Datasets:
city_id stringclasses 210
values | year int64 2.02k 2.03k | metric stringclasses 4
values | predicted_value float64 2.46 5.25k |
|---|---|---|---|
USC00042863 | 2,019 | annual_avg_temp | 66.638 |
USC00042863 | 2,019 | total_precip_in | 13.505 |
USC00042863 | 2,019 | hottest_day_temp | 104.031 |
USC00042863 | 2,019 | cdd | 1,695.252 |
USC00042863 | 2,020 | annual_avg_temp | 66.681 |
USC00042863 | 2,020 | total_precip_in | 13.475 |
USC00042863 | 2,020 | hottest_day_temp | 104.047 |
USC00042863 | 2,020 | cdd | 1,707.236 |
USC00042863 | 2,021 | annual_avg_temp | 66.725 |
USC00042863 | 2,021 | total_precip_in | 13.444 |
USC00042863 | 2,021 | hottest_day_temp | 104.064 |
USC00042863 | 2,021 | cdd | 1,719.275 |
USC00042863 | 2,022 | annual_avg_temp | 66.768 |
USC00042863 | 2,022 | total_precip_in | 13.414 |
USC00042863 | 2,022 | hottest_day_temp | 104.08 |
USC00042863 | 2,022 | cdd | 1,731.363 |
USC00042863 | 2,023 | annual_avg_temp | 66.811 |
USC00042863 | 2,023 | total_precip_in | 13.383 |
USC00042863 | 2,023 | hottest_day_temp | 104.096 |
USC00042863 | 2,023 | cdd | 1,743.509 |
USC00042863 | 2,024 | annual_avg_temp | 66.854 |
USC00042863 | 2,024 | total_precip_in | 13.353 |
USC00042863 | 2,024 | hottest_day_temp | 104.113 |
USC00042863 | 2,024 | cdd | 1,755.702 |
USC00042863 | 2,025 | annual_avg_temp | 66.897 |
USC00042863 | 2,025 | total_precip_in | 13.322 |
USC00042863 | 2,025 | hottest_day_temp | 104.129 |
USC00042863 | 2,025 | cdd | 1,767.954 |
USC00166584 | 2,019 | annual_avg_temp | 67.628 |
USC00166584 | 2,019 | total_precip_in | 55.488 |
USC00166584 | 2,019 | hottest_day_temp | 101.085 |
USC00166584 | 2,019 | cdd | 2,810.467 |
USC00166584 | 2,020 | annual_avg_temp | 67.704 |
USC00166584 | 2,020 | total_precip_in | 55.512 |
USC00166584 | 2,020 | hottest_day_temp | 101.096 |
USC00166584 | 2,020 | cdd | 2,822.995 |
USC00166584 | 2,021 | annual_avg_temp | 67.771 |
USC00166584 | 2,021 | total_precip_in | 55.535 |
USC00166584 | 2,021 | hottest_day_temp | 101.106 |
USC00166584 | 2,021 | cdd | 2,835.532 |
USC00166584 | 2,022 | annual_avg_temp | 67.833 |
USC00166584 | 2,022 | total_precip_in | 55.559 |
USC00166584 | 2,022 | hottest_day_temp | 101.116 |
USC00166584 | 2,022 | cdd | 2,848.078 |
USC00166584 | 2,023 | annual_avg_temp | 67.892 |
USC00166584 | 2,023 | total_precip_in | 55.583 |
USC00166584 | 2,023 | hottest_day_temp | 101.127 |
USC00166584 | 2,023 | cdd | 2,860.639 |
USC00166584 | 2,024 | annual_avg_temp | 67.949 |
USC00166584 | 2,024 | total_precip_in | 55.606 |
USC00166584 | 2,024 | hottest_day_temp | 101.137 |
USC00166584 | 2,024 | cdd | 2,873.212 |
USC00166584 | 2,025 | annual_avg_temp | 68.004 |
USC00166584 | 2,025 | total_precip_in | 55.63 |
USC00166584 | 2,025 | hottest_day_temp | 101.147 |
USC00166584 | 2,025 | cdd | 2,885.799 |
USC00280734 | 2,019 | annual_avg_temp | 51.456 |
USC00280734 | 2,019 | total_precip_in | 46.588 |
USC00280734 | 2,019 | hottest_day_temp | 94.729 |
USC00280734 | 2,019 | cdd | 784.459 |
USC00280734 | 2,020 | annual_avg_temp | 51.497 |
USC00280734 | 2,020 | total_precip_in | 46.618 |
USC00280734 | 2,020 | hottest_day_temp | 94.727 |
USC00280734 | 2,020 | cdd | 790.761 |
