Datasets:
city_id stringclasses 210
values | year int64 2.02k 2.03k | metric stringclasses 4
values | predicted_value float64 2.95 5.3k |
|---|---|---|---|
USC00042863 | 2,019 | annual_avg_temp | 67.005266 |
USC00042863 | 2,019 | cdd | 1,724.347 |
USC00042863 | 2,019 | hottest_day_temp | 104.131 |
USC00042863 | 2,019 | total_precip_in | 12.682057 |
USC00042863 | 2,020 | annual_avg_temp | 67.053841 |
USC00042863 | 2,020 | cdd | 1,736.512911 |
USC00042863 | 2,020 | hottest_day_temp | 104.146 |
USC00042863 | 2,020 | total_precip_in | 12.726249 |
USC00042863 | 2,021 | annual_avg_temp | 67.100349 |
USC00042863 | 2,021 | cdd | 1,748.44 |
USC00042863 | 2,021 | hottest_day_temp | 104.162 |
USC00042863 | 2,021 | total_precip_in | 12.679109 |
USC00042863 | 2,022 | annual_avg_temp | 67.145767 |
USC00042863 | 2,022 | cdd | 1,760.457 |
USC00042863 | 2,022 | hottest_day_temp | 104.177 |
USC00042863 | 2,022 | total_precip_in | 12.629385 |
USC00042863 | 2,023 | annual_avg_temp | 67.192066 |
USC00042863 | 2,023 | cdd | 1,772.514 |
USC00042863 | 2,023 | hottest_day_temp | 104.192 |
USC00042863 | 2,023 | total_precip_in | 12.593481 |
USC00042863 | 2,024 | annual_avg_temp | 67.237315 |
USC00042863 | 2,024 | cdd | 1,784.687411 |
USC00042863 | 2,024 | hottest_day_temp | 104.208 |
USC00042863 | 2,024 | total_precip_in | 12.607741 |
USC00042863 | 2,025 | annual_avg_temp | 67.283338 |
USC00042863 | 2,025 | cdd | 1,796.824 |
USC00042863 | 2,025 | hottest_day_temp | 104.223 |
USC00042863 | 2,025 | total_precip_in | 12.685537 |
USC00166584 | 2,019 | annual_avg_temp | 68.14485 |
USC00166584 | 2,019 | cdd | 2,841.092 |
USC00166584 | 2,019 | hottest_day_temp | 101.188 |
USC00166584 | 2,019 | total_precip_in | 57.445047 |
USC00166584 | 2,020 | annual_avg_temp | 68.225069 |
USC00166584 | 2,020 | cdd | 2,855.380095 |
USC00166584 | 2,020 | hottest_day_temp | 101.197 |
USC00166584 | 2,020 | total_precip_in | 57.506875 |
USC00166584 | 2,021 | annual_avg_temp | 68.293573 |
USC00166584 | 2,021 | cdd | 2,866.18 |
USC00166584 | 2,021 | hottest_day_temp | 101.206 |
USC00166584 | 2,021 | total_precip_in | 57.475832 |
USC00166584 | 2,022 | annual_avg_temp | 68.356577 |
USC00166584 | 2,022 | cdd | 2,878.621 |
USC00166584 | 2,022 | hottest_day_temp | 101.215 |
USC00166584 | 2,022 | total_precip_in | 57.43969 |
USC00166584 | 2,023 | annual_avg_temp | 68.41723 |
USC00166584 | 2,023 | cdd | 2,891.075 |
USC00166584 | 2,023 | hottest_day_temp | 101.225 |
USC00166584 | 2,023 | total_precip_in | 57.486809 |
USC00166584 | 2,024 | annual_avg_temp | 68.474623 |
USC00166584 | 2,024 | cdd | 2,905.224589 |
USC00166584 | 2,024 | hottest_day_temp | 101.234 |
USC00166584 | 2,024 | total_precip_in | 57.596529 |
USC00166584 | 2,025 | annual_avg_temp | 68.529838 |
USC00166584 | 2,025 | cdd | 2,916.056 |
USC00166584 | 2,025 | hottest_day_temp | 101.243 |
USC00166584 | 2,025 | total_precip_in | 58.33044 |
USC00280734 | 2,019 | annual_avg_temp | 51.729377 |
USC00280734 | 2,019 | cdd | 792.226 |
USC00280734 | 2,019 | hottest_day_temp | 94.729 |
USC00280734 | 2,019 | total_precip_in | 45.68048 |
USC00280734 | 2,020 | annual_avg_temp | 51.775076 |
