Dataset Viewer
Auto-converted to Parquet Duplicate
city_id
stringclasses
210 values
year
int64
2.02k
2.03k
metric
stringclasses
4 values
predicted_value
float64
3.06
5.17k
USC00042863
2,019
annual_avg_temp
66.819567
USC00042863
2,019
cdd
1,694.428684
USC00042863
2,019
hottest_day_temp
103.89
USC00042863
2,019
total_precip_in
13.031818
USC00042863
2,020
annual_avg_temp
66.979498
USC00042863
2,020
cdd
1,718.416711
USC00042863
2,020
hottest_day_temp
103.89
USC00042863
2,020
total_precip_in
13.031818
USC00042863
2,021
annual_avg_temp
66.928325
USC00042863
2,021
cdd
1,709.827683
USC00042863
2,021
hottest_day_temp
103.89
USC00042863
2,021
total_precip_in
13.031818
USC00042863
2,022
annual_avg_temp
66.737647
USC00042863
2,022
cdd
1,722.345955
USC00042863
2,022
hottest_day_temp
103.89
USC00042863
2,022
total_precip_in
13.031818
USC00042863
2,023
annual_avg_temp
66.787237
USC00042863
2,023
cdd
1,708.839627
USC00042863
2,023
hottest_day_temp
103.89
USC00042863
2,023
total_precip_in
13.031818
USC00042863
2,024
annual_avg_temp
66.836826
USC00042863
2,024
cdd
1,721.024505
USC00042863
2,024
hottest_day_temp
103.89
USC00042863
2,024
total_precip_in
13.031818
USC00042863
2,025
annual_avg_temp
66.886416
USC00042863
2,025
cdd
1,762.71671
USC00042863
2,025
hottest_day_temp
103.89
USC00042863
2,025
total_precip_in
13.031818
USC00166584
2,019
annual_avg_temp
67.580614
USC00166584
2,019
cdd
2,789.125669
USC00166584
2,019
hottest_day_temp
100.773077
USC00166584
2,019
total_precip_in
56.777241
USC00166584
2,020
annual_avg_temp
67.566547
USC00166584
2,020
cdd
2,779.414379
USC00166584
2,020
hottest_day_temp
100.773077
USC00166584
2,020
total_precip_in
56.777241
USC00166584
2,021
annual_avg_temp
67.395979
USC00166584
2,021
cdd
2,768.27531
USC00166584
2,021
hottest_day_temp
100.773077
USC00166584
2,021
total_precip_in
56.777241
USC00166584
2,022
annual_avg_temp
67.766574
USC00166584
2,022
cdd
2,817.567395
USC00166584
2,022
hottest_day_temp
100.773077
USC00166584
2,022
total_precip_in
56.777241
USC00166584
2,023
annual_avg_temp
67.816163
USC00166584
2,023
cdd
2,858.663413
USC00166584
2,023
hottest_day_temp
100.773077
USC00166584
2,023
total_precip_in
56.777241
USC00166584
2,024
annual_avg_temp
67.865753
USC00166584
2,024
cdd
2,800.38266
USC00166584
2,024
hottest_day_temp
100.773077
USC00166584
2,024
total_precip_in
56.777241
USC00166584
2,025
annual_avg_temp
67.915342
USC00166584
2,025
cdd
2,785.75852
USC00166584
2,025
hottest_day_temp
100.773077
USC00166584
2,025
total_precip_in
56.777241
USC00280734
2,019
annual_avg_temp
50.979228
USC00280734
2,019
cdd
745.975672
USC00280734
2,019
hottest_day_temp
95.3125
USC00280734
2,019
total_precip_in
44.915
USC00280734
2,020
annual_avg_temp
51.004022
USC00280734
2,020
cdd
749.127201
USC00280734
2,020
hottest_day_temp
95.3125
USC00280734
2,020
total_precip_in
44.915
USC00280734
2,021
annual_avg_temp
51.028817
USC00280734
2,021
cdd
752.278729
USC00280734
2,021
hottest_day_temp
95.3125
USC00280734
2,021
total_precip_in
44.915
USC00280734
2,022
annual_avg_temp
51.876224
USC00280734
2,022
cdd
755.430258
USC00280734
2,022
hottest_day_temp
95.3125
USC00280734
2,022
total_precip_in
44.915
USC00280734
2,023
annual_avg_temp
51.925813
USC00280734
2,023
cdd
758.581787
USC00280734
2,023
hottest_day_temp
95.3125
USC00280734
2,023
total_precip_in
44.915
USC00280734
2,024
annual_avg_temp
51.975403
USC00280734
2,024
cdd
761.733315
USC00280734
2,024
hottest_day_temp
95.3125
USC00280734
2,024
total_precip_in
44.915
USC00280734
2,025
annual_avg_temp
52.024992
USC00280734
2,025
cdd
764.884844
USC00280734
2,025
hottest_day_temp
95.3125
USC00280734
2,025
total_precip_in
44.915
USC00286055
2,019
annual_avg_temp
54.49724
USC00286055
2,019
cdd
1,126.673365
USC00286055
2,019
hottest_day_temp
97.5
USC00286055
2,019
total_precip_in
49.499
USC00286055
2,020
annual_avg_temp
54.641154
USC00286055
2,020
cdd
1,131.066334
USC00286055
2,020
hottest_day_temp
97.5
USC00286055
2,020
total_precip_in
49.499
USC00286055
2,021
annual_avg_temp
54.598394
USC00286055
2,021
cdd
1,148.843751
USC00286055
2,021
hottest_day_temp
97.5
USC00286055
2,021
total_precip_in
49.499
USC00286055
2,022
annual_avg_temp
54.847331
USC00286055
2,022
cdd
1,146.116698
USC00286055
2,022
hottest_day_temp
97.5
USC00286055
2,022
total_precip_in
49.499
End of preview. Expand in Data Studio