USC00280734 | 2,021 | annual_avg_temp | 51.538 |
USC00280734 | 2,021 | total_precip_in | 46.648 |
USC00280734 | 2,021 | hottest_day_temp | 94.725 |
USC00280734 | 2,021 | cdd | 797.064 |
USC00280734 | 2,022 | annual_avg_temp | 51.579 |
USC00280734 | 2,022 | total_precip_in | 46.678 |
USC00280734 | 2,022 | hottest_day_temp | 94.724 |
USC00280734 | 2,022 | cdd | 803.366 |
USC00280734 | 2,023 | annual_avg_temp | 51.62 |
USC00280734 | 2,023 | total_precip_in | 46.708 |
USC00280734 | 2,023 | hottest_day_temp | 94.722 |
USC00280734 | 2,023 | cdd | 809.669 |
USC00280734 | 2,024 | annual_avg_temp | 51.661 |
USC00280734 | 2,024 | total_precip_in | 46.738 |
USC00280734 | 2,024 | hottest_day_temp | 94.72 |
USC00280734 | 2,024 | cdd | 815.971 |
USC00280734 | 2,025 | annual_avg_temp | 51.703 |
USC00280734 | 2,025 | total_precip_in | 46.768 |
USC00280734 | 2,025 | hottest_day_temp | 94.718 |
USC00280734 | 2,025 | cdd | 822.288 |
USC00286055 | 2,019 | annual_avg_temp | 54.689 |
USC00286055 | 2,019 | total_precip_in | 49.982 |
USC00286055 | 2,019 | hottest_day_temp | 97.491 |
USC00286055 | 2,019 | cdd | 1,155.132 |
USC00286055 | 2,020 | annual_avg_temp | 54.752 |
USC00286055 | 2,020 | total_precip_in | 50.016 |
USC00286055 | 2,020 | hottest_day_temp | 97.49 |
USC00286055 | 2,020 | cdd | 1,162.473 |
USC00286055 | 2,021 | annual_avg_temp | 54.813 |
USC00286055 | 2,021 | total_precip_in | 50.049 |
USC00286055 | 2,021 | hottest_day_temp | 97.49 |
USC00286055 | 2,021 | cdd | 1,169.822 |
USC00286055 | 2,022 | annual_avg_temp | 54.872 |
USC00286055 | 2,022 | total_precip_in | 50.083 |
USC00286055 | 2,022 | hottest_day_temp | 97.489 |
USC00286055 | 2,022 | cdd | 1,177.181 |
Ceviche: Protostar Climate Challenge Submission
Negative result: no
Forecasts of yearly weather for 210 U.S. cities, 2019β2025, made as of the end of 2018 (no data that reveals 2019β2025 weather), for the Protostar Challenge: Climate Risk Intelligence.
In one line: each city's recent level, plus a warming rate pooled across cities, with CDD computed physically by warming each city's own daily weather instead of by extrapolating the CDD series; for average temperature, blended 75/25 with a pooled-trend ARIMA.
Team
| Name | Hugging Face username | University, Program |
|---|---|---|
| Arturo Arias Aguilar | @aaarias | Carnegie Mellon University, Heinz College MISM-BIDA |
| Alejandro Otero | @aotero22 | Carnegie Mellon University, Heinz College MSISPM |
| Joaquin Delgado | @Joawhi | Carnegie Mellon University, Heinz College MISM |
| Sebastian Leon | @sleonca | Carnegie Mellon University, Heinz College MISM |
Method
Submitted model: the "ARIMA ensemble B" (code/04_forecast_submission.py, team decision of 2026-11-08). It is our base model (code/final_model.py, table below) for precipitation, hottest day and CDD. For average temperature it is 0.75 Γ base model + 0.25 Γ pooled-trend ARIMA: an ARMA fitted to each city's 1949β2018 series after removing the warming rate pooled over all cities, with that pooled trend added back. It was proposed on the pre-2019 backtest (a pre-set rule: better mean Score and better 2011 fold) and confirmed on 2019β2021 (see Model selection).
