SPB Git
spboucher.ai

spb/wp2_uqo

Public

UQO Working Paper No. 2 — Decoding Real Estate Descriptions: text-based hedonic analysis of housing listings.

researcheconomicshousinghedonicnlpuqo
page 1

  1. Wed, Nov 4, 2026 2

    1. Stage G: final build of the revised manuscript (80 pages)
      …
      Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
      Simon-Pierre Boucher · 1 file
    2. Stage F B1/B2 + NB-01..11 — abstract/README: none survives over the grid and the 200 rephrasings together (B1), 414 FSA FE via \nFsaMain in abstract/§1/README (B2), distance-blocked increment −0.041 with CW p stated (NB-11); §1 main-sample language shares and labelled truncation shares via new macros (NB-04); §8.1 Needs Renovation 85 %/60 % (NB-09); §4.1/§8.7 text-free data not in the committed package, deposit = author action (NB-02); eq:embed/sim/modelB/modelC/modelFE referenced (NB-03); Fig. 1 n = 16,553 from sample_sizes.json (NB-05); Table 14 landscape footnotesize unscaled (NB-06); Table 10 star note under B = 999 (NB-10); survival_table fallback so 12 --quick completes + tests/test_quick_mode.py (NB-01); e9 artefacts removed, .gitignore deduplicated (NB-08); zhang2024describe left as the Crossref print record (NB-07); main.pdf rebuilt (80 pp., 0 errors, 0 undefined, overfull ≤ 1.6 pt)
      …
      Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
      Simon-Pierre Boucher · 23 files +134 −42

