# 🏠 Hedonic House Pricing with Language-Model and Vision Features **UQO Working Paper No. 14** — *Evidence from Quebec Real-Estate Listings* [-b31b1b?logo=latex&logoColor=white)](paper/main.pdf) [](requirements.txt) [](data/raw/README.md) [](scripts/) [](https://uqo.ca) [](mailto:simon-pierre.boucher@uqo.ca) > **TL;DR** — Much of a home's value is recorded only in the free text and the photographs of its listing. We use a > large language model to turn each description into structured specifications and a vision–language model to turn the > façade photo into objective visual attributes, both under strict JSON schemas. On 19,123 Quebec house listings the two > 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 > out-of-sample gain of any feature block for every estimator (OLS, gradient boosting, neural nets); the best ensemble > reaches out-of-sample R² 0.779 with a 14.0% median absolute error. Implicit prices are economically sensible (+2.3% per > point of curb appeal, +3–4% for brick or stone) and survive municipality-clustered inference. --- ## 📌 Status | Item | Value | |---|---| | Version | 2.0 (July 2026) — revised draft after an internal review round (`scripts/13_revision_extras.py`) | | SSRN | not yet submitted | | Journal | Elsevier-format highlights prepared (`docs/highlights_elsevier.tex`); target to be decided | | Code & data | reproducible from `scripts/` run from the repository root; raw crawls and LLM caches not committed | ## 🎯 What this paper does Hedonic models are bounded by the characteristics an analyst can measure. We treat general-purpose AI models as a **feature-engineering technology**: each listing description is parsed by an LLM into a fixed schema of property specifications (renovations, finishes, systems, extras), and each façade photograph is parsed by a vision–language model into objective attributes (cladding, storeys, garage, curb appeal). The outputs are ordinary econometric variables and enter a log-linear hedonic regression with city fixed effects: ``` ln P_i = α_city(i) + X_i'β + S_i'γ (LLM text layer) + V_i'δ (vision layer) + ε_i ``` We then measure, under identical 5-fold splits, how much each layer adds for OLS, histogram gradient boosting, a residual MLP with city/type embeddings, and a stacked blend. ## 🏆 Headline results | Object | Estimate | |---|---| | Base hedonic OLS (structural + location), in-sample R² | 0.717 | | + LLM text layer + vision layer | 0.755 | | Out-of-sample R², OLS: base → +text → +text+vision (5-fold) | 0.52 → 0.62 → … (see `paper/tables/ablation.tex`) | | Out-of-sample R², GBM: base → +text | 0.69 → 0.73 | | Best ensemble (robust NN + GBM blend) | R²_OOS = 0.779, median absolute error 14.0% | | Curb-appeal score (1–5) | +2.3% per point | | Brick / stone façade | +3–4% | | Inference | HC3 and municipality-clustered standard errors agree | ## 🗂 Repository layout ``` hedonic-llm-vision-features/ ├── data/ │ ├── raw/ # two Apify crawls (Realtor.ca, DuProprio), 2026-07-03 — not committed │ └── processed/ # LLM/vision caches, merged datasets, OOS predictions — not committed ├── scripts/ # run from the repository root, in order │ ├── 00a_test_openai_extraction.py / 00b_test_vision_claude.py # API smoke tests │ ├── 01_extract_specs_llm.py # description → JSON specs (parallel, resumable) │ ├── 02_facade_vision.py # façade photo → visual attributes │ ├── 03_merge_llm_features.py │ ├── 04_hedonic_ols_base.py / 05_hedonic_ols_v4_text.py │ ├── 06_hedonic_ml_cv.py / 07_torch_nn.py / 08_torch_nn_robust.py │ ├── 09_v5_ml_text_vision.py / 10_final_blend.py │ ├── 11_extras_v5_robustness.py / 12_ablation_extras.py / 13_revision_extras.py │ └── 14_figures_v5.py / 15_make_tables.py / 16_make_tables_v5.py / 17_gen_regression_table.py ├── results/ # hedonic summaries (txt), ablation_extras.json, revision_extras.json ├── docs/ # METHODOLOGIE.md + methodologie.tex/pdf (French data & method note), Elsevier highlights └── paper/ # main.tex + sections/ + tables/ + figures/ + references.bib → main.pdf ``` ## 🚀 Quick start ```bash python3 -m pip install -r requirements.txt cp .env.example .env # OPENAI_API_KEY=…, ANTHROPIC_API_KEY=… (only steps 00–02 call the APIs) # place the two Apify crawls in data/raw/ and build data/processed/maisons_avec_prix.json (docs/METHODOLOGIE.md §2–4), then: WORKERS=24 python3 scripts/01_extract_specs_llm.py python3 scripts/02_facade_vision.py python3 scripts/03_merge_llm_features.py python3 scripts/05_hedonic_ols_v4_text.py python3 scripts/08_torch_nn_robust.py && python3 scripts/09_v5_ml_text_vision.py && python3 scripts/10_final_blend.py python3 scripts/11_extras_v5_robustness.py && python3 scripts/12_ablation_extras.py && python3 scripts/13_revision_extras.py python3 scripts/14_figures_v5.py && python3 scripts/16_make_tables_v5.py && python3 scripts/17_gen_regression_table.py cd paper && make # latexmk → main.pdf ``` LLM extraction ≈ 600 tokens in / 110 out per listing (24 threads); vision ≈ 2,470 tokens in per image (16 threads); both are resumable. Econometrics and ML ≈ 20 minutes on an Apple-silicon laptop (PyTorch on MPS). ## 🗃 Data 27,302 Realtor.ca and 7,693 DuProprio listings crawled on 2026-07-03; houses (single-family and townhouse) with a valid price retained, outliers removed by the rules in `docs/METHODOLOGIE.md` §3, giving 19,123 observations. Structural variables (living area, lot, rooms, year built, type, city), text-derived specifications and façade attributes. See `data/raw/README.md` and `data/processed/README.md`. ## 🔬 Methodology 1. **AI feature layers** with strict JSON schemas and retry/backoff — `scripts/01_…`, `scripts/02_…`. 2. **Log-linear hedonic OLS** with city fixed effects, HC3 and municipality-clustered SEs — `scripts/05_…`, `scripts/13_…`. 3. **Ablation by feature block × estimator** under identical folds (OLS, GBM, NN) — `scripts/09_…`, `scripts/12_…`. 4. **Stacked blend** of the robust NN and HistGradientBoosting — `scripts/10_final_blend.py`. 5. **Robustness** by segment, extreme city effects, calibration and importance plots — `scripts/11_…`, `scripts/14_…`. ## 🖼 Figures & tables inventory | Figure | Content | |---|---| | `fig_ablation` | out-of-sample gain by feature block and estimator | | `fig_calibration` | predicted vs actual, blend | | `fig_importance` | feature importance | | `fig_premia` | location premia | | `fig_validation` | validation of AI-extracted attributes | | `nn_diagram` (TikZ) | architecture of the robust NN | Tables (`paper/tables/*.tex`): summary stats, regression (A/B panels), models, ablation, robustness, AI summary. ## ⚠️ Limitations - Asking prices from listings, single cross-section (July 2026). - Realtor.ca listings lack coordinates; location enters through city fixed effects and postal codes. - LLM/vision outputs are model-dependent; schemas and prompts are versioned in the scripts. - API costs make the extraction layers the expensive part of replication. ## 📝 Citation ```bibtex @techreport{boucher2026llmvisionhedonic, author = {Boucher, Simon-Pierre}, title = {Hedonic House Pricing with Language-Model and Vision Features: Evidence from Quebec Real-Estate Listings}, institution = {Universit\'e du Qu\'ebec en Outaouais, D\'epartement des sciences administratives}, type = {Working Paper}, number = {14}, year = {2026}, month = {July} } ``` ## 👤 Author & contact **Simon-Pierre Boucher** — Département des sciences administratives, Université du Québec en Outaouais (UQO), Gatineau, QC. 📧 [simon-pierre.boucher@uqo.ca](mailto:simon-pierre.boucher@uqo.ca) ## 🗒 Notes - Numbering: this paper was first issued as "Working Paper No. 11" (July 2026); number 11 belongs to *Half a Million Prices, Twenty Models*. Renumbered **14** on 2026-10-17; cover date fixed to July 2026 (was `\today`). - Harmonised 2026-10-17: flat scripts moved to `scripts/` (numbered), data to `data/`, logs removed, summaries to `results/`, methodology note to `docs/`; file paths inside the scripts now point to `data/…` and `results/…` (run them from the repository root). Bibliography switched from `plainnat` to `aer`. - The script `13_revision_extras.py` was written for a Journal of Housing Economics revision round (clustered SEs, in-sample ablation, extreme city effects).