# AUDIT — Original project immo-wp5-spb-20260519

Audit of ~/Desktop/UQO/UQO_WP/immo-wp5-spb-20260519 performed on 2026-08-05, before any restructuring. The original folder is left untouched; the clean repository is rebuilt at ~/Desktop/wp5_uqo.

# 1. Project summary

Hedonic and spatial econometric analysis of the relationship between Airbnb activity and residential rents in Quebec, Canada (UQO Working Paper No. 5, Simon-Pierre Boucher, May 2026). ~5,000 Airbnb listings + 8,356 Realtor.ca rental listings, spatial buffer merge (Haversine, 250 m–2 km), OLS/hedonic models, spatial models (SAR/SEM via spreg), quantile regressions, ML robustness (LASSO, Elastic Net, RF, GBM, SHAP).

# 2. Scripts (scripts/, 10 files, ~2,930 lines)

All paths are already relative to the project root via Path(__file__).resolve().parent.parent — good. Scripts are standalone monoliths with heavy duplication (see §5.4).

# 3. Data

• data_clean/ (5 parquet files, all present and readable): airbnb_clean.parquet (4,950 rows expected), rent_clean.parquet (8,303 rows), merged_spatial.parquet, merged_neighborhood.parquet, merged_analysis.parquet (8,258 rows × 42 cols).
• Raw data (airbnb.csv, rent.json) are absent from the folder and from the whole disk (searched ~/Desktop/UQO and Spotlight). See §5.1.

# 4. Figure inventory

Figures generated by the code (15) vs figures present in outputs/figures/ (10):

The 5 missing PDFs were evidently deleted (or the run predates those functions); the code that produces them is intact. Consequence: wp5/main.tex fails to compile (fatal error, no PDF); only wp5/main_web.tex compiles because it replaces missing figures with placeholder boxes. Re-running the pipeline regenerates all 15 figures and fixes the paper build.

wp5/figures/ contains a stale copy of the same 10 PDFs (duplicate of outputs/figures/).

# 5. Problems found (nothing was changed; fixes happen in the new repo)

# 5.1 Raw data missing (blocking for steps 01–03 only)

airbnb.csv and rent.json are referenced by scripts 01–03 and by the original README but do not exist anywhere on disk. The pipeline is therefore reproducible from data_clean/ onward (steps 04–10 + paper), which covers every number and figure in the paper. Steps 01–03 are kept in the new repo and fail with an explicit message if the raw files are absent. → Requires user review: locate/restore the raw files if full from-scratch reproduction is needed.

# 5.2 Two divergent copies of the paper

• paper/ — same sections, figures referenced as ../outputs/figures/…, no title page, compiled OK.
• wp5/ — adds UQO title page (uq_logo.jpg), Makefile, .latexmkrc, \graphicspath{{figures/}}, plus main_web.tex (variant with a "Figure indisponible" fallback macro).
• Only real content differences: figure paths in 03_data.tex, 05_results.tex, 06_robustness.tex.
• wp5/ is the canonical/most recent version (June 2026) → used as the base for the rewrite.

# 5.3 Figure caption vs content mismatch

03_data.tex captions the two distribution histograms as "(log scale)" but scripts plot levels (CAD). Caption corrected in the rewrite (editorial fix; figures unchanged).

# 5.4 Code-quality issues (fixed by refactor, results preserved)

• Significance-star helper duplicated in 4 scripts; LaTeX regression-table builder duplicated in 4 scripts; matplotlib style blocks duplicated in 3 scripts; city-FE OLS helper duplicated within script 10.
• Scripts 05 requests columns size_sqft, guests_count, amenities_count, num_images, quality_score that don't exist in the parquets (rent_clean has size_interior_sqft); the helper silently skips them, so the published summary tables lack those rows. Kept as-is to reproduce identical tables; noted for future work.
• Scripts 06–08 are top-level scripts (no main()); 02/03 likewise.
• scripts/__pycache__/ committed; .DS_Store files scattered.

