"""Experiment 2: plain average vs plain line vs line + random-forest hybrid (spec: exp02_line_forest_hybrid_spec.md). Run from the repo root: python code/exp02_line_forest.py [--charts] Writes outputs/exp02/: heldout_results.csv, decision.json (BEFORE the final forecast), submission_exp02.csv, run_log.json (+ charts/ with --charts). Then score it (analysis only): python code/starter_score.py obs_2018_2021.csv Data rule: only data up to 2018 is read. Nothing here reads observations after 2018; every choice is made on held-back years inside the training data. """ import argparse import json import subprocess import sys from datetime import datetime, timezone from importlib.metadata import version from pathlib import Path import numpy as np import pandas as pd from sklearn.ensemble import RandomForestRegressor sys.path.insert(0, str(Path(__file__).resolve().parent)) from shared import METRICS, assert_cutoff, dataset_revision, load_train, valid_flags ROOT = Path(__file__).resolve().parent.parent SEED = 42 WINDOW = 30 ORIGINS = [1998, 2004, 2011] H = 7 MIN_VALID = 10 MARGIN = 0.02 # hybrid must beat the better of line/average by 2 skill points N_TREES = 100 # "decide" = choose among hybrid / line / average with the held-back rule in decide(). METRIC_PLAN = {"annual_avg_temp": "decide", "total_precip_in": "line", "cdd": "decide", "hottest_day_temp": "avg"} CLIP0 = {"total_precip_in", "cdd"} # these can never be negative METHODS = ["avg", "line", "hybrid"] EXP02_LABEL = "Line + average (exp02)" # name of this experiment's row in the scorer table, charts and logs LABEL = {"avg": "Plain average", "line": "Plain line", "hybrid": "Line + Forest hybrid"} # ------------------------------------------------------------------ forecasting def _window_series(d, v, origin): """{city: Series(valid values, index=year)} over the 30 years ending at `origin`.""" lo = origin - WINDOW + 1 m = d["Year"].between(lo, origin) & v.notna() return {c: g.set_index("Year")["v"] for c, g in d.loc[m].assign(v=v[m]).groupby("ID")} def _line(s, years): slope, icpt = np.polyfit(s.index.values.astype(float), s.values, 1) return icpt + slope * np.asarray(years, float), slope, icpt def make_forecast(train_df, cov, metric_plan, origin=2018): """Returns {(city_id, metric): [v(origin+1), ..., v(origin+7)]} (2019..2025 when origin=2018). metric_plan maps metric -> 'avg' | 'line' | 'hybrid' (resolve 'decide' first, see decide()).""" assert_cutoff(train_df, "train_df") cities = sorted(train_df["ID"].unique()) d = train_df[train_df["Year"] <= origin] flags = valid_flags(d, cov) years = np.arange(origin + 1, origin + H + 1) out = {} for metric, method in metric_plan.items(): if method not in METHODS: raise ValueError(f"{metric}: method '{method}' unresolved") v = d[METRICS[metric]].where(flags[metric]) ser = _window_series(d, v, origin) pooled = float(np.concatenate([s.values for s in ser.values()]).mean()) if ser else 0.0 enough = {c for c, s in ser.items() if len(s) >= MIN_VALID} lo_clip = 0.0 if metric in CLIP0 else -np.inf fc = {} for c in cities: # plain-average fallback: whatever valid years exist, else the cross-city mean fc[c] = np.full(H, float(ser[c].mean()) if c in ser else pooled) if method in ("line", "hybrid"): lines = {c: _line(ser[c], years) for c in enough} for c, (vals, _, _) in lines.items(): fc[c] = vals if method == "hybrid" and enough: lw = np.arange(origin - WINDOW + 1, origin + 1) rows, last = [], {} for c in sorted(enough): _, slope, icpt = lines[c] resid = ser[c] - (icpt + slope * ser[c].index.values.astype(float)) full = resid.reindex(lw) # by year number: a missing year is NaN lag = pd.DataFrame({"y": full, **{f"lag_{k}": full.shift(k) for k in (1, 2, 3)}}).dropna() rows.append(lag) last[c] = (int(resid.index[-1]), list(resid.values[-3:])) tr = pd.concat(rows) feats = ["lag_1", "lag_2", "lag_3"] if len(tr): rf = RandomForestRegressor(n_estimators=N_TREES, random_state=SEED).fit(tr[feats], tr["y"]) hist = {c: list(h) for c, (_, h) in last.items()} start = {c: L + 1 for c, (L, _) in last.items()} pred = {c: [] for c in last} for t in range(min(start.values()), origin + H + 1): # step through gaps, one year at a time act = [c for c in sorted(last) if start[c] <= t] X = pd.DataFrame([hist[c][::-1][:3] for c in act], columns=feats) for c, p in zip(act, rf.predict(X)): hist[c].append(float(p)) if t > origin: pred[c].append(float(p)) for c in last: fc[c] = lines[c][0] + np.array(pred[c][-H:]) for c in cities: out[(c, metric)] = [float(x) for x in np.maximum(fc[c], lo_clip)] return out # ---------------------------------------------------------------- held-back test def heldout(train_df, cov, origins=ORIGINS): """MAE and skill (vs plain average) per metric, method and origin, pooled