Datasets:
city_id stringclasses 210
values | year int64 2.02k 2.03k | metric stringclasses 4
values | predicted_value float64 3.06 5.15k |
|---|---|---|---|
USC00042863 | 2,019 | annual_avg_temp | 66.543088 |
USC00042863 | 2,019 | total_precip_in | 13.031818 |
USC00042863 | 2,019 | hottest_day_temp | 103.89 |
USC00042863 | 2,019 | cdd | 1,599.872196 |
USC00042863 | 2,020 | annual_avg_temp | 66.594571 |
USC00042863 | 2,020 | total_precip_in | 13.031818 |
USC00042863 | 2,020 | hottest_day_temp | 103.89 |
USC00042863 | 2,020 | cdd | 1,610.287043 |
USC00042863 | 2,021 | annual_avg_temp | 66.646053 |
USC00042863 | 2,021 | total_precip_in | 13.031818 |
USC00042863 | 2,021 | hottest_day_temp | 103.89 |
USC00042863 | 2,021 | cdd | 1,620.701889 |
USC00042863 | 2,022 | annual_avg_temp | 66.697535 |
USC00042863 | 2,022 | total_precip_in | 13.031818 |
USC00042863 | 2,022 | hottest_day_temp | 103.89 |
USC00042863 | 2,022 | cdd | 1,631.116736 |
USC00042863 | 2,023 | annual_avg_temp | 66.749017 |
USC00042863 | 2,023 | total_precip_in | 13.031818 |
USC00042863 | 2,023 | hottest_day_temp | 103.89 |
USC00042863 | 2,023 | cdd | 1,641.531583 |
USC00042863 | 2,024 | annual_avg_temp | 66.800499 |
USC00042863 | 2,024 | total_precip_in | 13.031818 |
USC00042863 | 2,024 | hottest_day_temp | 103.89 |
USC00042863 | 2,024 | cdd | 1,651.94643 |
USC00042863 | 2,025 | annual_avg_temp | 66.851982 |
USC00042863 | 2,025 | total_precip_in | 13.031818 |
USC00042863 | 2,025 | hottest_day_temp | 103.89 |
USC00042863 | 2,025 | cdd | 1,662.361277 |
USC00166584 | 2,019 | annual_avg_temp | 67.350172 |
USC00166584 | 2,019 | total_precip_in | 56.777241 |
USC00166584 | 2,019 | hottest_day_temp | 100.773077 |
USC00166584 | 2,019 | cdd | 2,763.087304 |
USC00166584 | 2,020 | annual_avg_temp | 67.383576 |
USC00166584 | 2,020 | total_precip_in | 56.777241 |
USC00166584 | 2,020 | hottest_day_temp | 100.773077 |
USC00166584 | 2,020 | cdd | 2,770.107757 |
USC00166584 | 2,021 | annual_avg_temp | 67.41698 |
USC00166584 | 2,021 | total_precip_in | 56.777241 |
USC00166584 | 2,021 | hottest_day_temp | 100.773077 |
USC00166584 | 2,021 | cdd | 2,777.128209 |
USC00166584 | 2,022 | annual_avg_temp | 67.450385 |
USC00166584 | 2,022 | total_precip_in | 56.777241 |
USC00166584 | 2,022 | hottest_day_temp | 100.773077 |
USC00166584 | 2,022 | cdd | 2,784.148662 |
USC00166584 | 2,023 | annual_avg_temp | 67.483789 |
USC00166584 | 2,023 | total_precip_in | 56.777241 |
USC00166584 | 2,023 | hottest_day_temp | 100.773077 |
USC00166584 | 2,023 | cdd | 2,791.169115 |
USC00166584 | 2,024 | annual_avg_temp | 67.517193 |
USC00166584 | 2,024 | total_precip_in | 56.777241 |
USC00166584 | 2,024 | hottest_day_temp | 100.773077 |
USC00166584 | 2,024 | cdd | 2,798.189568 |
USC00166584 | 2,025 | annual_avg_temp | 67.550597 |
USC00166584 | 2,025 | total_precip_in | 56.777241 |
USC00166584 | 2,025 | hottest_day_temp | 100.773077 |
USC00166584 | 2,025 | cdd | 2,805.210021 |
USC00280734 | 2,019 | annual_avg_temp | 51.609382 |
USC00280734 | 2,019 | total_precip_in | 44.915 |
USC00280734 | 2,019 | hottest_day_temp | 95.3125 |
USC00280734 | 2,019 | cdd | 809.087388 |
