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
2.97
5.17k
USC00042863
2,019
annual_avg_temp
66.819567
USC00042863
2,019
cdd
1,694.428684
USC00042863
2,019
hottest_day_temp
103.908893
USC00042863
2,019
total_precip_in
12.713925
USC00042863
2,020
annual_avg_temp
66.979498
USC00042863
2,020
cdd
1,718.416711
USC00042863
2,020
hottest_day_temp
103.908893
USC00042863
2,020
total_precip_in
12.713925
USC00042863
2,021
annual_avg_temp
66.928325
USC00042863
2,021
cdd
1,709.827683
USC00042863
2,021
hottest_day_temp
103.908893
USC00042863
2,021
total_precip_in
12.713925
USC00042863
2,022
annual_avg_temp
66.737647
USC00042863
2,022
cdd
1,722.345955
USC00042863
2,022
hottest_day_temp
103.908893
USC00042863
2,022
total_precip_in
13.181966
USC00042863
2,023
annual_avg_temp
66.787237
USC00042863
2,023
cdd
1,708.839627
USC00042863
2,023
hottest_day_temp
103.908893
USC00042863
2,023
total_precip_in
13.181966
USC00042863
2,024
annual_avg_temp
66.836826
USC00042863
2,024
cdd
1,721.024505
USC00042863
2,024
hottest_day_temp
103.908893
USC00042863
2,024
total_precip_in
13.181966
USC00042863
2,025
annual_avg_temp
66.886416
USC00042863
2,025
cdd
1,762.71671
USC00042863
2,025
hottest_day_temp
103.908893
USC00042863
2,025
total_precip_in
13.181966
USC00166584
2,019
annual_avg_temp
67.580614
USC00166584
2,019
cdd
2,789.125669
USC00166584
2,019
hottest_day_temp
100.650164
USC00166584
2,019
total_precip_in
56.833182
USC00166584
2,020
annual_avg_temp
67.566547
USC00166584
2,020
cdd
2,779.414379
USC00166584
2,020
hottest_day_temp
100.650164
USC00166584
2,020
total_precip_in
56.833182
USC00166584
2,021
annual_avg_temp
67.395979
USC00166584
2,021
cdd
2,768.27531
USC00166584
2,021
hottest_day_temp
100.650164
USC00166584
2,021
total_precip_in
56.833182
USC00166584
2,022
annual_avg_temp
67.766574
USC00166584
2,022
cdd
2,817.567395
USC00166584
2,022
hottest_day_temp
100.650164
USC00166584
2,022
total_precip_in
56.505061
USC00166584
2,023
annual_avg_temp
67.816163
USC00166584
2,023
cdd
2,858.663413
USC00166584
2,023
hottest_day_temp
100.650164
USC00166584
2,023
total_precip_in
56.505061
USC00166584
2,024
annual_avg_temp
67.865753
USC00166584
2,024
cdd
2,800.38266
USC00166584
2,024
hottest_day_temp
100.650164
USC00166584
2,024
total_precip_in
56.505061
USC00166584
2,025
annual_avg_temp
67.915342
USC00166584
2,025
cdd
2,785.75852
USC00166584
2,025
hottest_day_temp
100.650164
USC00166584
2,025
total_precip_in
56.505061
USC00280734
2,019
annual_avg_temp
50.979228
USC00280734
2,019
cdd
745.975672
USC00280734
2,019
hottest_day_temp
95.335031
USC00280734
2,019
total_precip_in
44.046602
USC00280734
2,020
annual_avg_temp
51.004022
USC00280734
2,020
cdd
749.127201
USC00280734
2,020
hottest_day_temp
95.335031
USC00280734
2,020
total_precip_in
44.046602
USC00280734
2,021
annual_avg_temp
51.028817
USC00280734
2,021
cdd
752.278729
USC00280734
2,021
hottest_day_temp
95.335031
USC00280734
2,021
total_precip_in
44.046602
USC00280734
2,022
annual_avg_temp
51.876224
USC00280734
2,022
cdd
755.430258
USC00280734
2,022
hottest_day_temp
95.335031
USC00280734
2,022
total_precip_in
44.944225
USC00280734
2,023
annual_avg_temp
51.925813
USC00280734
2,023
cdd
758.581787
USC00280734
2,023
hottest_day_temp
95.335031
USC00280734
2,023
total_precip_in
44.944225
USC00280734
2,024
annual_avg_temp
51.975403
USC00280734
2,024
cdd
761.733315
USC00280734
2,024
hottest_day_temp
95.335031
USC00280734
2,024
total_precip_in
44.944225
USC00280734
2,025
annual_avg_temp
52.024992
USC00280734
2,025
cdd
764.884844
USC00280734
2,025
hottest_day_temp
95.335031
USC00280734
2,025
total_precip_in
44.944225
USC00286055
2,019
annual_avg_temp
54.49724
USC00286055
2,019
cdd
1,126.673365
USC00286055
2,019
hottest_day_temp
97.532076
USC00286055
2,019
total_precip_in
49.033335
USC00286055
2,020
annual_avg_temp
54.641154
USC00286055
2,020
cdd
1,131.066334
USC00286055
2,020
hottest_day_temp
97.532076
USC00286055
2,020
total_precip_in
49.033335
USC00286055
2,021
annual_avg_temp
54.598394
USC00286055
2,021
cdd
1,148.843751
USC00286055
2,021
hottest_day_temp
97.532076
USC00286055
2,021
total_precip_in
49.033335
USC00286055
2,022
annual_avg_temp
54.847331
USC00286055
2,022
cdd
1,146.116698
USC00286055
2,022
hottest_day_temp
97.532076
USC00286055
2,022
total_precip_in
48.836966
End of preview. Expand in Data Studio