USC00280734 | 2,020 | cdd | 798.648 |
USC00280734 | 2,020 | hottest_day_temp | 94.727 |
USC00280734 | 2,020 | total_precip_in | 45.715767 |
USC00280734 | 2,021 | annual_avg_temp | 51.818743 |
USC00280734 | 2,021 | cdd | 804.953 |
USC00280734 | 2,021 | hottest_day_temp | 94.725 |
USC00280734 | 2,021 | total_precip_in | 45.688074 |
USC00280734 | 2,022 | annual_avg_temp | 51.862335 |
USC00280734 | 2,022 | cdd | 811.255 |
USC00280734 | 2,022 | hottest_day_temp | 94.724 |
USC00280734 | 2,022 | total_precip_in | 45.686183 |
USC00280734 | 2,023 | annual_avg_temp | 51.905827 |
USC00280734 | 2,023 | cdd | 817.558 |
USC00280734 | 2,023 | hottest_day_temp | 94.722 |
USC00280734 | 2,023 | total_precip_in | 45.663152 |
USC00280734 | 2,024 | annual_avg_temp | 51.949265 |
USC00280734 | 2,024 | cdd | 823.884 |
USC00280734 | 2,024 | hottest_day_temp | 94.72 |
USC00280734 | 2,024 | total_precip_in | 45.759765 |
USC00280734 | 2,025 | annual_avg_temp | 51.993544 |
USC00280734 | 2,025 | cdd | 830.226 |
USC00280734 | 2,025 | hottest_day_temp | 94.718 |
USC00280734 | 2,025 | total_precip_in | 46.832809 |
USC00286055 | 2,019 | annual_avg_temp | 55.058689 |
USC00286055 | 2,019 | cdd | 1,178.625 |
USC00286055 | 2,019 | hottest_day_temp | 97.596 |
USC00286055 | 2,019 | total_precip_in | 49.671985 |
USC00286055 | 2,020 | annual_avg_temp | 55.126436 |
USC00286055 | 2,020 | cdd | 1,186.029 |
USC00286055 | 2,020 | hottest_day_temp | 97.594 |
USC00286055 | 2,020 | total_precip_in | 49.727517 |
USC00286055 | 2,021 | annual_avg_temp | 55.190037 |
USC00286055 | 2,021 | cdd | 1,193.307 |
USC00286055 | 2,021 | hottest_day_temp | 97.593 |
USC00286055 | 2,021 | total_precip_in | 49.708219 |
USC00286055 | 2,022 | annual_avg_temp | 55.250445 |
USC00286055 | 2,022 | cdd | 1,200.588 |
USC00286055 | 2,022 | hottest_day_temp | 97.591 |
USC00286055 | 2,022 | total_precip_in | 49.758432 |
Ceviche: Protostar Climate Challenge Submission
Negative result: no.
Forecasts of yearly weather for 210 U.S. cities, 2019–2025, made as of December 31, 2018 (no data that reveals 2019–2025 weather), for the Protostar Challenge: Climate Risk Intelligence. The 5,880 rows cover annual average temperature (°F), hottest-day Tmax (°F), cooling degree days above 65°F (°F-days) and total precipitation (inches).
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 |
What the forecast says
Compared with each city's 1989–2018 average (the 208 cities with at least 10 eligible baseline years), our 2019–2025 forecasts are:
| Target | Median change per city | Cities with an increase | Range across cities |
|---|---|---|---|
| Cooling degree days | +171 °F-days (+15%) | 100% | +15 to +365 °F-days |
| Annual average temperature | +1.4 °F | 100% | +0.5 to +2.2 °F |
| Hottest-day temperature | +0.1 °F | 83% | −0.1 to +0.5 °F |
| Total precipitation | +0.46 in (+1.3%) | 85% | −2.1 to +1.6 in |
The largest increases in cooling demand, about +345 to +365 °F-days per year, are in Gulf Coast and desert cities: Tampa, Corpus Christi, Galveston, Needles, Austin and Phoenix. Cooling demand is the clearest signal. Hottest-day and rainfall changes are small next to their year-to-year variability.