USC00280734 | 2,020 | annual_avg_temp | 51.63959 |
USC00280734 | 2,020 | total_precip_in | 44.915 |
USC00280734 | 2,020 | hottest_day_temp | 95.3125 |
USC00280734 | 2,020 | cdd | 813.42737 |
USC00280734 | 2,021 | annual_avg_temp | 51.669797 |
USC00280734 | 2,021 | total_precip_in | 44.915 |
USC00280734 | 2,021 | hottest_day_temp | 95.3125 |
USC00280734 | 2,021 | cdd | 817.767352 |
USC00280734 | 2,022 | annual_avg_temp | 51.700004 |
USC00280734 | 2,022 | total_precip_in | 44.915 |
USC00280734 | 2,022 | hottest_day_temp | 95.3125 |
USC00280734 | 2,022 | cdd | 822.107334 |
USC00280734 | 2,023 | annual_avg_temp | 51.730211 |
USC00280734 | 2,023 | total_precip_in | 44.915 |
USC00280734 | 2,023 | hottest_day_temp | 95.3125 |
USC00280734 | 2,023 | cdd | 826.447316 |
USC00280734 | 2,024 | annual_avg_temp | 51.760418 |
USC00280734 | 2,024 | total_precip_in | 44.915 |
USC00280734 | 2,024 | hottest_day_temp | 95.3125 |
USC00280734 | 2,024 | cdd | 830.787298 |
USC00280734 | 2,025 | annual_avg_temp | 51.790625 |
USC00280734 | 2,025 | total_precip_in | 44.915 |
USC00280734 | 2,025 | hottest_day_temp | 95.3125 |
USC00280734 | 2,025 | cdd | 835.12728 |
USC00286055 | 2,019 | annual_avg_temp | 54.473084 |
USC00286055 | 2,019 | total_precip_in | 49.499 |
USC00286055 | 2,019 | hottest_day_temp | 97.5 |
USC00286055 | 2,019 | cdd | 1,097.640874 |
USC00286055 | 2,020 | annual_avg_temp | 54.513336 |
USC00286055 | 2,020 | total_precip_in | 49.499 |
USC00286055 | 2,020 | hottest_day_temp | 97.5 |
USC00286055 | 2,020 | cdd | 1,102.523401 |
USC00286055 | 2,021 | annual_avg_temp | 54.553587 |
USC00286055 | 2,021 | total_precip_in | 49.499 |
USC00286055 | 2,021 | hottest_day_temp | 97.5 |
USC00286055 | 2,021 | cdd | 1,107.405929 |
USC00286055 | 2,022 | annual_avg_temp | 54.593839 |
USC00286055 | 2,022 | total_precip_in | 49.499 |
USC00286055 | 2,022 | hottest_day_temp | 97.5 |
USC00286055 | 2,022 | cdd | 1,112.289595 |
Zhoey: Protostar Climate Challenge Submission
Negative result: no
Forecasts of yearly weather for 210 U.S. cities, 2019–2025, made as of the end of 2018 (no data that reveals 2019–2025 weather), for the Protostar Challenge: Climate Risk Intelligence.
We locked our forecast (its fingerprint is in code/verify/PREREGISTRATION.md, committed on Nov 2 before we downloaded
anything), then scored it on 2019–2025 weather before the official scores came out; turns out we got 12.8 for
2019–2021, which the leaderboard also gave us later (12.8). Neither the model nor the forecast changed afterwards.
We also found that a large part of the score came from one number: how much warmer the whole country got. If we just added the real warming that happened, about 1.67 °F, to every city's 30-year average, that would score 24.0 on 2022–2025, way above our own 16.1. We raised this with the organizers in discussion #3.
Key findings
- The trend baseline did beat us on temperature, but only because it guessed more warming. However, if we take out each year's shared national error, our predictions can tell cities apart (a city-relative score we computed, not the challenge Score: 3.5), whereas the trend baseline cannot (0.0).