We found that, five to seven years ahead, almost all predictable signal is (a) where each city sits now and (b) how fast the climate warms. Year-to-year anomalies barely persist: adding a decaying share of the last year's anomaly gained under 1 % in year 1 and nothing after (code/experiments/persistence.py). So the model is about estimating those two quantities with as little noise as possible.
| Metric | Base model (code/final_model.py; in the submission, temperature is blended 75/25 with the ARIMA) |
|---|---|
annual_avg_temp |
Average of two forecasters: (1) mean of the last 15 valid years, extended at 1.25 Γ the warming rate pooled over all 210 cities (mean of city OLS slopes, 1969β2018, β 0.46 Β°F/decade β 0.58); (2) a regional-anomaly model: city 30-year mean + the city's offset from its 10 nearest neighbours + the neighbours' common anomaly trend. Pooling cancels most of the shared year-to-year noise that ruins single-city trend lines. |
cdd |
Shifted-daily-weather model. For each target year y, every valid day of the last N years (N = 10 and 20, averaged) is warmed by rate Γ (y β t) and CDD is recomputed from the shifted days (code/01_cdd_shift_table.py precomputes CDD for shifts of β2β¦+8 Β°F). This keeps CDD's nonlinear response to warming (more days cross 65 Β°F; humid vs. dry climates respond differently) that a linear CDD trend misses. |
hottest_day_temp |
Mean of the last 30 years + half of the regional (20 nearest cities) 70-year trend. |
total_precip_in |
Mean of the last 50 years + half of the regional 50-year trend. |
Rows follow the official validity rule throughout: years with > 10 missing days of the values a metric uses are excluded from training and scoring.
Considered and discarded: one model per metric (results/forecasts_frozen/per_metric.csv): ensemble B for temperature, a hierarchical MixedLM for precipitation, the base model for hottest day, and 0.8 Γ base + 0.2 Γ ETS for CDD. It ranked first on 2019β2021, but the team rejected it because its two changes beyond ensemble B did not hold up before 2019. The precipitation model had a negative backtest skill (S_m β0.015, vs +0.001 for the base model) and won only on a +0.02 S_m edge in 2019β2021. The CDD weights failed the out-of-sample check (fitted on 1990β2006, they score 0.231 on the 2011 fold vs 0.245 for the base model). On the backtest the per-metric proposal scores 5.43, below the base model (5.65) and ensemble B (5.78). Details in Model selection.
How we checked it: a rolling-origin backtest that never touches 2019+ data (code/02_backtest.py). For origins 1990, 1994, 1998, 2002, 2006 and 2011, every model sees data up to the origin only and forecasts the next 7 years, scored exactly like the challenge (MAE vs. the 30-year-average baseline, same 10-day and 10-valid-year rules). Our harness reproduces the organizers' practice run to the decimal (train β€ 2010, CDD MAE 211.7 / 174.6 Β°F-days for 30-year average / linear trend; official 212 / 175). All model designs and parameters were chosen on these folds. The final choice between our candidates also looked at the 2019β2021 score, as the live leaderboard allows (see Model selection); 2022β2025 was not used.
Composite Score (0β100) by forecast origin:
| Model | 1990 | 1994 | 1998 | 2002 | 2006 | 2011 | mean |
|---|---|---|---|---|---|---|---|
| Last year repeated | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| 30-year average | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| Linear trend | 0 | 0 | 0 | 0 | 0.2 | 2.3 | 0.4 |
| Base model | 3.0 | 3.9 | 5.8 | 4.5 | 5.8 | 10.9 | 5.6 |
| Submitted (ensemble B) | 3.2 | 4.0 | 6.3 | 4.6 | 5.5 | 11.0 | 5.8 |
MAE on the most recent fold (train β€ 2011, forecast 2012β2018):
| Model | Avg temp (Β°F) | CDD (Β°F-days) | Hottest day (Β°F) | Precip (in) |
|---|---|---|---|---|
| 30-year average | 1.71 | 201.3 | 2.42 | 5.98 |
| Linear trend | 1.49 | 165.7 | 2.61 | 6.79 |
| Ours | 1.41 | 152.0 | 2.39 | 5.98 |
Mean skill over folds, S_m = 1 β MAE/MAE_30yr: temperature 0.123, CDD 0.097, hottest day 0.005, precipitation 0.001. As the organizers predicted, hottest day and precipitation are essentially unforecastable beyond climatology; our gain there is only in not losing to it (the per-city linear trend loses 8β10 %).
What did not help (all on pre-2019 folds): per-city trend lines (too noisy); restricting each city's level to its current station after a relocation (Β± 0.001; the joined records look consistent); last-year anomaly persistence (< 1 % in year 1, β 0 after); medians instead of means (slightly worse).
Why it could be used for real: the method is transparent (two numbers per city: a level and a warming rate), cheap, needs only the city's own daily record, and the CDD component maps directly to cooling-energy planning.