Method
Each target has its own model; they are not refitted jointly. A city-year counts for a temperature target only if it has finite values and at most 10 missing Tmax and Tmin days. For precipitation it needs at most 10 missing precipitation days. Every fit uses only observations dated on or before its own issue date.
Annual average temperature
The forecast is a backbone plus small, bounded corrections, all added once without extra rounding:
- Backbone (75% / 25%):
- 75%, a local-level / pooled-trend recipe:
- a city's 15-year mean plus 1.25 × the national 50-year slope;
- blended equally with a regional-anomaly forecast (10 nearest cities; regional and national 30-year slopes; 10-year local offset).
- 25%, a pooled-trend ARMA forecast: the national trend is removed, a city ARMA(p,q) with p, q ≤ 3 is fitted by stepwise AICc, and the trend is added back.
- Fallback: cities with fewer than 15 observed years, more than 20% imputed values or a failed fit use climatology.
- 75%, a local-level / pooled-trend recipe:
- Residual calibration:
- weighted Ridge models (one per target) predict completed historical forecast errors from six features: forecast, distance from the 30-year mean, recent 5-year anomaly, trend distance, horizon and coverage;
- labels are used only after their full 7-year block has ended;
- the retained setting (30-year window, penalty 10, one-quarter of the fitted correction) was selected on earlier completed blocks.
- National residual state: 0.25 × [mean + φʰ × (latest − mean)], computed from the last 30 years of national average errors; persistence φ is clipped to ±0.5.
- Level adjustment:
- a fixed +0.20°F;
- a signed coverage-recentering term, clipped to ±0.5°F, which corrects for years with incomplete daily records;
- the earlier-block selector chose this over four alternatives at both the 2011 and 2018 origins.
- Pressure-conditioned trend:
- inputs: five EOFs of Northern Hemisphere sea-level pressure (NCEP/NCAR Reanalysis 1, 20–80°N, the 30 years ending November of the issue year);
- fit: a monthly temperature regression per city on year plus the five pressure components (Ridge penalty 4 × n on the components only);
- forecast term: the pressure-adjusted slope minus the raw slope (both pooled 50% city / 50% national) × horizon;
- interpretation: this removes the part of the historical trend associated with circulation. It is an empirical adjustment, not an attribution.
- Greenhouse-gas path:
- inputs: NOAA CO₂, CH₄ and N₂O concentrations, observed through two years before the issue, then projected with a bounded 10-year linear growth;
- fit: a pressure-controlled monthly regression of temperature on an approximate radiative-forcing index, compared with the same regression on calendar time;
- forecast term: 0.25 × (gas increment − time increment), clipped to ±0.5°F. It lets the trend follow the gas trajectory instead of a straight line.
- Ocean heat content:
- inputs: North Atlantic and North Pacific 0–700 m heat-content anomalies (NOAA NCEI), the 30 years ending two years before the issue;
- state: detrended and standardized, then projected forward by AR(1) relaxation;
- fit: the city's temperature-on-time regression plus both ocean terms (Ridge on the ocean terms), pooled 50/50 with national estimates;
- forecast term: 0.25 × (response to the projected ocean-state change + slope change × horizon), clipped to ±0.5°F.
Cooling degree days
- Base: observed daily temperature profiles are shifted by 1.25 × the national 50-year slope × elapsed years. CDD = Σ max(daily mean − 65°F, 0) is computed on the shifted profiles and averaged over 10- and 20-year donor windows (equal blend). This keeps the nonlinear threshold response.
- Residual calibration and national-state effect: the same as for average temperature, with the state passed through the daily-profile threshold, not a linear conversion.
- Leap-year correction: for 2020 and 2024, non-leap donor years receive an estimated February 29 (the mean of their shifted February 28 and March 1 contributions). It is a fixed calendar rule with no fitted coefficient, and only leap-year CDD changes.
- Not applied: the +0.20°F, pressure, gas and ocean corrections are not transferred to CDD.
Hottest-day temperature
- Level: the 30-year mean of the annual maximum Tmax.
- Trend: 70-year local slopes, averaged over the 20 nearest cities and halved.
- Calibration: a damped residual correction.
