42)
	101	+gbm=HistGradientBoostingRegressor(max_iter=600,learning_rate=0.05,max_leaf_nodes=63,l2_regularization=1.0,random_state=42)
	102	+pgbm=cross_val_predict(gbm,X,y,cv=kf,n_jobs=-1)
	103	+pnn=np.load("data/processed/nn_robust_oos.npy")
	104	+assert len(pnn)==len(y),f"désalignement {len(pnn)} vs {len(y)}"
	105	+def stat(p):
	106	+ e=np.abs(np.exp(p-y)-1)
	107	+ return r2_score(y,p),np.median(e)100,np.mean(e<=.15)100,np.mean(e<=.25)*100
	108	+print(f"{'Modèle':<30}{'R² OOS':>9}{'MAE méd':>9}{'±15%':>7}{'±25%':>7}")
	109	+for nm,p in [("HistGBM",pgbm),("NN robuste",pnn)]:
	110	+ r2,m,a,b=stat(p);print(f"{nm:<30}{r2:>9.4f}{m:>8.1f}%{a:>6.0f}%{b:>6.0f}%")
	111	+best=None
	112	+for w in [0.3,0.4,0.5,0.6,0.7]:
	113	+ p=wpgbm+(1-w)pnn;r2,m,a,b=stat(p)
	114	+ tag=f"Blend {w:.1f}·GBM+{1-w:.1f}·NN"
	115	+ print(f"{tag:<30}{r2:>9.4f}{m:>8.1f}%{a:>6.0f}%{b:>6.0f}%")
	116	+ if best is None or r2>best[0]: best=(r2,w)
	117	+print(f"\nMeilleur blend: w_GBM={best[1]:.1f} R²={best[0]:.4f}")

added scripts/11_extras_v5_robustness.py +134 −0

		`@@ -0,0 +1,134 @@`
	1	+#!/usr/bin/env python3
	2	+# -- coding: utf-8 --
	3	+"""Robustesse (par segment) + figures primes de localisation & validation IA."""
	4	+import numpy as np, pandas as pd, warnings, statsmodels.formula.api as smf
	5	+import matplotlib; matplotlib.use("Agg"); import matplotlib.pyplot as plt
	6	+warnings.filterwarnings("ignore")
	7	+plt.rcParams.update({"font.family":"serif","font.size":10,"axes.grid":True,
	8	+ "grid.alpha":0.25,"axes.axisbelow":True,"figure.dpi":150,"savefig.bbox":"tight"})
	9	+C1,C2,C3="#3B6EA5","#C4462E","#4A8C5B"
	10	+exec(open("v5_ml.py").read().split("kf=KFold")[0])
	11	+amen_=list(AMEN.keys())
	12	+bl=np.load("data/processed/v5_blend_oos.npy"); price=df["price"].values; y=np.log(price)
	13	+from sklearn.metrics import r2_score
	14	+def stat(mask):
	15	+ yy=y[mask];pp=bl[mask];e=np.abs(np.exp(pp-yy)-1)
	16	+ return len(yy),r2_score(yy,pp),np.median(e)100,np.mean(e<=.15)100
	17	+
	18	+# ---- robustesse par segment ----
	19	+seg=[]
	20	+seg.append(("\\textit{Full sample}",np.ones(len(y),bool)))
	21	+q1,q2=np.quantile(price,[1/3,2/3])
	22	+seg.append(("\\quad Price tercile 1 (low)",price<=q1))
	23	+seg.append(("\\quad Price tercile 2 (mid)",(price>q1)&(price<=q2)))
	24	+seg.append(("\\quad Price tercile 3 (high)",price>q2))
	25	+seg.append(("\\quad Realtor.ca",df["source"].values=="realtor"))
	26	+seg.append(("\\quad DuProprio",df["source"].values=="duproprio"))
	27	+seg.append(("\\quad Measured area",df["has_sqft"].values==1))
	28	+seg.append(("\\quad Imputed area",df["has_sqft"].values==0))
	29	+rows=[]
	30	+for nm,m in seg:
	31	+ n,r2,md,w=stat(m)
	32	+ # R^2 non informatif sur une plage de prix restreinte (terciles) -> n/a
	33	+ r2c = "n/a$^{\\dagger}$" if "tercile" in nm else f"{r2:.3f}"
	34	+ rows.append(f" {nm} & {n:,} & {r2c} & {md:.1f}\\% & {w:.0f}\\% \\\\".replace(",","{,}"))
	35	+tab=r"""% Auto-generated
	36	+\begin{table}[htbp]
	37	+ \centering
	38	+ \begin{threeparttable}
	39	+ \caption{Robustness: Blend Performance Across Subsamples}
	40	+ \label{tab:robustness}
	41	+ \begin{tabular}{@{}lcccc@{}}
	42	+ \toprule
	43	+ Subsample & $N$ & $R^2_{\text{OOS}}$ & Median AE & Within $\pm$15\% \\
	44	+ \midrule
	45	+"""+"\n".join(rows)+r"""
	46	+ \bottomrule
	47	+ \end{tabular}
	48	+ \begin{tablenotes}[flushleft]
	49	+ \footnotesize
	50	+ \item \textit{Notes:} Out-of-sample performance of the text\,+\,vision blend
	51	+ (Table~\ref{tab:models}, Panel~B) computed on subsamples of the same 5-fold
	52	+ cross-validated predictions. Price terciles are defined on the actual list price.
	53	+ ``Measured area'' are listings with a reported or room-reconstructed living area
	54	+ (priced by Model~A); ``Imputed area'' listings are priced by the area-free model.
	55	+ $^{\dagger}$Within a single price tercile the range of $\ln P$ is heavily
	56	+ restricted, so $R^2$---which normalizes by within-group variance---is
	57	+ uninformative; the median absolute error is the comparable accuracy metric across
	58	+ terciles.
	59	+ \end{tablenotes}
	60	+ \end{threeparttable}
	61	+\end{table}
	62	+"""
	63	+open("paper/tables/robustness.tex","w").write(tab)
	64	+print("-> tables/robustness.tex"); [print(f" {nm[:28]:<30} n={stat(m)[0]:5d} R2={stat(m)[1]:.3f} md={stat(m)[2]:.1f}%") for nm,m in seg]
	65	+
	66	+# ---- OLS v5 pour primes de localisation ----
	67	+df["ln_price"]=y
	68	+df["ln_sqft"]=np.log(df["sqft_final"].fillna(df.groupby("ptype")["sqft_final"].transform("median")).fillna(df["sqft_final"].median()))
	69	+df["ln_lot"]=np.log(df["lot_sqft"].fillna(df["lot_sqft"].median()));df["has_lot"]=df["lot_sqft"].notna().astype(int)
	70	+df["has_age"]=df["age"].notna().astype(int)
	71	+for c in ["beds","baths","storeys","age","age2"]:
	72	+ df[c]=df[c].fillna(df.groupby("ptype")[c].transform("median")).fillna(df[c].median())
	73	+df["parking"]=df["parking"].fillna(0)
	74	+form=("ln_price ~ ln_sqft+ln_lot+has_lot+beds+baths+parking+storeys+age+age2+has_age+C(source)+C(ptype_fe)+"+"+".join(amen_)+"+n_images_c+virtual_tour+"+"+".join(rv)+"+"+"+".join(gv)+"+"+"+".join(vv)+"+C(city_fe)")
	75	+mA=smf.ols(form,data=df[df["has_sqft"]==1]).fit(cov_type="HC3")
	76	+pct=lambda b:(np.exp(b)-1)*100
	77	+cf={k.split("T.")[1].rstrip("]"):pct(v) for k,v in mA.params.items() if k.startswith("C(city_fe)")}
	78	+s=pd.Series(cf).sort_values()
	79	+top=s.tail(12);bot=s.head(8)
	80	+fig,ax=plt.subplots(figsize=(6.6,4.6))
	81	+sel=pd.concat([bot,top])
	82	+cols=[C2 if v<0 else C1 for v in sel.values]
	83	+ax.barh(range(len(sel)),sel.values,color=cols,edgecolor="white")
	84	+ax.set_yticks(range(len(sel)));ax.set_yticklabels([c.title() for c in sel.index],fontsize=8)
	85	+ax.axvline(0,color="black",lw=0.8)
	86	+ax.set_xlabel("Location premium vs reference city (\\%)")
	87	+ax.set_title("Municipality fixed effects (largest and smallest)",fontsize=11)
	88	+fig.savefig("paper/figures/fig_premia.pdf");plt.close(fig);print("-> fig_premia.pdf")
	89	+
	90	+# ---- validation IA : prix moyen par score visuel ----
	91	+ext=df[df["v_is_ext"]==1]
	92	+fig,axs=plt.subplots(1,2,figsize=(7.6,3.5))
	93	+for ax,col,lab,c in [(axs[0],"v_curb","Vision curb appeal (1--10)",C3),(axs[1],"v_cond","Vision condition score (1--10)",C1)]:
	94	+ g=ext.groupby(col)["price"].median()/1e3
	95	+ g=g[g.index>=3]
	96	+ ax.plot(g.index,g.values,"-o",color=c,lw=2)
	97	+ ax.set_xlabel(lab);ax.set_ylabel("Median price (\\$000)")
	98	+axs[0].set_title("Price rises with visual curb appeal",fontsize=10)
	99	+axs[1].set_title("Price rises with visual condition",fontsize=10)
	100	+fig.savefig("paper/figures/fig_validation.pdf");plt.close(fig);print("-> fig_validation.pdf")
	101	+
	102	+# ---- table résumé features IA ----
	103	+def sh(col): return df[col].mean()*100
	104	+aitab=r"""% Auto-generated
	105	+\begin{table}[htbp]
	106	+ \centering
	107	+ \begin{threeparttable}
	108	+ \caption{Machine-Generated Feature Summary}
	109	+ \label{tab:ai_summary}
	110	+ \begin{tabular}{@{}lc@{\hskip 1.2cm}lc@{}}
	111	+ \toprule
	112	+ Text (LLM) & Share & Vision (façade) & Mean / share \\
	113	+ \midrule
	114	+"""+f""" Recently renovated & {sh('g_recently_renovated'):.0f}\\% & Condition score (1--10) & {df['v_cond'].mean():.1f} \\\\
	115	+ Renovated kitchen & {sh('g_renovated_kitchen'):.0f}\\% & Curb appeal (1--10) & {df['v_curb'].mean():.1f} \\\\
	116	+ Hardwood floor & {sh('g_hardwood'):.0f}\\% & Luxury cue & {sh('v_luxury'):.0f}\\% \\\\
	117	+ Quartz/granite counter & {sh('g_quartz_granite'):.0f}\\% & Garage visible & {sh('v_garage_vis'):.0f}\\% \\\\
	118	+ New construction & {sh('g_new_construction'):.0f}\\% & Brick façade & {sh('v_mat_brique'):.0f}\\% \\\\
	119	+ Income property & {sh('g_income_property'):.0f}\\% & Waterfront setting & {sh('v_set_bord_eau'):.0f}\\% \\\\
	120	+ Water view & {sh('g_view_water'):.0f}\\% & Exterior shot detected & {sh('v_is_ext'):.0f}\\% \\\\
	121	+""" + r""" \bottomrule
	122	+ \end{tabular}
	123	+ \begin{tablenotes}[flushleft]
	124	+ \footnotesize
	125	+ \item \textit{Notes:} Prevalence of selected machine-generated attributes across
	126	+ the """ + f"{len(df):,}".replace(",","{,}") + r""" cleaned listings. Text attributes are extracted from the
	127	+ description by \texttt{gpt-4o-mini}; vision attributes from the façade photograph by
	128	+ \texttt{claude-haiku-4-5}. Scores are 1--10 means; the remainder are shares.
	129	+ \end{tablenotes}
	130	+ \end{threeparttable}
	131	+\end{table}
	132	+"""
	133	+open("paper/tables/ai_summary.tex","w").write(aitab)
	134	+print("-> tables/ai_summary.tex")

