Datasets:
city_id stringclasses 210
values | year int64 2.02k 2.03k | metric stringclasses 4
values | predicted_value float64 2.05 5.23k |
|---|---|---|---|
USC00042863 | 2,019 | annual_avg_temp | 66.855283 |
USC00042863 | 2,019 | cdd | 1,643.75625 |
USC00042863 | 2,019 | hottest_day_temp | 103.89 |
USC00042863 | 2,019 | total_precip_in | 13.031818 |
USC00042863 | 2,020 | annual_avg_temp | 66.907991 |
USC00042863 | 2,020 | cdd | 1,643.75625 |
USC00042863 | 2,020 | hottest_day_temp | 103.89 |
USC00042863 | 2,020 | total_precip_in | 13.031818 |
USC00042863 | 2,021 | annual_avg_temp | 66.960698 |
USC00042863 | 2,021 | cdd | 1,643.75625 |
USC00042863 | 2,021 | hottest_day_temp | 103.89 |
USC00042863 | 2,021 | total_precip_in | 13.031818 |
USC00042863 | 2,022 | annual_avg_temp | 67.013405 |
USC00042863 | 2,022 | cdd | 1,643.75625 |
USC00042863 | 2,022 | hottest_day_temp | 103.89 |
USC00042863 | 2,022 | total_precip_in | 13.031818 |
USC00042863 | 2,023 | annual_avg_temp | 67.066113 |
USC00042863 | 2,023 | cdd | 1,643.75625 |
USC00042863 | 2,023 | hottest_day_temp | 103.89 |
USC00042863 | 2,023 | total_precip_in | 13.031818 |
USC00042863 | 2,024 | annual_avg_temp | 67.11882 |
USC00042863 | 2,024 | cdd | 1,643.75625 |
USC00042863 | 2,024 | hottest_day_temp | 103.89 |
USC00042863 | 2,024 | total_precip_in | 13.031818 |
USC00042863 | 2,025 | annual_avg_temp | 67.171527 |
USC00042863 | 2,025 | cdd | 1,643.75625 |
USC00042863 | 2,025 | hottest_day_temp | 103.89 |
USC00042863 | 2,025 | total_precip_in | 13.031818 |
USC00166584 | 2,019 | annual_avg_temp | 67.812712 |
USC00166584 | 2,019 | cdd | 2,791.72161 |
USC00166584 | 2,019 | hottest_day_temp | 100.773077 |
USC00166584 | 2,019 | total_precip_in | 56.777241 |
USC00166584 | 2,020 | annual_avg_temp | 67.86542 |
USC00166584 | 2,020 | cdd | 2,803.08505 |
USC00166584 | 2,020 | hottest_day_temp | 100.773077 |
USC00166584 | 2,020 | total_precip_in | 56.777241 |
USC00166584 | 2,021 | annual_avg_temp | 67.918127 |
USC00166584 | 2,021 | cdd | 2,814.449112 |
USC00166584 | 2,021 | hottest_day_temp | 100.773077 |
USC00166584 | 2,021 | total_precip_in | 56.777241 |
USC00166584 | 2,022 | annual_avg_temp | 67.970835 |
USC00166584 | 2,022 | cdd | 2,825.817828 |
USC00166584 | 2,022 | hottest_day_temp | 100.773077 |
USC00166584 | 2,022 | total_precip_in | 56.777241 |
USC00166584 | 2,023 | annual_avg_temp | 68.023542 |
USC00166584 | 2,023 | cdd | 2,837.221716 |
USC00166584 | 2,023 | hottest_day_temp | 100.773077 |
USC00166584 | 2,023 | total_precip_in | 56.777241 |
USC00166584 | 2,024 | annual_avg_temp | 68.076249 |
USC00166584 | 2,024 | cdd | 2,848.751593 |
USC00166584 | 2,024 | hottest_day_temp | 100.773077 |
USC00166584 | 2,024 | total_precip_in | 56.777241 |
USC00166584 | 2,025 | annual_avg_temp | 68.128957 |
USC00166584 | 2,025 | cdd | 2,860.289125 |
