Datasets:
File size: 22,717 Bytes
d79a78d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 | """Backtest harness and model selection.
Folds simulate the real task (train up to an origin, forecast the next 7 years):
selection folds <=2011 -> 2012-2018 and <=2004 -> 2005-2011 (candidates are ranked here)
confirmation folds <=1997 -> 1998-2004 and <=1990 -> 1991-1997 (veto only, never ranking)
practice fold <=2010 -> 2011-2018 (organizers' published numbers)
Selection climbs a ladder per metric: climatology -> trend (pooled slope; for cdd also the
regression on forecast temperature, competing on the same rung) -> blend. The best candidate of a rung (mean skill over
the two selection folds) replaces the incumbent only if
1. its gain is positive on BOTH selection folds (i.e. larger than one fold-to-fold
standard error) and larger than 2 standard errors of a paired bootstrap over cities, and
2. it does not lose to the 30-year average on the confirmation folds (mean skill >= 0).
Otherwise the simpler model stays ("stop adding complexity when gains < noise").
Writes reports/backtest/{candidates.csv, table.csv, selected.json, BACKTEST.md}.
Run: PYTHONPATH=code python -m protostar.backtest
"""
import json
import numpy as np
import pandas as pd
from .common import (BASELINE_WINDOW, CONFIRMATION_FOLDS, METRICS, MIN_VALID_YEARS, PRACTICE_FOLD,
REPORTS, SEED, SELECTION_FOLDS, UNITS)
from .data import Panel, cities, load_annual, long_table, panels
from .models import BASELINES, neighbor_order, predict, window_mean
from .scoring import challenge_score, skill
OUT = REPORTS / "backtest"
N_BOOT = 2000
SE_MULT = 2.0 # a rung must gain more than this many standard errors
K_GRID = (8, 16, 32) # nearest cities pooled for the slope
LAM_GRID = (0.0, 0.25, 0.5, 0.75, 1.0) # weight of the city's own slope
W_STEP = 0.1 # blend weights on a simplex grid
def key(spec: dict) -> str:
return json.dumps(spec, sort_keys=True)
class Fold:
"""One origin: training panels, truth, scored rows, and cached per-city errors."""
def __init__(self, P: dict[str, Panel], origin: int, first: int, last: int,
min_valid: int = MIN_VALID_YEARS):
self.origin, self.first, self.last = origin, first, last
self.name = f"<={origin} -> {first}-{last}"
self.years = np.arange(first, last + 1)
self.train = {m: p.upto(origin) for m, p in P.items()}
truth = {m: p.between(first, last).y for m, p in P.items()}
self.truth = truth
# scored = valid test year of a city-metric that has a 30-year-average baseline
self.scored = {m: (np.isfinite(window_mean(self.train[m], BASELINE_WINDOW, min_valid))[:, None]
& np.isfinite(truth[m])) for m in P}
# the reference of the skill: the 30-year average under this fold's minimum of valid years
self.ref = {"type": "mean", "n": BASELINE_WINDOW, "min_valid": min_valid}
self._cache = {}
def predict(self, spec: dict, metric: str, order: np.ndarray) -> np.ndarray:
"""Forecast of `spec` for the test years, from the training panels only."""
return predict(spec, self.train, metric, order, self.years)
def errors(self, spec: dict, metric: str, order: np.ndarray):
"""(sum of |error| per city, scored rows per city) of `spec` on this fold."""
k = (metric, key(spec))
if k not in self._cache:
pred = self.predict(spec, metric, order)
mask = self.scored[metric]
err = np.abs(np.where(mask, pred - self.truth[metric], 0.0))
assert np.isfinite(err).all(), f"non-finite prediction on a scored row: {spec}"
self._cache[k] = (err.sum(axis=1), mask.sum(axis=1))
return self._cache[k]
def mae(self, spec, metric, order) -> float:
s, n = self.errors(spec, metric, order)
return float(s.sum() / n.sum())
def skill(self, spec, metric, order) -> float:
return skill(self.mae(spec, metric, order), self.mae(self.ref, metric, order))
def bias(self, spec, metric, order) -> float:
"""Mean signed error (forecast - actual) over the scored rows, as a share of the mean actual."""