  2. Tue, Nov 3, 2026 10

    1. Stage E integration (Agent 10 after Agent 9's hand-off) — figure floats \input from paper/figures/, Table 1 notes block, README v2.0, final build
      …
      Sections: every inline \includegraphics replaced by \input{figures/figN_*} (14 floats, same labels) with a '% Generated by' line; fig8 (LMG panel) placed in §5.2, fig6 (illustrative scatters) in Appendix B; the three prose descriptions of Figs 13–15 corrected (fig13 = per-dimension specification curves, not block gains); '% Generated by' line above all 30 table \input; Table 1 rebuilt as a threeparttable with a notes block.
Exhibits (Agent 9 finished; generators patched with a minimal change): src/tex.py figure() gains a height cap; fig13 capped at 0.78\textheight (float was 34 pt too tall); tab_extended_specs_v2 refactored to estimate/CI side by side in scriptsize and split into a 4-specification main table + tab_extended_specs_v2_lang (French/English subsamples, Appendix C) — the 12-column table was 162 pt too tall for the page.
Preamble: the threeparttable/adjustbox shim is KEPT — without it the build has 0 errors but threeparttable measures the unscaled tabular and the notes paragraph overflows the text width (17 overfull boxes up to 107 pt, Table 1 included); with it 0 overfull > 5 pt.
README: TL;DR, status, headline results, methodology and build lines rewritten from the macro values (v2.0, 79 pages).
Final build: latexmk -pdf main.tex — 0 errors, 0 undefined references/citations, 0 'Float too large', max overfull 1.6 pt, 79 pages, metadata set; paper/main.pdf committed.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
      Simon-Pierre Boucher · 18 files +280 −298
    2. Stage E exhibits (Agent 9) — rebuild of every table and figure to the Part X standard: src/tex.py + src/figstyle.py design system (threeparttable notes, Okabe-Ito, viridis, vector PDF at text width), 31 generated tables with SE/CI, named inference, N/clusters/R2 rows, economic-magnitude notes and source lines; 16 figures redrawn with self-contained floats in paper/figures/; tab_references and tab_robustness_summary now generated; tab_block_tests from block_tests_v2 (F(q,G-1), Holm, preferred spec) — WP02-EDIT-03 EDIT-13 EDIT-14 ADVE-10 ADVE-12 ADVE-17 ADVE-18 ECON-11 ECON-13 ECON-14 THEO-12 THEO-13 DATA-09 LITE-06
      …
      Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
      Simon-Pierre Boucher · 88 files +2,561 −1,731
    3. §5 — alias label sec:extended kept on the Results section so that the generated table notes still pointing to the first-version 'extended specifications' section resolve
      …
      Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
      Simon-Pierre Boucher · 1 file +2
    4. §1, §5.3, inline macros — sentences no longer open with a numeral macro; values that round to zero carry no sign (Motivated Seller 0.0 % in the preferred specification)
      …
      Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
      Simon-Pierre Boucher · 4 files +5 −3
    5. §4.5, §5.5, abstract — truncation sentence made macro-based (no typed fraction), n of the LLM-coded label regression from the CSV (\nLabelRegression), abstract trimmed to 148 words (ADVE-15, ADVE-03)
      …
      Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
      Simon-Pierre Boucher · 5 files +11 −4
    6. Stage E writing — abstract, §1–§9, appendices A–C rewritten on the Stage D evidence (WP02 clusters A–I)
      …
      Abstract/§1/§9: headline re-anchored on the preferred specification (ΔR² 0.030 [0.026, 0.035], F(20,413) = 48.2; OOS +0.030/+0.026/+0.015/−0.041 by fold design) — EDIT-08 THEO-11 ADVE-02 EDIT-04 ECON-03 THEO-04 ADVE-06 ADVE-09 ADVE-12 THEO-14; English contrasts per unit (+44.3 → +19.9 → +15.0 MiniLM; +8.4 [5.8, 11.1] preferred) — ECON-02 THEO-01 EDIT-15 THEO-06; 'five robust dimensions' withdrawn, Luxury only under the pre-stated criterion, none over grid × phrasing — EDIT-01 EDIT-02 EDIT-05 ADVE-01 ADVE-05 DATA-06 ECON-01 ECON-04; LMG attribution replaces the order-dependent 63 % — ECON-10 THEO-07; symmetric 61 % truncation and chunk-and-pool — EDIT-07 ADVE-07 THEO-02; span invariance stated — THEO-03 EDIT-16; literature repositioned, Table 1 rebuilt, competitors and precedents cited, sibling SSRN reconciled — LITE-01..08 EDIT-18 ADVE-18; data audit, DAS, LLM-coded validation labelled everywhere — DATA-01 DATA-03 DATA-05 DATA-07 DATA-08 DATA-10 DATA-13 EDIT-06 EDIT-09 EDIT-10 EDIT-11 ADVE-03 ADVE-04 THEO-05 THEO-08 THEO-09 THEO-15 ECON-06 ECON-07 ECON-12; spatial inference and Conley rule — EDIT-12 ECON-05 ECON-11; Model E post-selection, causal nouns and U8 markers removed — EDIT-17 ECON-08 THEO-10 ADVE-14 ADVE-10 ECON-13 THEO-12 ADVE-11 ADVE-15 LITE-06 ADVE-16 DATA-12.
Structure: sections/extended.tex and sections/appendix.tex replaced by sections/oos.tex and appendix/{a_references,b_baseline,c_audit}.tex; first-version Models A–E moved to Appendix B; cleveref added; PDF subject/keywords set; title change proposed in a comment only.
All numbers are macros from paper/inline_macros.tex (paper/make_inline_macros.py, 1,085 macros). Compile status: builds cleanly (0 errors, 75 pages, no undefined references or citations) in a scratch copy where tables/tab_oos_v2.tex's malformed '$p$ < 0.001$' cells are corrected; in the worktree the build is blocked by that generator bug (Agent 9's scripts/18_tables_v2.py) — the threeparttable/adjustbox 'Missing \endgroup' failure is neutralised by the shim added to preamble.tex.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