- When we decide where cooling help goes first, the forecast didn't really add anything. Ranking cities with the forecast and without it gave almost the same result.
- State AC averages don't always point help to the right places. Metro surveys and state averages differ by 4.3 percentage points beyond survey error. Vancouver, WA: 21% of homes have no AC, but Washington's figure says 48%.
- We reproduced the source paper: 37 of its 50 statistics matched exactly and 11 closely. Its main warning, that a trend depends on the years you fit it to, turned out to matter for forecasting too: trends over a city's whole history couldn't predict 4–7 years ahead, but 30–45-year trends could. An LSTM and two pretrained models lost to a basic 10-year average, and one of them (TimesFM) had been trained on weather up to 2024, which breaks the challenge's rules.
Full write-up, with every experiment: docs/TECHNICAL_REPORT.md. Explorer:
ProtoStar Climate Risk Explorer.
Team
| Name | Hugging Face username | University, Program |
|---|---|---|
| Kevin Zhou | @zhoism | Georgia Institute of Technology |
| Arav Mestry | @aravmestry | Cornell University |
Method
Model: We fit 3 lines through the city's past temperature 3 ways: over the last 45 years, the last 30, and since 1975, since this is when the paper lists global warming taking off. We then pull these trends halfway towards the national trend, to reduce the effect of anomaly years. We then average these 3 trends with the city's last 10 years, aiming to average a bunch of reasonable guesses to reduce noise.
For cooling degree days (basically, how much AC weather a city gets), we take the city's actual daily temperatures from the last 30 years, raise each day by our forecast warming, and recount. This matters because a couple of degrees of warming adds much more AC demand in hot cities than in mild ones, and a straight trend line would miss that.
For rain and the hottest day of the year, we just use the 1989–2018 average. These jump around too much from year to year, and nothing we tried, trend lines included, beat the plain average.
Features: only each city's own history up to 2018: yearly values, plus daily temperatures for cooling degree days. Years with more than 10 missing days are skipped.
How we checked it: a backtest. We pretend it's an earlier year (1995–2011), forecast 4–7 years ahead, and score
that the way the challenge does. We picked this combination because it scored best of everything we tried
(19 statistical variants, an LSTM, and two pretrained models), but the gain over a plain trend is small. A
separate check (code/check_leakage.py) scrambles all data after the cutoff and confirms the forecast doesn't change,
so nothing leaks.
Backtest (uv run code/backtest.py): train to year T, forecast T+1…T+7, score as the challenge does, origins
1995–2011, lead years 4–7.
| model | temp S | precip S | hot S | CDD S | Score | sd across origins |
|---|---|---|---|---|---|---|
| 30-year mean (baseline) | 0 | 0 | 0 | 0 | 0.0 | 0.0 |
| 30-year trend (organizers' baseline) | 0.106 | −0.077 | −0.081 | 0.058 | 0.2 | 4.3 |
| our model | 0.153 | 0.0 | 0.0 | 0.136 | 7.2 | 2.8 |
Checked against real weather (code/verify/): over the full 2019–2025 period, which the official Final Score uses,
our own score is 14.8 (trend baseline 12.0). The organizers' published scoring code gives the same numbers as
ours (12.80 for 2019–2021 and 16.05 for 2022–2025). Our 2022–2025 score is against public NOAA records, not the hidden answer key. On the live board, 12.8 put us
last of the 9 teams scored on Nov 6 (leader 16.0), though above all three baselines.
The cooling gap. Joining the forecasts to the American Housing Survey (metro AC access, 39 cities), EIA RECS 2020
(state AC access, the other 171) and Census population flags Vancouver, WA (1.00), Spokane, WA (0.95), Portland, ME (0.83), Seattle, WA (0.82), Fresno, CA (0.71), Helena, MT (0.65), Bakersfield, CA (0.60), Missoula, MT (0.57), Cheyenne, WY (0.56), Portland, OR (0.52), Escondido, CA (0.51) as likely to combine above-median CDD
growth with ≥ 15% of homes without AC (3.08 M residents without AC, 90% interval 0.62–5.20 M). How the flag holds up
under other rules: code/policy/robustness.md.
