# Dr-CiK forecast cache for the self-evolving time-series pipeline

`champion_forecasts_gpt56sol_high_toto_balanced_v3_20260906.tar.gz` (19 MB compressed, 199 MB / 49,340 JSON files unpacked) is the ForecastStore cache that `scripts/nrd_precompute.py` in
[kaichen-z/time-series @ self-evolving-nrd](https://github.com/kaichen-z/time-series/tree/self-evolving-nrd/experiments/self_evolving/final_method) reads with `cache_only=True`.

```bash
mkdir -p runs/champion_forecasts
tar -xzf champion_forecasts_gpt56sol_high_toto_balanced_v3_20260906.tar.gz -C runs/champion_forecasts
```

Each file is one cached forecast (fields `forecast`, `status`, `key`, `metric_policy`, …), keyed by method, input history, horizon, frequency and portfolio identity. The pipeline uses the future forecasts of the 31 methods listed in `inputs/full_methods.json` on the 199 Dr-CiK public tasks; the remaining entries come from other experiments. No evaluation labels are stored.

sha256 of the tarball: `30d2e1b56971682e1de1bd106e3028c35d67b00f60e1a5f520134e7d8f2adc46`