# Limitations and reproducibility notes — WP22 ## Data - One date per sale (feed publication of the deed); signing/registration/publication not separable. - Provider floor: no sale below $50,000. - Characteristics from a single current roll vintage (post-sale); renovations invisible; 2025–2026 new builds may be coded as land → excluded. - No bedrooms/bathrooms/parking/quality grade/arm's-length/financing/listing/image data. - `unite_voisinage` unusable (unit-level code); location = metric grid. - Statistics Canada CMA boundaries not obtainable (bot check); CMM/CMQ used for external comparisons. - Feed start-up: January–February 2021 volumes are a third of 2022 levels; the 2021 base averages over thin months. - 2026 volumes 8–14 % below 2025; whether market or coverage cannot be separated. ## Model - Sequential two-stage estimation; stage-2 conditions on stage-1 residuals; intervals omit hedonic-parameter and hyper-parameter uncertainty and cross-level covariance (bootstrap ratio ≈ 1.5 for mid-size municipalities, < 1 for the province and the thinnest cells). - Administrative hierarchy (MRC parent) regardless of market similarity; spatial-neighbour or similarity parents not implemented (listed as future work). - Not seasonally adjusted; monthly profiles contain unobserved-composition seasonality. - Robustness via Huber IRLS / innovation down-weighting, not a full Student-t likelihood. - Repeat-sales benchmark uses parcel identifiers; condominium projects and shared parcels excluded; renovation bias remains. ## Reproducibility - Python 3.14.4, package versions in `requirements.txt`; seeds in `config.py` and scripts; all figures/tables regenerated by `make figures tables`. - Raw microdata and `SDA.gpkg` are not committed; the index panel, vintages, validation outputs and literature verification files are. - Monte Carlo: 80 replications × 9 scenarios (≈ 3.6 h on 12 cores); results carry Monte Carlo standard errors.