USC00166584 | 2,025 | hottest_day_temp | 100.773077 |
USC00166584 | 2,025 | total_precip_in | 56.777241 |
USC00280734 | 2,019 | annual_avg_temp | 50.985 |
USC00280734 | 2,019 | cdd | 739.333333 |
USC00280734 | 2,019 | hottest_day_temp | 95.3125 |
USC00280734 | 2,019 | total_precip_in | 44.915 |
USC00280734 | 2,020 | annual_avg_temp | 50.985 |
USC00280734 | 2,020 | cdd | 739.333333 |
USC00280734 | 2,020 | hottest_day_temp | 95.3125 |
USC00280734 | 2,020 | total_precip_in | 44.915 |
USC00280734 | 2,021 | annual_avg_temp | 50.985 |
USC00280734 | 2,021 | cdd | 739.333333 |
USC00280734 | 2,021 | hottest_day_temp | 95.3125 |
USC00280734 | 2,021 | total_precip_in | 44.915 |
USC00280734 | 2,022 | annual_avg_temp | 50.985 |
USC00280734 | 2,022 | cdd | 739.333333 |
USC00280734 | 2,022 | hottest_day_temp | 95.3125 |
USC00280734 | 2,022 | total_precip_in | 44.915 |
USC00280734 | 2,023 | annual_avg_temp | 50.985 |
USC00280734 | 2,023 | cdd | 739.333333 |
USC00280734 | 2,023 | hottest_day_temp | 95.3125 |
USC00280734 | 2,023 | total_precip_in | 44.915 |
USC00280734 | 2,024 | annual_avg_temp | 50.985 |
USC00280734 | 2,024 | cdd | 739.333333 |
USC00280734 | 2,024 | hottest_day_temp | 95.3125 |
USC00280734 | 2,024 | total_precip_in | 44.915 |
USC00280734 | 2,025 | annual_avg_temp | 50.985 |
USC00280734 | 2,025 | cdd | 739.333333 |
USC00280734 | 2,025 | hottest_day_temp | 95.3125 |
USC00280734 | 2,025 | total_precip_in | 44.915 |
USC00286055 | 2,019 | annual_avg_temp | 54.718325 |
USC00286055 | 2,019 | cdd | 1,169.232441 |
USC00286055 | 2,019 | hottest_day_temp | 97.5 |
USC00286055 | 2,019 | total_precip_in | 49.499 |
USC00286055 | 2,020 | annual_avg_temp | 54.771033 |
USC00286055 | 2,020 | cdd | 1,175.942885 |
USC00286055 | 2,020 | hottest_day_temp | 97.5 |
USC00286055 | 2,020 | total_precip_in | 49.499 |
USC00286055 | 2,021 | annual_avg_temp | 54.82374 |
USC00286055 | 2,021 | cdd | 1,182.653329 |
USC00286055 | 2,021 | hottest_day_temp | 97.5 |
USC00286055 | 2,021 | total_precip_in | 49.499 |
USC00286055 | 2,022 | annual_avg_temp | 54.876447 |
USC00286055 | 2,022 | cdd | 1,189.363773 |
USC00286055 | 2,022 | hottest_day_temp | 97.5 |
USC00286055 | 2,022 | total_precip_in | 49.499 |
Team Unicorns: 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.
submission.csv holds our forecasts: 210 cities × 7 years (2019–2025) × 4 metrics = 5,880 rows. It is the Experiment 5B forecast. We developed the work in a GitHub repository (JessCai06/-Protostar-Challenge-CMU-2026-Climate-Risk-Intelligence), and this dataset holds the same files at its commit cb90e92. The commit hashes below refer to that repository: Experiment 5B was frozen in 0fbbe5b. The earlier submissions were Experiment 5 (caccb04) and Experiment 3 (295c10a).
Policy slide deck: Will Everyone Endure the Heat? (live page; its source is index.html and presentation/ in this dataset).