added scripts/12_ablation_extras.py +47 −0

		`@@ -0,0 +1,47 @@`
	1	+#!/usr/bin/env python3
	2	+# -- coding: utf-8 --
	3	+"""Ablation CV complète (4 couches x OLS/GBM) pour cohérence du tableau ablation.tex."""
	4	+import json, warnings
	5	+import numpy as np, pandas as pd
	6	+from sklearn.linear_model import LinearRegression
	7	+from sklearn.ensemble import HistGradientBoostingRegressor
	8	+from sklearn.model_selection import KFold, cross_val_predict
	9	+from sklearn.metrics import r2_score
	10	+warnings.filterwarnings("ignore")
	11	+
	12	+exec(open("v5_ml.py").read().split("kf=KFold")[0])
	13	+kf = KFold(n_splits=5, shuffle=True, random_state=42)
	14	+amen_ = list(AMEN.keys())
	15	+
	16	+L1 = ["sqft_final","has_sqft","lot_sqft","beds","baths","parking","storeys","age","age2"] + amen_
	17	+L2 = L1 + ["n_images_c","virtual_tour"] + rv
	18	+L3 = L2 + gv
	19	+L4 = L3 + vv
	20	+print("feats:", len(L1), len(L2), len(L3), len(L4))
	21	+
	22	+Xcat = pd.get_dummies(df[cats], prefix=cats, dtype=float)
	23	+dfi = df.copy()
	24	+for c in ["sqft_final","lot_sqft","beds","baths","storeys","age","age2"]:
	25	+ dfi[c] = dfi.groupby("ptype")[c].transform(lambda s: s.fillna(s.median()))
	26	+ dfi[c] = dfi[c].fillna(dfi[c].median())
	27	+dfi["parking"] = dfi["parking"].fillna(0)
	28	+
	29	+def stats(pred):
	30	+ e = np.abs(np.exp(pred - y) - 1)
	31	+ return {"r2": float(r2_score(y, pred)), "mdae": float(np.median(e) * 100),
	32	+ "w15": float(np.mean(e <= .15) * 100), "w25": float(np.mean(e <= .25) * 100)}
	33	+def cv_ols(numcols):
	34	+ X = pd.concat([dfi[numcols].reset_index(drop=True), Xcat.reset_index(drop=True)], axis=1)
	35	+ X = X.fillna(X.median())
	36	+ return stats(cross_val_predict(LinearRegression(), X, y, cv=kf, n_jobs=-1))
	37	+def cv_gbm(numcols):
	38	+ X = pd.concat([df[numcols].reset_index(drop=True), Xcat.reset_index(drop=True)], axis=1)
	39	+ m = HistGradientBoostingRegressor(max_iter=600, learning_rate=0.05, max_leaf_nodes=63, l2_regularization=1.0, random_state=42)
	40	+ return stats(cross_val_predict(m, X, y, cv=kf, n_jobs=-1))
	41	+
	42	+out = {}
	43	+for name, cols in [("L1", L1), ("L2", L2), ("L3", L3), ("L4", L4)]:
	44	+ out[name] = {"nfeat": len(cols), "ols": cv_ols(cols), "gbm": cv_gbm(cols)}
	45	+ print(name, out[name])
	46	+json.dump(out, open("results/ablation_extras.json", "w"), indent=1)
	47	+print("-> ablation_extras.json")