MKDir: ProtoStar Climate Challenge Submission

Negative result: no

Forecasts of annual average temperature, total precipitation, hottest-day temperature and cooling degree days for 210 U.S. cities, 2019-2025, built only from observations through 2018-12-31.

Team

Name Hugging Face username University, Program
Mauricio Posadas @MekkoMau Carnegie Mellon University
Deen Ojo @DeenOjo Carnegie Mellon University
Luke Corbett @corbe19 Carnegie Mellon University

Method

Each city's forecast starts from its own recent average and adds one warming trend that all 210 cities share. Cooling degree days (CDD) are the average of two models.

Metric How we forecast it
Average temperature The city's recent average (the mean of its last 10 and last 15 valid years) plus a warming trend of 0.050 °F per year, fitted to all cities together over 1979-2018. For 2019-2021 we also average in the ridge model described in the next row. From 2022 on we use the trend alone.
CDD The mean of two forecasts. The first uses the same shared trend, scaled for each city: a hot city gains more CDD from one extra degree than a cool one. The second is a ridge regression fitted to all cities at once, with one model per forecast year, using each city's recent values, longer averages, latitude, longitude and region. It follows Montero-Manso & Hyndman (2021) and was built by Mauricio Posadas.
Precipitation and hottest day The 1989-2018 average. Nothing we tried beat it consistently.

Why

A trend fitted to a single city mostly picks up year-to-year noise. One trend fitted to all cities at once is much steadier. The two CDD models work in very different ways and each did better in different test periods, so their average beat both of them.

The ridge model looks at each city's last few years. That helps for the next year or two, then fades. In our tests, mixing it into the temperature forecast raised the score for years 1-3 (1.32 vs 0.70) but lowered it for years 4-7 (6.05 vs 7.12). So we use it only for 2019-2021. Our forecasts for 2022-2025 are the same as in our previous submission.

