# QHPI v2 — monthly, robust: research record and design decisions ## Why revisit v1 The v1 weekly index estimated ONE pooled hedonic over 2021–2026 (fixed coefficients for 5.5 years) and published weekly points whose thin-cell noise had to be repaired by heavy shrinkage. Two structural weaknesses: 1. **Fixed hedonic coefficients over the full pool.** The implicit prices of floor area, age and location are not constant across a boom, a rate-hike correction and a recovery. The literature flags the fixed-parameter restriction as the main drawback of the full-pool time-dummy method. 2. **Full-sample estimation revises history on every run** — each refresh re-estimates every β, so past index points move. ## What the literature recommends - **Rolling-Time-Dummy (RTD)**: estimate the time-dummy hedonic on a rolling window of recent periods only, then link (splice) each new estimate onto the existing series. Used by official HPIs; requires less data, adapts to parameter drift, and is "well suited for computing higher frequency HPIs". Hill, Scholz, Shimizu & Steurer (2022, *Journal of Official Statistics* 38(1):127–151) treat the three design choices: window length, linking variant (movement / window / half / **mean splice**), and low-volume robustness. - **Chained/spliced time dummies produce a revision-free series** and relax the fixed-parameter assumption (Eurostat/IMF/OECD *RPPI Handbook*, ch. 5). - **Mean splice** (geometric mean over all feasible link positions) is the robust default in the splicing literature (Diewert & Fox 2021 for multilateral updating): no single anomalous overlap period drives the link. - **Robust estimation**: OLS hedonics chase influential observations; M-estimation (Huber weights) or documented influence-trimming is standard practice for official filters (cf. IMF WP/16/213, MNB HPI outlier procedures). - **Monthly frequency**: weekly cells below the metro level are too thin for publication-grade points without aggressive shrinkage; monthly multiplies cell counts ~4.3× and is the modal frequency of official HPIs. Sources: - Hill, Scholz, Shimizu & Steurer (2022) — https://journals.sagepub.com/doi/abs/10.2478/jos-2022-0007 - Eurostat et al., Handbook on Residential Property Prices Indices — https://www.oecd.org/content/dam/oecd/en/publications/reports/2013/04/handbook-on-residential-property-price-indices_g1g2e251/9789264197183-en.pdf - RPPI Handbook ch.5 (hedonic methods) — https://www.elibrary.imf.org/display/book/9789279259845/ch005.xml - IMF WP/16/213, "How to Better Measure Hedonic RPPIs" — https://www.imf.org/external/pubs/ft/wp/2016/wp16213.pdf - IMF RPPI Practical Guide — https://www.imf.org/-/media/files/data/guides/rppi/rppi-guide.pdf ## v2 design (implemented) 1. **Frequency: monthly** (labels `YYYY-MM`, continuous grid, trailing month flagged partial while registrations arrive). 2. **Stage 1 — RTD, 13-month rolling windows, monthly step.** Within each window: hedonic `log P ~ log_fa + fa_missing + age bins + building type | FSA FE + (month × type) FE`, estimated by **iteratively reweighted least squares with Huber weights** (k = 1.345 on MAD-standardized residuals, 2 reweighting passes) — outliers are downweighted inside the regression instead of silently trimmed. A 13-month window spans every calendar month once (seasonal balance) and holds ~120k transactions. 3. **Linking: mean splice.** For each new window the province path per type is extended by the geometric mean of the growth implied at every feasible overlap position. Movement and window splices are computed as published sensitivity diagnostics. The published history never revises. 4. **Stage 2 — unchanged hierarchy, monthly grid.** Each transaction's quality-adjusted residual comes from the window in which its month is the NEWEST month (fully real-time, β local in time). Cell deviations (region → municipality) follow the heteroskedastic local-level Kalman model; thin months shrink toward the parent path. 5. **Growth horizons**: 1m, 3m, 6m, 12m (YoY). Reliability grades recalibrated by a monthly downsampling experiment. 6. **New diagnostics**: hedonic coefficient drift across windows (the empirical justification for RTD), RTD vs full-pool comparison, splice variant sensitivity. 7. Cleaning screens, geography, `indéterminé` exclusion, fixed-share "all" composition, assessment-gap module: unchanged from v1 (already validated). Author: Simon-Pierre Boucher — contact@spboucher.ai