over scored city-years.""" flags_all = valid_flags(train_df, cov) rows = [] for o in origins: res = {m: make_forecast(train_df, cov, {k: m for k in METRICS}, o) for m in METHODS} for metric, col in METRICS.items(): v = train_df[col].where(flags_all[metric]) win = _window_series(train_df[train_df["Year"] <= o], v[train_df["Year"] <= o], o) ok = {c for c, s in win.items() if len(s) >= MIN_VALID} t = train_df.assign(v=v)[train_df["Year"].between(o + 1, o + H)].dropna(subset=["v"]) t = t[t["ID"].isin(ok)] err = {m: [] for m in METHODS} for r in t.itertuples(): for m in METHODS: err[m].append(abs(res[m][(r.ID, metric)][r.Year - o - 1] - r.v)) mae = {m: float(np.mean(err[m])) for m in METHODS} for m in METHODS: rows.append({"metric": metric, "method": m, "origin": o, "mae": mae[m], "skill": 1 - mae[m] / mae["avg"], "n": len(err[m])}) df = pd.DataFrame(rows) avg = df.groupby(["metric", "method"], sort=False)[["mae", "skill", "n"]].mean().reset_index() avg.insert(2, "origin", "average") return pd.concat([df.assign(origin=df.origin.astype(str)), avg], ignore_index=True) def decide(res, plan=METRIC_PLAN, margin=MARGIN): """The decision rule of spec section 5. Returns {metric: {...numbers and chosen method...}}.""" dec = {} for metric, how in plan.items(): r = res[res.metric == metric] sk = {m: float(r[(r.method == m) & (r.origin == "average")].skill.iloc[0]) for m in METHODS} per = {m: {o: float(x) for o, x in zip(r[(r.method == m) & (r.origin != "average")].origin, r[(r.method == m) & (r.origin != "average")].skill)} for m in METHODS} info = {"plan": how, "avg_skill": sk, "per_origin_skill": per} if how == "decide": best_other = "line" if sk["line"] >= sk["avg"] else "avg" wins = sum(per["hybrid"][o] > per["line"][o] and per["hybrid"][o] > per["avg"][o] for o in per["hybrid"]) checks = {"avg_skill_higher_than_line_and_avg": sk["hybrid"] > max(sk["line"], sk["avg"]), "beats_both_in_at_least_2_of_3_origins": wins >= 2, f"margin_ge_{margin}": sk["hybrid"] - max(sk["line"], sk["avg"]) >= margin} info.update(checks=checks, origins_won=wins, margin_achieved=sk["hybrid"] - max(sk["line"], sk["avg"]), method="hybrid" if all(checks.values()) else best_other, hybrid_dropped=not all(checks.values())) else: info["method"] = how dec[metric] = info return dec # ------------------------------------------------------------------------ main def git(*a): try: return subprocess.check_output(["git", *a], cwd=ROOT, text=True, stderr=subprocess.DEVNULL).strip() except Exception: return "" def to_rows(fc): return pd.DataFrame([{"city_id": c, "year": 2019 + i, "metric": m, "predicted_value": x} for (c, m), vals in fc.items() for i, x in enumerate(vals)] ).sort_values(["city_id", "year", "metric"]).reset_index(drop=True) def main(argv=None): ap = argparse.ArgumentParser() ap.add_argument("--out-dir", default=str(ROOT / "outputs" / "exp02")) ap.add_argument("--charts", action="store_true", help="also draw the charts (code/exp02_charts.py)") a = ap.parse_args(argv) out = Path(a.out_dir) (out / "charts").mkdir(parents=True, exist_ok=True) train_df, cov = load_train() # 1. held-back test -> heldout_results.csv, then the decision -> decision.json (before any final forecast) res = heldout(train_df, cov) res.to_csv(out / "heldout_results.csv", index=False) dec = decide(res) (out / "decision.json").write_text(json.dumps(dec, indent=2)) # 2. final forecast with the decided plan plan = {m: d["method"] for m, d in dec.items()} fc = make_forecast(train_df, cov, plan, 2018) sub = to_rows(fc) assert len(sub) == 210 * 7 * 4 and np.isfinite(sub.predicted_value).all() assert (sub[sub.metric.isin(CLIP0)].predicted_value >= 0).all() sub.to_csv(out / "submission_exp02.csv", index=False) (out / "run_log.json").write_text(json.dumps({ "experiment": "exp02", "scorer_label": EXP02_LABEL, "timestamp_utc": datetime.now(timezone.utc).isoformat(), "config": {"METRIC_PLAN": METRIC_PLAN, "resolved_plan": plan, "MARGIN": MARGIN, "WINDOW": WINDOW, "ORIGINS": ORIGINS, "N_TREES": N_TREES, "SEED": SEED}, "dataset": "CMU-OSPO/protostar-climate-train-to-2018", "dataset_revision": dataset_revision(), "packages": {p: version(p) for p in ["pandas", "numpy", "scikit-learn", "matplotlib", "huggingface_hub"]}, "git_commit": git("rev-parse", "HEAD"), "git_dirty": bool(git("status", "--porcelain"))}, indent=2)) show = res[res.origin == "average"].pivot(index="metric", columns="method", values="skill") print("Held-back average skill (vs plain average):\n", show.round(3).to_string()) for m, d in dec.items(): print(f"{m}: {d['method']}" + (" (hybrid dropped)" if d.get("hybrid_dropped") else "")) print("wrote", out) if a.charts: from exp02_charts import make_charts make_charts(out, train_df, cov) if __name__ == "__main__": main()