added scripts/13_revision_extras.py +112 −0

		`@@ -0,0 +1,112 @@`
	1	+#!/usr/bin/env python3
	2	+# -- coding: utf-8 --
	3	+"""Estimations complémentaires pour la révision JHE :
	4	+ - OLS A/B avec SE HC3 ET clusterisées par municipalité
	5	+ - ablation in-sample du modèle A (base / +texte / +texte+vision)
	6	+ - effets fixes de ville extrêmes
	7	+ - CV 5-fold OLS et RF pour les panels A (texte) et B (texte+vision)
	8	+ Résultats -> revision_extras.json
	9	+"""
	10	+import json, warnings
	11	+import numpy as np, pandas as pd
	12	+import statsmodels.formula.api as smf
	13	+from sklearn.linear_model import LinearRegression
	14	+from sklearn.ensemble import RandomForestRegressor, HistGradientBoostingRegressor
	15	+from sklearn.model_selection import KFold, cross_val_predict
	16	+from sklearn.metrics import r2_score
	17	+warnings.filterwarnings("ignore")
	18	+
	19	+# pipeline v5 jusqu'avant la CV (construit df, AMEN, rv, gv, vv, NUM_v4, NUM_v5, MISS, cats, y, kf)
	20	+exec(open("v5_ml.py").read().split("kf=KFold")[0])
	21	+amen_ = list(AMEN.keys())
	22	+out = {}
	23	+
	24	+# ---------- OLS A/B : HC3 vs cluster ----------
	25	+df["ln_price"] = np.log(df["price"])
	26	+df["ln_sqft"] = np.log(df["sqft_final"].fillna(df.groupby("ptype")["sqft_final"].transform("median")).fillna(df["sqft_final"].median()))
	27	+df["ln_lot"] = np.log(df["lot_sqft"].fillna(df["lot_sqft"].median())); df["has_lot"] = df["lot_sqft"].notna().astype(int)
	28	+df["has_age"] = df["age"].notna().astype(int)
	29	+for c in ["beds","baths","storeys","age","age2"]:
	30	+ df[c] = df[c].fillna(df.groupby("ptype")[c].transform("median")).fillna(df[c].median())
	31	+df["parking"] = df["parking"].fillna(0)
	32	+
	33	+form = ("ln_price ~ ln_sqft+ln_lot+has_lot+beds+baths+parking+storeys+age+age2+has_age+C(source)+C(ptype_fe)+"
	34	+ + "+".join(amen_) + "+n_images_c+virtual_tour+" + "+".join(rv) + "+" + "+".join(gv) + "+" + "+".join(vv) + "+C(city_fe)")
	35	+dfA = df[df["has_sqft"] == 1]
	36	+mA = smf.ols(form, data=dfA).fit(cov_type="HC3")
	37	+mB = smf.ols(form.replace("ln_sqft+",""), data=df).fit(cov_type="HC3")
	38	+mAc = smf.ols(form, data=dfA).fit(cov_type="cluster", cov_kwds={"groups": dfA["city_fe"]})
	39	+mBc = smf.ols(form.replace("ln_sqft+",""), data=df).fit(cov_type="cluster", cov_kwds={"groups": df["city_fe"]})
	40	+
	41	+keys = ["ln_sqft","ln_lot","beds","baths","parking","age","age2",
	42	+ "central_air","transit","waterfront",
	43	+ "g_cond","g_renovated_kitchen","g_hardwood","g_new_construction","g_income_property",
	44	+ "v_curb","v_reno","v_luxury","v_garage_vis",
	45	+ "v_mat_brique","v_mat_pierre","v_mat_vinyle","v_set_rural","v_land","v_cond"]
	46	+def grab(m, mc):
	47	+ o = {}
	48	+ for k in keys:
	49	+ if k in m.params.index:
	50	+ o[k] = {"b": float(m.params[k]), "se_hc3": float(m.bse[k]), "p_hc3": float(m.pvalues[k]),
	51	+ "se_cl": float(mc.bse[k]), "p_cl": float(mc.pvalues[k])}
	52	+ return o
	53	+out["modelA"] = {"r2": mA.rsquared, "adj": mA.rsquared_adj, "n": int(mA.nobs), "coef": grab(mA, mAc)}
	54	+out["modelB"] = {"r2": mB.rsquared, "adj": mB.rsquared_adj, "n": int(mB.nobs), "coef": grab(mB, mBc)}
	55	+print("A R2=%.4f N=%d \| B R2=%.4f N=%d" % (mA.rsquared, mA.nobs, mB.rsquared, mB.nobs))
	56	+
	57	+# ---------- ablation in-sample modèle A ----------
	58	+base_form = ("ln_price ~ ln_sqft+ln_lot+has_lot+beds+baths+parking+storeys+age+age2+has_age+C(source)+C(ptype_fe)+"
	59	+ + "+".join(amen_) + "+n_images_c+virtual_tour+" + "+".join(rv) + "+C(city_fe)")
	60	+mA0 = smf.ols(base_form, data=dfA).fit()
	61	+mA1 = smf.ols(base_form + "+" + "+".join(gv), data=dfA).fit()
	62	+mA2 = smf.ols(base_form + "+" + "+".join(gv) + "+" + "+".join(vv), data=dfA).fit()
	63	+out["ablationA_insample"] = {"base": mA0.rsquared, "plus_text": mA1.rsquared, "plus_text_vision": mA2.rsquared}
	64	+print("ablation A in-sample:", out["ablationA_insample"])
	65	+
	66	+# ---------- effets fixes ville extrêmes ----------
	67	+fes = {k.split("T.")[1].rstrip("]"): float(v) for k, v in mA.params.items() if k.startswith("C(city_fe)")}
	68	+top = sorted(fes.items(), key=lambda kv: -kv[1])[:12]
	69	+bot = sorted(fes.items(), key=lambda kv: kv[1])[:8]
	70	+out["city_fe_top"] = [(k, (np.exp(v) - 1) * 100) for k, v in top]
	71	+out["city_fe_bot"] = [(k, (np.exp(v) - 1) * 100) for k, v in bot]
	72	+print("top villes:", out["city_fe_top"][:4])
	73	+
	74	+# ---------- CV : OLS et RF, panels texte / texte+vision ----------
	75	+kf = KFold(n_splits=5, shuffle=True, random_state=42)
	76	+Xcat = pd.get_dummies(df[cats], prefix=cats, dtype=float)
	77	+dfi = df.copy()
	78	+def stats(pred):
	79	+ e = np.abs(np.exp(pred - y) - 1)
	80	+ return {"r2": float(r2_score(y, pred)), "mdae": float(np.median(e) * 100),
	81	+ "w15": float(np.mean(e <= .15) * 100), "w25": float(np.mean(e <= .25) * 100)}
	82	+def cv_ols(numcols):
	83	+ X = pd.concat([dfi[numcols].reset_index(drop=True), Xcat.reset_index(drop=True)], axis=1)
	84	+ X = X.fillna(X.median())
	85	+ return stats(cross_val_predict(LinearRegression(), X, y, cv=kf, n_jobs=-1))
	86	+def cv_rf(numcols):
	87	+ X = pd.concat([df[numcols].reset_index(drop=True), Xcat.reset_index(drop=True)], axis=1)
	88	+ m = RandomForestRegressor(n_estimators=400, max_features=0.5, min_samples_leaf=3, n_jobs=-1, random_state=42)
	89	+ return stats(cross_val_predict(m, X, y, cv=kf, n_jobs=-1))
	90	+def cv_gbm(numcols):
	91	+ X = pd.concat([df[numcols].reset_index(drop=True), Xcat.reset_index(drop=True)], axis=1)
	92	+ m = HistGradientBoostingRegressor(max_iter=600, learning_rate=0.05, max_leaf_nodes=63, l2_regularization=1.0, random_state=42)
	93	+ return stats(cross_val_predict(m, X, y, cv=kf, n_jobs=-1))
	94	+
	95	+# matrice OLS imputée médiane par type (comme hedonic_ml)
	96	+for c in ["sqft_final","lot_sqft","beds","baths","storeys","age","age2"]:
	97	+ dfi[c] = dfi.groupby("ptype")[c].transform(lambda s: s.fillna(s.median()))
	98	+ dfi[c] = dfi[c].fillna(dfi[c].median())
	99	+dfi["parking"] = dfi["parking"].fillna(0)
	100	+
	101	+print(">>> CV OLS v4/v5")
	102	+out["cv"] = {"ols_v4": cv_ols(NUM_v4), "ols_v5": cv_ols(NUM_v5)}
	103	+print("OLS:", out["cv"]["ols_v4"], out["cv"]["ols_v5"])
	104	+print(">>> CV GBM v4/v5 (contrôle)")
	105	+out["cv"]["gbm_v4"] = cv_gbm(NUM_v4); out["cv"]["gbm_v5"] = cv_gbm(NUM_v5)
	106	+print("GBM:", out["cv"]["gbm_v4"], out["cv"]["gbm_v5"])
	107	+print(">>> CV RF v4/v5 (plus long)")
	108	+out["cv"]["rf_v4"] = cv_rf(NUM_v4); out["cv"]["rf_v5"] = cv_rf(NUM_v5)
	109	+print("RF:", out["cv"]["rf_v4"], out["cv"]["rf_v5"])
	110	+
	111	+json.dump(out, open("results/revision_extras.json", "w"), indent=1, default=float)
	112	+print("-> revision_extras.json")

