"""Rolling-origin backtests on data up to 2018 only. For each origin year o, every model sees years <= o and forecasts o+1..o+7, scored with the challenge formula against the 30-year average. Origins 1997-2004 are used to choose settings ("tune"); 2005-2011 are held out to check that the choice generalises ("check"). python code/backtest.py compare # every model we tried -> code/experiments.csv (~10 min) python code/backtest.py grid # settings grid of the level + pooled-trend model (~5 min) """ import sys import pandas as pd import climate_lib as cl TUNE = list(range(1997, 2005)) CHECK = list(range(2005, 2012)) def summarise(name, f, panel, meta): rows = [] for split, origins in [("tune", TUNE), ("check", CHECK)]: r = cl.backtest(f, panel, meta, origins).drop(columns="origin").mean() rows.append({"model": name, "split": split, **r.to_dict()}) print(f"{name:34s} {split:5s} " + " ".join(f"{m}={r[m]:+.3f}" for m in cl.METRICS), flush=True) return rows def compare(panel, meta): import learned from forecast import PARAMS, model as final B = {m: learned.FeatureBuilder(panel, meta, m) for m in cl.METRICS} def lt(**kw): return lambda p, mt, o, y: cl.f_level_trend(p, mt, o, y, **kw) def ml(make, span=None): return lambda p, mt, o, y: learned.f_learned(p, mt, o, y, builders=B, make_model=make, train_span=span) T = PARAMS["annual_avg_temp"] def cdd_temp(p, mt, o, y): return cl.f_cdd_from_temp(p, mt, o, y, temp_params=T) def temp(**kw): return lambda p, mt, o, y: cl.f_level_trend(p, mt, o, y, metrics=["annual_avg_temp"], params={"annual_avg_temp": {**T, **kw}}) def tmax_tmin(p, mt, o, y): a = cl.f_level_trend(p, mt, o, y, metrics=["tmax", "tmin"], params={"tmax": T, "tmin": T}) a = a.groupby(["city_id", "year"], as_index=False)["predicted_value"].mean() return a.assign(metric="annual_avg_temp") def gbm_cdd(p, mt, o, y): return learned.f_learned(p, mt, o, y, metrics=["cdd"], builders=B, make_model=learned.make_gbm, train_span=30) models = [ ("baseline: last year", lambda p, mt, o, y: cl.f_last(p, mt, o, y)), ("baseline: city linear trend 30y", lambda p, mt, o, y: cl.f_city_trend(p, mt, o, y, k=30)), ("mean of last 10 years", lambda p, mt, o, y: cl.f_mean(p, mt, o, y, k=10)), ("mean of last 15 years", lambda p, mt, o, y: cl.f_mean(p, mt, o, y, k=15)), ("level15 + pooled trend (one setting)", lt(level_k=15, trend_k=50, radius_km=1000)), ("level+trend, per-metric PARAMS", lt(params=PARAMS)), ("ridge, all origins since 1940", ml(learned.make_ridge)), ("ridge, last 30 origins", ml(lambda: learned.make_ridge(100), 30)), ("GBM, all origins since 1940", ml(learned.make_gbm)), ("GBM, last 30 origins", ml(learned.make_gbm, 30)), ("temp: trend fitted since 1975", temp(trend_start=1975)), ("temp: trend x1.5 (amplified)", temp(trend_start=1975, damp=1.5)), ("temp: Tmax and Tmin forecast separately", tmax_tmin), ("CDD from temperature forecast", cdd_temp), ("CDD: from temperature + GBM (previous final)", lambda *a: cl.blend(cdd_temp(*a), gbm_cdd(*a))), ("CDD: delta method, 15 years", lambda p, mt, o, y: cl.f_cdd_delta(p, mt, o, y, temp_params=T, k=15)), ("CDD: delta method, 30 years", lambda p, mt, o, y: cl.f_cdd_delta(p, mt, o, y, temp_params=T, k=30)), ("FINAL (forecast.py)", final), ] rows = [] for name, f in models: rows += summarise(name, f, panel, meta) out = pd.DataFrame(rows) out.round(4).to_csv(cl.ROOT / "code" / "experiments.csv", index=False) return out def grid(panel, meta): rows = [] for adj in (False, True): for lk in (10, 15, 30): for tk in (30, 50): for r in (500, 1000, 5000): for d in (0.5, 1.0): name = f"adj={adj:d} L={lk} T={tk} R={r} d={d}" f = (lambda p, mt, o, y, kw=dict(level_k=lk, trend_k=tk, radius_km=r, damp=d, adjust_stations=adj): cl.f_level_trend(p, mt, o, y, **kw)) rows += summarise(name, f, panel, meta) pd.DataFrame(rows).round(4).to_csv(cl.ROOT / "code" / "grid_results.csv", index=False) if __name__ == "__main__": panel, meta = cl.load_panel() {"compare": compare, "grid": grid}[sys.argv[1] if len(sys.argv) > 1 else "compare"](panel, meta)