Tonic's picture
Protostar Climate submission: v2 forecast + v3/DCPP amendment audit (29 gated candidates)
49f3de8 verified
Raw History Blame Contribute Delete
4.26 kB
"""Rolling-origin backtests on pre-2019 data only + pre-registered selection.
Origins 1998/2004/2011, forecast origin+1..origin+7, valid rows only.
Reference = mean30 at the same origin. Selection rule (fixed before running):
highest mean skill across origins; must beat mean30 in >=2 of 3 origins;
if within 0.01 of the runner-up, take the simpler method (SIMPLICITY order).
"""
from __future__ import annotations
import json
import sys
from pathlib import Path
import numpy as np
import pandas as pd
sys.path.insert(0, str(Path(__file__).parent))
from dataset import METRICS, download, load_annual, load_coverage, add_validity, long_valid
from models import CANDIDATES, SIMPLICITY, forecast_metric
from scorer import score_forecasts
ORIGINS = [1998, 2004, 2011]
HORIZON = 7
def run_backtest(long_df: pd.DataFrame) -> tuple[pd.DataFrame, pd.DataFrame]:
detail_rows, skill_rows = [], []
for origin in ORIGINS:
years = list(range(origin + 1, origin + HORIZON + 1))
truth = long_df[long_df["Year"].isin(years)][["city_id", "Year", "metric", "value"]]
ref = pd.concat([
forecast_metric(long_df[long_df["metric"] == m], m, origin, years, "mean30")
for m in METRICS])
ref_scored = score_forecasts(truth, ref, ref)
for method in CANDIDATES:
pred = pd.concat([
forecast_metric(long_df[long_df["metric"] == m], m, origin, years, method)
for m in METRICS])
res = score_forecasts(truth, pred, ref)
for m in METRICS:
detail_rows.append({"origin": origin, "method": method, "metric": m,
"mae": res["mae"].get(m), "mae_ref": res["mae_ref"].get(m),
"skill": res["skill"].get(m), "n": res["n_rows"]})
skill_rows.append({"origin": origin, "method": method, "score": res["score"]})
return pd.DataFrame(detail_rows), pd.DataFrame(skill_rows)
def select_methods(detail: pd.DataFrame) -> dict:
"""Per-metric selection following the pre-registered rule."""
chosen = {}
summary = {}
for metric in METRICS:
sub = detail[detail["metric"] == metric]
piv = sub.pivot_table(index="method", columns="origin", values="skill")
mean_skill = piv.mean(axis=1)
beats = (piv > piv.loc["mean30"]).sum(axis=1)
ranked = mean_skill.sort_values(ascending=False)
best, best_val = ranked.index[0], float(ranked.iloc[0])
# Near-tie -> simpler method among those within 0.01 of best and eligible.
eligible = [m for m in ranked.index
if best_val - float(ranked[m]) < 0.01 and (m == "mean30" or beats.get(m, 0) >= 2)]
if not eligible or (best != "mean30" and beats.get(best, 0) < 2):
pick = "mean30"
else:
pick = min(eligible, key=lambda m: SIMPLICITY.index(m))
chosen[metric] = pick
summary[metric] = {
"chosen": pick,
"mean_skill_by_method": {k: float(v) for k, v in mean_skill.items()},
"origins_beating_mean30": {k: int(v) for k, v in beats.items()},
}
return {"methods": chosen, "summary": summary}
def main() -> None:
root = Path(__file__).resolve().parent.parent
data_root = download(root / "data" / "_source")
annual = add_validity(load_annual(data_root), load_coverage(data_root))
long_df = long_valid(annual)
print(f"valid obs: {len(long_df)} rows, {long_df['city_id'].nunique()} cities, "
f"{long_df['Year'].min()}-{long_df['Year'].max()}")
detail, _ = run_backtest(long_df)
out = root / "outputs"
out.mkdir(exist_ok=True)
detail.to_csv(out / "backtest_detail.csv", index=False)
decision = select_methods(detail)
(out / "decision.json").write_text(json.dumps(decision, indent=2))
print("\nPer-metric mean skill across origins (skill vs mean30 at same origin):")
piv = detail.pivot_table(index="method", columns="metric", values="skill", aggfunc="mean")
print(piv.round(3).to_string())
print("\nChosen methods:", decision["methods"])
if __name__ == "__main__":
main()