mask = self.scored[metric]
diff = np.where(mask, self.predict(spec, metric, order) - self.truth[metric], 0.0)
return float(diff.sum() / np.where(mask, self.truth[metric], 0.0).sum())
class Harness:
def __init__(self, P: dict[str, Panel], order: np.ndarray):
self.P, self.order = P, order
self.sel = [Fold(P, *f) for f in SELECTION_FOLDS]
self.hold = [Fold(P, *f) for f in CONFIRMATION_FOLDS]
rng = np.random.default_rng(SEED)
n = len(order)
self.boot = rng.multinomial(n, np.full(n, 1.0 / n), size=N_BOOT).astype(float) # [B, C]
# mean skill over the selection folds, and its per-fold values
def skills(self, spec, metric, folds=None):
return [f.skill(spec, metric, self.order) for f in (folds or self.sel)]
def mean_skill(self, spec, metric) -> float:
return float(np.mean(self.skills(spec, metric)))
def paired(self, spec_a, spec_b, metric):
"""Gain of a over b in mean selection-fold skill: (gain, bootstrap SE, per-fold gains)."""
per_fold, boot = [], np.zeros(N_BOOT)
for f in self.sel:
(sa, n), (sb, _), (sr, _) = (f.errors(s, metric, self.order) for s in (spec_a, spec_b, f.ref))
per_fold.append(float((sb.sum() - sa.sum()) / sr.sum()))
boot += (self.boot @ (sb - sa)) / (self.boot @ sr) / len(self.sel)
return float(np.mean(per_fold)), float(boot.std(ddof=1)), per_fold
def best(self, specs, metric):
return max(specs, key=lambda s: self.mean_skill(s, metric))
# ── Candidate families ───────────────────────────────────────────────────────
RECENT = {"type": "recent", "n": 10} # 10-year mean (2009-2018 for complete records)
LONG = {"type": "mean", "n": BASELINE_WINDOW} # 30-year mean (1989-2018)
def climatology_specs():
return [LONG, RECENT]
def sensitivity_specs(metric):
"""Not used for selection: how much the choices of window, k and pooling matter."""
out = [{"type": "mean", "n": n, "min_valid": min(MIN_VALID_YEARS, n // 2)} for n in (15, 20, 40, 50)]
out += [{"type": "ptrend", "n": n, "k": k, "lam": lam}
for n, k, lam in ((30, 64, 0.5), (30, 0, 0.5), (30, 0, 0.0), (40, 32, 0.5), (40, 0, 0.5), (50, 32, 0.5))]
if metric == "cdd":
out += [{"type": "blend", "parts": [[0.5, {"type": "ptrend", "n": 30, "k": 32, "lam": 0.5}],
[0.5, {"type": "cdd_via_temp", "n": 30,
"temp": {"type": "ptrend", "n": 30, "k": 32, "lam": 0.5}}]]}]
return out
def trend_specs(k_grid=K_GRID):
return [{"type": "ptrend", "n": BASELINE_WINDOW, "k": k, "lam": lam}
for k in k_grid for lam in LAM_GRID if not (lam == 1.0 and k != k_grid[0])]
def blend_specs(recent, long, trend):
steps = int(round(1 / W_STEP))
out = []
for i in range(steps + 1):
for j in range(steps + 1 - i):
w = (i / steps, j / steps, (steps - i - j) / steps)
parts = [[wi, s] for wi, s in zip(w, (recent, long, trend)) if wi > 0]
if len(parts) > 1:
out.append({"type": "blend", "parts": parts})
return out
def label(spec: dict) -> str:
t = spec["type"]
if t == "mean":
return f"{spec['n']}-year mean"
if t == "recent":
return f"mean of last {spec['n']} valid years"
if t == "last":
return "last year repeated"
if t == "ols":
return f"per-city OLS line ({spec['n']} y)"
if t == "ptrend":
k = "all" if spec["k"] == 0 else spec["k"]
window = "" if spec["n"] == BASELINE_WINDOW else f"{spec['n']} y, "
return f"pooled trend ({window}k={k}, own-slope weight {spec['lam']:g})"
if t == "cdd_via_temp":
window = "" if spec["n"] == BASELINE_WINDOW else f" ({spec['n']} y)"
return f"cdd ~ temp regression{window} on [{label(spec['temp'])}]"
if t == "blend":
return "blend: " + " + ".join(f"{w:g} x [{label(s)}]" for w, s in spec["parts"])
return t
# ── Selection ladder ─────────────────────────────────────────────────────────
def climb(h: Harness, metric: str, rungs: list[tuple[str, list[dict]]], log: list[dict]) -> dict:
"""Climb the ladder for one metric; returns the selected spec."""