How we chose

We tested each idea by pretending it was 1980, 1995 or 2011, forecasting the next seven years, and scoring the result with the challenge's formula. Before running anything, we wrote down the candidates and the rule for picking one (both are in code/trial.yaml). We never used the leaderboard to choose. This time we compared two options: switching the ridge off after 2021, or fading it out over three years. They scored almost the same, so we kept the simpler one.

These did not help (details in code/evidence/): separate trends by region, a trend 1.5 times steeper, trends for precipitation or the hottest day, a hottest-day trend scaled to each city, rebuilding CDD from shifted daily temperatures, and using data from each city's current station only.

Caveats

  • The shared trend is not allowed to go negative. We added this rule after the first version failed in the 1980 test, when mid-century cooling made the 40-year trend negative, so the 1980 result is not an independent check of it. It does not affect these forecasts, because the trend through 2018 is positive.
  • We compared about 40 versions over the project, so small score differences between them are probably noise. The clear gain comes from using a shared trend at all.
  • Most of the gain over the 30-year average is in 2022-2025. The gain for 2019-2021 is smaller. In a wider check with seven start years, this version did slightly worse than our previous one in three of them (1985, 2005 and 2011).

Test scores

The submitted forecaster is lead_split3 in code/trial.yaml. It was picked from the candidates in lc_lead_blend.yaml using only these tests: seven-year forecasts made as if it were 1980, 1995 or 2011, scored with the challenge's formula against the 30-year average. The 2019-2025 data and the leaderboard were not used to choose.

"Score" is the average official score over the 3 tests, without the official floor at zero. Skill is the percent cut in error compared with the 30-year average. "Tests won" counts the tests where the forecaster beat the 30-year average. Years 4-7 match the Final Score.

Forecast years Forecaster Score Temperature skill % Precipitation skill % Hottest day skill % CDD skill % Tests won
All 7 years lead_split3 4.69 9.86 0.0 0.0 8.91 3
All 7 years thirty_year_mean (reference) 0.0 0.0 0.0 0.0 0.0 0
Years 1-3 lead_split3 1.32 -0.09 0.0 0.0 5.35 3
Years 1-3 thirty_year_mean (reference) 0.0 0.0 0.0 0.0 0.0 0
Years 4-7 lead_split3 7.12 17.73 0.0 0.0 10.74 2
Years 4-7 thirty_year_mean (reference) 0.0 0.0 0.0 0.0 0.0 0

AI tools and agents

  • Tools and models: Claude Code (mostly Opus and Sonnet models) and Warp.
  • What they did: Wrote and refactored code, tests and documentation drafts under the rules in AGENTS.md.
  • What a person reviewed: The team reviewed the agents' output and workflow, and validated the results.

Agent rules for this project are in code/AGENTS.md: they enforce the 2018-12-31 cut-off and forbid agents from submitting or supplying forecast values.

Policy idea

Rolling climate normals for cooling-season planning

1. What risk did we notice? Utilities, regulators and cities plan around "normal" weather, usually a fixed 30-year average. In every test we ran, that average was a poor guide to temperature and cooling. A city's recent average plus a shared trend cut forecast error by about 9% (8.7% for temperature, 8.9% for CDD, in seven-year tests from 1980, 1995 and 2011).

For 2019-2025, our forecast puts yearly CDD above the 1989-2018 normal in 206 of the 208 scored cities, by a median of 9.7% (119 °F-days a year). Eleven cities, including Chicago and San Diego, move above 1,000 CDD, and eight, including Atlanta, move above 2,000. So planning from the 30-year normal understates cooling needs almost everywhere. The gap is largest in places that used to need little cooling: Seattle's CDD rise 41%, from a low starting point.

2. What would we do, and who would do it? State public utility commissions should require utilities to replace the fixed 30-year normal with a rolling one wherever they use it today: summer peak-load forecasts, reserve planning and efficiency targets. The rolling normal is a recent average plus a trend shared across cities, updated every year and checked against past years the way we did. Utilities would publish it for each city every year. When a city's normal crosses 1,000 or 2,000 CDD, the utility and the city's sustainability office would jointly fund efficiency and cooling programs (heat pumps, cool roofs, shade) before the summer. Local data on who is most at risk, which this dataset does not have, would decide where the money goes.