ARIMA experiments
Following the course workflow (CMU 95-835: EDA β ADF/KPSS β differencing β ACF/PACF β fit β Ljung-Box β backtest), we fitted one ARIMA per city and metric (statsmodels 0.15.0, pmdarima 2.1.1) on the same pre-2019 folds. Full report: results/arima_report.md.
- Tried: fixed orders from ACF/PACF, auto-ARIMA (d by KPSS or ADF, stepwise AICc), d=0 + linear trend, ARMA around the pooled warming rate, ARIMA temperature β CDD through shifted daily weather, 40/50/70-year windows, and monthly SARIMA (m = 12, 20 cities).
- Stationarity: 63 % of temperature series are trend-stationary and only 2 % have a unit root, so d=1 over-differences (lag-1 ACF of Ξy β β0.45). auto-ARIMA with ADF differenced 82 % of temperature series and scored worst.
- Result: no ARIMA beats our model. Best single ARIMA (
pooled_auto): mean Score 2.7 vs 5.6 (6.3 vs 10.9 at origin 2011); best ARIMA per metric: 3.86. ARIMA under-forecasts warming (bias β0.3 Β°F vs β0.05 Β°F), and for precipitation and hottest day every ARIMA has S_m β€ 0. SARIMA: 3.1 vs 7.7 on its 20 cities. - Why: after the trend, annual anomalies are nearly white noise (Ljung-Box), so ARMA terms help only at 1 year ahead; what matters at 5β7 years is the current level and the warming rate.
- Ensemble: 0.75 Γ our model + 0.25 Γ pooled-trend ARIMA for temperature passed our pre-set rule (better mean Score and 2011 fold): 5.78 vs 5.65, 11.03 vs 10.88. The gain is small and the ensemble was picked from 12 options, so it is kept as a candidate in
results/submission_arima.csv;submission.csvis unchanged.
Post-hoc evaluation (frozen forecasts, SHA-256 in results/forecasts_frozen/MANIFEST.csv)
The organizers clarified at the Hugging Face meeting of 7 November 2026 that forecasts may be compared with the observed 2019β2025 weather as long as those observations are not used to fit or retrain any model (the FAQ already allowed scoring a frozen, declared forecast). We first froze 15 candidate forecasts (origin 2018, SHA-256 recorded) and only then downloaded NOAA GHCN-Daily 2019β2025 into analysis_posthoc/, which no script in code/ reads. We recomputed the metrics with the challenge definitions (validated against the official 2018 values: 94.7β100 % within a small tolerance), re-verified every hash and scored them with the official formula. Details: analysis_posthoc/README.md.
| Frozen forecast | Backtest mean | Live 2019β2021 | Final 2022β2025 |
|---|---|---|---|
ARIMA ensemble B (now submission.csv) |
5.78 | 14.89 | 18.77 |
Base model (former submission.csv) |
5.65 | 14.38 | 18.63 |
Best single ARIMA (pooled_auto@1949) |
2.49 | 12.70 | 15.65 |
| Linear trend / 30-year average | 0.42 / 0 | 9.46 / 0 | 12.93 / 0 |
The pre-2019 backtest ranked the models correctly (Spearman 0.92 over 15 forecasts) but understated the level: 2019β2025 ran 1.44 Β°F above the 1989β2018 average, which rewards any warming term. No forecast or model choice was changed after this comparison.
Note: this first post-hoc step (15 candidates) did compute 2022β2025 once. Those numbers were not used to choose the submission; the full 2022β2025 comparison of all candidates, computed only after the decision, is in Post-hoc, not used for selection.
Model selection
Decision (team, 2026-11-08): submit the ARIMA ensemble B. submission.csv = results/forecasts_frozen/arima_ensemble_B.csv (SHA-256 7b90fecfβ¦, as in the MANIFEST), reproduced byte for byte by bash code/run_all.sh (code/04_forecast_submission.py).
Ensemble B was proposed on the pre-2019 backtest alone. It is the only candidate that beats the base model on all three checks: backtest mean Score 5.78 vs 5.65, 2011 fold 11.03 vs 10.88, and 2019β2021 15.1 vs 14.6.
The per-metric proposal below scored higher on 2019β2021 (15.8). The team discarded it because its two extra changes are not supported before 2019: the hierarchical precipitation model has negative backtest skill, and the CDD weights fail the 1990β2006 β 2011 out-of-sample check. Without them, the proposal reduces to ensemble B.
