QWHPI validation report
1. Repeat-sales (BMN) vs hedonic index
- quebec/unifamilial: Δcorr=0.691, end levels RS=161.5 vs hedonic=157.9
- quebec/condo: Δcorr=0.322, end levels RS=133.8 vs hedonic=140.0
- quebec/plex: Δcorr=0.454, end levels RS=149.6 vs hedonic=158.5
- montreal/condo: Δcorr=0.000, end levels RS=121.8 vs hedonic=128.7
- montreal/unifamilial: Δcorr=0.264, end levels RS=125.6 vs hedonic=132.8
- quebec-city/unifamilial: Δcorr=0.073, end levels RS=165.7 vs hedonic=166.0
2. Downsampling experiment — Montréal condo lab
- 100 tx/wk: RMSE=0.317% bias=-0.046% vol_ratio=1.051 turn_agree=0.861 CI95_cov=1.0
- 50 tx/wk: RMSE=0.636% bias=-0.177% vol_ratio=0.987 turn_agree=0.799 CI95_cov=0.992
- 25 tx/wk: RMSE=0.799% bias=0.002% vol_ratio=0.762 turn_agree=0.771 CI95_cov=0.977
- 15 tx/wk: RMSE=1.15% bias=-0.154% vol_ratio=0.449 turn_agree=0.778 CI95_cov=0.881
- 10 tx/wk: RMSE=1.361% bias=-0.18% vol_ratio=0.358 turn_agree=0.735 CI95_cov=0.828
- 5 tx/wk: RMSE=1.633% bias=0.189% vol_ratio=0.457 turn_agree=0.685 CI95_cov=0.856
Reading: RMSE vs the full-sample (~145 tx/wk) smoothed path; CI coverage of the deviation posterior; turning points on 4-week MA.
3. Composition-shock simulation (province unifamilial)
- Shock: drop 70% of below-median floorArea sales in 2024-H1 (12,406 of 439,118 rows removed)
- Raw weekly median moves by +11.74% in shocked weeks (must move)
- Hedonic time-dummy moves by -0.74% (must stay ~0)
4. Weekly vs monthly (same methodology)
- quebec/unifamilial: weekly signal SD=3.665%, noise SD=0.0%, monthly ΔSD=3.198%, real-time revision SD=0.0%
- quebec/condo: weekly signal SD=3.778%, noise SD=0.0%, monthly ΔSD=2.993%, real-time revision SD=0.0%
- montreal/condo: weekly signal SD=3.782%, noise SD=3.583%, monthly ΔSD=3.031%, real-time revision SD=3.999%