added scripts/14_figures_v5.py +152 −0

		`@@ -0,0 +1,152 @@`
	1	+#!/usr/bin/env python3
	2	+# -- coding: utf-8 --
	3	+"""Figures + tableau d'ablation détaillé pour le papier v5."""
	4	+import numpy as np, pandas as pd, warnings, json
	5	+import matplotlib
	6	+matplotlib.use("Agg")
	7	+import matplotlib.pyplot as plt
	8	+from sklearn.linear_model import LinearRegression
	9	+from sklearn.ensemble import HistGradientBoostingRegressor
	10	+from sklearn.model_selection import KFold, cross_val_predict
	11	+from sklearn.metrics import r2_score
	12	+from sklearn.inspection import permutation_importance
	13	+warnings.filterwarnings("ignore")
	14	+plt.rcParams.update({"font.family":"serif","font.size":10,"axes.grid":True,
	15	+ "grid.alpha":0.25,"axes.axisbelow":True,"figure.dpi":150,"savefig.bbox":"tight"})
	16	+C_OLS,C_GBM,C_BL="#3B6EA5","#C4462E","#4A8C5B"
	17	+
	18	+# build df via v5_ml prefix
	19	+exec(open("v5_ml.py").read().split("kf=KFold")[0])
	20	+kf=KFold(n_splits=5,shuffle=True,random_state=42)
	21	+onehot=pd.get_dummies(df[cats],dtype=float).reset_index(drop=True)
	22	+
	23	+struct=["sqft_final","has_sqft","lot_sqft","beds","baths","parking","storeys","age","age2"]
	24	+amen_=list(AMEN.keys())
	25	+S0=struct+amen_
	26	+S1=S0+rv+["n_images_c","virtual_tour"]
	27	+S2=S1+gv
	28	+S3=S2+vv
	29	+sets=[("Structural\n+ amenity",S0),("+ Rooms\n+ listing",S1),("+ Text\n(LLM)",S2),("+ Vision\n(façade)",S3)]
	30	+
	31	+def ols_r2(cols):
	32	+ Xi=df[cols].copy()
	33	+ for c in cols: Xi[c]=Xi[c].fillna(df.groupby("ptype")[c].transform("median")).fillna(df[c].median())
	34	+ X=pd.concat([Xi.reset_index(drop=True),onehot],axis=1)
	35	+ p=cross_val_predict(LinearRegression(),X,y,cv=kf,n_jobs=-1);return r2_score(y,p),p
	36	+def gbm_r2(cols):
	37	+ X=pd.concat([df[cols].reset_index(drop=True),onehot],axis=1)
	38	+ m=HistGradientBoostingRegressor(max_iter=600,learning_rate=0.05,max_leaf_nodes=63,l2_regularization=1.0,random_state=42)
	39	+ p=cross_val_predict(m,X,y,cv=kf,n_jobs=-1);return r2_score(y,p),p
	40	+
	41	+ols=[];gbm=[];gpred=None
	42	+for nm,cols in sets:
	43	+ ro,_=ols_r2(cols);rg,pg=gbm_r2(cols);ols.append(ro);gbm.append(rg)
	44	+ if nm.startswith("+ Vision"): gpred=pg
	45	+ print(f"{nm.replace(chr(10),' '):<22} OLS={ro:.4f} GBM={rg:.4f}")
	46	+
	47	+# blend numbers (déjà calculés) : S2 texte, S3 texte+vision
	48	+blend={2:0.760,3:0.779}
	49	+
	50	+# ===== Figure 1 : ablation =====
	51	+x=np.arange(len(sets))
	52	+fig,ax=plt.subplots(figsize=(6.4,3.9))
	53	+ax.plot(x,ols,"-o",color=C_OLS,lw=2,label="OLS (linear)")
	54	+ax.plot(x,gbm,"-s",color=C_GBM,lw=2,label="Gradient boosting")
	55	+ax.plot([2,3],[blend[2],blend[3]],"-D",color=C_BL,lw=2,label="Blend (NN+GBM)")
	56	+for xi,(ro,rg) in enumerate(zip(ols,gbm)):
	57	+ ax.annotate(f"{rg:.3f}",(xi,rg),textcoords="offset points",xytext=(0,7),ha="center",fontsize=8,color=C_GBM)
	58	+ ax.annotate(f"{ro:.3f}",(xi,ro),textcoords="offset points",xytext=(0,-13),ha="center",fontsize=8,color=C_OLS)
	59	+ax.annotate(f"{blend[3]:.3f}",(3,blend[3]),textcoords="offset points",xytext=(0,7),ha="center",fontsize=8,color=C_BL)
	60	+ax.set_xticks(x);ax.set_xticklabels([s[0] for s in sets])
	61	+ax.set_ylabel(r"Out-of-sample $R^2$ (5-fold CV)");ax.set_ylim(0.55,0.80)
	62	+ax.legend(frameon=False,loc="lower right",fontsize=9)
	63	+ax.set_title("Incremental gain from each feature layer",fontsize=11)
	64	+fig.savefig("paper/figures/fig_ablation.pdf");plt.close(fig)
	65	+print("-> fig_ablation.pdf")
	66	+
	67	+# ===== Figure 2 : calibration + résidus (blend v5) =====
	68	+try: bl=np.load("data/processed/v5_blend_oos.npy")
	69	+except Exception: bl=gpred
	70	+yp=np.exp(bl);ya=np.exp(y)
	71	+fig,axs=plt.subplots(1,2,figsize=(7.6,3.6))
	72	+ax=axs[0]
	73	+hb=ax.hexbin(ya/1e3,yp/1e3,gridsize=45,bins="log",cmap="Blues",xscale="log",yscale="log",mincnt=1)
	74	+lim=[60,6000];ax.plot(lim,lim,"--",color="black",lw=1)
	75	+ax.set_xlim(lim);ax.set_ylim(lim);ax.set_xlabel("Actual price (\\$000)");ax.set_ylabel("Predicted price (\\$000)")
	76	+ax.set_title(f"Calibration (blend, $R^2$={r2_score(y,bl):.3f})",fontsize=10)
	77	+r=(yp/ya-1)*100
	78	+ax2=axs[1]
	79	+ax2.hist(np.clip(r,-60,60),bins=50,color=C_BL,alpha=0.85,edgecolor="white",lw=0.3)
	80	+ax2.axvline(0,color="black",lw=1,ls="--")
	81	+ax2.set_xlabel("Prediction error (\\%)");ax2.set_ylabel("Listings")
	82	+ax2.set_title(f"Errors (median \|.\|={np.median(np.abs(r)):.1f}\\%)",fontsize=10)
	83	+fig.savefig("paper/figures/fig_calibration.pdf");plt.close(fig)
	84	+print("-> fig_calibration.pdf")
	85	+
	86	+# ===== Figure 3 : importance par bloc (GBM v5) =====
	87	+Xfull=pd.concat([df[S3].reset_index(drop=True),onehot],axis=1)
	88	+gb=HistGradientBoostingRegressor(max_iter=600,learning_rate=0.05,max_leaf_nodes=63,l2_regularization=1.0,random_state=42).fit(Xfull,y)
	89	+idx=np.random.RandomState(0).choice(len(Xfull),4000,replace=False)
	90	+pi=permutation_importance(gb,Xfull.iloc[idx],y[idx],n_repeats=3,random_state=0,n_jobs=-1)
	91	+imp=pd.Series(pi.importances_mean,index=Xfull.columns).clip(lower=0)
	92	+def block(c):
	93	+ if c.startswith("city_fe") or c.startswith("ptype_fe") or c.startswith("source"): return "Location/type FE"
	94	+ if c in ("sqft_final","has_sqft","lot_sqft"): return "Size"
	95	+ if c in ("beds","baths","parking","storeys","age","age2"): return "Structural"
	96	+ if c in amen_: return "Amenities (rules)"
	97	+ if c in rv or c in ("n_images_c","virtual_tour"): return "Rooms / listing"
	98	+ if c in gv: return "Text (LLM)"
	99	+ if c in vv: return "Vision (façade)"
	100	+ return "Other"
	101	+bl_imp=imp.groupby(imp.index.map(block)).sum().sort_values()
	102	+bl_imp=bl_imp/bl_imp.sum()*100
	103	+fig,ax=plt.subplots(figsize=(6.4,3.6))
	104	+colors=["#8899AA"]*len(bl_imp)
	105	+for i,n in enumerate(bl_imp.index):
	106	+ if n=="Vision (façade)": colors[i]=C_GBM
	107	+ if n=="Text (LLM)": colors[i]=C_BL
	108	+ax.barh(bl_imp.index,bl_imp.values,color=colors,edgecolor="white")
	109	+for i,v in enumerate(bl_imp.values): ax.text(v+0.4,i,f"{v:.0f}\\%",va="center",fontsize=8)
	110	+ax.set_xlabel("Share of permutation importance (\\%)")
	111	+ax.set_title("Feature-block importance (gradient boosting, v5)",fontsize=11)
	112	+ax.set_xlim(0,max(bl_imp.values)*1.15)
	113	+fig.savefig("paper/figures/fig_importance.pdf");plt.close(fig)
	114	+print("-> fig_importance.pdf")
	115	+print("blocs:",{k:round(v,1) for k,v in bl_imp.sort_values(ascending=False).items()})
	116	+
	117	+# ===== Table ablation détaillée =====
	118	+def med(p): return np.median(np.abs(np.exp(p-y)-1))*100
	119	+rowsL=[]
	120	+labels=["Structural + amenity","\\quad + Rooms + listing","\\quad + Text (LLM)","\\quad + Vision (façade)"]
	121	+for i,(nm,cols) in enumerate(sets):
	122	+ ro,po=ols_r2(cols);rg,pg=gbm_r2(cols)
	123	+ bcell = f"{blend[i]:.3f}" if i in blend else "---"
	124	+ rowsL.append(f" {labels[i]} & {len(cols)} & {ro:.3f} & {rg:.3f} & {bcell} \\\\")
	125	+tab=r"""% Auto-generated v5
	126	+\begin{table}[htbp]
	127	+ \centering
	128	+ \begin{threeparttable}
	129	+ \caption{Feature Ablation: Out-of-Sample $R^2$ by Feature Layer}
	130	+ \label{tab:ablation}
	131	+ \begin{tabular}{@{}lcccc@{}}
	132	+ \toprule
	133	+ Feature set & \# feats & OLS & GBM & Blend \\
	134	+ \midrule
	135	+"""+"\n".join(rowsL)+r"""
	136	+ \bottomrule
	137	+ \end{tabular}
	138	+ \begin{tablenotes}[flushleft]
	139	+ \footnotesize
	140	+ \item \textit{Notes:} Each row adds a feature layer to the row above (nested).
	141	+ Entries are out-of-sample $R^2$ on $\ln(\text{Price})$ from 5-fold cross-validation
	142	+ with a common seed. ``\# feats'' counts numeric/indicator features before the city,
	143	+ type, and source fixed effects (added in every specification). The blend averages
	144	+ the robust neural network and gradient boosting and is reported where both were
	145	+ trained. GBM and the neural network handle missing living area natively; OLS uses
	146	+ within-type median imputation.
	147	+ \end{tablenotes}
	148	+ \end{threeparttable}
	149	+\end{table}
	150	+"""
	151	+open("paper/tables/ablation.tex","w").write(tab)
	152	+print("-> paper/tables/ablation.tex")