Total precipitation
- Climate-ratio reference:
- each city's observed monthly rainfall (30-year anchors with 90% daily coverage) is multiplied by the ratio of 30-year ensemble-mean precipitation at the target year to that at the issue year;
- the ensemble is 16 CMIP3 SRES A1B models with equal weight, nearest valid 2° cell within 250 km;
- these are simulations produced in 2007–2009;
- unsupported cities use their 30-year mean.
- LightGBM anomaly model:
- learns departures from each city's rainfall climatology from 327 origin-local features: multiscale levels, anomalies, variability, slopes, lags, coverage, regional and national summaries, location, seasonal summaries and horizon;
- fixed settings: L2 loss, 15 leaves, depth 4, 250 trees, learning rate 0.025, seed 20261110;
- the predicted anomaly is clipped to ±max(30-year SD, 1 inch) and blended 50/50 with the reference;
- the earlier-block selector compared 37 candidates over 60 fitted boosters and chose this configuration and weight.
- Regional precision-weighted trend:
- each city's 50-year relative rainfall slope is shrunk toward an inverse-variance-weighted average of its 40 nearest cities;
- the result is clipped to ±1% per year and applied to a median-based issue-year level;
- the final rainfall forecast is 75% of the blend above plus 25% of this regional forecast; unsupported cities keep the blend unchanged.
How we checked and chose it
- Backtests: chronological, each forecasting seven years ahead from the data available at its start year only.
- Automatic settings: chosen on earlier completed blocks with conservative tie rules (residual heads, national state, level adjustment and rainfall booster).
- Final package choice: made by the team on the 2012–2018 block (origin 2011). That includes the pressure, gas and ocean corrections, the regional rainfall trend and the unchanged base CDD settings. The regional rainfall trend was not selected by the earlier-block comparison; the team adopted it on the 2012–2018 block.
- What was never used: every target used for any choice ends in 2018 or earlier. Leaderboard scores and post-2018 data played no role.
- Limits:
- neighbouring cities and years are correlated, so one favourable 7-year block is limited evidence of general superiority;
- the pressure, gas and ocean terms are empirical predictors, not causal estimates;
- the fixed weights (75/25, 0.25, ±0.5°F caps) are design choices, not physical constants.
How we arrived at Ceviche
We started with a target-specific model rather than one universal annual trend. Average temperature needs a stable local level and a pooled warming signal; CDD needs daily profiles because days cross the 65°F threshold nonlinearly; the hottest day is a noisy extreme; rainfall has large local variability. We kept that backbone, including its limited ARMA contribution, and asked which remaining errors could be corrected without replacing it.
First, we learned from completed historical forecast errors. Strongly regularized, quarter-strength Ridge corrections restrained the influence of noisy residuals. We then pooled the error shared across cities into a damped national temperature state, passing its CDD effect through the daily threshold calculation. Finally, we separated the temperature level from its trend: a fixed +0.20°F and a bounded, signed coverage correction addressed a cool starting level and differences in which historical years supplied daily profiles. Together, these first three additions account for about 1.79 of the total 1.96 historical score-point gain.
Rainfall required a different approach. We anchored sixteen older CMIP3 simulations to observed monthly rainfall, using climate-model ratios for slow change rather than individual storms. A pooled LightGBM model then learned bounded anomalies around local climatology. Shrinking noisy city rainfall slopes toward precision-weighted neighbouring slopes supplied a final, small regional blend.
For average temperature, pressure patterns offered a way to condition the extrapolated trend on historical circulation. Projected greenhouse-gas concentrations supplied a small departure from a straight time trend; detrended Atlantic and Pacific ocean heat content supplied a gradually relaxing state. We added differences from matched reference fits, with damping and caps on the gas and ocean terms, to avoid adding another complete warming trend. These are empirical adjustments, not causal attribution. Their measured incremental gains were much smaller than those of the initial calibration steps.
The last change was a calendar correction: non-leap donor profiles otherwise omit an estimated February 29 contribution when forecasting leap-year CDD. It adds no fitted parameter. We deliberately retained the original CDD donor windows and did not transfer the average-temperature physical corrections to CDD.
This is the story of the retained sequence, not every experiment attempted. Earlier completed blocks selected the automatic policies; the team chose the final package using 2012–2018. The regional rainfall candidate did not win its earlier-block automatic comparison. The table below makes both the improvements and their limits visible.
