"""Pre-registered gate for the EO branch. Question: does the static OlmoEarth-2018 city embedding add information -- beyond lat/lon (+elevation-free geography) -- about (a) city climate normals 1989-2018 and (b) city trend slopes, in leave-cities-out CV? Gate (fixed): EO ships only if, for annual_avg_temp or cdd slopes, the lat/lon+EO model beats lat/lon-only MAE by >=3% in 5-fold city CV. Otherwise the EO branch is dropped and reported as a negative result. """ from __future__ import annotations import json import sys from pathlib import Path import numpy as np import pandas as pd from sklearn.decomposition import PCA from sklearn.linear_model import Ridge from sklearn.model_selection import KFold from sklearn.pipeline import make_pipeline from sklearn.preprocessing import StandardScaler sys.path.insert(0, str(Path(__file__).parent)) from dataset import (METRIC_COLUMNS, add_validity, download, load_annual, load_coverage, long_valid) from models import _ols_slope def city_stats(long_df: pd.DataFrame) -> pd.DataFrame: rows = [] for (city, metric), g in long_df.groupby(["city_id", "metric"]): w = g[(g["Year"] >= 1989) & (g["Year"] <= 2018)] if len(w) < 10: continue rows.append({"city_id": city, "metric": metric, "normal": float(w["value"].mean()), "slope": _ols_slope(w["Year"].to_numpy(float), w["value"].to_numpy(float)), "Lat": float(g["Lat"].iloc[0]), "Lon": float(g["Lon"].iloc[0])}) wide = pd.DataFrame(rows).pivot(index="city_id", columns="metric", values=["normal", "slope"]) wide.columns = [f"{a}_{b}" for a, b in wide.columns] geo = pd.DataFrame(rows).drop_duplicates("city_id").set_index("city_id")[["Lat", "Lon"]] return wide.join(geo).reset_index() def cv_mae(X: np.ndarray, y: np.ndarray, seed: int = 0) -> float: kf = KFold(5, shuffle=True, random_state=seed) errs = [] for tr, te in kf.split(X): m = make_pipeline(StandardScaler(), Ridge(alpha=10.0)) m.fit(X[tr], y[tr]) errs.append(np.abs(m.predict(X[te]) - y[te]).mean()) return float(np.mean(errs)) def main() -> None: root = Path(__file__).resolve().parent.parent emb_path = root / "outputs" / "eo" / "city_embeddings.csv" if not emb_path.exists(): raise SystemExit("run eo_model.py first") emb = pd.read_csv(emb_path, dtype={"city_id": str}) data_root = download(root / "data" / "_source") annual = add_validity(load_annual(data_root), load_coverage(data_root)) stats = city_stats(long_valid(annual)) df = stats.merge(emb, on="city_id", how="inner") print(f"gate sample: {len(df)} cities with embeddings + stats") emb_cols = [c for c in emb.columns if c.startswith("emb_")] # PCA to 8 dims on all cities (descriptor compression; target-free). pca = PCA(n_components=min(8, len(emb_cols))) pcs = pca.fit_transform(StandardScaler().fit_transform(df[emb_cols])) for i in range(pcs.shape[1]): df[f"pc{i}"] = pcs[:, i] targets = [f"normal_{m}" for m in METRIC_COLUMNS] + \ [f"slope_{m}" for m in METRIC_COLUMNS] results = {} for t in targets: sub = df.dropna(subset=[t]) geo = sub[["Lat", "Lon"]].to_numpy() geo_poly = np.column_stack([geo, geo ** 2, geo[:, :1] * geo[:, 1:]]) both = np.column_stack([geo_poly, sub[[f"pc{i}" for i in range(pcs.shape[1])]].to_numpy()]) mae_geo, mae_both = cv_mae(geo_poly, sub[t].to_numpy()), cv_mae(both, sub[t].to_numpy()) results[t] = {"mae_latlon": mae_geo, "mae_latlon_plus_eo": mae_both, "improvement_pct": 100 * (mae_geo - mae_both) / mae_geo, "n": len(sub)} print(f"{t:28s} lat/lon MAE {mae_geo:9.3f} +EO {mae_both:9.3f} " f"({results[t]['improvement_pct']:+.1f}%)") gate_targets = ["slope_annual_avg_temp", "slope_cdd"] passed = any(results[t]["improvement_pct"] >= 3.0 for t in gate_targets) decision = {"gate_passed": bool(passed), "gate": ">=3% MAE improvement on temp or cdd slope", "results": results} (root / "outputs" / "eo" / "gate_decision.json").write_text(json.dumps(decision, indent=2)) print("\nEO GATE:", "PASS" if passed else "FAIL (EO dropped; negative result)") if __name__ == "__main__": main()