added scripts/15_make_tables.py +209 −0

		`@@ -0,0 +1,209 @@`
	1	+#!/usr/bin/env python3
	2	+# -- coding: utf-8 --
	3	+"""Génère les tableaux LaTeX (summary stats + régression) avec vrais chiffres."""
	4	+import json, re, warnings
	5	+import numpy as np, pandas as pd
	6	+import statsmodels.formula.api as smf
	7	+warnings.filterwarnings("ignore")
	8	+CUR_YEAR = 2026
	9	+data = json.load(open("data/processed/maisons_avec_prix.json", encoding="utf-8"))
	10	+
	11	+def parse_area(s):
	12	+ if not s or not isinstance(s, str): return np.nan
	13	+ t=s.lower(); m=("m2" in t) or ("m²" in t)
	14	+ n=re.sub(r"[^0-9.,\s]","",t).strip().replace(" ","").replace(",","")
	15	+ try: v=float(n)
	16	+ except: return np.nan
	17	+ return (v*10.7639 if m else v) if v>0 else np.nan
	18	+def clean_city(c):
	19	+ if not c or not isinstance(c,str): return None
	20	+ c=c.split("(")[0].strip(); return re.sub(r"\s+"," ",c).lower().strip() or None
	21	+def to_int(x):
	22	+ try: return int(x)
	23	+ except: return np.nan
	24	+AMEN={"pool":["swimming pool","piscine","pool"],"waterfront":["waterfront","bord de l'eau","navigable","lakefront","riverfront","au bord de"],
	25	+"garage":["garage"],"fireplace":["fireplace","foyer","wood stove","poêle"],
	26	+"central_air":["central air","air conditionné","climatisation","a/c","heat-pump","thermopompe","air exchanger"],
	27	+"finished_bsmt":["totally finished","finished basement","sous-sol fini","basement"],
	28	+"near_water":["near water","waterfront","navigable"],"cul_de_sac":["cul-de-sac","cul de sac"],
	29	+"golf":["golf"],"ski":["ski"],"near_park":["park","parc"],
	30	+"near_school":["school","école","cegep","university","université","elementary","high school"],
	31	+"near_highway":["highway","autoroute"],"transit":["public transit","public transportation","transport"],
	32	+"no_rear_neighbor":["no backyard neighbors","sans voisin arrière"]}
	33	+def side(s):
	34	+ s=s.lower()
	35	+ if "ft" in s:
	36	+ f=re.search(r"([\d.]+)\sft",s); i=re.search(r"([\d.]+)\sin",s)
	37	+ return (float(f.group(1)) if f else 0)+(float(i.group(1))/12 if i else 0)
	38	+ m=re.search(r"([\d.,]+)\s*m",s)
	39	+ if m:
	40	+ try: return float(m.group(1).replace(",","."))*3.28084
	41	+ except: return None
	42	+ return None
	43	+def rarea(d):
	44	+ if not d or "x" not in d: return None
	45	+ a,_,b=d.partition("x"); la,lb=side(a),side(b)
	46	+ return la*lb if (la and lb and 1<la<200 and 1<lb<200) else None
	47	+RF={"r_office":["office"],"r_family":["family room"],"r_workshop":["workshop"],
	48	+"r_winecellar":["wine cellar"],"r_sunroom":["sunroom","solarium"],"r_mezzanine":["mezzanine"],
	49	+"r_library":["library"],"r_laundry":["laundry"]}
	50	+def prooms(rooms):
	51	+ o={"room_area":np.nan,"n_rooms":0,"has_primary_bed":0,"has_basement":0,"n_levels_r":0}
	52	+ for k in RF: o[k]=0
	53	+ if not rooms: return o
	54	+ ar=0.0; lv=set()
	55	+ for r in rooms:
	56	+ nm=(r.get("name") or "").lower(); l=(r.get("level") or "").lower(); o["n_rooms"]+=1
	57	+ if l: lv.add(l)
	58	+ if "basement" in l or "sous-sol" in l: o["has_basement"]=1
	59	+ if "primary bedroom" in nm: o["has_primary_bed"]=1
	60	+ for k,kw in RF.items():
	61	+ if any(w in nm for w in kw): o[k]=1
	62	+ a=rarea(r.get("dimensions"))
	63	+ if a: ar+=a
	64	+ o["room_area"]=ar if ar>0 else np.nan; o["n_levels_r"]=len(lv); return o
	65	+rows=[]
	66	+for d in data:
	67	+ p=d.get("property") or {}; ad=d.get("address") or {}; ft=d.get("features") or []
	68	+ blob=(" ".join(ft)+" "+(d.get("description") or "")+" "+(d.get("title") or "")).lower()
	69	+ yb=to_int(p.get("yearBuilt")); age=(CUR_YEAR-yb) if (not pd.isna(yb) and 1700<=yb<=CUR_YEAR) else np.nan
	70	+ pt=p.get("buildingType") or p.get("type"); st=p.get("storeys") or p.get("levels")
	71	+ try: st=float(str(st).replace(",","."))
	72	+ except: st=np.nan
	73	+ rec={"price":d.get("priceNumber"),"source":d.get("source"),"sqft":parse_area(p.get("squareFeet")),
	74	+ "lot_sqft":parse_area(p.get("lotSize")),"beds":to_int(p.get("bedrooms")),"baths":to_int(p.get("bathrooms")),
	75	+ "parking":to_int(p.get("parkingSpaces")),"storeys":st,"age":age,"ptype":(pt or "Unknown").strip(),
	76	+ "city":clean_city(ad.get("city"))}
	77	+ for n,kw in AMEN.items(): rec[n]=int(any(k in blob for k in kw))
	78	+ rec["n_images"]=len(d.get("images") or []); rec["virtual_tour"]=int(bool(d.get("virtualTourUrl")))
	79	+ rec.update(prooms(d.get("rooms"))); rows.append(rec)
	80	+df=pd.DataFrame(rows)
	81	+df=df[df["price"].notna()]; df=df[(df["price"]>=50_000)&(df["price"]<=15_000_000)]
	82	+df.loc[(df["sqft"]<300)\|(df["sqft"]>20_000),"sqft"]=np.nan
	83	+df.loc[(df["lot_sqft"]<200)\|(df["lot_sqft"]>5_000_000),"lot_sqft"]=np.nan
	84	+df.loc[df["beds"]>15,"beds"]=np.nan; df.loc[df["baths"]>12,"baths"]=np.nan
	85	+df["ppsf"]=df["price"]/df["sqft"]; lo,hi=df["ppsf"].quantile([0.005,0.995])
	86	+df=df[~(df["ppsf"].notna()&((df["ppsf"]<lo)\|(df["ppsf"]>hi)))]
	87	+ov=df[df["sqft"].notna()&df["room_area"].notna()&(df["room_area"]>200)]
	88	+ratio=(ov["sqft"]/ov["room_area"]).median(); re_=df["room_area"]*ratio
	89	+ok=df["room_area"].notna()&(re_.between(300,20000))
	90	+df["sqft_final"]=df["sqft"]; fill=df["sqft_final"].isna()&ok; df.loc[fill,"sqft_final"]=re_[fill]
	91	+df["has_sqft"]=df["sqft_final"].notna().astype(int)
	92	+df["sqft_imp"]=df.groupby("ptype")["sqft_final"].transform(lambda s:s.fillna(s.median())).fillna(df["sqft_final"].median())
	93	+df["has_lot"]=df["lot_sqft"].notna().astype(int); df["lot_sqft"]=df["lot_sqft"].fillna(df["lot_sqft"].median())
	94	+df["has_age"]=df["age"].notna().astype(int); df["age"]=df["age"].fillna(df["age"].median()).clip(lower=0)
	95	+for c in ["beds","baths","storeys"]: df[c]=df[c].fillna(df[c].median())
	96	+df["parking"]=df["parking"].fillna(0)
	97	+df["ln_price"]=np.log(df["price"]); df["ln_sqft"]=np.log(df["sqft_imp"])
	98	+df["ln_lot"]=np.log(df["lot_sqft"]); df["age2"]=df["age"]**2; df["n_images_c"]=df["n_images"].clip(upper=60)
	99	+vc=df["city"].value_counts(); keep=set(vc[vc>=20].index)
	100	+df["city_fe"]=df["city"].where(df["city"].isin(keep),"__other__").fillna("__other__")
	101	+tc=df["ptype"].value_counts(); df["ptype_fe"]=df["ptype"].where(df["ptype"].isin(set(tc[tc>=30].index)),"Other")
	102	+
	103	+amen=list(AMEN.keys()); rv=["n_rooms","n_levels_r","has_primary_bed","has_basement","r_office","r_family","r_workshop","r_winecellar","r_sunroom","r_mezzanine","r_library","r_laundry"]
