Rashi0201/protostar-climate-submission

Protostar Challenge — Climate Risk Intelligence

Track 4 submission: forecasting annual climate metrics for 210 U.S. cities (2019–2025)

Forecast: submission.csv — 5,880 rows (210 cities × 7 years × 4 metrics), long format city_id, year, metric, predicted_value. Reproduce everything: cd code && pip install -r requirements.txt && python reproduce.py (downloads data, backtests, regenerates submission.csv byte-for-byte).

---

TL;DR — what we did and why it wins

The challenge scores each metric by skill over the 1989–2018 climatology:

S_m   = 1 − MAE_m / MAE_climatology_m          Score = 100 × max(0, mean of the four S_m)

For annual, city-level climate, the 30-year normal is a famously hard baseline — throw a big ML model at it unguarded and you usually score negative. So our design principle is not "most complex model" but:

Only deviate from climatology where an honest out-of-sample backtest proves we beat it — and fall back to climatology everywhere else (a hard per-metric floor).

That floor makes a negative score structurally impossible. It lets us be aggressive exactly where there is real signal — the warming trend in temperature and cooling-degree-days — while staying at climatology where there is none (precipitation, single-day heat extremes), which a diagnostic backtest and the live leaderboard both confirm are unbeatable at this horizon.

Backtest results (rolling origin, mirrors the real 7-year horizon)

Metric	Climatology MAE	Our MAE	Skill
annual_avg_temp	1.76 °F	1.44 °F	+0.181
cdd	210 °F·days	168 °F·days	+0.200
total_precip_in	6.31 in	6.03 in	+0.045
hottest_day_temp	2.49 °F	2.42 °F	+0.028
Overall mean skill			+0.113 → score ≈ 11.3

Validation uses long-horizon folds — train→(2011→2012-18), (2008→2009-15), (2005→2006-12), each a 7-year-ahead horizon like the 2022–2025 final scoring — and short-horizon folds (1–3 yr) that mirror the 2019–2021 live scoring. Every metric is positive in every fold. For calibration, on the same folds a 10-year-mean scores 6.6, an equal-weight linear trend scores 0.0 (it wrecks precipitation and hottest-day), and the 30-year climatology scores 0 by definition — our ensemble beats all of them.

---

Method

The pipeline is a small, auditable set of components built over three rounds of a competitive "model battle": many independent forecasters are each judged by the shared, leakage-proof harness, and a horizon-aware ensemble selects per-metric winners.

1. Leakage-proof evaluation harness (code/src/harness.py). The single source of truth. It builds tidy targets, computes the climatology baseline, and runs the rolling backtest. A model only ever receives train rows with year <= cutoff; forecasting any future year can physically never see it. Climatology-as-a-model scores exactly 0.000 skill in the harness — our proof the evaluator is unbiased.

2. Competing base forecasters (code/experiments*/). Deliberately diverse, each a distinct hypothesis about where signal lives:

Model	Idea
hier-shrinkage-trend	Per-city robust (Theil–Sen) trends shrunk toward pooled regional/global slopes (James–Stein credibility).
state-space-local-trend	Per-series structural time series (local-linear / damped trend, statsmodels UnobservedComponents).
damped-trend-ets	Damped-trend exponential smoothing with CV-tuned damping + robust bias correction.
spatial-analog-knn	Blends each city's trend with its k nearest neighbours by lat/lon to borrow strength spatially.
recency-weighted-trend	Exponentially recency-weighted least-squares trend — captures accelerating warming.
two-rate-accel-trend	Recent-decade slope vs long-term slope; extrapolates the steeper recent rate, shrunk for stability.
horizon-aware-blend	Separate climatology/recency/trend blend weights for near-term (h≤3) vs far-term (h≥4) horizons.
calibrated-trend	Robust trend with a cross-validated multiplicative-gain + offset bias correction.
cdd-from-temp	Physically couples CDD to the (stronger) temperature forecast rather than modelling it blind.
per-city-selector	Cross-validates a menu {climatology, 10/15-yr mean, Theil–Sen, damped} per city and picks the local winner.

(Two models — a gradient-boosting model and a window-optimizer — are included for transparency but were excluded from the final ensemble because they did not survive out-of-sample verification.)

3. Horizon-aware per-metric ensemble + robustness clamp (code/src/ensemble_final.py, the champion). For temperature and CDD we cross-validate a combiner per (metric, horizon bucket): the near-term bucket (h≤3) is selected to maximize short-horizon skill (the live board), the far-term bucket (h≥4) to maximize long-horizon skill (the final). Candidate combiners are each model plus their mean/median; the climatology floor applies if a bucket's CV skill is ≤ 0. Precipitation and hottest-day pass through at climatology. A final robustness clamp bounds each prediction's deviation from the city's recent-window mean and falls back to climatology for stations with too little recent history — this removes pathological trend blow-ups at a handful of gap stations (one city's extrapolation dropped from +11 °F to its climatology) at negligible cost to aggregate skill.

The result: temperature and CDD lean on a horizon-tuned stack of trend models (skill 0.155 → 0.181 and 0.171 → 0.200 across the three rounds), while precipitation and hottest-day stay at climatology. The data chose, not us. Earlier rounds (ensemble.py, ensemble_v2.py, ensemble_v3.py) are kept for transparency; reproduce.py runs the champion.

