WP23 — plan (2026-10-21)
Inputs (WP22, read-only)
analytic_transactions.parquet(632,006 headline sales, roll characteristics, coordinates, nested grid cells, classes).index_panel.parquet(retrospective SHARP-HPI, 2,696 cells × 69 months) — used only as the target (final estimate of the quality-adjusted market movement), never as a predictor.index_vintages.parquet(real-time SHARP-HPI at each monthly publication date, province / regions / municipalities ≥ 1,500 sales, filtered and smoothed modes) — the only admissible index predictors at an origin.- SDA boundaries and municipal centroids (graph construction), external indices (not used as predictors: current-vintage).
Variables actually available (WP22 data dictionary)
Floor area, lot area, year built / age, storeys, building genre, physical link / subtype, number of units (plex), coordinates, municipality / MRC / region, assessment at sale (used only as a feature for Task A with care: it is in force at the sale date and known before the sale). Not available: bedrooms, bathrooms, parking, basement, condition, renovations, quality grade, listing attributes.
Tasks
- A property-level price at sale: origin = month o; train on sales dated ≤ o (features at origin: hedonic + municipal real-time index level/momentum/volume at o + spatial neighbours' past prices ≤ o); test on sales in months o+1…o+h grouped by horizon. Selection: sold-property population only; selection documented via comparison of sold vs stock characteristics (roll) and a Heckman-style diagnostic where feasible.
- B quality-adjusted index growth: cell × origin × horizon; target = final smoothed log index change over h; predictors from the vintage at o.
- C probabilistic versions of A and B (quantile / distributional heads, Bayesian posterior predictive, conformal calibration on past origins only).
Architectures (torch 2.12 available; PyMC/JAX/catboost/hmmlearn absent → own implementations)
HSHN (embedding-based hierarchical MLP with partial pooling by shrinkage of municipality embeddings to MRC/region), DSGFN (graph attention over municipality graph with distance + co-movement + composition edges, temporal convolution), RAREF (Gaussian-HMM regimes on province/region growth estimated at origin + gated mixture of experts), BHPF (hierarchical Bayesian linear-Gaussian model with conjugate/Laplace posteriors; posterior predictive), QANSS (neural hedonic encoder + per-cell Kalman latent market state with learned transition), QREFA (integration with ablations). Benchmarks: naive family, AR/ARIMA/local linear trend, VAR/BVAR (small, regions), DFM (PCA factors), hedonic OLS/ridge/EN, SAR-lag, RF, XGBoost, LightGBM, SVR (subsample), MLP, LSTM/GRU/TCN/TFT-lite.
Evaluation
Expanding window, monthly origins 2022-12 → 2025-09 (12-month targets must exist by 2026-09); metrics MAE/RMSE/MedAE/MASE/bias; CRPS/pinball/WIS/coverage; DM (HAC with h−1 lags), GW conditional test, MCS (bootstrap), moving-block bootstrap; heterogeneity by region/size/type/horizon/regime.
Phases
0 discovery & plan → 1 datasets & leakage audit → 2 benchmarks → 3 architectures → 4 simulation → 5 empirical runs → 6 tests/heterogeneity/interpretability/ablations → 7 manuscript.