Basis. The organizers clarified at the Hugging Face meeting of 7 November 2026 that the observed weather may be used for comparison as long as it is not used to fit or retrain a model. The live leaderboard reports each submission's 2019β2021 score, which teams may use when choosing what to submit. We therefore ran a per-metric selection using the pre-2019 backtest and the 2019β2021 score (below), and the team made the final choice with both in view. Every forecast is still made only from data up to 2018-12-31, no model is fitted on 2019+ data, and 2022β2025 is not used.
Honest note. Unlike the original submission, whose every choice came from pre-2019 folds, this selection does use 2019β2021 observations (as the leaderboard allows); ensemble B itself had been proposed before any 2019+ data was downloaded. Its 2019β2021 score is therefore optimistic as an estimate of future skill. On the pre-2019 backtest alone the proposal scores 5.43, slightly below the current model's 5.65; the precipitation choice causes all of the difference.
Rule (written in results/arima_report.md before it was applied; run by analysis_posthoc/select_per_metric.py):
for each metric, rank every frozen candidate by its mean backtest S_m (6 origins) and by its 2019β2021 S_m (denominator = the leaderboard's 30-year-average MAE). Combined score = 0.5 Γ backtest rank + 0.5 Γ 2019β2021 rank, lowest wins, ties go to the simpler model, and identical forecasts are merged first. A winner on 2019β2021 that sits in the lower half of the backtest is flagged as possible overfitting (no flag was triggered).
Candidates: the current model, the 3 reference models, 12 ARIMA variants, the ARIMA ensembles A and B, and the new models of code/models_extra.py (ETS Holt / damped / pooled-trend, a hierarchical MixedLM, Tmax and Tmin forecast separately) with a per-metric weighted ensemble. 24 frozen files in total, all SHA-256 verified.
| Metric | Winner | S_m backtest | S_m 2011 | S_m 2019β21 | Runner-up (combined score) | Current model: bt / 2019β21 |
|---|---|---|---|---|---|---|
| Avg temp | ARIMA ensemble B | 0.129 | 0.183 | 0.304 | weighted ensemble (tie 2.5, less simple) | 0.124 / 0.284 |
| Precip | hierarchical | β0.015 | 0.005 | 0.017 | ETS + pooled trend (4.5) | 0.001 / β0.003 |
| Hottest day | current model | 0.005 | 0.013 | 0.011 | weighted ensemble (2.0) | β |
| CDD | weighted ensemble | 0.099 | 0.237 | 0.300 | current model (2.0) | 0.097 / 0.293 |
| Forecast | Backtest Score mean | Backtest 2011 | 2019β2021 Score (our estimate) | Live-leaderboard position (2026-11-07) |
|---|---|---|---|---|
| Per-metric proposal (discarded) | 5.43 | 10.93 | 15.8 | 3rd (1st 16.0, 2nd 15.9) |
| ARIMA ensemble B (submitted) | 5.78 | 11.03 | 15.1 | 4th |
Base model (former submission.csv) |
5.65 | 10.88 | 14.6 | 5th |
Our 2019β2021 estimates are about 0.3 points optimistic: our recomputed baseline MAEs are 0.1β0.4 % below the leaderboard's, so the proposal would land near 15.5 and ensemble B near 14.9. The official Score is the one the Space returns. Full tables: analysis_posthoc/results/selection_per_metric.csv, analysis_posthoc/leaderboard_comparison.csv, and results/arima_report.md (sections "Regla de selecciΓ³n por mΓ©trica" and "Resultado de la selecciΓ³n").
Post-hoc, not used for selection
Computed after the team replaced submission.csv (2026-11-08) and not used for any choice; submission.csv will not be changed whatever these numbers say. Every frozen candidate (23 forecasts, SHA-256 verified against the MANIFEST) is scored on the observed 2022β2025 weather (NOAA GHCN-Daily, analysis_posthoc/posthoc_final.py), with the official formula. There is no leaderboard for 2022β2025, so the denominator is our recomputed 30-year-average MAE. For 2019β2021 the leaderboard's denominator is used, as in the selection (about 0.25 points higher than our own denominator would give). Basis: the organizers' clarification at the Hugging Face meeting of 7 November 2026 (comparison allowed; no fitting on observations).
