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
2.43
5.46k
USC00042863
2,019
annual_avg_temp
66.812335
USC00042863
2,019
total_precip_in
12.462235
USC00042863
2,019
hottest_day_temp
104.269892
USC00042863
2,019
cdd
1,730.681876
USC00042863
2,020
annual_avg_temp
66.891076
USC00042863
2,020
total_precip_in
12.457456
USC00042863
2,020
hottest_day_temp
104.303051
USC00042863
2,020
cdd
1,752.71631
USC00042863
2,021
annual_avg_temp
66.970245
USC00042863
2,021
total_precip_in
12.452697
USC00042863
2,021
hottest_day_temp
104.33621
USC00042863
2,021
cdd
1,774.169589
USC00042863
2,022
annual_avg_temp
67.04986
USC00042863
2,022
total_precip_in
12.447958
USC00042863
2,022
hottest_day_temp
104.369367
USC00042863
2,022
cdd
1,796.988442
USC00042863
2,023
annual_avg_temp
67.129932
USC00042863
2,023
total_precip_in
12.443239
USC00042863
2,023
hottest_day_temp
104.402525
USC00042863
2,023
cdd
1,820.475565
USC00042863
2,024
annual_avg_temp
67.210471
USC00042863
2,024
total_precip_in
12.43854
USC00042863
2,024
hottest_day_temp
104.435683
USC00042863
2,024
cdd
1,844.52373
USC00042863
2,025
annual_avg_temp
67.2915
USC00042863
2,025
total_precip_in
12.433862
USC00042863
2,025
hottest_day_temp
104.46884
USC00042863
2,025
cdd
1,869.125412
USC00166584
2,019
annual_avg_temp
67.725057
USC00166584
2,019
total_precip_in
57.046637
USC00166584
2,019
hottest_day_temp
101.015952
USC00166584
2,019
cdd
2,835.785298
USC00166584
2,020
annual_avg_temp
67.79937
USC00166584
2,020
total_precip_in
57.117585
USC00166584
2,020
hottest_day_temp
101.038862
USC00166584
2,020
cdd
2,870.682353
USC00166584
2,021
annual_avg_temp
67.874154
USC00166584
2,021
total_precip_in
57.188568
USC00166584
2,021
hottest_day_temp
101.061773
USC00166584
2,021
cdd
2,897.66174
USC00166584
2,022
annual_avg_temp
67.949418
USC00166584
2,022
total_precip_in
57.259589
USC00166584
2,022
hottest_day_temp
101.084681
USC00166584
2,022
cdd
2,922.301886
USC00166584
2,023
annual_avg_temp
68.025186
USC00166584
2,023
total_precip_in
57.330644
USC00166584
2,023
hottest_day_temp
101.107592
USC00166584
2,023
cdd
2,947.004257
USC00166584
2,024
annual_avg_temp
68.101477
USC00166584
2,024
total_precip_in
57.401737
USC00166584
2,024
hottest_day_temp
101.130502
USC00166584
2,024
cdd
2,972.283594
USC00166584
2,025
annual_avg_temp
68.178308
USC00166584
2,025
total_precip_in
57.472866
USC00166584
2,025
hottest_day_temp
101.153411
USC00166584
2,025
cdd
3,001.812486
USC00280734
2,019
annual_avg_temp
52.063119
USC00280734
2,019
total_precip_in
46.324381
USC00280734
2,019
hottest_day_temp
94.703509
USC00280734
2,019
cdd
858.877191
USC00280734
2,020
annual_avg_temp
52.139653
USC00280734
2,020
total_precip_in
46.402589
USC00280734
2,020
hottest_day_temp
94.705944
USC00280734
2,020
cdd
879.29147
USC00280734
2,021
annual_avg_temp
52.217341
USC00280734
2,021
total_precip_in
46.480828
USC00280734
2,021
hottest_day_temp
94.70838
USC00280734
2,021
cdd
891.600396
USC00280734
2,022
annual_avg_temp
52.295267
USC00280734
2,022
total_precip_in
46.559098
USC00280734
2,022
hottest_day_temp
94.710816
USC00280734
2,022
cdd
904.355548
USC00280734
2,023
annual_avg_temp
52.373781
USC00280734
2,023
total_precip_in
46.637399
USC00280734
2,023
hottest_day_temp
94.71325
USC00280734
2,023
cdd
917.239145
USC00280734
2,024
annual_avg_temp
52.452971
USC00280734
2,024
total_precip_in
46.715729
USC00280734
2,024
hottest_day_temp
94.715685
USC00280734
2,024
cdd
930.484091
USC00280734
2,025
annual_avg_temp
52.532869
USC00280734
2,025
total_precip_in
46.794092
USC00280734
2,025
hottest_day_temp
94.718121
USC00280734
2,025
cdd
944.189702
USC00286055
2,019
annual_avg_temp
54.877477
USC00286055
2,019
total_precip_in
50.759315
USC00286055
2,019
hottest_day_temp
97.860107
USC00286055
2,019
cdd
1,153.98296
USC00286055
2,020
annual_avg_temp
54.956514
USC00286055
2,020
total_precip_in
50.846525
USC00286055
2,020
hottest_day_temp
97.871165
USC00286055
2,020
cdd
1,177.763202
USC00286055
2,021
annual_avg_temp
55.036763
USC00286055
2,021
total_precip_in
50.93377
USC00286055
2,021
hottest_day_temp
97.882222
USC00286055
2,021
cdd
1,192.683795
USC00286055
2,022
annual_avg_temp
55.117257
USC00286055
2,022
total_precip_in
51.021051
USC00286055
2,022
hottest_day_temp
97.893276
USC00286055
2,022
cdd
1,208.01866
End of preview. Expand in Data Studio

