# Data windows and splits. Dates are chosen from the actual coverage of the Vrai Prix snapshot
# (registered sales 2021-01-04 .. 2026-10-07, snapshot taken 2026-10-21) — see data/raw/README.md.
snapshot:
  source: "WP22 processed layer of vraiprix-20261021.db (sha256 2816d8dd… pinned in the WP22 repo)"
  transactions_parquet: "data/raw/wp22_clean_transactions.parquet"
  units_stock_parquet:  "data/raw/wp22_raw_units_stock.parquet"
  market_index_parquet: "data/raw/wp22_raw_market_index.parquet"
sample:
  property_classes: [house, condo, plex]      # headline residential classes
  price_min: 50000
  price_max: 5000000
  require_roll_match: true
  require_coordinates: true
splits:
  train:      {start: "2021-01-01", end: "2024-06-30"}
  validation: {start: "2024-07-01", end: "2025-03-31"}
  test:       {start: "2025-04-01", end: "2025-12-31"}
  holdout:    {start: "2026-01-01", end: "2026-10-07"}     # final untouched temporal holdout, evaluated once
  prospective: {note: "sales registered in the live DB after the protocol freeze (fetched by scripts/60_fetch_prospective.py); ids absent from the snapshot"}
spatial_holdout:
  # entire regions withheld from training in the spatial-generalisation experiment
  regions: ["Outaouais", "Abitibi-Témiscamingue", "Bas-Saint-Laurent"]
comparables:
  radius_km: [1.0, 3.0, 10.0]      # expanding search
  lookback_days: 540
  max_comps: 12
  min_comps: 3
  same_class_required: true
llm_eval_subsample:
  # LLM arms are evaluated on stratified random subsamples (cost); baselines on the full partitions
  validation_n: 240
  test_n: 600
  holdout_n: 300
  strata: [region_group, property_class, price_tercile]
seed: 20261027