incumbent = BASELINES["30-year average"]
for rung, specs in rungs:
cand = h.best(specs, metric)
gain, se, per_fold = h.paired(cand, incumbent, metric)
confirm = float(np.mean(h.skills(cand, metric, h.hold)))
consistent = bool(gain > SE_MULT * se and min(per_fold) > 0)
accept = bool(consistent and confirm >= 0)
log.append({"metric": metric, "rung": rung, "candidate": label(cand), "spec": key(cand),
"n_candidates": len(specs), "incumbent": label(incumbent),
"gain_mean_skill": gain, "se_gain": se,
**{f"gain_{f.name}": g for f, g in zip(h.sel, per_fold)},
"confirmation_mean_skill": confirm,
"verdict": "accepted" if accept else
("vetoed: loses to the 30-year average on the confirmation folds" if consistent
else "rejected: worse than the incumbent" if gain <= 0
else "rejected: gain within noise"),
"accepted": accept})
if accept:
incumbent = cand
return incumbent
def select(h: Harness):
log, chosen, rung_best = [], {}, {}
def ladder(metric, trend_like):
rung_best[(metric, "climatology")] = RECENT
rung_best[(metric, "trend")] = comp = h.best(trend_like, metric)
blends = blend_specs(RECENT, LONG, comp)
rung_best[(metric, "blend")] = h.best(blends, metric)
rungs = [("1 climatology (10-year mean)", [RECENT]),
("2-3 trend" if metric == "cdd" else "2 pooled trend", trend_like),
("4 blend (recent, long, trend)", blends)]
return climb(h, metric, rungs, log)
chosen["annual_avg_temp"] = ladder("annual_avg_temp", trend_specs())
# cdd: the direct pooled trend and the regression on forecast temperature compete on one rung
temps = {key(s): s for s in (chosen["annual_avg_temp"], rung_best[("annual_avg_temp", "trend")])}
via = [{"type": "cdd_via_temp", "n": BASELINE_WINDOW, "temp": t} for t in temps.values()]
chosen["cdd"] = ladder("cdd", trend_specs() + via)
chosen["total_precip_in"] = ladder("total_precip_in", trend_specs())
chosen["hottest_day_temp"] = ladder("hottest_day_temp", trend_specs())
return chosen, pd.DataFrame(log), rung_best
# ── Reporting ────────────────────────────────────────────────────────────────
def fold_row(f: Fold, name: str, system: dict[str, dict], order) -> dict:
row = {"model": name, "fold": f.name}
skills = {}
for m in METRICS:
row[f"mae_{m}"] = f.mae(system[m], m, order)
skills[m] = f.skill(system[m], m, order)
row[f"skill_{m}"] = skills[m]
row[f"bias_{m}"] = f.bias(system[m], m, order)
row["score"] = challenge_score(skills)
row["n_rows"] = int(sum(f.scored[m].sum() for m in METRICS))
return row
def systems(chosen: dict, rung_best: dict, h: Harness) -> dict[str, dict[str, dict]]:
"""Named full systems (one spec per metric) for the model x fold table."""
same = lambda s: {m: s for m in METRICS}
out = {name: same(spec) for name, spec in BASELINES.items()}
out["1 climatology: 10-year mean"] = same(RECENT)
for k in K_GRID:
out[f"2 pooled trend k={k}, own-slope weight 0.5"] = same({"type": "ptrend", "n": 30, "k": k, "lam": 0.5})
direct = {m: h.best(trend_specs(), m) for m in METRICS}
out["2 pooled trend, best (k, weight) per metric"] = direct
via = {"type": "cdd_via_temp", "n": BASELINE_WINDOW, "temp": chosen["annual_avg_temp"]}
out["3 as 2, but cdd ~ temp regression on forecast temp"] = {**direct, "cdd": via}
out["4 best blend (recent, long, trend) per metric"] = {m: rung_best[(m, "blend")] for m in METRICS}
out["FINAL (selected per metric)"] = chosen
return out
def practice_check(P, order) -> pd.DataFrame:
"""Reproduce the organizers' practice-run numbers (train <= 2010, test 2011-2018).
official rule: baselines need >= 10 valid years in 1981-2010 (About tab, Space discussion #3)
notebook rule: no minimum; the line needs 2 valid years. The notebook scores its 10-year mean
on every valid test row that has a forecast, 6 rows more than its reference.