| Frozen forecast | Backtest (rank) | 2019β2021 (rank) | 2022β2025 (rank) |
|---|---|---|---|
| Weighted ensemble | 5.91 (1) | 15.00 (3) | 18.97 (1) |
| Tmax/Tmin (1.25 Γ pooled rate; others = base) | 5.45 (4) | 14.70 (4) | 18.78 (2) |
| ARIMA ensemble B (submitted) | 5.78 (2) | 15.14 (2) | 18.77 (3) |
| Base model (former submission) | 5.65 (3) | 14.64 (5) | 18.63 (4) |
| Tmax/Tmin (1.0 Γ pooled rate) | 5.38 (6) | 14.17 (6) | 18.07 (5) |
| Per-metric proposal (discarded) | 5.43 (5) | 15.80 (1) | 17.68 (6) |
| hierarchical | 3.37 (8) | 12.85 (9) | 17.10 (7) |
| ARIMA A (best ARIMA per metric) | 3.86 (7) | 11.67 (11) | 16.30 (8) |
| ets_pooled | 3.03 (9) | 13.09 (7) | 15.87 (9) |
| pooled_auto@1949 | 2.49 (13) | 12.97 (8) | 15.65 (10) |
| pooled_auto | 2.73 (10) | 10.96 (17) | 13.38 (11) |
| arima_100_ct | 2.56 (11) | 10.99 (16) | 13.16 (12) |
| arima_101_ct | 2.55 (12) | 10.61 (18) | 13.11 (13) |
| auto_d0_ct | 2.43 (14) | 10.06 (20) | 13.09 (14) |
| linear_trend | 0.42 (20) | 9.73 (21) | 12.93 (15) |
| arima_111_drift | 2.18 (16) | 11.41 (13) | 12.35 (16) |
| ets_damped | 2.07 (17) | 11.26 (15) | 12.25 (17) |
| arima_011_drift | 2.23 (15) | 12.14 (10) | 12.19 (18) |
| ets_holt | 1.59 (18) | 11.44 (12) | 12.06 (19) |
| auto_kpss | 1.21 (19) | 10.45 (19) | 11.07 (20) |
| auto_adf | 0.11 (21) | 11.34 (14) | 8.00 (21) |
| avg_30yr | 0.00 (22) | 0.28 (22) | 0.00 (22) |
| last_year | 0.00 (23) | 0.00 (23) | 0.00 (23) |
Which predicted 2022β2025 better, the backtest or 2019β2021? The pre-2019 backtest did.
| Spearman rank correlation with 2022β2025 | Backtest | 2019β2021 |
|---|---|---|
| Composite Score, all 23 candidates | 0.97 | 0.81 |
| Composite Score, without the two zero-score references | 0.96 | 0.75 |
| S_m avg temp | 0.69 | 0.69 |
| S_m precipitation | 0.94 | 0.19 |
| S_m hottest day | 0.43 | 0.11 |
| S_m CDD | 0.68 | 0.82 |
- The backtest's best candidate was also best on 2022β2025. The weighted ensemble ranked 1st on the backtest (5.91) and 1st on 2022β2025 (18.97), but only 3rd on 2019β2021. The 2019β2021 favourite, the per-metric proposal, fell to 6th (17.68), below the submitted ensemble B (3rd, 18.77) and the base model (4th, 18.63).
- The two discarded changes are exactly what hurt the per-metric proposal. Its precipitation model (hierarchical) has S_m β0.061 on 2022β2025 vs β0.025 for the base model, and its CDD blend 0.380 vs 0.387. This matches their negative backtest skill and the failed out-of-sample check. 2019β2021 was misleading for precipitation (rank correlation 0.19; backtest 0.94) and hottest day (0.11 vs 0.43). Three years of a noisy metric are too few to rank models.
- 2019β2021 was more informative only for CDD (0.82 vs 0.68), where the warm recent years carry real signal. For temperature both were equally informative (0.69).
- The submitted ensemble B keeps its small edge over the base model (18.77 vs 18.63), entirely from temperature (S_m 0.372 vs 0.367). The weighted ensemble (18.97) and Tmax/Tmin (18.78) did marginally better on 2022β2025. This is not used for any decision.
- All Scores are much higher than in the backtest, because 2019β2025 was far warmer than the 1989β2018 baseline. The leading models still under-forecast temperature (bias β0.5 to β0.7 Β°F on 2022β2025).
Files: analysis_posthoc/results/posthoc_final_scores.csv, posthoc_final_by_metric.csv, posthoc_final_predictiveness.csv.
AI Tools and Agents
- Tools and models: Claude Code (Claude Opus 5.5), running as an agent in the Claude desktop app.