Component ablation: cumulative additions
Historical development comparison only: issue December 31, 2011; targets 2012–2018. Each row retains the previous row and adds or replaces the named component. This is a sequential addition study, not a leave-one-out test: effects depend on order and interactions. Every target used in this table ends by 2018. Lower mean absolute error (MAE) is better; higher score is better.
| Step | Model change | Average temperature MAE °F | Hottest-day MAE °F | CDD MAE °F-days | Rainfall MAE inches | Overall score | Gain |
|---|---|---|---|---|---|---|---|
| 1 | Original team baseline | 1.3971 | 2.385 | 151.98 | 5.9752 | 11.027 | — |
| 2 | Forecast-error calibration | 1.3722 | 2.393 | 148.96 | 5.9654 | 11.731 | +0.704 |
| 3 | Shared annual temperature correction | 1.3496 | 2.393 | 146.88 | 5.9654 | 12.319 | +0.588 |
| 4 | Mean-temperature level adjustment | 1.3154 | 2.393 | 146.88 | 5.9654 | 12.819 | +0.500 |
| 5 | Climate-informed rainfall | 1.3154 | 2.393 | 146.88 | 5.9552 | 12.862 | +0.043 |
| 6 | Pooled rainfall anomaly model | 1.3154 | 2.393 | 146.88 | 5.9519 | 12.876 | +0.014 |
| 7 | Pressure-adjusted temperature trend | 1.3145 | 2.393 | 146.88 | 5.9519 | 12.889 | +0.013 |
| 8 | Three-gas forcing adjustment | 1.3143 | 2.393 | 146.88 | 5.9519 | 12.893 | +0.004 |
| 9 | Ocean heat-content adjustment | 1.3140 | 2.393 | 146.88 | 5.9519 | 12.897 | +0.005 |
| 10 | Regional rainfall blend | 1.3140 | 2.393 | 146.88 | 5.9316 | 12.982 | +0.085 |
| 11 | Leap-year CDD correction | 1.3140 | 2.393 | 146.86 | 5.9316 | 12.985 | +0.003 |
Every row uses the same 1,424 eligible city-years per temperature/CDD target and 1,434 for rainfall. The 30-year-climatology MAEs are 1.7103 °F, 2.415 °F, 201.28 °F-days and 5.9785 inches. The score uses the challenge formula, 100 × max(0, mean over four targets of (1 − model MAE / climatology MAE)), and each score in the table matches that formula applied to its MAEs.
What this supports: the package raises the local score from 11.03 to 12.99 (+1.96 points). MAE falls by 0.083 °F for average temperature, 5.1 °F-days for CDD and 0.044 inches for rainfall. Hottest-day MAE rises by 0.007 °F. Most of the gain comes from the three calibration steps. The physical-predictor corrections (pressure, gases, ocean) add small, consistent improvements, and the regional rainfall blend is the largest later step. The +0.20 °F and coverage terms were evaluated together, so their separate contributions are not isolated. For rainfall, the end-2011 row uses the head selected at that origin (L1); the 2018 submission uses the head selected at 2018 (L2).
Limits: the figures come from the team's historical experiment runs on this block; the full experiment logs are not part of this minimal package. A comparable component table for the earlier blocks (1991–1997, 1998–2004, 2005–2011) is not available. Because the team developed on 2012–2018 repeatedly, and cities and years are correlated, these gains are evidence rather than independent confirmation.
AI Tools and Agents
- Tools and models: ChatGPT Dots, ChatGPT (6.1 Sol Pro), Claude Code (Claude Opus 5.5), Codex.
- What they did: all four tools contributed to both coding (data preparation, model reconstruction, experiments, historical backtests and reproducibility testing) and documentation (including drafting this README and the data card). Forecast values are generated by code, not edited manually.
- What a person reviewed: the team defined which ideas to attempt and which data to consider, reviewed the data exploration, decided on the data cleaning, reviewed the results of the final run and of all previous runs, and decided on the policy recommendations.
- Agent instructions file:
code/AGENTS.md.
Policy Idea
Cooling-ready public buildings: a pilot of efficient cooling and accessible cooling refuges in existing libraries and community centres.