City Climate Ensemble

Negative result: no

Annual temperature, precipitation, hottest day, and cooling degree day forecasts for 210 U.S. cities. The forecast origin is December 31, 2018. The 2019–2025 years are forecast labels; observations and leaderboard results for those years were excluded from modeling and selection.

Method

We model each station's climate level and smooth trend, using the supplied station history to avoid treating a site move as climate change. Related stations share trend information through geographic weighting. Seasonal temperature trends are mapped through the fraction of days above 65 °F to estimate changes in cooling degree days. We tested nonlinear warming trends, NOAA temperature and ocean predictors, seasonal extreme temperature models, and precipitation medians. We fit separate convex blends by metric and combined the nationwide weights with a smaller regional component.

Model selection compared 1,007 candidates on 14 historical forecast origins from 1976 through 2011, with all observed targets ending by 2018. The challenge mask removes years with more than ten missing days for a required variable and city–metrics with fewer than ten usable baseline years. On the 2011–2018 practice window, we reproduced the published CDD baselines: 211.7 °F-days MAE for the 30-year average and 174.6 for the plain linear trend.

2011–2018 practice evaluation Overall skill Annual temperature MAE Precipitation MAE Hottest day MAE CDD MAE
30-year average 0.00 1.66 °F 6.18 in 2.55 °F 211.7 °F-days
Chronological replay: blend weights fit before 2011 9.14 1.40 °F 6.23 in 2.54 °F 167.4 °F-days
Final blend configuration, scored on this tuning window 14.45 1.28 °F 6.03 in 2.51 °F 145.9 °F-days

The replay only used completed earlier folds to fit blend weights. The final configuration uses all eligible historical folds, including the practice window, so the last row is descriptive rather than an independent performance estimate. reports/validation.json contains the other historical replay and tuning scores. The 2019–2025 score is unknown; year-to-year variability dominates precipitation and annual extremes.

AI Tools and Agents

OpenAI Codex (GPT-6) helped implement the cutoff checks and forecasting code, ran the historical experiments, and drafted the documentation. All forecast values are computed by code/forecast.py from local data. No LLM forecast values or pretrained weather models are used. The repository and code/AGENTS.md specify the cutoff, and the forecast process rejects network access. Human review of this package has not yet been recorded.

Policy Idea

Pittsburgh's airport record rose from an average of 774 cooling degree days in 1989–2003 to 849 in 2004–2018, an increase of about 10% across eligible years. That is a signal to plan for higher cooling demand, although the airport record cannot identify which neighborhoods have the highest indoor exposure.

Pittsburgh and its utility could pilot summer heat retrofits in affordable multifamily buildings: install exterior shade and reflective roofing where suitable, then provide efficient heat pumps when existing equipment is replaced. Prioritize buildings using resident reports and indoor temperature measurements, not the airport series alone. Track indoor hours above a health-based threshold, electricity peak demand, and tenant bills before and after the pilot. Passive cooling can reduce both dangerous indoor heat and peak electricity demand, while bill protections can keep retrofit costs from shifting to tenants.

Feedback to the Ecosystem

The annual missing-day counts, paired temperature coverage, and station changes materially affect how a model should be trained and scored. The published CDD benchmark was a useful integration test. A machine-readable scoring fixture with the exact eligibility mask for a historical practice split would make it easier to verify how the 10-day and 10-year rules interact, without exposing any competition targets. Providing a stable link to a pinned data revision and a downloadable field dictionary would also make it easier to reproduce the forecast and audit the effects of site moves.

Data and Reproduction

The CMU-OSPO cutoff dataset is used under CC0. Historical sources include Lai and Dzombak's daily station compilation and annual indices. External sources are NOAA PSL ERSSTv5 (1900–2018) and NOAA NCEI Climate at a Glance (1880–2018); both were acquired with date bounds and tested as forecast inputs. The data card records the exact sources, dates, preparation, and hashes.

Install code/requirements.txt, then run python code/forecast.py and python code/check_local.py submission.csv. The included curated inputs and frozen weights reproduce submission.csv offline. See code/README.md for backtesting and data preparation.

Downloads last month
122