      Simon-Pierre Boucher · 18 files +439 −375
    7. references.bib + inline macros — WP02-LITE-01 LITE-02 LITE-04 LITE-05 LITE-07: 32 DOIs added, Hansen 2018 / Zietz 2008, NeurIPS versions of Lundberg-Lee and MiniLM, 11 verified references added (Bajari 2025, Zhang 2024, Kolbe 2021, Faxvaag 2025, Kozlowski 2019, Grand 2022, Moschella 2023, Rodriguez 2023, Boucher 2026 SSRN, Opitz 2025, Wegrzyn 2025); paper/make_inline_macros.py generates paper/inline_macros.tex (1,082 macros) from results/inline_numbers.json and results CSVs so that no number is typed in the prose
      …
      Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
      Simon-Pierre Boucher · 3 files +2,026 −41
    8. WP02-ADVE-16 DATA-12 — robustness_results.csv regenerated by scripts/04_robustness.py --legacy: all 163 v1.2 rows identical (max relative difference 0), one new row Winsorized_1pct / R2_adj = 0.486 (full-sample 0.511)
      …
      Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
      Simon-Pierre Boucher · 1 file +1
    9. Stage D batches N2/N3/N4/N5/L2 — WP02-ECON-02 THEO-01 EDIT-15 THEO-06 ADVE-05 ECON-07 ECON-10 THEO-07 ECON-11 EDIT-12 ECON-05 ADVE-03 EDIT-01 EDIT-02 ADVE-01 EDIT-04 ECON-03 THEO-14 ADVE-02 THEO-04 ADVE-06 ADVE-12 ADVE-09 ADVE-13 DATA-02 ADVE-08 DATA-08 DATA-07 DATA-09 DATA-11 DATA-01 DATA-03 EDIT-08 THEO-11 ADVE-18
      …
      N2: scripts/12_specifications_v2.py (11 designs x 5 encoder variants x {raw, centred contrasts, Löwdin}
x {S_main, S_v11, S_nodup} x {rule, langdetect} = 530 cells; unscaled dummies; per-0.10-cosine scale;
HC3 / FSA-clustered t and F(q, G-1); Holm, BH, Romano–Wolf (FSA-cluster bootstrap B=999) on the 55
main-grid cells with a pre-stated 90 % survival criterion; LMG decomposition; domain tests with Holm;
Models A–E on the v2 design; leverage of the raw lot field; --legacy reproduces the 14 v1.2 rows at 1e-14),
scripts/15_spatial_inference.py (Moran's I k=10 with permutation p; Conley 5/10/20 km; decision rule).
N3: scripts/13_multiverse.py (200 phrasing draws x 2 encoders x 3 specs x {raw, centred} x {S_main, S_v11};
the three v1.2 variants reproduced). N4: scripts/14_oos_v2.py (random x20, FSA x20, leave-one-region-out,
GroupKFold by region, k-means distance blocks + 10 km buffer, duplicate-grouped, S_nodup; training-fold FE
with municipality/region/grand-mean fallback; Clark–West, Diebold–Mariano, bootstrap CIs, MCS; price-masked
encoders; --legacy reproduces revision_oos.csv at 1e-16). N5: scripts/16_pca_benchmark.py (Table 9 from a
script; PC1 correlations), scripts/17_inline_numbers.py (results/inline_numbers.json, 2,607 keys),
run_all.sh, tests/ (data, legacy results at rtol 1e-6, embeddings at 1e-5, leakage), environment/
(lock with transitive pins, pip freeze, hardware, HF revisions, SHA-256), data/DATA_AVAILABILITY.md
(users-table exclusion, scrape window 7–11 Jan 2026, shareable extract), LICENSE (MIT), data citation and
model-card entries in references.bib, README runtimes. L2: scripts/18_tables_v2.py (16 tables with labels,
Table 9 regenerated from CSV without self-ratings) and scripts/19_figures_v2.py (fig13–fig16).
results/v2/*.csv|json committed; no data, embeddings or texts.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
      Simon-Pierre Boucher · 87 files +110,028 −18
    10. Stage D batches L1/N0/N1 — WP02-EDIT-05 ECON-04 ECON-01 DATA-06 ECON-09 EDIT-03 ECON-14 EDIT-13 ADVE-17 THEO-13 EDIT-14 ADVE-10 ECON-13 THEO-12 ADVE-16 DATA-12 EDIT-17 ADVE-14 ECON-08 THEO-10 DATA-05 EDIT-09 THEO-09 EDIT-06 ADVE-04 DATA-10 ECON-12 THEO-05 EDIT-10 THEO-15 EDIT-11 DATA-13 ECON-06 THEO-08 ADVE-03 EDIT-07 ADVE-07 THEO-02 THEO-03 EDIT-16 DATA-04 DATA-07
      …
      L1: scripts/20_stored_checks.py (survival over 14 specs + 3 phrasing variants, Holm/BH counts,
per-unit dummy contrasts, CIs, MDE from the stored v1.2 CSVs); 06/08 generators (escape fix,
association wording, Model E post-selection label, SE/CI companion table, lot-area columns,
Fig. 12 error bars, Table 15 baseline row from CSV); 04 (R2_adj, Trimmed label, --legacy);
05 (Fig. 2 labels beyond CI, Fig. 9 clip stated, similarity_correlations.csv export).
N0: src/data.py legacy=False (float storeys, missing indicators, ids, FSA, lot/floor m2,
boilerplate, price_mention, geo_price_dup; legacy=True byte-identical to houses.parquet);
scripts/09_data_audit.py (sample_audit.csv, municipality->region map from MAMH MUN.csv,
langdetect 1.0.9 second rater, 300-listing language and dimension samples coded by two
independent LLM passes — labelled LLM-coded, not human-coded).
N1: src/embeddings.py (pinned HF revisions, token_lengths, chunk-and-pool encoder),
scripts/10_embeddings_v2.py (truncation audit; caches minilm_256/minilm_chunk/ml_128/
ml_512/ml_chunk + price-masked; SHA-256), scripts/11_reference_geometry.py (Gram, residual
norms, nearest listings, span-invariance check, Löwdin/centred similarity caches).
src/inference.py (Holm, BH, Romano–Wolf cluster bootstrap, LMG, Moran, Conley, CW, DM, MCS),
src/dataset.py (v2 frame, unscaled-dummy design), src/multiverse.py.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
      Simon-Pierre Boucher · 47 files +4,332 −152

