# Limitations — WP23 - 69 months of data, 34–45 forecast origins per horizon: tests have low power; MCS p-values with T = 12 origins (Task A) are indicative only. - Task A assumes the sold dwelling's characteristics are known at the origin (valuation of a known dwelling h months ahead); the identity of future sellers is not forecast; transaction selection documented, not corrected. - Characteristics limited to the assessment roll (no bedrooms/bathrooms/condition/listing data). - Task B targets are the WP22 final smoothed index, itself an estimate; targets embed the revision between the real-time and final series. - Neural models trained with fixed hyper-parameters and small epochs; no architecture search; results could change with tuning (by design, to avoid test leakage). - DSGFN graph over 64 municipal nodes + regions; graph methods designed for hundreds of nodes with long histories. - Simulation forecasters run with 10 (6) replications; Monte Carlo s.e. reported. - Tree and torch families run in separate processes; QREFA does not include LightGBM.