MKDir: ProtoStar Climate Challenge Submission

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

Forecaster r4_winners: hottest_day_temp = 0.5 x [final_hottest_nudge: Hottest day: 30-year mean nudged by 0.3 x the city's mean departure over the last 10 years from its 31 nearest cities (Tatakae)]; 0.5 x [final_hottest_second_high: Hottest day: mean second-highest daily maximum over complete years (>= 360 days) of the 30-year window + tail spacing pooled by a 300 km Gaussian kernel (PstarAlu); 30-year mean if < 10 complete years]; other metrics from [swap_metric: total_precip_in = 0.5 x [final_precip_nudge: Precipitation: 30-year mean nudged by 0.3 x the city's mean departure over the last 10 years from its 31 nearest cities (Tatakae), as a ratio]; 0.5 x [final_precip_two_half: Precipitation: 0.5 x 30-year median (median/mean ratio shrunk 0.5 toward the 300 km kernel mean) + 0.5 x complete-year mean (30 years for leads 1-3, 45 for later leads if >= 10/15 complete years) (PstarAlu); leads > 3: Fay-Herriot shrinkage of the 45-year local deviation vs the 90-year record (300 km leave-one-out kernel)]; other metrics from [lead_blend: Pooled trend (levels [10, 15], 40-year slope, CDD sensitivity) blended with the pooled P09/P14 ridge (alpha 1): temperature trend weight by lead 0.5/0.5/0.5/1 (last holds), CDD trend weight 0.5; 30-year mean for precipitation and hottest day]].