AI Tools and Agents
- Tools and models: Claude Code (Anthropic, Claude Opus 5.5) and Codex (OpenAI, GPT-6.1 Sol).
- What they did: wrote and ran the code and experiments, and drafted the documentation and the feedback posts.
- What a person reviewed: We made all the directional calls on approach and managed the agents. There were a lot of paths to choose from, and sorting what was useful from what wasn't was our main job. We set the data rule they had to follow, decided to lock the forecast before scoring it against real weather, and chose the policy focus. Every public-facing decision and post was approved by us, and checks guard the data's integrity: one fails if a forecast reads any data after its cutoff, and another fails if a number in the README doesn't match the results.
- Agent instructions file:
code/AGENTS.md(it states the data rule).
No forecast value came from an AI model; code/check_leakage.py fails if a forecast reads data after its origin; the
only post-2018 weather data was downloaded after the forecast was frozen, and only to score it. code/ reproduces
submission.csv without any agent.
Policy Idea
- What risk did we notice? In the 2021 Northwest heat wave, people died at home without working air conditioning. Of 72 heat deaths in Multnomah County (Portland), 94% were at home. The Northwest has some of the fewest homes with AC (46.6% of Seattle-area homes have none), and its cooling need rose fastest in 2022–2025: up 44%, when we had forecast 12%.
- What would we do about it, and who would do it? Cooling money is handed out by formulas that count heat and income, but not whether homes already have AC. Oregon's heat-pump grants send 40% of their money by "cooling zone × energy-burdened households", and Washington's energy-aid plan budgets nothing for cooling, so AC only comes as emergency aid. At its next rule update, Oregon's Department of Energy should multiply that cooling share by each area's share of homes without AC. Washington's Department of Commerce, whose rules already allow AC for at-risk residents, should send its emergency AC units to those areas first. Rank places by residents without AC times today's cooling need; the forecast only tells you which way things are heading.
- Why is it sustainable? It adds no new cost: it redirects money these programs already spend. Heat pumps replace gas or oil furnaces while adding cooling, so cooling can grow without raising emissions, and the Northwest grid is largely hydropower. No new data is needed: the American Housing Survey measures AC every two years. The weak spot is that only 39 of 210 cities have their own survey, so the rest use state numbers, which can be way off (Vancouver, WA: 21% without AC, against Washington's 48%). We'd keep Oregon's rule that 30% of the money is split equally, so places where the numbers are wrong still get a share.
Feedback to the Ecosystem
What got in the way:
- The data stores "missing" in 3 different ways: blank, infinity, and 0.00 inches of rain. This messes with average computations; one city's normal rainfall came out as 2 inches per year instead of about 60 because of this.
- 4 cities' records were missing decades of data that NOAA had.
- Some scoring rules weren't written down anywhere public.
- Popular AI forecasting models were trained on the years we were trying to predict, so anyone using them would break the rules without knowing.
- The rules changed after launch without an official announcement, which caused a little confusion.
- A widely used dataset-metadata tool silently turned county code 06073 into 6073, which would break matching against the Census.
What would fix it (we posted these in the appropriate channels):
- Store missing values one way, and explain it on the data card: our card fix is an open pull request.
- Fill the missing decades from NOAA, or at least flag them: posted.
- Write the scoring rules down: the organizers answered, added them to the About tab and shared our thread with the judges.
- List the AI models known to have seen the test years: posted.
- Keep a dated "What changed" list: the organizers adopted it and credited us.
- Read county codes as text: our fix is an open pull request, and the maintainer ran the tests and says they pass.
We also sent a fix so the organizers' format checker works offline (pull request). FRICTION_LOG.md contains
more details.
Reuse of Lai & Dzombak (2019)
We redid the original researchers' analysis from the challenge's copy of their data and got their numbers back (key
finding 4). We also packaged our model so anyone can run it on their own cities: it works on any NOAA weather station,
and we tested it on 48 stations outside the challenge. How to run it: analysis/docs/TUTORIAL.md. The full
reproduction: code/reuse/RESULTS_reuse.md.
Interoperability
The challenge data line up exactly with CMU's original release, but matching them to NOAA's own records is messy: old
stations have no NOAA IDs, and station moves aren't marked. We still linked our results to five government datasets
(NOAA's station and county records, the Census Bureau's housing survey, the federal energy survey and Census population
data), and the standard metadata file we wrote for our dataset (croissant.json) is how we caught the county-code bug.