Team
| Name | Hugging Face username | University, Program |
|---|---|---|
| Jessica Cai | @Jessiter | Carnegie Mellon University, Undergraduate, Information Systems |
| Pisit Tangwongsiri | @Riceee293 | Carnegie Mellon University, Undergraduate, Statistics and Machine Learning |
| Iris Yin | @catsvsdogslala | Carnegie Mellon University, Undergraduate, Statistics and Machine Learning |
| Carina Zheng | @cariniaz | Carnegie Mellon University, Undergraduate, Computer Science and Art |
Method
- Model: a separate simple model for each metric, each chosen by a rule we wrote before running it.
annual_avg_tempuses the mean of the city's last 15 valid years plus one warming slope shared by all cities.cddtakes the city's actual days from its last 10 valid years, warms each daily high and low by the shared slope of annual mean Tmax and Tmin, and recomputes cdd.total_precip_inandhottest_day_tempuse the city's 30-year average, because no trend method beat it on held-back years. - Features: only the challenge dataset, up to 2018 (
annual_indices,coverage_and_stations,daily). A year counts only if the station missed at most 10 days of the data that metric needs. - How we checked it: we held back years inside the training data. At the origins 1998, 2004 and 2011 we trained on data up to the origin, forecast the next 7 years and scored them. Skill = 1 − MAE / MAE of the plain average. A candidate had to beat the plain average in 2 of 3 origins, and the simpler idea won any lead under 0.01. The result for Experiment 5B: temperature skill +0.177, cdd +0.126, precipitation and hottest day 0 (the plain average).
Experiment 5B (the submission)
Experiment 5B is Experiment 5 (below) plus two candidates built from each city's daily highs and lows (the challenge's daily table, up to 2018). Everything else is unchanged: Experiment 5's candidates, the held-back origins, and the decision rule. The new candidates count as more complex than all of Experiment 5's for the 0.01 tie rule. outputs/exp05b/decision.json was written before the final forecast.
- cdd from daily highs and lows (chosen). cdd sums
max(0, (Tmax + Tmin)/2 - 65)over days, so the same warming adds many degree days in Miami and almost none in Eureka. Experiment 5 adds one cdd slope to every city. Experiment 5B instead takes the city's actual days from its last 10 valid cdd years, warms every daily high by the shared slope of annual mean Tmax (+0.035 °F/yr at origin 2018) and every daily low by the shared slope of annual mean Tmin (+0.047 °F/yr; nights warm faster), times (year − those years' middle year), and recomputes cdd. With no warming this equals Experiment 5's 10-year level, so the only change is how warming turns into cdd. A 15-year version was also tested. - Hottest day from the 95th percentile of daily highs (not chosen). The single hottest day is noisy, so we used the shared slope of each year's 95th-percentile daily high (much steadier) with a 30- or 15-year level. It did not beat the plain average (−0.007 and −0.011). The daily data shows why: the hot tail of daily highs barely moved before 2018 (shared slope −0.005 to +0.014 °F/yr depending on origin), while mean highs and especially lows rose. See
outputs/exp05b/slopes.csv.
| Metric | Method used | Held-back skill: chosen (Exp05's pick) | Per origin 1998 / 2004 / 2011 |
|---|---|---|---|
annual_avg_temp |
15-year level + shared slope (as Exp05) | +0.177 | 0.196 / 0.167 / 0.169 |
cdd |
daily delta method, last 10 years | +0.126 (+0.112) | −0.003 / 0.139 / 0.244 |
total_precip_in |
A plain average (as Exp05) | 0 | nothing beat the plain average |
hottest_day_temp |
A plain average (as Exp05) | 0 | nothing beat the plain average |
The cdd choice is not uniform: the delta method loses to Exp05's pick at origin 1998 (−0.003 vs +0.022) and wins at 2004 and 2011. Its forecast differs most where a single slope fits worst: cool coastal cities drop to near zero (e.g., USW00024213 from about 80 to 4 cdd), and the hottest cities rise about 2%. Code: code/exp05b_daily_distribution.py; outputs: outputs/exp05b/.
Bug fix in the held-back test (2026-11-08). shared.valid_flags subtracted a merged column with a fresh 0..n−1 index from a column that kept the data's own index. When the data is cut at a held-back origin the two no longer line up, so about one cdd validity flag in four in the 30-year windows was wrong (1,725 of 6,300 at origin 1998; 0 at origin 2018). The fix (.to_numpy()) leaves every submitted forecast unchanged, since those use origin 2018. Experiment 5's picks are also the same with it, but its cdd held-back skill is +0.112, not the +0.086 in outputs/exp05/. The held-back cdd numbers in outputs/exp03/ to outputs/exp07/ were made before the fix and are not regenerated. Rerun with the fix, Experiment 3 would pick the half slope for cdd. Experiments 5B, 9 and 10 use the fixed code.