Chosen from the candidates in mp_final_round.yaml using only backtests on the team's fixed folds: seven-year forecasts from cut-offs [1980, 1995, 2011], scored with the official skill score against the 30-year average. Test and leaderboard data were not used for selection. Mean raw_score (the official score without its floor at zero) and per-metric skill in percent:

window method raw_score annual_avg_temp_skill total_precip_in_skill hottest_day_temp_skill cdd_skill origins_beating_baseline
all r4_winners 4.85 9.86 0.21 0.43 8.91 2
all thirty_year_mean 0.0 0.0 0.0 0.0 0.0 0
years_1_3 r4_winners 1.46 -0.09 -0.15 0.73 5.35 3
years_1_3 thirty_year_mean 0.0 0.0 0.0 0.0 0.0 0
years_4_7 r4_winners 7.3 17.73 0.51 0.23 10.74 2
years_4_7 thirty_year_mean 0.0 0.0 0.0 0.0 0.0 0

origins_beating_baseline counts the folds (out of 3) where the method beats thirty_year_mean. The Final Score uses all seven years (2019-2025) with equal weight, so all is the matching window.

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

Policy idea: Cooling-resilience planning from recent CDD normals

What risk did we notice?

Cooling degree days translate temperature into the annual intensity of air-conditioning demand. The model-selection evidence shows that recent ten-year conditions are the most reliable simple signal for CDD and annual temperature across the historical seven-year tests. Planning from a thirty-year normal alone can therefore understate the cooling conditions a city or utility must be ready to serve.

What should happen, and who should do it?

City sustainability offices and electric utilities should publish an annual Cooling Resilience Trigger based on each city's backtest-validated CDD and annual-temperature forecast, compared with its 30-year baseline. When the trigger is elevated, they should jointly:

  1. reserve summer peak capacity and expand heat-emergency communications;
  2. prioritize cool roofs, shade trees, and building-envelope retrofits in high-energy-burden areas using a separate local vulnerability assessment;
  3. report avoided peak demand and program coverage before the next cooling season.

The station-level forecasts set the city-scale planning signal. Local agencies, not this dataset, must supply neighborhood vulnerability and siting data before targeting interventions.

Why is it sustainable?

Avoiding cooling load through shade, reflective roofs, and efficiency reduces both peak-grid stress and lifetime energy use. Reserving capacity earlier is less costly and less emissions-intensive than emergency generation during heat events. The proposal also avoids treating airport weather stations as a proxy for neighborhood heat exposure, preserving the boundary between a city-scale forecast and an equitable local implementation.

Feedback to the ecosystem

Feedback to the ProtoStar ecosystem

What got in the way

  1. Validity is metric-specific but not machine-labeled. The CMU annual table exposes Miss Tmax, Miss Tmin, and Miss Prcp; participants must infer that average temperature and CDD need both temperature completeness checks, while precipitation and hottest-day values use different checks. Applying one generic valid-year filter would silently bias some targets.

  2. Cross-repository station linkage needs a stated contract. CMU's city_info.ID can join directly to NOAA's GHCN inventory, and our reproducible report matched all 210 IDs. That capability was discoverable only by inspecting both schemas and fixed-width NOAA documentation.

  3. Data provenance needs freeze-friendly references. NOAA's station inventory is a living file. A URL alone is insufficient to reproduce the exact metadata integration after an update.

  4. No-data years look like real measurements. In years with no observations, several columns hold 0 or ±inf instead of a blank: ToT Prcp, Wet Days, Max 5-day P, the four warm and cold percentage columns, and the max/min columns. Max 5-day P even has a nonzero value in 12 such years. Our master table blanks 13,160 of these cells.

What would fix it

  1. Publish a valid_<metric> boolean for each annual index using the challenge's scoring completeness rule, plus a short data contract that identifies the applicable missing-day columns.
  2. Publish a compact station crosswalk with the intended join key, metadata source, and any known station-thread/location caveats.
  3. Include a release manifest with source URLs, retrieval date, checksum, cutoff date, units, and the exact evaluation-validity rule. That would make the dataset directly usable by both people and automated research agents.
  4. Store blanks for every index in a year with no observations of that variable.

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
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

Built from commit 07af4ea (with uncommitted changes) on 2026-11-10. In the repository the same folder is produced by uv run proto-star climate-package mp_final_round.yaml.

Downloads last month
312