Details: code/interop/RESULTS_interop.md.
Limitations
- Our forecasts are only modestly better than the baseline: for rain and the hottest day we don't beat the 30-year average at all, and we under-guessed how much warming would happen.
- Each city is one weather station, usually at an airport, sometimes moved over the years, so these are regional numbers, not neighbourhood ones.
- The cooling-gap results are estimates: 171 of 210 cities rely on statewide AC data, and cooling degree days measure how much cooling is needed, not health outcomes.
Files
| Path | Content |
|---|---|
submission.csv |
The forecasts (from code/predict.py) |
data/ |
The pinned input files the code reads; data card in data/README.md |
croissant.json |
Croissant 1.0 metadata for the forecasts, their ranges and the cooling-gap table |
docs/TECHNICAL_REPORT.md |
The full write-up |
code/ |
Model, forecasts, backtest, leakage check, experiments, paper reproduction, lineage tests, prospective check |
analysis/ |
Toolkit and CLI, data joins, cooling gap, prediction ranges, explorer app, policy brief, reviewer kit |
FRICTION_LOG.md, DEVLOG.md |
Problems met, with evidence and fixes; dead ends |
GIT_HISTORY.txt |
Commit list with dates, so the order of choices can be checked |
How to Reproduce
code/ and data/ alone rebuild submission.csv byte for byte, with no download:
pip install -r requirements.txt && python code/predict.py (writes out/submission.csv). The full pipeline needs
uv:
uv sync --group analysis
uv run --group analysis python analysis/reviewer/run_review.py # ~20 s: checks the published numbers
bash scripts/ship.sh # ~5 min: data -> submission -> official checker, leakage check
uv run code/backtest.py # ~15 s: the backtest table
uv run --group analysis python analysis/run_all.py --full # ~1 GB of public data: joins, cooling gap, ranges, brief
The analysis inputs are archived, with hashes, in zhoism/protostar-climate-2026-inputs;
analysis/fetch_external.py falls back to it if a source moves or changes a file (--mirror uses only the archive).
Run from the repository root. The forecasts are deterministic (no random seeds).
Data
Forecast inputs: only CMU-OSPO/protostar-climate-train-to-2018
at commit 43e98eb5d1a9bc65ccdb7401ffc422524b98b157, with the rows of the
submission template at
38f74f79feadd7e5c8292cf31f298c4269f84cfb; copies are in data/.
Analysis only: CMU KiltHub v1 and v6; NOAA GHCN-Daily and nClimDiv (read up to 2018); Census/HUD American
Housing Survey 2019, 2021, 2023; EIA RECS 2020; Census SAIPE 2018, Population Estimates 2018, boundaries 2018 and the
2013 CBSA delineation (URLs and hashes in analysis/integration/provenance.json); pretrained models
amazon/chronos-bolt-small,
amazon/chronos-bolt-base,
google/timesfm-2.5-200m-pytorch.
Policy sources: Multnomah County, 2021 heat deaths report (https://multco.us/node/272339); Oregon OAR 330-270-0020,
Community Heat Pump allocation (https://www.law.cornell.edu/regulations/oregon/Or-Admin-Code-SS-330-270-0020);
Washington LIHEAP state plan FY2025 (https://liheapch.ncat.org/docs/2025/state-plans/WA_Plan_2025.pdf); Washington
weatherization manual 2026 (https://nascsp.org/wp-content/uploads/2026/07/2026-Washington-State-Wx-Manual.pdf); Census
AHS methodology (https://census.gov/programs-surveys/ahs/about/methodology.html).
Scoring only: NOAA GHCN-Daily 2009–2025 for the 210 stations, downloaded after the forecast was frozen.
Citations
- Lai, Y., & Dzombak, D. A. (2019). Compiled historical daily temperature and precipitation data for selected 210 U.S. cities. Carnegie Mellon University. Dataset. 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. Dataset. 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), 4299–4320. https://doi.org/10.1175/JCLI-D-18-0630.1
Licences: code MIT; forecasts, results and documentation CC BY 4.0.
- Downloads last month
- 1,059