"""
rows = []
for rule, mv, line_mv in (("official (>= 10 valid years in the window)", MIN_VALID_YEARS, MIN_VALID_YEARS),
("example notebook (no minimum)", 1, 2)):
f = Fold(P, *PRACTICE_FOLD, min_valid=mv)
for name, spec in (("30-year average", f.ref),
("Linear trend", {"type": "ols", "n": BASELINE_WINDOW, "min_valid": line_mv}),
("10-year mean", RECENT)):
rows.append({"rule": rule, "model": name, **{m: f.mae(spec, m, order) for m in METRICS}})
f = Fold(P, *PRACTICE_FOLD, min_valid=1)
own_rows = {}
for m in METRICS:
pred = predict(RECENT, f.train, m, order, f.years, fill=False)
mask = np.isfinite(f.truth[m]) & np.isfinite(pred)
own_rows[m] = float(np.abs(pred - f.truth[m])[mask].mean())
rows.append({"rule": "example notebook, scored on the rows the model covers", "model": "10-year mean", **own_rows})
return pd.DataFrame(rows)
def md_table(df: pd.DataFrame, fmt: dict | None = None) -> str:
fmt = fmt or {}
cols = list(df.columns)
lines = ["| " + " | ".join(cols) + " |", "|" + "|".join("---" for _ in cols) + "|"]
for _, r in df.iterrows():
cells = []
for c in cols:
v = r[c]
cells.append(fmt[c].format(v) if c in fmt and pd.notna(v) else str(v))
lines.append("| " + " | ".join(cells) + " |")
return "\n".join(lines)
def main():
OUT.mkdir(parents=True, exist_ok=True)
annual = load_annual()
P = panels(long_table(annual))
geo = cities(annual)
assert list(geo["city_id"]) == list(P["cdd"].city_ids)
order = neighbor_order(geo["lat"].to_numpy(), geo["lon"].to_numpy())
h = Harness(P, order)
practice = practice_check(P, order)
chosen, log, rung_best = select(h)
# every candidate on every fold (for the record)
cands = []
for m in METRICS:
used = (climatology_specs() + trend_specs() + list(BASELINES.values()) + [chosen[m]]
+ [json.loads(r["spec"]) for _, r in log[log["metric"] == m].iterrows()]
+ blend_specs(RECENT, LONG, rung_best[(m, "trend")]))
specs = {key(s): ("selection", s) for s in used}
specs.update({key(s): ("sensitivity only", s) for s in sensitivity_specs(m) if key(s) not in specs})
for role, s in specs.values():
for kind, folds in (("selection", h.sel), ("confirmation", h.hold)):
for f in folds:
cands.append({"metric": m, "model": label(s), "role": role, "fold": f.name, "fold_kind": kind,
"n_rows": int(f.scored[m].sum()), "mae": f.mae(s, m, order),
"mae_30y_average": f.mae(f.ref, m, order),
"skill": f.skill(s, m, order), "spec": key(s)})
cands = pd.DataFrame(cands)
assert not cands.duplicated(["metric", "model", "fold"]).any(), "two specs share a label"
cands.to_csv(OUT / "candidates.csv", index=False)
table = pd.DataFrame([fold_row(f, name, system, order)
for name, system in systems(chosen, rung_best, h).items()
for f in h.sel + h.hold])
table["fold_kind"] = np.where(table["fold"].isin([f.name for f in h.sel]), "selection", "confirmation")
table.to_csv(OUT / "table.csv", index=False)
log.to_csv(OUT / "ladder.csv", index=False)
practice.to_csv(OUT / "practice_run.csv", index=False)
# gate: the final system must beat both baselines' Score on both selection folds
final = table[table["model"] == "FINAL (selected per metric)"].set_index("fold")
gate = {}
for f in h.sel:
others = table[(table["fold"] == f.name) & table["model"].isin(["30-year average", "Linear trend"])]
gate[f.name] = bool((final.loc[f.name, "score"] > others["score"]).all())
(OUT / "selected.json").write_text(json.dumps(
{"selected": chosen, "labels": {m: label(s) for m, s in chosen.items()},
"gate_beats_both_baselines_on_both_selection_folds": gate}, indent=1) + "\n")
if not all(gate.values()):
raise SystemExit(f"GATE FAILED: the final system does not beat both baselines on both folds: {gate}")
write_report(table, log, practice, chosen, gate, h, cands)
pd.set_option("display.width", 250, "display.max_columns", 30, "display.max_colwidth", 70)
print(practice.round(3).to_string(index=False), "\n")
print(log.drop(columns=["spec", "incumbent"]).round(4).to_string(index=False), "\n")
print(table.drop(columns=[c for c in table if c.startswith("mae_")]).round(4).to_string(index=False))
print("\nselected:", json.dumps({m: label(s) for m, s in chosen.items()}, indent=1))
print("gate:", gate)
def write_report(table, log, practice, chosen, gate, h, cands):
sk = {f"skill_{m}": "{:+.1%}" for m in METRICS}
mae = {"mae_annual_avg_temp": "{:.3f}", "mae_total_precip_in": "{:.3f}",
"mae_hottest_day_temp": "{:.3f}", "mae_cdd": "{:.1f}"}
t = table[["model", "fold_kind", "fold", "score", *sk, *mae, "n_rows"]]
lad = log[["metric", "rung", "candidate", "n_candidates", "incumbent", "gain_mean_skill", "se_gain",
*[f"gain_{f.name}" for f in h.sel], "confirmation_mean_skill", "verdict"]]
sens = (cands.pivot(index=["metric", "role", "model"], columns="fold", values="skill")
.reindex(pd.MultiIndex.from_frame(cands[["metric", "role", "model"]].drop_duplicates()))
[[f.name for f in h.sel + h.hold]].reset_index())
sens["mean selection"] = sens[[f.name for f in h.sel]].mean(axis=1)
sens["mean confirmation"] = sens[[f.name for f in h.hold]].mean(axis=1)
sens = sens[~sens["model"].str.startswith("blend: 0.")