- What risk did we notice? Cooling demand rises everywhere in our forecast: median +15% CDD versus 1989–2018 across all 208 cities, with the largest absolute increases in Gulf Coast and desert cities. Residents without reliable home cooling bear most of that risk. Cooling demand is also the target our model predicts best (CDD error 27% below the 30-year average on the 2012–2018 test), which makes it a sound basis for annual planning.
- What would we do about it, and who would do it?
- City sustainability office: rank candidate public buildings using the CDD outlook together with building-condition, energy and neighbourhood-access data, publish the selection criteria and pick a small pilot set.
- Facilities team and local utility: audit the pilot buildings and assess shading, envelope improvements and efficient cooling.
- Heat-health team: plan staffed refuge hours and outreach. Day-to-day activation follows operational heat alerts, since annual CDD is a planning input, not a warning system.
- Why is it sustainable?
- It reuses existing public buildings, and cuts heat gains before adding cooling capacity.
- The budget and success thresholds are set before procurement; the pilot expands only after a cooling season's review of weather-normalized energy use, refuge use and operating cost.
- The outlook is regenerated each January from free public data with the included code.
See the one-page policy brief for an implementation and evaluation plan.
Feedback to the Ecosystem
These are specific reuse barriers found in the pinned training revision and our reproduction work. Dataset counts below were checked against the bundled raw files; they describe the source, not errors remaining in our forecast file.
| Barrier and evidence | Why it matters / our response | Requested upstream fix and acceptance check |
|---|---|---|
Missing hottest-day values look numerical: 1,184 of 28,052 annual rows contain Max Tmax = -inf. |
Nonfinite labels can break fitting. Our temperature eligibility excludes nonfinite annual values and years exceeding the missing-day limit. | Dataset maintainers: publish null instead of infinity and a metric-specific eligibility flag. Check that hottest-day fields contain no infinities and the 1,184 source rows remain explicitly missing. |
No observed rainfall looks like a dry year: 932 rows have ToT Prcp = 0 with at least 365 missing precipitation days. |
A numerical zero can pull fitted rainfall levels downward. We require at most 10 missing precipitation days; a zero in a sufficiently observed year remains valid. | Dataset maintainers: make totals null when there are no valid observations; publish valid-day counts and eligibility. Check all 932 rows are missing/ineligible without relabelling genuinely observed dry years. |
Inconsistent daily temperatures: 9 daily rows have tmin > tmax among 10,245,754 records. |
Physically inconsistent pairs affect paired-temperature calculations. Our pressure/monthly and donor-coverage calculations require Tmax >= Tmin; this is not a claim that every inherited calculation applies the same mask. |
Dataset maintainers: add a quality flag and document whether derived indices exclude flagged pairs. Check all 9 are flagged and that daily-to-annual recomputation uses the declared rule. |
Station discontinuities: 247 city-years have Station Changes > 0. |
A station change may move the measured level; it does not prove all 247 are biased. We retain the provided coverage/station metadata and document the risk. | Dataset maintainers: publish machine-readable segment boundaries, station coordinates and homogenization status. Check joins preserve one row per city-year and every flagged year links to its station history. |
Validation still needs a live service: check_submission.py reads the city roster from the Hub, although data/city_ids.txt and raw city metadata are bundled. |
The raw-source build is offline after environment setup, but the current quick launcher ends with a network-dependent validator. Bundling inputs has not removed that last dependency. | Template maintainers: validate against a pinned bundled roster, with an explicit optional refresh. Acceptance: validate a complete file and reject a missing/unknown city with networking disabled. |
| Byte equality is machine-sensitive: a second Linux/WSL rebuild changed average temperature by at most 0.09°F (mean 0.0014°F; 446/1,470 values) and CDD by at most 0.013°F-days; Windows line endings can also differ. | Our exact checks report these mismatches rather than substituting reference values. Quick inference reproduces the reference bytes on tested Windows and Linux environments. | Challenge/tool maintainers: distinguish an exact-artifact check from numerical reproduction; standardize text line endings and publish per-target absolute/relative tolerance rules with machine/version metadata. Any tolerance must be explicit; a 0.09°F drift deserves diagnosis rather than being dismissed as identical. |