- What they did: downloaded the challenge data and template, wrote all code in
code/(data loading, backtest harness, models, forecast script), ran the backtests and parameter searches, and drafted this README and the data card. The agent ran with the template'scode/AGENTS.mdas its instructions file (copied toCLAUDE.md), downloaded only the challenge dataset (which ends 2018-12-31), and did not search the web or use any weather data or values from 2019 onward to build or fit a model (the final choice between frozen candidates also looked at their 2019β2021 score, as the leaderboard allows; see Model selection). Everypredicted_valueis produced bycode/04_forecast_submission.py(on top ofcode/final_model.py); none came from the LLM.code/common.pystops if any weather input is dated after 2018. - ARIMA experiments: Claude Code (Claude Opus 5.5) also wrote and ran the ARIMA study (
code/arima_*.py,code/models_arima.py) from a step-by-step prompt by the team. It worked phase by phase and stopped after each phase for a team member's review. It used only the challenge data (β€ 2018), kept the post-2018 guard incommon.py, made every model choice on the pre-2019 backtest, wrote its forecasts toresults/submission_arima.csvand did not modifysubmission.csv. It did not use git or upload anything. After freezing the candidate forecasts with SHA-256, and at the team's request (the Nov 7 clarification), it downloaded NOAA GHCN-Daily 2018β2025 intoanalysis_posthoc/for the post-hoc comparison only; no file incode/reads it and no forecast or model choice was changed afterwards. - New models and per-metric selection: at the team's request (a second step-by-step prompt), Claude Code wrote
code/models_extra.py(ETS, hierarchical MixedLM, Tmax/Tmin), backtested them (code/extra_backtest.py), froze their origin-2018 forecasts with SHA-256 (code/extra_freeze.py), scored all frozen forecasts on 2019β2021 only (analysis_posthoc/score_live.py), applied the pre-written per-metric rule (analysis_posthoc/select_per_metric.py) and wrotecode/03b_forecast_per_metric.py. Here 2019β2021 observations were used to choose models, as the leaderboard allows. They were never used to fit any model or forecast, and 2022β2025 was neither computed nor used. It downloaded no new data, did not use git and did not modifysubmission.csv; the replacement is left to the team. - Submission replacement and post-hoc scoring: on the team's decision, Claude Code copied the frozen ensemble B over
submission.csv(SHA-256 checked against the MANIFEST), wrotecode/04_forecast_submission.pyand updatedcode/run_all.sh. It then verified in a clean copy of the repository thatrun_all.sh(including the fresh dataset download) reproducessubmission.csvbyte for byte. Only after that did it score 2022β2025 for all frozen candidates (analysis_posthoc/posthoc_final.py);submission.csvwas not touched afterwards. - What a person reviewed: Sebastian Leon reviewed the model choices, the backtest results and the final forecasts.
- Agent instructions file:
code/AGENTS.md(alsoCLAUDE.mdat the root).
Policy Idea
- What risk did we notice? Cooling demand is rising fastest, in relative terms, where buildings were never designed for it. Our 2025 CDD forecast exceeds the 1961β1990 average by a median of 25 % across the 210 cities, but cooler, northern and western cities show the largest relative jumps: Reno about 3 times its 1961β1990 CDD, and Vancouver WA, Salem, Sault Ste. Marie, Seattle, Helena, Manchester NH and Missoula up 60β170 % (
results/cdd_change_1961_1990_to_2025.csv). Hot cities (Phoenix, Miami, Tucson) add the most degree-days in absolute terms, but they already have near-universal air conditioning. In the newly warm cities, many homes, rentals and schools have none, so the same heat does more harm. - What would we do about it, and who would do it? A "Cool-Ready Heat Pump" program run by electric utilities with state energy offices, targeted at cities whose CDD has risen by more than 50 %: on-bill-financed, rebated swaps of furnace or boiler for a cold-climate heat pump, prioritizing renters, older adults and schools. Cities pair it with a summer cooling standard for rental housing (a maximum indoor temperature, as many already have for winter heating) and with tree-canopy and cool-roof programs in the hottest neighbourhoods. The CDD shift model can be rerun each year to update which cities qualify.
- Why is it sustainable? One machine provides cooling and replaces fossil heating, at 2β4 times the efficiency of resistance or gas heat, so the region's emissions fall even as cooling is added. It avoids a wave of cheap, inefficient window units that would raise summer peak demand and electricity bills. On-bill financing repays from energy savings, and canopy and cool roofs cut cooling load without using any energy.
Feedback to the Ecosystem
- Practice-run numbers depended on an unstated rule. The 10-valid-year baseline rule changed the published CDD baselines (213/177 β 212/175). Fix: ship a tiny reference scorer (
score.py) with the dataset, so every team scores identically; ourcode/common.py::scoreandcode/backtest.pyreproduce the official numbers and could serve as a starting point. - CDD and average temperature use different day sets.
