{
  "tag": "real-weather-forecasts-aug10",
  "goal": "Recover the perfect-foresight gap using REAL archived day-ahead weather forecasts as a causal future covariate. Honest RTO-7 test MASE is 0.572; the oracle (perfect weather + renewables) ceiling is 0.468. Close at least half that gap without leakage and without degrading held-out BAs.",
  "metric": "mase",
  "metric_direction": "lower",
  "baseline": {
    "note": "7 RTOs, 2025 test split, causal covariates, deduplicated store.",
    "no_future_weather": 0.5937,
    "plus_peer_covariates": 0.5804,
    "plus_causal_readaptation": 0.572,
    "chronos2_zeroshot_plus_peers": 0.5908,
    "oracle_ceiling_not_a_forecast": 0.4683
  },
  "target": 0.52,
  "success": {
    "primary": "RTO-7 test MASE <= 0.52 (>= 50% of the 0.572 -> 0.468 gap closed)",
    "stretch": "<= 0.50",
    "must_not_regress": "holdout_gap worse by more than 0.02; cov_pi80 below 0.75",
    "must_beat": "Chronos-2 zero-shot given the SAME forecast covariates \u2014 otherwise the gain belongs to the covariate, not to surge."
  },
  "run_command": "bash scripts/run_remote_experiment.sh $EXP_NAME $CONFIG_JSON",
  "eval_contract": {
    "entrypoint": "experiments.research_eval",
    "split": "validation (calendar 2024). Test (2025) is reserved for a single final confirmation via {\"confirm\": true}.",
    "objective_field": "mase",
    "generalisation_field": "holdout_gap",
    "rejects": "future_mode='oracle', and any setup whose future covariates change when post-origin actuals are perturbed"
  },
  "modifiable_files": [
    "experiments/features.py",
    "experiments/eval_c2.py",
    "experiments/models.py",
    "experiments/finetune_c2.py",
    "experiments/ensemble.py"
  ],
  "readonly_files": [
    "experiments/causal_guard.py",
    "experiments/research_eval.py",
    "src/surge/store.py",
    "src/surge/scrapers/eia.py",
    "src/surge/scrapers/asos.py"
  ],
  "hard_constraints": [
    "The forecast covariate at hour t must be the value forecast at or before t-horizon. Never substitute the observed series or the reanalysis - that is the oracle leak.",
    "Train and serve on the SAME covariate kind. Training on observed temperature and serving a forecast is the train/serve mismatch that made every causal proxy lose to withholding temperature entirely.",
    "Because the forecast archive starts 2021-03, retrain on 2021-2023 and keep val=2024 and test=2025 unchanged so results stay comparable to results.tsv.",
    "Never score on test during search. Validation only; one final confirmation.",
    "Always report Chronos-2 zero-shot with the same covariates as the control."
  ],
  "ideas": [
    {
      "id": "forecast_temp_covariate",
      "priority": 1,
      "rationale": "Day-ahead forecast temp has MAE ~1.3C and corr ~0.99 against the analysis, so it should capture most of the 0.084 MASE that oracle temperature was worth. This is the single largest measured lever.",
      "try": "ingest temperature_2m_previous_day1 for the 7 RTO centroids 2021-03..2026-04; declare it known-future; retrain so the model learns to trust it appropriately."
    },
    {
      "id": "forecast_renewables",
      "priority": 2,
      "rationale": "Oracle wind/solar was worth a further 0.043 MASE. Open-Meteo exposes shortwave radiation and 100m wind speed as forecast variables, which proxy solar and wind output causally.",
      "try": "add shortwave_radiation_previous_day1 and wind_speed_100m_previous_day1."
    },
    {
      "id": "asos_backfill_46",
      "priority": 2,
      "rationale": "46 of 53 BAs have zero-filled weather. Forecast covariates can cover all 53 from Open-Meteo centroids without needing the ASOS backfill at all.",
      "try": "fetch forecast temp for all 53 BA centroids, not just the RTOs."
    },
    {
      "id": "continuous_retraining",
      "priority": 3,
      "rationale": "Measured 8% degradation from Jan-Apr 2025 to Jan-Apr 2026 for both surge and zero-shot. Literature (arXiv 2607.15705) finds continuous retraining materially important and commonly overlooked.",
      "try": "rolling monthly LoRA re-adaptation; 37s per run makes this nearly free."
    },
    {
      "id": "tabpfn_ts",
      "priority": 3,
      "rationale": "TabPFN-TS (11M params) claims SOTA on covariate-informed forecasting on GIFT-Eval and fev-bench, and reportedly captures simple covariate-target relations better than Chronos-2 at short horizons.",
      "try": "evaluate with identical covariates; consider cross-family output ensembling."
    },
    {
      "id": "public_benchmark",
      "priority": 3,
      "rationale": "SOTA cannot be claimed on a self-defined BA set with a self-computed denominator. Needs a benchmark with published numbers.",
      "try": "GEFCom2014-L (probabilistic, pinball loss), fev-bench/GIFT-Eval covariate tracks, or the frozen German live-challenge leaderboard (arXiv 2608.05018)."
    }
  ],
  "timeout_seconds": 1800,
  "no_new_dependencies": false,
  "stop_conditions": {
    "wall_clock_minutes": 360,
    "no_improvement_streak": 6
  },
  "measured_inputs": {
    "weather_forecast_source": "Open-Meteo Previous Runs API, temperature_2m_previous_day1 (GFS seamless) \u2014 the forecast issued ~24h before valid time",
    "archive_start": "2021-03 for GFS 2m temperature; ~2024-01 for most other models",
    "day_ahead_forecast_skill_2025": {
      "mae_degC": {
        "PJM": 1.3,
        "CISO": 1.06,
        "ERCO": 1.25,
        "MISO": 1.38,
        "NYIS": 1.31,
        "ISNE": 1.55,
        "SWPP": 1.22
      },
      "corr": "0.983 - 0.990",
      "bias_degC": "-0.40 .. +0.20"
    }
  }
}