AlecAlameddine's picture
Upload 105 files
d79a78d verified
Raw History Blame Contribute Delete
8.35 kB
"""Regenerate the number tables of README.md from the reports, so the write-up cannot drift.
Blocks between `<!-- BEGIN:name -->` and `<!-- END:name -->` in README.md are replaced:
backtest Score by model and fold; MAE and skill by metric (reports/backtest/table.csv)
policy-numbers the forecast numbers quoted in the Policy Idea (reports/forecast/forecast_summary.csv)
Also writes reports/forecast/POLICY_NUMBERS.md (all cities behind the policy numbers).
Run: PYTHONPATH=code python -m protostar.docs
"""
import json
import re
import pandas as pd
from .common import BASELINE_WINDOW, METRICS, REPORTS, ROOT, TRAIN_END, UNITS
from .data import load_coverage
SHORT = {"annual_avg_temp": "avg temp (°F)", "total_precip_in": "precip (in)",
"hottest_day_temp": "hottest day (°F)", "cdd": "CDD (°F-days)"}
ROWS_B = ["30-year average", "Linear trend", "1 climatology: 10-year mean", "FINAL (selected per metric)"]
MIN_CDD_FOR_PCT = 500 # relative changes are only ranked for cities with a real cooling season
LEVER_PCT, LEVER_ABS = 10, 200 # Policy Idea lever (b): >= 10 % or >= 200 °F-days above the normal, cities with >= 500 CDD
MIN_YEARS_RANKED = 25 # city rankings: at least 25 valid years in 1989-2018 and no documented station change there
def backtest_block() -> str:
t = pd.read_csv(REPORTS / "backtest" / "table.csv")
folds = list(dict.fromkeys(t["fold"]))
kind = dict(zip(t["fold"], t["fold_kind"]))
head = "| Model | " + " | ".join(f"{f}<br>({kind[f]})" for f in folds) + " |"
lines = ["**Score (0–100, the challenge's formula) by model and fold**", "", head,
"|---|" + "---:|" * len(folds)]
for model in dict.fromkeys(t["model"]):
g = t[t["model"] == model].set_index("fold")
name = f"**{model}**" if model.startswith("FINAL") else model
lines.append(f"| {name} | " + " | ".join(f"{g.loc[f, 'score']:.2f}" for f in folds) + " |")
lines += ["", "**MAE (skill vs the 30-year average) by metric**", "",
"| Fold | Model | " + " | ".join(SHORT[m] for m in METRICS) + " |", "|---|---|" + "---:|" * len(METRICS)]
for f in folds:
for model in ROWS_B:
r = t[(t["fold"] == f) & (t["model"] == model)].iloc[0]
cells = []
for m in METRICS:
digits = 1 if m == "cdd" else 3
cells.append(f"{r[f'mae_{m}']:.{digits}f} ({100 * r[f'skill_{m}']:+.1f} %)")
name = "**FINAL**" if model.startswith("FINAL") else model.replace("1 climatology: ", "")
lines.append(f"| {f} | {name} | " + " | ".join(cells) + " |")
sel = json.loads((REPORTS / "backtest" / "selected.json").read_text())
lines += ["", "Selected per metric: " + "; ".join(f"`{m}` → {sel['labels'][m]}" for m in METRICS) + "."]
return "\n".join(lines)
def policy_frames():
s = pd.read_csv(REPORTS / "forecast" / "forecast_summary.csv")
s = s[s["n_valid_1989_2018"] >= 10] # city-metrics with a baseline
c = s[s["metric"] == "cdd"].copy()
c["change"] = c["forecast_mean_2019_2025"] - c["mean_1989_2018"]
c["change_pct"] = 100 * c["change"] / c["mean_1989_2018"]
cov = load_coverage()
moved = set(cov.loc[(cov["Station Changes"] > 0) & (cov["Year"] > TRAIN_END - BASELINE_WINDOW), "ID"])
c["station_change_1989_2018"] = c["city_id"].isin(moved)
c["ranked"] = (c["n_valid_1989_2018"] >= MIN_YEARS_RANKED) & ~c["station_change_1989_2018"]
return s, c
def bias_row() -> str:
"""Signed CDD error (share of actual CDD) of the 30-year normal and of our model, per backtest fold."""