Earlier submission: Experiment 5
Experiment 5 keeps Experiment 3's shared slope but changes the city's level. Experiment 3 starts from the 30-year mean, which is centered on about 2003, so every forecast leans on the slope for 16–22 years. Experiment 5 also tries the mean of the city's last 10 or 15 valid years, with the shared slope continued from those years' middle year. Those are the window lengths of NOAA's Optimal Climate Normals (Huang, van den Dool & Barnston 1996, J. Climate 9:809–817). It also tried a station-aware level (years at the city's latest station only), a cdd slope scaled by each city's own cdd-per-°F sensitivity, and an equal-weight ensemble.
| Metric | Method used | Held-back skill: chosen (Exp03's pick) | Beat Exp03's pick in |
|---|---|---|---|
annual_avg_temp |
15-year level + shared slope | +0.177 (+0.157) | 3 of 3 origins |
cdd |
10-year level + shared slope | +0.086 (+0.070) | 3 of 3 origins |
total_precip_in |
A plain average | 0 (0) | nothing beat the plain average |
hottest_day_temp |
A plain average | 0 (0) | nothing beat the plain average |
Full results for all ten candidates: outputs/exp05/heldout_results.csv.
Earlier submission: Experiment 3
We forecast each metric from the city's last 30 valid years (1989–2018) and compared four ideas: A plain average; B a line through the city's own years; C one slope shared by all cities, each city keeping its own level (city mean + shared slope × (year − city mean year)); D half the city's own slope. Precipitation and cdd are clipped at 0. A city with fewer than 10 valid years uses the plain average of the valid years it has, or the all-city mean if it has none.
| Metric | Method used | Average held-back skill (A / B / C / D) |
|---|---|---|
annual_avg_temp |
C shared slope | 0 / +0.135 / +0.157 / +0.129 |
cdd |
C shared slope | 0 / +0.045 / +0.070 / +0.054 |
total_precip_in |
A plain average | 0 / −0.083 / −0.020 / −0.020 |
hottest_day_temp |
A plain average | 0 / −0.083 / −0.017 / −0.011 |
Experiments not chosen
Each was run with the same held-back test and a rule fixed beforehand. None changed the submission.
- Experiment 2 (line + random-forest hybrid): dropped by its own held-back rule.
code/exp02_*.py,outputs/exp02/. - Experiment 4 (city locations): regional slopes from cities within 500 or 1,000 km, and yearly gaps filled from neighboring challenge cities. No candidate led the shared slope by 0.01, so the tie rule kept the simpler one for cdd. For temperature it picked the 1,000 km regional slope (+0.158 vs +0.157), and Experiment 5 then replaced that pick.
code/exp04_*.py,outputs/exp04/. - Experiment 5C (recent warming rate): shared slopes fitted on only the last 15 or 20 years. This was declared: the idea came from a post-hoc check showing our temperature forecast runs cold against 2019–2025 observations. The candidates were fixed beforehand and judged only on the held-back origins. The best (+0.185 for temperature) led Experiment 5B's pick by 0.008, under the 0.01 tie rule, so
outputs/exp05c/submission_exp05c.csvis identical tosubmission.csv.code/exp05c_recent_slope.py,outputs/exp05c/. - Experiment 6 (nearby stations): up to 8 NOAA GHCN-Daily stations within 100 km of each city (1,396 stations, 1969–2018, every download capped at 2018-12-31). Their averaged anomalies were used for a regional level or slope. The best result fell short of Experiment 5 (temperature +0.162 vs +0.177, cdd +0.080 vs +0.086). Our reading: part of a station's recent shift is local (site changes, growth around the airport), and averaging with neighbors dilutes it.
code/exp06_*.py,outputs/exp06/. - Experiment 7 (medians and partially pooled slopes): medians fit the MAE score, and the pooled slopes blended each city's own slope with the shared one using empirical-Bayes weights. Neither beat Experiment 5 (temperature +0.172 / +0.174 vs +0.177). For temperature the data put only about 14% weight on each city's own slope (
outputs/exp07/pooling.json).code/exp07_median_pooled.py,outputs/exp07/. - Experiment 9 (fill missing days from nearby stations): an accurate fill (hiding a 30-day block, the yearly error drops from 1.14 to 0.05 °F and from 126 to 8.5 °F-days;
outputs/exp09/mask_test.csv). The held-back gain was below 0.01 for every metric, so we kept the original data.code/exp09_daily_gapfill.py,outputs/exp09/. - Experiment 10 (weighted blend): rule written first (
docs/exp10_blend_rule.md). Per metric, a blend of the Experiment 5 and 5B candidates, tested leave-one-origin-out. No metric passed (temperature +0.163 vs +0.177, cdd +0.111 vs +0.127), sooutputs/exp10/submission_exp10.csvis identical tosubmission.csv.code/exp10_blend.py,outputs/exp10/.