  3. Sat, Oct 17, 2026 4

    1. révision scientifique complète (editorial board, 2026-10-17) — citations complétées (94/94 rendues)
      …
      Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
      Simon-Pierre Boucher · 3 files +3 −3
    2. révision scientifique complète (editorial board, 2026-10-17)
      …
      v1.1 → v1.2 : analyse de la langue des annonces (40 % anglais, η² = 0,86 sur la
similarité moyenne), effets fixes RTA/municipalité + erreurs groupées, terrain en m²,
sous-échantillon superficie, tests par domaine, validation croisée aléatoire et
spatiale, sensibilité au libellé des références, encodeur multilingue ; sections
réécrites, nouvelle section 7, 8 références vérifiées ajoutées (86 → 94),
docs/REVIEW-REPORT.md + docs/CHANGELOG-REVISION.md, README et data/README mis à jour.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
      Simon-Pierre Boucher · 44 files +2,436 −351
    3. PDF de référence avant révision scientifique
      …
      Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
      Simon-Pierre Boucher · 1 file
    4. harmonisation standard UQO WP (2026-10-17)
      …
      Page titre du gabarit, natbib + aer, docs/ pour les rapports d'audit, README au standard (Status, citation WP 2), .gitignore et Makefile/.latexmkrc du gabarit, main.pdf recompilé.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
      Simon-Pierre Boucher · 11 files +196 −297

  4. Wed, Aug 5, 2026 3

    1. Complete bibliography audit (100% verified) + README refresh
      …
      - Final sweep of the 38 remaining entries: 3 fixes (Halvorsen-Pollakowski 1981
  JUE replaces mismatched Halvorsen-Palmquist 1980 AER metadata; Irwin 2002
  Land Economics open-space paper replaces Irwin-Bockstael; Goodman-Thibodeau
  1995 heteroskedasticity paper replaces 1998 segmentation mismatch).
- Des Rosiers et al. (2002) landscaping premium range verified (4-12%) and
  stated precisely in the discussion.
- ~700-character description truncation confirmed empirically in the raw DB
  (pile-up at 701-703, mid-word cuts before GST/QST boilerplate).
- README: 53-page badge, '86 verified references' badge, new docs in tree.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
      simon-pierre boucher · 7 files +68 −39
    2. Scholarly upgrade: verified bibliography (59→86 refs), expanded framing, 53-page paper
      …
      - Audit found 8 fabricated citations in the original bib (Demers-Eisfeldt,
  Goodwin 2020, Huang-Ni, Lam et al., Li et al. 2023, Ozdogan, Hong 'LDA',
  Bayer 2016) — replaced with OpenAlex-verified real papers; 4 more corrected
  (Nowak-Smith venue, Shen->Shen-Ross JUE 2021, Gibbons 2015, Cheshire-Sheppard).
- 28 new verified references with DOI: text-as-data (Gentzkow-Kelly-Taddy,
  Ash-Hansen), hedonic identification (Ekeland-Heckman-Nesheim, Bajari-Benkard,
  Kuminoff et al.), quantile hedonics (Zietz et al., McMillen), listing language
  (Pryce-Oates, Haag et al., Grossman), spatial/Quebec (Anselin, Dubin, Can,
  Des Rosiers et al., Dube-Legros), vision (Glaeser et al.), interpretable ML
  (Rudin, SHAP, LIME), NLP (MTEB, zero-shot anchoring).
- Restructured literature review (new Text-as-Data section, two-sided gap),
  deepened intro, cited methodology choices, new 'Relation to Prior Findings'
  discussion subsection, softened unverifiable claims.
- Results, tables, figures, data untouched. PAPER_REVIEW.md + UPGRADE_REPORT.md.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
      simon-pierre boucher · 10 files +817 −220
    3. Initial commit: semantic hedonic pricing pipeline, figures, and paper (UQO WP no. 2)
      …
      Restructured from immo-wp2-spb-20260519: modular src/ package, numbered
reproduction pipeline (01-06), machine-generated LaTeX tables, regenerated
figures (incl. previously missing fig2/fig3), corrected paper statistics
(see CHANGES.md), 47-page PDF compiling clean.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
      simon-pierre boucher · 72 files +4,180
Menu

Theme
↑↓ navigate ↵ open esc close / or ⌘K anywhere