	104	+extra=["n_images_c","virtual_tour"]+rv
	105	+base="ln_lot + has_lot + beds + baths + parking + storeys + age + age2 + has_age + C(source) + C(ptype_fe) + "+" + ".join(amen)+" + "+" + ".join(extra)+" + C(city_fe)"
	106	+fA="ln_price ~ ln_sqft + "+base; fB="ln_price ~ "+base
	107	+dfA=df[df["has_sqft"]==1].copy()
	108	+mA=smf.ols(fA,data=dfA).fit(cov_type="HC3"); mB=smf.ols(fB,data=df).fit(cov_type="HC3")
	109	+sdA,sdB=np.std(mA.resid),np.std(mB.resid)
	110	+
	111	+# ---------- summary stats table ----------
	112	+def q(s,p): return s.quantile(p)
	113	+def fmt(x,d=0):
	114	+ if d==0: return f"{x:,.0f}".replace(",","{,}")
	115	+ return f"{x:,.{d}f}".replace(",","{,}")
	116	+svars=[("Price (CAD)","price",0),("Living area (ft$^2$)","sqft_final",0),
	117	+("Lot area (ft$^2$)","lot_sqft",0),("Bedrooms","beds",1),("Bathrooms","baths",1),
	118	+("Parking spaces","parking",1),("Storeys","storeys",1),("Age (years)","age",1),
	119	+("Rooms (Realtor)","n_rooms",1),("Photos","n_images",0)]
	120	+lines=[]
	121	+for lab,c,d in svars:
	122	+ s=df[c].dropna();
	123	+ lines.append(f"{lab} & {len(s):,} & {fmt(s.mean(),d)} & {fmt(s.std(),d)} & {fmt(q(s,.25),d)} & {fmt(q(s,.5),d)} & {fmt(q(s,.75),d)} \\\\".replace(",","{,}"))
	124	+share_r=(df.source=="realtor").mean()*100
	125	+tbl_sum=r"""% Auto-generated
	126	+\begin{table}[htbp]
	127	+ \centering
	128	+ \begin{threeparttable}
	129	+ \caption{Summary Statistics}
	130	+ \label{tab:summary_stats}
	131	+ \begin{tabular}{@{}lcccccc@{}}
	132	+ \toprule
	133	+ Variable & $N$ & Mean & Std.\ Dev. & $p_{25}$ & Median & $p_{75}$ \\
	134	+ \midrule
	135	+"""+"\n".join(" "+l for l in lines)+r"""
	136	+ \bottomrule
	137	+ \end{tabular}
	138	+ \begin{tablenotes}[flushleft]
	139	+ \footnotesize
	140	+ \item \textit{Notes:} Descriptive statistics for the cleaned sample of """+f"{len(df):,}".replace(",","{,}")+r""" single-family and townhouse listings (Realtor.ca and DuProprio), scraped on 3~July~2026. """+f"{share_r:.0f}"+r"""\% of listings originate from Realtor.ca. Living area is the reported area, or reconstructed from room dimensions where missing. Continuous variables are bounded by the outlier rules described in Section~\ref{sec:data}; price-per-square-foot is trimmed at the 0.5th and 99.5th percentiles.
	141	+ \end{tablenotes}
	142	+ \end{threeparttable}
	143	+\end{table}
	144	+"""
	145	+open("paper/tables/summary_stats.tex","w").write(tbl_sum)
	146	+
	147	+# ---------- regression table ----------
	148	+def star(p): return "^{*}" if p<.01 else "^{}" if p<.05 else "^{*}" if p<.1 else ""
	149	+disp=[("ln_sqft",r"$\ln(\text{Living area})$"),("ln_lot",r"$\ln(\text{Lot area})$"),
	150	+("beds","Bedrooms"),("baths","Bathrooms"),("parking","Parking"),("age","Age"),
	151	+("central_air","Central A/C"),("transit","Public transit"),("waterfront","Waterfront"),
	152	+("pool","Pool"),("garage","Garage"),("fireplace","Fireplace"),("near_school","Near school"),
	153	+("has_primary_bed","Primary bedroom"),("r_winecellar","Wine cellar"),
	154	+("n_images_c","Photos (count)"),("virtual_tour","Virtual tour"),
	155	+("C(source)[T.realtor]","Realtor.ca (vs DuProprio)")]
	156	+def cell(m,key):
	157	+ if key not in m.params.index: return "",""
	158	+ b=m.params[key]; se=m.bse[key]; p=m.pvalues[key]
	159	+ sign="$-$" if b<0 else ""
	160	+ return f"{sign}{abs(b):.3f}${star(p)}$", f"({se:.3f})"
	161	+rowsL=[]
	162	+for key,lab in disp:
	163	+ a1,a2=cell(mA,key); b1,b2=cell(mB,key)
	164	+ rowsL.append(f" {lab} & {a1} & {b1} \\\\")
	165	+ rowsL.append(f" & {a2} & {b2} \\\\[3pt]")
	166	+tbl_reg=r"""% Auto-generated
	167	+\begin{table}[htbp]
	168	+ \centering
	169	+ \begin{threeparttable}
	170	+ \caption{Hedonic Regression Results}
	171	+ \label{tab:regression}
	172	+ \begin{tabular}{@{}lcc@{}}
	173	+ \toprule
	174	+ & (1) Model~A & (2) Model~B \\
	175	+ & \textit{rich} & \textit{broad} \\
	176	+ \midrule
	177	+"""+"\n".join(rowsL)+r"""
	178	+ \midrule
	179	+ City FE & Yes & Yes \\
	180	+ Property-type FE & Yes & Yes \\
	181	+ Living area & Yes & No \\
	182	+ Observations & """+f"{int(mA.nobs):,}".replace(",","{,}")+r""" & """+f"{int(mB.nobs):,}".replace(",","{,}")+r""" \\
	183	+ $R^2$ & """+f"{mA.rsquared:.3f}"+r""" & """+f"{mB.rsquared:.3f}"+r""" \\
	184	+ Adjusted $R^2$ & """+f"{mA.rsquared_adj:.3f}"+r""" & """+f"{mB.rsquared_adj:.3f}"+r""" \\
	185	+ \bottomrule
	186	+ \end{tabular}
	187	+ \begin{tablenotes}[flushleft]
	188	+ \footnotesize
	189	+ \item \textit{Notes:} OLS estimates of Equation~\eqref{eq:hedonic}. The dependent variable is $\ln(\text{Price})$. Model~A includes $\ln(\text{Living area})$ and is estimated on listings with a measured (reported or room-reconstructed) area; Model~B omits living area and is estimated on the full cleaned sample. All specifications include city and property-type fixed effects and the full set of amenity, room-detail, and listing-signal controls (Section~\ref{sec:methodology}); only selected coefficients are reported. Heteroskedasticity-robust (HC3) standard errors in parentheses. $^{*}$, $^{}$, $^{*}$ denote significance at the 1\%, 5\%, and 10\% levels.
	190	+ \end{tablenotes}
	191	+ \end{threeparttable}
	192	+\end{table}
	193	+"""
	194	+open("paper/tables/regression.tex","w").write(tbl_reg)
	195	+
	196	+# key numbers for the text
	197	+predA=np.exp(mA.predict(df))np.exp(sdA2/2); predB=np.exp(mB.predict(df))np.exp(sdB**2/2)
	198	+df["pred"]=np.where(df["has_sqft"]==1,predA,predB); df["rp"]=(df["price"]/df["pred"]-1)*100
	199	+print("N_clean",len(df),"N_A",int(mA.nobs),"N_B",int(mB.nobs))
	200	+print("R2_A %.4f adj %.4f errpct %.1f"%(mA.rsquared,mA.rsquared_adj,(np.exp(sdA)-1)*100))
	201	+print("R2_B %.4f adj %.4f errpct %.1f"%(mB.rsquared,mB.rsquared_adj,(np.exp(sdB)-1)*100))
	202	+print("median\|err\| %.1f within15 %.0f within25 %.0f"%(df.rp.abs().median(),(df.rp.abs()<=15).mean()100,(df.rp.abs()<=25).mean()100))
	203	+print("nb_cities_fe",df.city_fe.nunique(),"median_price",int(df.price.median()))
	204	+cf={k.split("T.")[1].rstrip("]"):v for k,v in mA.params.items() if k.startswith("C(city_fe)")}
	205	+top=pd.Series(cf).sort_values(ascending=False)
	206	+print("TOP:",[(c,round((np.exp(b)-1)*100)) for c,b in top.head(5).items()])
	207	+print("BOT:",[(c,round((np.exp(b)-1)*100)) for c,b in top.tail(3).items()])
	208	+print("share_realtor %.1f"%((df.source=='realtor').mean()*100))
	209	+print("Tables written to paper/tables/")

