"""Phase 4 contested candidates: precip wet-day decomposition, hottest EVT. Each gets evaluated on tune era; adoption additionally requires a positive check-era result in the Phase-5 confirmation (one shot, no iteration). """ from __future__ import annotations import sys from pathlib import Path import numpy as np import pandas as pd from scipy import stats sys.path.insert(0, str(Path(__file__).parent)) from candidates import level_trend from dataset import download from harness import evaluate, skill_table from panel import Panel def show(metric: str, detail: pd.DataFrame) -> None: tab = skill_table(detail) piv = tab.pivot_table(index="candidate", columns="era", values="skill") print(f"\n== {metric} [phase4] ==") print(piv.sort_values("tune", ascending=False).round(4).to_string()) def precip_wetday_arrays(root: Path, panel: Panel): data_root = download(root / "data" / "_source") daily = pd.read_parquet(data_root / "daily" / "train-00000-of-00001.parquet", columns=["ID", "Date", "prcp"]) daily["Year"] = daily["Date"].astype(str).str.slice(0, 4).astype(int) daily = daily[(daily["Year"] >= panel.years[0]) & daily["prcp"].notna()] daily["wet"] = (daily["prcp"] > 0.01).astype(float) g = daily.groupby(["ID", "Year"]).agg(total=("prcp", "sum"), wet=("wet", "sum"), ndays=("prcp", "size")).reset_index() wet_arr = np.full((len(panel.cities), len(panel.years)), np.nan) int_arr = np.full_like(wet_arr, np.nan) ci = {c: i for i, c in enumerate(panel.cities)} for r in g.itertuples(): if r.ID in ci and int(r.Year) in panel.year_idx and r.ndays >= 355 and r.wet > 0: wet_arr[ci[r.ID], panel.year_idx[int(r.Year)]] = r.wet int_arr[ci[r.ID], panel.year_idx[int(r.Year)]] = r.total / r.wet return wet_arr, int_arr def main() -> None: root = Path(__file__).resolve().parent.parent panel = Panel(root) out = root / "outputs" / "v2" wet_arr, int_arr = precip_wetday_arrays(root, panel) def wetday_fc(p: Panel, origin: int) -> np.ndarray: i1 = p.year_idx[origin] w = wet_arr[:, i1 - 29:i1 + 1] inten = int_arr[:, i1 - 29:i1 + 1] w_fc = np.nanmean(w, axis=1) i_fc = np.nanmean(inten, axis=1) fc = w_fc * i_fc fb = np.nanmean(p.values["total_precip_in"][:, i1 - 29:i1 + 1], axis=1) fc = np.where(np.isfinite(fc), fc, fb) return np.repeat(np.maximum(fc, 0)[:, None], 7, axis=1) detail_p = evaluate(panel, "total_precip_in", { "wetday_decomp": wetday_fc, "L30mean": lambda p, o: level_trend(p, "total_precip_in", o, 30, "mean", "none"), }) detail_p.to_csv(out / "detail_precip_phase4.csv", index=False) show("total_precip_in", detail_p) def gev_fc(p: Panel, origin: int) -> np.ndarray: vals = p.values["hottest_day_temp"] i1 = p.year_idx[origin] out_fc = np.full((len(p.cities), 7), np.nan) base = np.nanmean(vals[:, i1 - 29:i1 + 1], axis=1) for ci in range(len(p.cities)): series = vals[ci, i1 - 29:i1 + 1] series = series[np.isfinite(series)] if len(series) >= 15: try: c, loc, scale = stats.genextreme.fit(series) out_fc[ci] = stats.genextreme.ppf(0.5, c, loc=loc, scale=scale) continue except Exception: pass out_fc[ci] = base[ci] return out_fc def top2_fc(p: Panel, origin: int) -> np.ndarray: vals = p.values["hottest_day_temp"] i1 = p.year_idx[origin] window = vals[:, i1 - 29:i1 + 1] second = np.partition(window, -2, axis=1)[:, -2] base = np.nanmean(window, axis=1) fc = np.where(np.isfinite(second), second, base) return np.repeat(fc[:, None], 7, axis=1) detail_h = evaluate(panel, "hottest_day_temp", { "gev_median": gev_fc, "top2_second": top2_fc, "L30mean": lambda p, o: level_trend(p, "hottest_day_temp", o, 30, "mean", "none"), "L30_reghalf": lambda p, o: level_trend(p, "hottest_day_temp", o, 30, "mean", "reg", 50, trend_mult=0.5), }) detail_h.to_csv(out / "detail_hottest_phase4.csv", index=False) show("hottest_day_temp", detail_h) if __name__ == "__main__": main()