3. Why is it sustainable? It is cheap. The method runs in minutes on data agencies already have, and the code is open. Knowing the real load early lets utilities meet it with efficiency and demand response instead of emergency generation, which costs more and pollutes more. And because the normal is refitted and retested every year, it adjusts if warming speeds up or slows down instead of locking in a new fixed number.

Limits: the forecasts are for weather stations, mostly at airports, not for neighborhoods. They carry the errors shown in code/evidence/.

Feedback to the ecosystem

What got in the way, and what would fix it

  1. Which years count as valid depends on the metric, but the data doesn't say so. The annual table has Miss Tmax, Miss Tmin and Miss Prcp. You have to work out that temperature and CDD need both temperature counts, while precipitation and hottest day each use a different one. A single valid-year filter quietly skews some targets. Fix: add a validity column for each metric that follows the scoring rule, and a short note saying which missing-day columns each metric uses. Our cleaned table (data/annual_master.parquet) does this.

  2. Years with no data look like real measurements. In those years, ToT Prcp, Wet Days, Max 5-day P, the four warm/cold percentage columns and the max/min columns hold 0 or ±inf instead of a blank. Max 5-day P is even nonzero in 12 of them. Our master table blanks 13,160 such cells. Fix: leave the value blank whenever a year has no observations of that variable.

  3. A gated dataset looked public. Hugging Face turned on its access gate for our first dataset automatically. The README opened without a login, but submission.csv returned HTTP 401, and the Space reads files without a token. Nothing on the dataset page made this clear. Fix: have the Space check the dataset's gated flag and say "turn off access requests" in its error message, and say "public and ungated" in the submission instructions.

  4. "Latest submission counts" was easy to misread. At first our team thought that submitting several datasets would keep several entries scored. Only the team's most recent submission counts, whatever dataset it points to. Fix: put "this replaces your team's current entry" next to the submit button.

  5. The baseline rule changed partway through. On November 6 the organizers clarified that the 30-year baseline also needs 10 valid years in its window. Our scorer had matched the earlier practice-run numbers (212.9 and 177.4 against the published 213 and 177 °F-days) and then no longer matched the corrected 212 and 175. Fix: publish the scoring and baseline code, plus an answer key for a practice period before 2019, so teams can check their scorer exactly.

  6. How stations link to NOAA isn't written down. city_info.ID joins directly to NOAA's GHCN-Daily station list (all 210 IDs matched; see scripts/build_station_metadata_report.py in our GitHub repository). We only found this by reading both schemas and NOAA's file documentation. Fix: publish a station lookup table with the join key, where the metadata comes from, and known station moves.

  7. Files that change are hard to cite. NOAA updates its station list in place, so a URL alone can't reproduce an earlier result. Fix: publish a release manifest with source URLs, download date, checksums, cut-off date, units and the validity rule. Our data/annual_master.manifest.json is an example.

Data

  • External data: none. The forecasts use only the challenge training data, through 2018-12-31.
  • Curated data: see the data card in data/README.md: each file, its source, license, preparation and checksum.

Training data: CMU-OSPO/protostar-climate-train-to-2018, from:

Files

File Content
submission.csv 5,880 forecasts: 210 cities x 7 years x 4 metrics
code/ Pipeline, trial config trial.yaml, requirements.txt, reproduce.py, backtest.py
code/evidence/ Trial write-ups: every candidate tried, with its scores
data/ Curated inputs; each file is described in data/README.md

How to reproduce

pip install -r code/requirements.txt
python code/reproduce.py        # rebuilds the forecasts from data/ and compares with submission.csv
python code/backtest.py         # reruns the backtest table above (a few minutes)

Built from commit d7e6eb0 on 2026-11-09. In the repository the same folder is produced by uv run proto-star climate-package lc_lead_blend.yaml.

Downloads last month
151