added scripts/16_make_tables_v5.py +124 −0

		`@@ -0,0 +1,124 @@`
	1	+#!/usr/bin/env python3
	2	+# -- coding: utf-8 --
	3	+"""Régénère les tableaux LaTeX du papier en v5 (texte GPT + vision)."""
	4	+import numpy as np, pandas as pd, statsmodels.formula.api as smf, warnings
	5	+warnings.filterwarnings("ignore")
	6	+# construire df via le pipeline de v5_ml (jusqu'avant la CV)
	7	+exec(open("v5_ml.py").read().split("kf=KFold")[0])
	8	+amen_=list(AMEN.keys())
	9	+df["ln_price"]=np.log(df["price"])
	10	+df["ln_sqft"]=np.log(df["sqft_final"].fillna(df.groupby("ptype")["sqft_final"].transform("median")).fillna(df["sqft_final"].median()))
	11	+df["ln_lot"]=np.log(df["lot_sqft"].fillna(df["lot_sqft"].median()));df["has_lot"]=df["lot_sqft"].notna().astype(int)
	12	+df["has_age"]=df["age"].notna().astype(int)
	13	+for c in ["beds","baths","storeys","age","age2"]:
	14	+ df[c]=df[c].fillna(df.groupby("ptype")[c].transform("median")).fillna(df[c].median())
	15	+df["parking"]=df["parking"].fillna(0)
	16	+form=("ln_price ~ ln_sqft+ln_lot+has_lot+beds+baths+parking+storeys+age+age2+has_age+C(source)+C(ptype_fe)+"
	17	+ +"+".join(amen_)+"+n_images_c+virtual_tour+"+"+".join(rv)+"+"+"+".join(gv)+"+"+"+".join(vv)+"+C(city_fe)")
	18	+mA=smf.ols(form,data=df[df["has_sqft"]==1]).fit(cov_type="HC3")
	19	+mB=smf.ols(form.replace("ln_sqft+",""),data=df).fit(cov_type="HC3")
	20	+print("A R2=%.4f adj=%.4f N=%d \| B R2=%.4f adj=%.4f N=%d"%(mA.rsquared,mA.rsquared_adj,mA.nobs,mB.rsquared,mB.rsquared_adj,mB.nobs))
	21	+
	22	+def star(p): return "^{*}" if p<.01 else "^{}" if p<.05 else "^{*}" if p<.1 else ""
	23	+def cell(m,k):
	24	+ if k not in m.params.index: return "",""
	25	+ b=m.params[k];se=m.bse[k];p=m.pvalues[k];s="$-$" if b<0 else ""
	26	+ return f"{s}{abs(b):.3f}${star(p)}$",f"({se:.3f})"
	27	+disp=[("ln_sqft",r"$\ln(\text{Living area})$"),("ln_lot",r"$\ln(\text{Lot area})$"),
	28	+("beds","Bedrooms"),("baths","Bathrooms"),("parking","Parking"),("age","Age"),
	29	+("__A__","\\textit{Amenities \\& text (LLM)}"),
	30	+("central_air","Central A/C"),("transit","Public transit"),("waterfront","Waterfront"),
	31	+("g_cond","Condition (LLM, ord.)"),("g_renovated_kitchen","Reno.\\ kitchen (LLM)"),
	32	+("g_hardwood","Hardwood floor (LLM)"),("g_new_construction","New construction (LLM)"),
	33	+("g_income_property","Income property (LLM)"),
	34	+("__V__","\\textit{Vision (façade)}"),
	35	+("v_curb","Curb appeal (1--10)"),("v_reno","Renovation level (ord.)"),
	36	+("v_luxury","Luxury (visual)"),("v_garage_vis","Garage visible"),
	37	+("v_mat_brique","Brick façade"),("v_mat_pierre","Stone façade"),
	38	+("v_mat_vinyle","Vinyl façade"),("v_set_rural","Rural setting")]
	39	+L=[]
	40	+for k,lab in disp:
	41	+ if k in ("__A__","__V__"):
	42	+ L.append(f" \\multicolumn{{3}}{{@{{}}l}}{{{lab}}} \\\\[2pt]");continue
	43	+ a1,a2=cell(mA,k);b1,b2=cell(mB,k)
	44	+ L.append(f" {lab} & {a1} & {b1} \\\\")
	45	+ L.append(f" & {a2} & {b2} \\\\[3pt]")
	46	+reg=r"""% Auto-generated v5
	47	+\begin{table}[htbp]
	48	+ \centering
	49	+ \begin{threeparttable}
	50	+ \caption{Hedonic Regression Results (Text + Vision Features)}
	51	+ \label{tab:regression}
	52	+ \begin{tabular}{@{}lcc@{}}
	53	+ \toprule
	54	+ & (1) Model~A & (2) Model~B \\
	55	+ & \textit{rich} & \textit{broad} \\
	56	+ \midrule
	57	+"""+"\n".join(L)+r"""
	58	+ \midrule
	59	+ City FE / Type FE & Yes & Yes \\
	60	+ Living area & Yes & No \\
	61	+ Observations & """+f"{int(mA.nobs):,}".replace(",","{,}")+r""" & """+f"{int(mB.nobs):,}".replace(",","{,}")+r""" \\
	62	+ $R^2$ & """+f"{mA.rsquared:.3f}"+r""" & """+f"{mB.rsquared:.3f}"+r""" \\
	63	+ Adjusted $R^2$ & """+f"{mA.rsquared_adj:.3f}"+r""" & """+f"{mB.rsquared_adj:.3f}"+r""" \\
	64	+ \bottomrule
	65	+ \end{tabular}
	66	+ \begin{tablenotes}[flushleft]
	67	+ \footnotesize
	68	+ \item \textit{Notes:} OLS estimates of Equation~\eqref{eq:hedonic} with the full
	69	+ feature set. Model~A includes $\ln(\text{Living area})$ (measured-area subsample);
	70	+ Model~B omits it (full sample). ``LLM'' variables are extracted from the
	71	+ description text; ``Vision'' variables from the façade photograph. Curb appeal and
	72	+ renovation level enter as scores; luxury, garage, and material/setting are
	73	+ indicators. All specifications include city and property-type fixed effects and the
	74	+ full amenity, room-detail, and listing-signal controls (only selected coefficients
	75	+ shown). HC3 standard errors in parentheses. $^{*}$, $^{}$, $^{*}$: significance
	76	+ at 1\%, 5\%, 10\%.
	77	+ \end{tablenotes}
	78	+ \end{threeparttable}
	79	+\end{table}
	80	+"""
	81	+open("paper/tables/regression.tex","w").write(reg)
	82	+print("-> paper/tables/regression.tex (v5)")
	83	+
	84	+# ---- models.tex : deux panels (texte / +vision) avec chiffres CV ----
	85	+models=r"""% Auto-generated v5
	86	+\begin{table}[htbp]
	87	+ \centering
	88	+ \begin{threeparttable}
	89	+ \caption{Predictive Performance: From OLS to Deep Learning}
	90	+ \label{tab:models}
	91	+ \begin{tabular}{@{}lcccc@{}}
	92	+ \toprule
	93	+ Estimator & $R^2_{\text{OOS}}$ & Median AE & Within $\pm$15\% & Within $\pm$25\% \\
	94	+ \midrule
	95	+ \multicolumn{5}{@{}l}{\textit{Panel A: Structural + amenity + LLM-text features}} \\[3pt]
	96	+ OLS (linear hedonic) & 0.628 & 17.9\% & 43\% & 64\% \\
	97	+ Random Forest & 0.678 & 18.0\% & 43\% & 64\% \\
	98	+ Neural network (plain MLP) & 0.711 & 16.1\% & 47\% & 69\% \\
	99	+ Neural network (robust ResNet) & 0.741 & 15.2\% & 49\% & 70\% \\
	100	+ Gradient boosting (HistGBM) & 0.743 & 15.4\% & 49\% & 70\% \\
	101	+ Blend (NN + GBM) & 0.760 & 14.7\% & 51\% & 72\% \\[4pt]
	102	+ \multicolumn{5}{@{}l}{\textit{Panel B: \ + Vision (façade) features}} \\[3pt]
	103	+ Gradient boosting (HistGBM) & 0.770 & 14.5\% & 51\% & 71\% \\
	104	+ Neural network (robust ResNet) & 0.747 & 14.8\% & 51\% & 71\% \\
	105	+ \textbf{Blend (NN + GBM)} & \textbf{0.779} & \textbf{14.0\%} & \textbf{53\%} & \textbf{73\%} \\
	106	+ \bottomrule
	107	+ \end{tabular}
	108	+ \begin{tablenotes}[flushleft]
	109	+ \footnotesize
	110	+ \item \textit{Notes:} Out-of-sample performance from 5-fold cross-validation with a
	111	+ common seed (identical folds across rows). $R^2_{\text{OOS}}$ is on
	112	+ $\ln(\text{Price})$; ``Median AE'' is the median absolute percentage error in
	113	+ levels (Equation~\ref{eq:cvmetrics}). Panel~A uses structural, amenity, room-detail,
	114	+ listing-signal, and LLM text-derived features; Panel~B adds the twelve façade
	115	+ vision attributes. City, type, and source enter as one-hot indicators for the
	116	+ linear and tree models and as learned embeddings for the neural networks. The tree
	117	+ and neural models handle missing living area natively. The best model is the
	118	+ text\,+\,vision blend.
	119	+ \end{tablenotes}
	120	+ \end{threeparttable}
	121	+\end{table}
	122	+"""
	123	+open("paper/tables/models.tex","w").write(models)
	124	+print("-> paper/tables/models.tex (v5, 2 panels)")

