"""Lightweight scoring helpers (numpy/pandas only). Kept separate from evaluate.py (which needs scikit-learn and matplotlib for plots and permutation importance) so the browser app can import them without downloading those packages. evaluate.py re-exports both functions. Formulas match sklearn.metrics exactly (tests/test_portable.py checks it). """ import numpy as np import pandas as pd from src import config def metrics(y, pred) -> dict: """Overall RMSE/MAE/R² plus RMSE/MAE on storm hours (Kp ≥ 5-).""" y, pred = np.asarray(y, dtype=float), np.asarray(pred, dtype=float) err = pred - y storm = y >= config.STORM_THRESHOLD ss_tot = np.sum((y - y.mean()) ** 2) return { "rmse": float(np.sqrt(np.mean(err ** 2))), "mae": float(np.mean(np.abs(err))), "r2": float(1 - np.sum(err ** 2) / ss_tot) if ss_tot > 0 else np.nan, "storm_rmse": float(np.sqrt(np.mean(err[storm] ** 2))) if storm.any() else np.nan, "storm_mae": float(np.mean(np.abs(err[storm]))) if storm.any() else np.nan, "n": len(y), } def attach_block_kp(pred: pd.DataFrame, kp_blocks: pd.Series, name: str) -> pd.DataFrame: """Add the Kp of the 3-h block containing each hour (e.g. NOAA estimated or GFZ Kp).""" block = pred.index.floor(f"{config.KP_BLOCK_HOURS}h") return pred.assign(**{name: kp_blocks.reindex(block).to_numpy()})