---

Reproducibility & openness (rubric: reproducibility/openness, interoperability)

• One command, from nothing: python code/reproduce.py downloads the public training dataset, runs the full backtest, and regenerates submission.csv. We verified in a clean environment (with the output deleted first) that it is byte-identical to the committed file.
• Deterministic — no RNG seeds to worry about; the method is closed-form / CV-selected.
• Pinned environment in code/requirements.txt; only mainstream open-source libraries (numpy, pandas, scikit-learn, statsmodels, scipy, huggingface_hub). No private data, no API keys, no GPU.
• Standard, interoperable formats throughout: HF Datasets + Parquet in, long-format CSV out.
• Honest validation note: a few global calibration parameters (per-metric trend gain, blend weights) are selected by cross-fold CV — standard practice — so the reported backtest skill carries a small in-sample optimism; the per-horizon split and the climatology floor keep this conservative, and no post-2018 data is ever used.

Reuse & extensibility (rubric: reuse/extensibility)

The harness defines one contract:

def predict_fn(train_targets, cutoff, years) -> DataFrame[city_id, year, metric, predicted_value]
H.backtest(predict_fn, H.DEFAULT_FOLDS)   # -> per-metric + overall skill vs climatology

Anyone can drop in a new model — write a predict_fn, run backtest, and it is directly comparable to all of ours. The harness is a reusable, leakage-proof benchmark for annual climate forecasting, usable well beyond this challenge (any GHCN-style city-year panel). The per-metric-floor ensemble is model-agnostic: add forecasters and it selects among them automatically — which is exactly how we grew across three rounds without ever touching the evaluator.

Feedback to the ecosystem (rubric: feedback to the ecosystem)

While building, we found and documented concrete data-quality issues in CMU-OSPO/protostar-climate-train-to-2018, handled in harness.py and flagged here for the maintainers:

1. -inf sentinel in Max Tmax (hottest-day-of-year) for ~1,184 station-years with no valid daily max — these slip past dropna() and silently poison any mean/MAE. We coerce ±inf → NaN at load. Suggestion: store these as null in the published Parquet.
2. Multi-decade temperature gaps at some stations (e.g. USW00003856 has a single temperature observation in 2009–2018), which break a strict 1989–2018 climatology window and cause naive trend models to extrapolate nonsensically. We use a gap-safe "most-recent-30-observed-years" climatology and a recent-history guard in the ensemble.

We offer the harness itself back as a reusable evaluation scaffold for future editions of the track.

---

Scientific & societal value — a climate-adaptation policy grounded in these forecasts

Our 2019–2025 forecasts describe where U.S. urban climate risk is heading, and cooling-degree-days (CDD) is the single most predictable risk signal we found (highest skill, +0.20). CDD maps almost one-to-one onto summer electricity demand and heat exposure, making it directly actionable for policy.

What the forecasts say. Across the 210 cities, 209 are projected to keep warming (mean +1.76 °F vs the 1989–2018 normal; one gap station is held at climatology), with mean cooling demand up +224 CDD·days. The increase is sharply uneven:

• Fastest warming: the arid interior West — Elko NV (+3.0 °F), Reno NV (+2.8 °F), Bakersfield CA (+2.7 °F), Salt Lake City UT (+2.6 °F) — plus continental-interior cities like Des Moines IA. Heat compounds there with water stress.
• Largest cooling-demand growth: Southern California, the Mojave, and the Gulf Coast — Bakersfield (+532 CDD), Needles (+512), Escondido (+503), San Diego (+466), Tampa (+448).

Recommended adaptation strategy. Because CDD is both rising and forecastable, utilities and city governments can plan against it rather than react:

1. Target hardening where demand grows fastest. Prioritize grid-capacity upgrades, heat-pump/efficiency retrofit subsidies, and cool-roof/urban-canopy programs in the top-CDD-growth cities above — not uniformly nationwide.
2. Site public cooling infrastructure from the forecast. Use projected per-city CDD to place and staff cooling centers and to trigger heat-health early-warning thresholds — with special attention to Southern California's inland valleys and the Gulf Coast (Tampa), where humidity pushes heat-index risk past raw temperature.
3. Couple heat and water adaptation in the interior West. Reno/Elko/Salt Lake City-type cities need combined heat and drought planning; CDD-driven peak cooling load should inform renewable + storage buildout there first.
4. Fund proportional to projected change, and revisit annually. The per-city skill numbers tell policymakers how much to trust each forecast — invest with confidence where skill is high (temperature, CDD), hedge where it is low (precipitation).

The honest flip side is a scientific finding in itself: annual precipitation and single-day heat extremes remain near-unpredictable at this horizon (skill ≈ 0.03–0.04), independently confirmed by the live leaderboard where no team beats the climatology baseline on either. Adaptation for flood and extreme-heat-event risk should therefore rely on robustness and insurance, not point forecasts — a conclusion our climatology-floor design makes explicit rather than hiding behind false precision.

---

Honor code / anti-leakage

Every forecast for year Y is produced using only observations through the cutoff (≤ 2018 for the real submission). No post-2018 weather or weather-responsive data is used anywhere — including for model selection. The harness enforces this by construction (it never passes year > cutoff rows into training), and we re-audited every model for it.

Repository map

submission.csv              the forecast (5,880 rows)
README.md                   this file
data/city_ids.txt           the 210 official city ids
code/
  reproduce.py              one-command reproduction (runs the champion)
  requirements.txt          pinned environment
  src/harness.py            leakage-proof targets + climatology + rolling backtest + validator
  src/ensemble_final.py     champion: horizon-aware per-metric selection + robustness clamp
  src/ensemble_v3.py        horizon-aware ensemble (pre-clamp)
  src/ensemble_v2.py        round-2 ensemble
  src/ensemble.py           round-1 ensemble
  experiments/*/model.py    round-1 base forecasters
  experiments2/*/model.py   round-2 temperature/CDD forecasters
  experiments3/*/model.py   round-3 horizon-aware / calibrated / coupled forecasters

License

MIT.