| Observation dates and publication vintages differ: the ocean extract is a January 2026 revision of observations ending in 2016. | Applying an observation cutoff cannot demonstrate what was available to an end-2018 forecaster. We disclose source dates and retain exact bytes/provenance. | Data providers/challenge maintainers: retain release snapshots and record observation_end, release_date, retrieval date, revision and checksum separately. Acceptance: a user can select and verify an issue-date release, or see a clear revised-product limitation. |
Files
| File | Content |
|---|---|
policy_brief.pdf |
One-page cooling-ready public buildings policy proposal |
submission.csv |
Our forecasts: city_id,year,metric,predicted_value |
run.py |
Launcher: setup, quick and full modes |
code/forecast.py |
Fast inference from the fitted inputs in data/ |
code/build_inputs.py |
Full rebuild of every fitted input from sources/, with exact comparisons |
code/ (other files) |
Training, selection, rainfall, pressure, gas, ocean, regional-rainfall and leap-day modules; check_submission.py (template validator); AGENTS.md; pinned requirements |
data/ |
Fitted inputs and configuration; see data/README.md |
sources/ |
Original raw inputs with URLs and provenance; see sources/README.md |
How to Reproduce
Use Python 3.12. From this folder (on Windows, use py -3.12 instead of python):
python run.py setup
python run.py quick
python run.py full
- setup creates two isolated, pinned environments, one for the main model and one for the pressure/gas fits. It needs internet access.
- quick evaluates the fitted inputs in
data/, rebuildssubmission.csv(no existing file needed), prints PASS, and runs the template validator. The validator needs internet access to read the official city list, and prints OK. - full works in an empty temporary workspace and refits every component from the raw files in
sources/, with no download needed. It then compares each regenerated input and the final file with the committed ones and prints PASS/FAIL for each. Add--work-dir new-empty-folderto keep its outputs. It took 18 minutes in the reference Linux environment and 36 minutes on a 12-core laptop under WSL; runtime depends on hardware.
Reproduction results. The quick path reproduces submission.csv byte for byte; we tested it on Windows and Linux. The full rebuild compares every regenerated file byte by byte, and its ARMA temperature fits are sensitive to the CPU and numerical library. On a second Linux machine (WSL, Python 3.12.16), the rainfall, gas-path and hottest-day components were reproduced exactly. Average temperature differed by at most 0.09°F (mean 0.0014°F, 446 of 1,470 values) and CDD by at most 0.013°F-days, so its exact-match checks report FAIL there. On Windows, line endings also make text files differ byte by byte.
Random seeds and numerical-thread settings are fixed in code. The reference submission.csv SHA-256 is 684eb7860a43f6923e9db66ddc534d2e62ea149618e144b5218bbcf7cd63ed33.
How to reuse this
Intended use: a transparent annual planning baseline and a set of reusable forecast components. It is not a seasonal outlook, daily heat warning, storm/flood design model, peak-load forecast or causal estimate of policy benefits. Annual rainfall does not identify storm intensity; CDD is not electricity demand without a separate building/energy model. New locations need compatible units, enough historical support, and local validation.
Run these Python snippets from the repository folder, after python run.py setup, with the main environment interpreter (.venv-main/bin/python on Linux/macOS or .venv-main/Scripts/python.exe on Windows). All example outputs go outside this package.
1. Generate the fitted forecast for one city
import sys
from pathlib import Path
sys.dont_write_bytecode = True
sys.path.insert(0, str(Path('code').resolve()))
from forecast import forecast
city = forecast(Path('data')).query("city_id == 'USW00094823'") # Pittsburgh
city.to_csv('../pittsburgh_forecast.csv', index=False) # 7 years x 4 metrics
This evaluates the bundled model and then filters its output; it does not refit Pittsburgh alone. Replace the ID with another entry in data/city_ids.txt.
2. Change the historical issue date and refit the thermal base
import sys
from pathlib import Path
sys.dont_write_bytecode = True
sys.path.insert(0, str(Path('code').resolve()))
from raw_inputs import prepare
from training.thermal import train
work = Path('../ceviche_origin_2011').resolve() # new external workspace
prepare(work / 'data/raw', raw_dir=Path('sources/raw'))
base = train(work, work / 'training', origins=(2011,), jobs=2, resume=False)
This produces the thermal base for 2012–2018, with completed earlier blocks supplying calibration. Supported origins are 1976, 1983, 1990, 1997, 2004, 2011 and 2018. It is a substantial refit, not a quick inference call. It does not regenerate the full final package: pressure, gases, ocean, rainfall and calendar components must also be rebuilt at the new issue. The frozen inference files and forecast.py encode origin 2018; changing a year column is insufficient. The current training loader rejects weather after 2018.
