# Phase 0 — evidence-based research plan (2026-10-21)

# What exists

Data: Vrai Prix production SQLite (M3U96a), 769,995 registry transactions 2021-01-04 → 2026-10-07, geocoded, provider type, matched to the MAMH assessment roll (3.75 M units: CUBF, units, link, floor/lot area, year, storeys, neighbourhood unit, values). Four earlier extractions (Aug 6, Aug 9, Oct 10, Oct 17 2026) allow measuring arrival lags.
Prior code: QWHPI (PhD/recherche/QWHPI) — a working monthly/weekly hedonic index engine (RTD + Huber + hierarchical Kalman shrinkage), used here as benchmark K and as a source of lessons (building-type backfill, time-correlated missingness, mean splice).
Boundaries: SDA V2026-07 (municipality, MRC, region, metropolitan community, borough); MAMH register.
Tooling: Python 3.14 with polars, duckdb, pyarrow, statsmodels, pyfixest, lightgbm, geopandas; LaTeX (latexmk).

No deed-type, arm's-length, financing, listing, bedroom/bathroom, renovation or image data.
Roll characteristics are a single current vintage (post-sale for most sales).
Statistics Canada CMA boundaries blocked (bot check) → CMM/CMQ used.
No independent monthly municipal index exists for Quebec to validate against; external comparison limited to province/CMA-level public indices (to be obtained and harmonised in Phase 6, if accessible).

Data inventory, provenance, cleaning — scripts 01–06 (done 2026-10-21).
Benchmarks A–K at province × type and region × type; liquid municipalities for the feasible subset.
SHARP-HPI: robust hedonic with fine-location FE and smoothly time-varying quality prices → nested hierarchical local-level state-space on quality-adjusted residual means (province → region → MRC → municipality) with empirical-Bayes calibrated innovation variances and an observable reliability score; adaptive window / pooling.
Monte Carlo on synthetic Quebec-like markets (known latent paths), including counterexamples.
Full monthly panel 2021-01 → 2026-10 (partial), schema of §17, parquet + CSV.
Validation: forward-chaining backtests, repeat-sales agreement, placebo & composition shocks, real-time vintages and revision statistics, ablations, aggregation sensitivity.
Manuscript (LaTeX, UQO template), figures/tables generated from results, reproducibility audit.