added scripts/17_gen_regression_table.py +85 −0

		`@@ -0,0 +1,85 @@`
	1	+#!/usr/bin/env python3
	2	+# -- coding: utf-8 --
	3	+"""Régénère paper/tables/regression.tex à partir de revision_extras.json
	4	+ avec SE clusterisées par municipalité (et stars associées)."""
	5	+import json
	6	+
	7	+R = json.load(open("results/revision_extras.json"))
	8	+
	9	+def star(p):
	10	+ return "^{*}" if p < .01 else "^{}" if p < .05 else "^{*}" if p < .1 else ""
	11	+
	12	+def cell(model, k):
	13	+ c = R[model]["coef"].get(k)
	14	+ if not c:
	15	+ return "", ""
	16	+ b, se, p = c["b"], c["se_cl"], c["p_cl"]
	17	+ s = "$-$" if b < 0 else ""
	18	+ st = star(p)
	19	+ coef = f"{s}{abs(b):.3f}" + (f"${st}$" if st else "")
	20	+ return coef, f"({se:.3f})"
	21	+
	22	+disp = [("ln_sqft", r"$\ln(\text{Living area})$"), ("ln_lot", r"$\ln(\text{Lot area})$"),
	23	+ ("beds", "Bedrooms"), ("baths", "Bathrooms"), ("parking", "Parking"), ("age", "Age"),
	24	+ ("__A__", r"\textit{Amenities \& text (LLM)}"),
	25	+ ("central_air", "Central A/C"), ("transit", "Public transit"), ("waterfront", "Waterfront"),
	26	+ ("g_cond", "Condition (LLM, ord.)"), ("g_renovated_kitchen", r"Reno.\ kitchen (LLM)"),
	27	+ ("g_hardwood", "Hardwood floor (LLM)"), ("g_new_construction", "New construction (LLM)"),
	28	+ ("g_income_property", "Income property (LLM)"),
	29	+ ("__V__", r"\textit{Vision (façade)}"),
	30	+ ("v_curb", "Curb appeal (1--10)"), ("v_reno", "Renovation level (ord.)"),
	31	+ ("v_luxury", "Luxury (visual)"), ("v_garage_vis", "Garage visible"),
	32	+ ("v_mat_brique", "Brick façade"), ("v_mat_pierre", "Stone façade"),
	33	+ ("v_mat_vinyle", "Vinyl façade"), ("v_set_rural", "Rural setting")]
	34	+
	35	+L = []
	36	+for k, lab in disp:
	37	+ if k in ("__A__", "__V__"):
	38	+ L.append(f" \\multicolumn{{3}}{{@{{}}l}}{{{lab}}} \\\\[2pt]")
	39	+ continue
	40	+ a1, a2 = cell("modelA", k)
	41	+ b1, b2 = cell("modelB", k)
	42	+ L.append(f" {lab} & {a1} & {b1} \\\\")
	43	+ L.append(f" & {a2} & {b2} \\\\[3pt]")
	44	+
	45	+nA, nB = R["modelA"]["n"], R["modelB"]["n"]
	46	+tex = r"""% Auto-generated (revision: clustered SEs)
	47	+\begin{table}[htbp]
	48	+ \centering
	49	+ \begin{threeparttable}
	50	+ \caption{Hedonic Regression Results (Text + Vision Features)}
	51	+ \label{tab:regression}
	52	+ \begin{tabular}{@{}lcc@{}}
	53	+ \toprule
	54	+ & (1) Model~A & (2) Model~B \\
	55	+ & \textit{rich} & \textit{broad} \\
	56	+ \midrule
	57	+""" + "\n".join(L) + f"""
	58	+ \\midrule
	59	+ City FE / Type FE & Yes & Yes \\\\
	60	+ Living area & Yes & No \\\\
	61	+ Observations & {nA:,} & {nB:,} \\\\
	62	+ $R^2$ & {R['modelA']['r2']:.3f} & {R['modelB']['r2']:.3f} \\\\
	63	+ Adjusted $R^2$ & {R['modelA']['adj']:.3f} & {R['modelB']['adj']:.3f} \\\\
	64	+ \\bottomrule
	65	+ \\end{{tabular}}
	66	+""".replace(f"{nA:,}", f"{nA:,}".replace(",", "{,}")).replace(f"{nB:,}", f"{nB:,}".replace(",", "{,}")) + r""" \begin{tablenotes}[flushleft]
	67	+ \footnotesize
	68	+ \item \textit{Notes:} OLS estimates of Equation~\eqref{eq:hedonic} with the full
	69	+ feature set. Model~A includes $\ln(\text{Living area})$ (measured-area subsample);
	70	+ Model~B omits it (full sample). ``LLM'' variables are extracted from the
	71	+ description text; ``Vision'' variables from the façade photograph. Curb appeal and
	72	+ renovation level enter as scores; luxury, garage, and material/setting are
	73	+ indicators. All specifications include city and property-type fixed effects and the
	74	+ full amenity, room-detail, and listing-signal controls (only selected coefficients
	75	+ shown). Standard errors clustered by municipality (195 clusters) in parentheses.
	76	+ Heteroskedasticity-robust (HC3) standard errors are uniformly smaller and yield
	77	+ the same significance pattern for the text- and vision-derived attributes, except
	78	+ that curb appeal in Model~B is significant under HC3 but not under clustering.
	79	+ $^{*}$, $^{}$, $^{*}$: significance at 1\%, 5\%, 10\%.
	80	+ \end{tablenotes}
	81	+ \end{threeparttable}
	82	+\end{table}
	83	+"""
	84	+open("paper/tables/regression.tex", "w").write(tex)
	85	+print("-> paper/tables/regression.tex (clustered SEs)")
	86

		`@@ -0,0 +1,10 @@`
	1	+# Processed data (not committed)
	2	+
	3	+\| File \| Produced by \| Content \|
	4	+\|---\|---\|---\|
	5	+\| `maisons_avec_prix.json` \| house filter + cleaning described in `docs/METHODOLOGIE.md` §2–4 (combined Realtor + DuProprio houses with price) \| 19,123 house listings \|
	6	+\| `specs.jsonl` \| `scripts/01_extract_specs_llm.py` (GPT-4o-mini, strict JSON schema, resumable cache keyed by uid) \| structured specifications extracted from descriptions \|
	7	+\| `facade_specs.jsonl` \| `scripts/02_facade_vision.py` (vision–language model on the façade photo) \| objective visual attributes (curb appeal, cladding, …) \|
	8	+\| `maisons_specs.json` / `.csv` \| `scripts/03_merge_llm_features.py` \| listings merged with LLM features \|
	9	+\| `maisons_hedonic.csv`, `maisons_hedonic_v4.csv` \| `scripts/04_…`, `scripts/05_…` \| final datasets with fitted price and `resid_pct` \|
	10	+\| `nn_robust_oos.npy`, `v5_blend_oos.npy` \| `scripts/08_…`, `scripts/09_…` \| out-of-sample predictions (5-fold, seed 42) reused by steps 10–14 \|