3. Reuse the regional rainfall component independently
import sys
from pathlib import Path
import pandas as pd
sys.dont_write_bytecode = True
sys.path.insert(0, str(Path('code').resolve()))
from regional_rainfall import eligible, fit, predict
annual = pd.read_parquet('sources/raw/annual_indices.parquet')
origin = 2011
annual = annual.loc[annual.Year <= origin].copy()
d = fit(annual, origin, window=50, pool=40, weighting='precision')
normal = eligible(annual.query('Year >= @origin - 29')).groupby('ID')['ToT Prcp'].mean()
b = d[['origin', 'city_id']].merge(pd.DataFrame({'year': range(origin+1, origin+8)}), how='cross')
b['predicted_value'] = b.city_id.map(normal) # replace with your own base forecast
b = b.dropna(subset=['predicted_value']).copy() # omit cities without an observed base
b['predicted_value'] = predict(d, b, alpha=0.25)
b.to_csv('../regional_rainfall_example.csv', index=False)
To use another dataset, replace annual with a table containing ID, Year, Lat, Lon, Miss Prcp and ToT Prcp (annual inches; one row per city-year). Keep input observations at or before the issue. The example uses a simple historical-mean base, not the submission's CMIP3/LightGBM reference, and omits cities without a finite base. A local slope needs 25 eligible years spanning 20 years, at least 10 supported neighbours, and 10 recent level years; unsupported cities retain the supplied base. To cover every city, supply a finite base for each and evaluate on completed historical blocks before adopting it.
Reuse and attribution: the code, derived features and forecasts are CC BY 4.0. Cite this team and the original CMU study; retain NOAA/LLNL and software terms listed in the data cards. The raw sources, fitted state and modular interfaces let another researcher replace one component while retaining the others, provided they preserve the issue-date rules and rebuild affected state.
Data
Training data:
CMU-OSPO/protostar-climate-train-to-2018, revisionbbbdebdab2a18e00f97ef87bd3f35a71e222180b(CC0); observations end 2018-12-31.External data (details in
sources/README.mdanddata/README.md):- Sea-level pressure: NOAA PSL NCEP/NCAR Reanalysis 1 monthly mean, January 1948 – November 2018.
- Greenhouse gases: NOAA GML global CO₂ and CH₄ annual means and HATS N₂O monthly means, observations through 2016.
- Ocean heat content: NOAA NCEI North Atlantic and North Pacific 0–700 m annual anomalies, observations 1955–2016.
- Climate projections: CMIP3 SRES A1B monthly precipitation from 16 models (LLNL archive), created 2007, packaged 2009. These are simulations, not observed weather.
All external observations are dated before 2019. Each historical fit uses only data up to its own issue date (pressure through November; gases and ocean with a two-year lag). Values after the observed period are model projections, never later observations.
Citations and credits:
- 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). Use of historical data to assess regional climate change. Journal of Climate, 32(14), 4299–4320. https://doi.org/10.1175/JCLI-D-18-0630.1
- Kalnay, E., et al. (1996). The NCEP/NCAR 40-Year Reanalysis Project. Bulletin of the American Meteorological Society, 77(3), 437–471. https://doi.org/10.1175/1520-0477(1996)077%3C0437:TNYRP%3E2.0.CO;2
- NCEP Reanalysis data provided by NOAA PSL, Boulder, Colorado, USA (https://psl.noaa.gov/). Greenhouse-gas data from NOAA Global Monitoring Laboratory (https://gml.noaa.gov/). Ocean heat content from NOAA National Centers for Environmental Information (https://www.ncei.noaa.gov/access/global-ocean-heat-content/). CMIP3 projections from the LLNL archive.
- Base model: the team's earlier version, revision b47743bd887b4d23088ba63f9f299138bd78ba7f, CC BY 4.0.
Team-authored code, derived features and forecasts are CC BY 4.0. Third-party data and software keep their own terms.
- Downloads last month
- 275