sleonca's picture
Upload folder using huggingface_hub
b47743b verified
Raw History Blame Contribute Delete
4.84 kB
"""Step 10 of the ARIMA experiments: choose one ARIMA-based configuration per metric from the
pre-2019 backtest only (cached predictions of arima_backtest.py; no data after 2018 is involved).
Selection rule, fixed before looking at the results:
A. best pure ARIMA per metric = ARIMA-family model with the highest mean S_m over the 6 origins;
B. ensemble per metric = w * ARIMA + (1 - w) * current model, w in {0.25, 0.5}, for the ARIMA models
listed in CANDIDATES. An ensemble is accepted for a metric only if it beats the current model on
BOTH the mean S_m and S_m at origin 2011; among accepted ones, the highest mean S_m wins.
Otherwise the metric keeps the current model.
The resulting configurations are scored with the unchanged backtest.run() and compared with the
current model on the mean composite Score and at origin 2011.
Outputs: results/arima_select/{ensembles.csv, configs.csv, score_by_origin.csv}
Run: python code/arima_select.py
"""
import warnings
warnings.filterwarnings("ignore")
import pandas as pd
import backtest
from arima_backtest import MODELS, PREDS, REFERENCE
from backtest import ORIGINS
from common import METRICS, ROOT
OUT = ROOT / "results" / "arima_select"
WEIGHTS = (0.25, 0.5)
CANDIDATES = {
"annual_avg_temp": ["arima_100_ct", "arima_101_ct", "arima_111_drift", "auto_d0_ct", "pooled_auto", "pooled_auto@1949"],
"cdd": ["arima_100_ct", "arima_101_ct", "arima_111_drift", "auto_d0_ct", "pooled_auto", "cdd_hybrid_N10"],
"total_precip_in": ["arima_100_c", "auto_d0_c", "pooled_auto", "pooled_auto@1949"],
"hottest_day_temp": ["arima_100_c", "auto_d0_c", "pooled_auto", "arima_100_ct"],
}
def load(name):
return pd.read_parquet(PREDS / f"{name.replace('@', '_at_')}.parquet").set_index(["ID", "metric", "Year", "origin"]).pred
def run_cached(pred: pd.Series, metric):
p = pred.rename("pred").reset_index()
def fn(lv, o, yrs, coords, m):
return p[(p.origin == o) & (p.metric == m) & p.Year.isin(yrs)][["ID", "metric", "Year", "pred"]]
return backtest.run(fn, metric)
def main():
OUT.mkdir(parents=True, exist_ok=True)
ours = load("ours")
arima_names = [n for n in MODELS if n not in REFERENCE]
rows = []
for m in METRICS:
r = run_cached(ours, m)
rows.append(dict(metric=m, model="ours", w=0.0, S_mean=r.S.mean(), S_2011=r.S.iloc[-1]))
for name in arima_names:
if m not in MODELS[name]["metrics"]:
continue
a = load(name)
r = run_cached(a, m)
rows.append(dict(metric=m, model=name, w=1.0, S_mean=r.S.mean(), S_2011=r.S.iloc[-1]))
if name in CANDIDATES[m]:
for w in WEIGHTS:
e = (w * a + (1 - w) * ours).dropna()
r = run_cached(e, m)
rows.append(dict(metric=m, model=name, w=w, S_mean=r.S.mean(), S_2011=r.S.iloc[-1]))
T = pd.DataFrame(rows)
T.to_csv(OUT / "ensembles.csv", index=False)
configs = {"ours": {}, "A_best_arima": {}, "B_ensemble_rule": {}}
for m in METRICS:
t = T[T.metric == m]
base = t[t.model == "ours"].iloc[0]
configs["ours"][m] = ("ours", 0.0)
a = t[t.w == 1.0].sort_values("S_mean", ascending=False).iloc[0]
configs["A_best_arima"][m] = (a.model, 1.0)
ens = t[(t.w > 0) & (t.w < 1) & (t.S_mean > base.S_mean) & (t.S_2011 > base.S_2011)]
if len(ens):
b = ens.sort_values("S_mean", ascending=False).iloc[0]
configs["B_ensemble_rule"][m] = (b.model, b.w)
else:
configs["B_ensemble_rule"][m] = ("ours", 0.0)
score_rows, cfg_rows = [], []
for cname, cfg in configs.items():
Ss = []
for m, (name, w) in cfg.items():
pred = ours if w == 0 else (w * load(name) + (1 - w) * ours).dropna()
r = run_cached(pred, m)
Ss.append(r.set_index("origin").S.rename(m))
cfg_rows.append(dict(config=cname, metric=m, model=name, w=w, S_mean=r.S.mean(), S_2011=r.S.iloc[-1]))
S = pd.concat(Ss, axis=1)
sc = S.mean(1).clip(lower=0) * 100
score_rows.append(dict(config=cname, **{str(o): sc[o] for o in ORIGINS}, mean=sc.mean()))
C = pd.DataFrame(cfg_rows)
C.to_csv(OUT / "configs.csv", index=False)
SC = pd.DataFrame(score_rows).set_index("config")
SC.to_csv(OUT / "score_by_origin.csv")
pd.set_option("display.width", 220); pd.set_option("display.max_rows", 200)
for m in METRICS:
t = T[T.metric == m].sort_values("S_mean", ascending=False)
print(f"\n{m} (top 12)\n", t.head(12).round(4).to_string(index=False))
print("\nConfigurations:\n", C.round(4).to_string(index=False))
print("\nComposite Score by origin:\n", SC.round(2).to_string())
if __name__ == "__main__":
main()