| sens["model"].isin([label(s) for s in chosen.values()])
| (sens["role"] == "sensitivity only")]
text = f"""# Backtest report
Generated by `python -m protostar.backtest` (do not edit). Scoring is the challenge's:
MAE per metric over valid years (10-day rule) of city-metrics with a baseline (>= {MIN_VALID_YEARS}
valid years in the 30-year window), skill `S = 1 - MAE / MAE(30-year average)`, and
`Score = 100 * max(0, mean of the four S)`. Every fold trains on years <= origin only.
## Harness check against the organizers' practice run (train <= 2010, test 2011-2018)
Published (temp / precip / hottest day / cdd):
- official rule, CDD only: 30-year average 212 and linear trend 175 (About tab); 211.7 and 174.6
(organizer, Space discussion #3);
- example notebook: 30-year average 1.657 / 6.198 / 2.563 / 212.636 and its 10-year mean
1.513 / 6.207 / 2.568 / 184.059. The notebook scores its model on every valid test row that has
a forecast, 6 rows more than its reference (last table row); on the reference's rows the same
model gives the "example notebook (no minimum) / 10-year mean" row.
Our linear trend is a per-city least-squares line on the valid years of the window (x = calendar
year), not floored at zero; the organizers' fit is not published beyond these numbers.
{md_table(practice, {m: "{:.3f}" for m in METRICS})}
## Model x fold x metric
`selection` folds rank the candidates; `confirmation` folds can only veto (see the ladder).
Rows 1-4 are the rungs of the plan applied to all four metrics; FINAL is the per-metric choice.
{md_table(t, {"score": "{:.2f}", **sk, **mae})}
## Selection ladder
The best candidate of a rung (mean skill over the two selection folds) replaces the incumbent only
if its gain is positive on both selection folds and larger than {SE_MULT:g} standard errors (paired
bootstrap over cities, {N_BOOT} resamples, seed {SEED}), and its mean skill on the confirmation
folds is not negative. Gains are in skill units (0.01 = 1 % of the 30-year-average MAE).
Caveats. (1) Cities are spatially correlated, so a bootstrap over cities gives a lower bound on
the standard error; that is why a gain must also hold on both folds. (2) The test is applied to
the best of `n_candidates` with no adjustment for that choice; the accepted rungs clear it by a
wide margin and the one marginal case was vetoed. (3) The veto compares with the 30-year average,
not with the incumbent: `confirmation_mean_skill` of an accepted rung can be below the
incumbent's (compare the rows of the table above). (4) Because the veto decided the hottest-day
model, the confirmation folds are a clean holdout only for the other three metrics.
{md_table(lad, {c: "{:+.4f}" for c in lad.columns if c.startswith("gain_")} | {"se_gain": "{:.4f}", "confirmation_mean_skill": "{:+.4f}"})}
## Selected
{chr(10).join(f"- `{m}` ({UNITS[m]}): {label(s)}" for m, s in chosen.items())}
Gate (final system beats both the 30-year average and the linear trend on the Score, on both
selection folds): {gate}
## Candidates, skill per fold (and choices we did not let the backtest make)
Every candidate is in `candidates.csv`; this table leaves out the 63 blends of each metric (the
best one of each is in the ladder above). `sensitivity only` rows were never candidates: they
show how much the window length (40 or 50 years), the number of pooled neighbors (k=64, all
cities) and other climatology windows would matter. Pooled-trend rows whose label names no
window use the 30-year window.
{md_table(sens, {c: "{:+.1%}" for c in sens.columns if c not in ("metric", "role", "model")})}
"""
(OUT / "BACKTEST.md").write_text(text)
if __name__ == "__main__":
main()
|