# 5.5 Dead / duplicate / unused files

• paper/ entire directory superseded by wp5/ (kept only as reference in the original folder).
• wp5/main_web.tex + main_web.* build artifacts: web variant with placeholder hack — obsolete once figures are regenerated.
• LaTeX build artifacts (.aux, .log, .out, .bbl, .blg, .fls, .fdb_latexmk, .toc, .synctex.gz).
• scripts/__pycache__/, .DS_Store (×6).
• outputs/figures/ vs wp5/figures/: duplicated PDFs.

# 5.6 Environment

Python 3.14.4 (Homebrew) with pandas 3.0.2, numpy 2.4.4, statsmodels 0.14.6, scikit-learn 1.6.1, scipy 1.17.1, matplotlib 3.10.9, pyarrow 24.0.0, shap 0.48.0, libpysal 4.14.1, spreg 1.9.0. TeX Live (pdflatex, latexmk, bibtex) available. Note: with libpysal+spreg installed, script 07 takes the SAR/SEM branch — consistent with the published spatial_models.tex.

# 6. Paper (wp5/)

• main.tex (184 lines): clean preamble (newtx, booktabs, natbib/apalike, fancyhdr, hyperref), metadata macros, inputs 8 sections + title page + 3 appendices; 24 BibTeX entries, all cited.
• Sections: 01 introduction, 02 literature, 03 data, 04 methodology (models 1–6 + identification), 05 results, 06 robustness, 07 discussion, 08 conclusion; appendices: data, methods, robustness.
• Prose already in good academic English; rewrite pass = flow/consistency polish, caption fix (§5.3), figure/table path normalization to the new layout, and making every figure resolvable.
• Tables are \input from ../outputs/tables/*.tex → becomes ../results/tables/ in the new repo.

# 7. Verification plan (Phase 2b)

1. Copy the 5 parquets to data/processed/ (bit-identical, checksummed).
2. Re-run refactored steps 04→10; compare regenerated merged_*.parquet (DataFrame equality) and every .tex/.csv table against the originals (diff). Figures: confirm regeneration and spot-check values; PDF bytes differ (timestamps) so tables are the numeric ground truth.
3. Any discrepancy is recorded below — never silently corrected.

# 8. Discrepancies found during verification

Full pipeline (steps 04–10) re-run in the new repo on 2026-08-05 and compared to the originals:

• Merged parquets (3/3): identical — merged_spatial, merged_neighborhood, merged_analysis regenerated from the committed clean parquets are value-identical to the originals (pandas.testing.assert_frame_equal).
• Tables: 14/16 byte-identical after aligning the generators with the published fragments (the original outputs/tables/*.tex had been hand-stripped of their \begin{table}…\end{table} wrappers after generation; the new scripts emit those fragments directly — see CHANGES.md). Residual differences, both preserved and NOT corrected:
  1. quantile_regression.tex — in the τ=0.10 column only, three control coefficients differ in the 4th decimal (bathrooms 0.2410→0.2411, bt_House 0.0607→0.0608, bt_Row/Townhouse 0.1302→0.1310, SE 0.0469→0.0468). The Airbnb coefficient, all other quantiles, and the OLS column are identical. Cause: IRLS convergence jitter of statsmodels.QuantReg at the extreme quantile; no significance level or claim in the paper changes.
  2. ml_comparison.csv — two values differ at the 16th significant digit (~1e-16 relative; float representation). The 4-decimal ml_comparison.tex is byte-identical.
• Figures: all 15 regenerated, including the 5 missing from the original (map_airbnb, map_rent, quantile_coefficients, rent_by_city, rent_by_airbnb_bins), which un-breaks the paper build. PDF bytes differ from the surviving originals (embedded timestamps/IDs); the numeric ground truth is the table set above, plus the underlying data verified identical.