t = pd.read_csv(REPORTS / "backtest" / "table.csv")
cells = []
for fold in dict.fromkeys(t["fold"]):
normal = t[(t["fold"] == fold) & (t["model"] == "30-year average")]["bias_cdd"].iloc[0]
ours = t[(t["fold"] == fold) & t["model"].str.startswith("FINAL")]["bias_cdd"].iloc[0]
cells.append(f"{fold.split('-> ')[1].replace('-', '–')}: {100 * normal:+.1f} % → {100 * ours:+.1f} %")
return "| CDD bias in backtests, 30-year normal → our model (forecast minus actual, share of actual) | " \
+ "; ".join(cells) + " |"
def policy_block() -> str:
s, c = policy_frames()
t = s[s["metric"] == "annual_avg_temp"]
h = s[s["metric"] == "hottest_day_temp"]
big = c[c["mean_1989_2018"] >= MIN_CDD_FOR_PCT]
top_abs = c[c["ranked"]].sort_values("change", ascending=False).head(8)
top_pct = big[big["ranked"]].sort_values("change_pct", ascending=False).head(10)
ratio = 100 * (c["forecast_mean_2019_2025"].mean() / c["mean_1989_2018"].mean() - 1)
recent = 100 * (c["mean_last10_valid"].mean() / c["mean_1989_2018"].mean() - 1)
lever = int(((big["change_pct"] >= LEVER_PCT) | (big["change"] >= LEVER_ABS)).sum())
lines = [
"| Quantity (our forecasts vs each city's 1989–2018 mean) | Value |", "|---|---|",
f"| Cities with a baseline | {len(c)} of 210 |",
f"| CDD, mean over cities: 1989–2018 → forecast 2019–2025 | {c['mean_1989_2018'].mean():,.0f} → "
f"{c['forecast_mean_2019_2025'].mean():,.0f} °F-days ({ratio:+.1f} %) |",
f"| CDD change, median city | {c['change_pct'].median():+.1f} % |",
f"| CDD, last 10 valid years (observed, up to 2018) vs the 1989–2018 mean | {recent:+.1f} % |",
f"| Cities with CDD ≥ +10 % / ≥ +15 % | {int((c['change_pct'] >= 10).sum())} / {int((c['change_pct'] >= 15).sum())} |",
f"| Cities with ≥ {MIN_CDD_FOR_PCT} CDD and CDD ≥ +{LEVER_PCT} % or ≥ +{LEVER_ABS} °F-days (lever b) | {lever} |",
bias_row(),
f"| Average temperature change, mean over cities | {(t['forecast_mean_2019_2025'] - t['mean_1989_2018']).mean():+.2f} °F |",
"| Hottest day of the year | no trend survived the backtests: forecast = the 1989–2018 mean "
f"(largest difference {(h['forecast_mean_2019_2025'] - h['mean_1989_2018']).abs().max():.2f} °F, rounding) |",
"| Largest CDD increases, °F-days* | " + ", ".join(
f"{r['name']} {r['change']:+.0f}" for _, r in top_abs.iterrows()) + " |",
f"| Largest relative CDD increases (cities with ≥ {MIN_CDD_FOR_PCT} CDD)* | " + ", ".join(
f"{r['name']} {r['change_pct']:+.0f} %" for _, r in top_pct.iterrows()) + " |",
"",
f"\\* Rankings use cities with at least {MIN_YEARS_RANKED} valid years in 1989–2018 and no documented station "
f"change in that window ({int(c['ranked'].sum())} of {len(c)}); all cities are in "
"`reports/forecast/policy_cdd_by_city.csv`.",
]
return "\n".join(lines)
def write_policy_report():
_, c = policy_frames()
cols = ["name", "city_id", "lat", "lon", "n_valid_1989_2018", "station_change_1989_2018", "mean_1989_2018",
"mean_last10_valid", "forecast_2019", "forecast_2025", "forecast_mean_2019_2025", "change", "change_pct"]
out = c.sort_values("change_pct", ascending=False)[cols].round(1)
out.to_csv(REPORTS / "forecast" / "policy_cdd_by_city.csv", index=False)
text = ["# Numbers behind the Policy Idea", "",
"Generated by `python -m protostar.docs` from `reports/forecast/forecast_summary.csv` "
"(our forecasts and the 1989–2018 valid years of the training data; no data after 2018).", "",
policy_block(), "",
"All cities, sorted by relative change: `reports/forecast/policy_cdd_by_city.csv` "
f"(columns in {UNITS['cdd']}; `change` = forecast mean 2019–2025 minus the 1989–2018 mean).", ""]
(REPORTS / "forecast" / "POLICY_NUMBERS.md").write_text("\n".join(text))
def replace_block(text: str, name: str, body: str) -> str:
pattern = re.compile(rf"(<!-- BEGIN:{name} -->\n).*?(<!-- END:{name} -->)", re.S)
if not pattern.search(text):
raise SystemExit(f"README.md has no block named {name!r}")
return pattern.sub(lambda m: m.group(1) + body + "\n" + m.group(2), text)
def main():
write_policy_report()
readme = ROOT / "README.md"
text = readme.read_text()
for name, body in (("backtest", backtest_block()), ("policy-numbers", policy_block())):
text = replace_block(text, name, body)
readme.write_text(text)
print("README.md tables regenerated from reports/")
if __name__ == "__main__":
main()