Avg Tempis missing in 1,193 city-years butcddin 1,206, because it is computed from Tmax and Tmin separately while CDD needs both on the same day. This is only discoverable by recomputing fromdaily. Fix: a one-line per-column definition in the dataset card ("days used: days with both Tmax and Tmin"). - The validity rule lives in the Space, not in the data. Fix: add boolean
valid_<metric>columns toannual_indices, so the 10-day rule is machine-actionable rather than re-implemented by every team. - Homogenization status is unclear. Records join several stations (
city_info,Station Segment), but the card does not say whether the joins were adjusted for offsets. We tested it indirectly (masking pre-relocation years changed backtest skill by < 0.002) and found no strong breaks, but users should not have to infer this. Fix: state it in the card and link the method from Lai & Dzombak (2019). - Versioning. The dataset changed on Nov 5 (subset added, typos fixed). Fix: tag releases (e.g.,
v1.1) and keep a changelog in the card, so a submission can cite an exact version. We record the commit we used:bbbdebdab2a18e00f97ef87bd3f35a71e222180b. - What worked well: a CC0 license, Parquet subsets that load in one line,
coverage_and_stations, and an agent instructions file in the template, which made the data rule easy to enforce.
Files
| File | Content |
|---|---|
submission.csv |
Our forecasts: city_id,year,metric,predicted_value |
code/00_download.py |
Downloads the challenge dataset to data/raw/ |
code/01_cdd_shift_table.py |
Precomputes CDD under daily temperature shifts β data/processed/shift_table.parquet |
code/02_backtest.py |
Pre-2019 rolling-origin backtest of our model and the 3 baselines β results/backtest.csv |
code/03_forecast.py |
Base model alone (data β€ 2018) β results/submission_current_model.csv |
code/04_forecast_submission.py |
Submitted ARIMA ensemble B (data β€ 2018) β submission.csv, checked against the MANIFEST |
code/common.py, models.py, final_model.py, backtest.py |
Data loading and validity rules, the models, the submitted configuration, the backtest harness |
code/experiments/ |
The parameter searches used to choose the configuration (pre-2019 folds only) |
results/ |
Backtest scores and the CDD-change table used in the policy idea |
code/arima_*.py, code/models_arima.py |
ARIMA experiments (optional; see code/README.md) |
code/models_extra.py, code/extra_*.py, code/candidates_2018.py, code/03b_forecast_per_metric.py |
New candidates (ETS, hierarchical, Tmax/Tmin, weighted ensemble) and the per-metric proposal β results/submission_per_metric.csv |
results/forecasts_frozen/ |
Frozen origin-2018 forecasts of every candidate, incl. per_metric.csv, with SHA-256 in MANIFEST.csv |
analysis_posthoc/ |
Post-hoc comparison with observed weather and the live leaderboard; per-metric selection (see its README) |
results/arima_report.md, results/arima_* |
ARIMA report, figures and tables; results/submission_arima.csv is the ARIMA-ensemble candidate (not submitted) |
data/ |
The curated data our code uses; described in its data card, data/README.md |
Metric codes: annual_avg_temp (Β°F), total_precip_in (inches), hottest_day_temp (Β°F), cdd (Β°F-days).
City IDs: the station ID from the city_info table of the training dataset (e.g., Pittsburgh = USW00094823).
How to Reproduce
pip install -r code/requirements.txt(Python 3.12).bash code/run_all.sh: downloads the data (or usesdata/raw/), builds the shift table, runs the backtest, writessubmission.csv(ARIMA ensemble B,code/04_forecast_submission.py, which also checks its SHA-256 againstresults/forecasts_frozen/MANIFEST.csv) and checks it. This takes a few minutes on a laptop (the ARIMA part of the submission fits one model per city). No agent or network access is needed beyond the dataset download.
Data
- External data: none. Forecasts use only the challenge dataset.
- Curated data: see the data card in
data/README.md(each file, its source, license and preparation).
Training data: CMU-OSPO/protostar-climate-train-to-2018 (commit bbbdebd), from:
- Lai, Y., & Dzombak, D. A. (2019). Compiled historical daily temperature and precipitation data for selected 210 U.S. cities. Carnegie Mellon University. https://doi.org/10.1184/R1/7890488.v6
- Lai, Y., & Dzombak, D. A. (2019). Historical changes of annual temperature and precipitation indices at selected 210 U.S. cities. Carnegie Mellon University. https://doi.org/10.1184/R1/7961012.v6
- Lai, Y., & Dzombak, D. A. (2019). Use of historical data to assess regional climate change. Journal of Climate, 32(14), 4299β4320. https://doi.org/10.1175/JCLI-D-18-0630.1
- Downloads last month
- 86