{
  "_comment": "Pre-registered evaluation protocol. Frozen before the first backtest run; do not adapt to results. Values marked start_value have no evidence behind them.",
  "frozen_on": "2026-10-13",
  "target": {
    "indicator": "are",
    "series": "00+",
    "unit": "per_100k"
  },
  "pandemic_exclusion": {
    "from": "2020-W10",
    "to": "2022-W39",
    "reason": "Testing, closure and quarantine rules distort the series; not used for baselines, training or scoring."
  },
  "horizons": [
    1,
    2,
    3,
    4
  ],
  "seasons": {
    "definition": "ISO week 40 to week 39 of the following year",
    "tuning": [
      "2023/24",
      "2024/25"
    ],
    "test": "2025/26",
    "test_rule": "The test season is evaluated once; `backtest test` refuses to overwrite its result without --force."
  },
  "availability_lags_weeks": {
    "target": 0,
    "wastewater_sars": 1,
    "wastewater_flu": 1,
    "wastewater_rsv": 1,
    "tmax": 0,
    "tmin": 0
  },
  "vintage_note": "Real data versions of ARE exist since 2023-09 (reconstructed from the RKI's public git history, data/long/are.csv.gz). The tuning and test backtests still use today's corrected data; a separate descriptive comparison with the data versions known at each origin is in forecasts/backtest/vintage.json.",
  "models": {
    "baseline": {
      "name": "baseline-seasonal",
      "window_weeks": 1,
      "min_samples": 2,
      "calendar_align": true
    },
    "persistence": {
      "name": "persistence",
      "min_samples": 20
    },
    "regional": {
      "ridge": 1.0,
      "ridge_status": "start_value",
      "min_train_rows": 300,
      "max_iter": 60,
      "calendar_adjust": true
    },
    "ensemble": {
      "model_weight": 0.5,
      "baseline_weight": 0.5,
      "weight_status": "start_value (v1: 0.5); v2 searches the grid on the tuning seasons",
      "model_weight_grid": [
        0.5,
        0.6,
        0.75
      ]
    },
    "feature_candidates": [
      {
        "name": "dyn",
        "groups": [
          "dynamics"
        ]
      },
      {
        "name": "dyn-season",
        "groups": [
          "dynamics",
          "season"
        ]
      },
      {
        "name": "dyn-season-holiday",
        "groups": [
          "dynamics",
          "season",
          "holiday"
        ]
      },
      {
        "name": "dyn-season-weather",
        "groups": [
          "dynamics",
          "season",
          "weather"
        ]
      },
      {
        "name": "dyn-season-wastewater",
        "groups": [
          "dynamics",
          "season",
          "wastewater"
        ]
      },
      {
        "name": "all",
        "groups": [
          "dynamics",
          "season",
          "holiday",
          "weather",
          "wastewater"
        ]
      }
    ],
    "spread_factor_grid": [
      1.0,
      1.1,
      1.2,
      1.35,
      1.5,
      1.75,
      2.0
    ],
    "selection_rule": "Feature set and ensemble weight/spread with the lowest ensemble WIS over the weight x spread grid on the tuning seasons."
  },
  "metrics": [
    "mae",
    "wis",
    "coverage50",
    "coverage80"
  ],
  "breakdowns": [
    "region",
    "horizon",
    "wave_phase",
    "season"
  ],
  "peak_scoring": "v2: per horizon, argmax of the centred 3-week mean over W45..W15 (week between the years and the week after removed) of the median forecast versus the actual values; median |timing error| in weeks and share within one week; persistence lags by about h weeks by construction (reference). v1 pooled all horizons and weeks and was discontinuous.",
  "gate": {
    "reference": "baseline-seasonal",
    "min_rel_wis_improvement": 0.1,
    "min_rel_wis_improvement_status": "start_value",
    "evaluated_on": "test season, all regions and horizons",
    "if_failed": "ship the baseline",
    "reference_note": "v2: 'baseline-seasonal' is the calendar-aligned baseline."
  },
  "bootstrap": {
    "block_weeks": 4,
    "n": 1000,
    "seed": 20261013
  },
  "external_comparison": [
    "RESPINOW",
    "RespiCast"
  ],
  "external_comparison_note": "RESPINOW comparison (national, absolute counts converted to per 100k) is in forecasts/backtest/external.json; RespiCast was not reachable.",
  "amendments": [
    {
      "version": 2,
      "date": "2026-10-13",
      "reason": "A diagnosis on the tuning seasons only (2023/24, 2024/25) found a large, regular care-seeking calendar effect in ARE (week between the years about 0.4x, each closed weekday about 0.81x; absent in GrippeWeb) that the v1 model could not explain, a flawed peak-timing metric, and an inconsistency of the lag of wastewater features between training and prediction.",
      "changes": [
        "Calendar adjustment: fixed log effect -1.175*closed_share - 0.356*[week contains 28 Dec], estimated on national ARE before 2023-W36 only; applied to the regional model (dynamics features, training target, forecast) and to the seasonal baseline (samples aligned to the target week).",
        "Regional model: availability lags are now applied identically in training and prediction.",
        "Ensemble weight of the regional model and spread factor are chosen jointly on the tuning seasons from grids (weights 0.5/0.6/0.75, the existing spread grid); the feature candidates are unchanged.",
        "Peak timing: per horizon, winter window W45..W15, centred 3-week mean, week between the years and the week after removed; share within one week reported; the persistence lag is the reference.",
        "Gate reference is now the calendar-aligned seasonal baseline (stricter and also what is shipped on fallback); the plain v1 baseline is reported as 'baseline-plain'.",
        "Vacation periods now cover 2012/13 onward (KMK PDFs), so the Christmas dips of earlier winters enter training with their calendar features."
      ],
      "test_season_status": "Season 2025/26 was evaluated once under protocol v1 (forecasts/backtest/test-v1.json, untouched at that time). Under v2 it is evaluated again for description only and is no longer an untouched test season. The next untouched test season is 2026/27: evaluate it once after ISO week 2027-W39.",
      "selection_data": "Tuning seasons only; the v1 test-season results were not used to select anything in v2 (the diagnosis explicitly excluded season 2025/26)."
    }
  ]
}