Post-hoc scoring declaration
After each forecast was frozen and committed, we scored it against NOAA GHCN-Daily observations, for analysis only. These observations were never a forecast input and never changed any choice: every decision above was made on the held-back years 1998, 2004 and 2011. We built the truth values from NOAA because the official answer key is not public. Our reproduction of the leaderboard's "10-year mean" reference row gives the same Score (10.4), with MAE values within about 1–2%, so treat these numbers as approximate. Scoring code: code/fetch_truth.py, code/truth.py, code/starter_score.py, code/score_all_submissions.py; the model code never reads them.
| Model | Score 2019–2021 | Score 2022–2025 |
|---|---|---|
| 30-year average (reference) | 0.0 | 0.0 |
| Exp03 (earlier submission) | 13.8 | 16.3 |
| Exp05 (earlier submission) | 14.1 | 17.8 |
| Exp05B (this submission) | 13.8 | 18.4 |
Experiment 5B was frozen even though it scores below Experiment 5 on 2019–2021: the scores did not change the choice. One station, USW00094014, has no data after 2020. The Exp05 and Exp05B differences are small next to year-to-year swings. Full tables: outputs/exp05/scorer_posthoc.txt, outputs/exp05b/scorer_posthoc.txt, outputs/exp10/scorer_posthoc.txt, outputs/score_all/all_scores.csv.
AI Tools and Agents
- Tools and models: Claude Code (Anthropic) running Claude Sonnet 5.5 (
claude-sonnet-5-5) in VS Code, for Experiments 2–3 and the first policy analysis. Claude Code in the Claude desktop app running Claude Opus 5.5 (claude-opus-5-5) for Experiments 5–10, the grid-strain and affordability essay, the slide deck, and moving the GitHub repository into this dataset. - What they did: wrote the code in
code/from the team's written specs (docs/specs/) and instructions. They suggested candidate ideas (recent level, station-aware level, scaled cdd slope, nearby stations, the daily-distribution and gap-fill methods), and a team member chose which to test. They ran the held-back tests and the post-hoc scorer, and found and fixed thevalid_flagsindex bug. They downloaded the GHCN-Daily station data with a server-side end date of 2018-12-31, pinned the dataset revision, and checked that a clean rebuild matchessubmission.csvbyte for byte. For the policy idea they found and tested the data sources, wrotecode/policy_*.py,code/census_acs.pyandcode/export_deck_data.py, and compileddata/policy_rules.csvfrom web searches (each row cites its source). They drafteddocs/policy_essay_outline.md(revised after a team-requested review for contradictions) and builtpresentation/from the team's Figma layout. They also drafted this README and the data card. - What they did not do: they supplied no forecast values: every
predicted_valuecomes from the code and the dataset. They downloaded no weather after 2018 for any forecast. Post-2018 NOAA data was used twice, both times approved by a team member and for analysis only: the post-hoc scoring and the policy analysis. - What a person reviewed: the team wrote the specs and decided the rules and settings. We decided each iteration of the experiments and what to change to improve accuracy; Claude Code implemented our changes and informed our decisions. We used Claude to explore general policy directions and to make incremental pivots that tested our thesis. The team chose the policy question, the definitions of "affordable" and "livable" and the national focus, and made the base deck design in Figma for Claude to iterate on.
- Agent instructions file:
code/AGENTS.md(holds the data rule). The code reproducessubmission.csvwithout the agent.
Policy Idea
Plan on the weather that's coming: keep grids reliable and cooling affordable as U.S. summers warm. Full reasoning, with every source: docs/policy_essay_outline.md. Slides: the deck.
- What risk did we notice? Grid reserves (NERC's 50/50 peak forecasts) and household budgets are planned on past weather. Our forecast, made only with data up to 2018, compares 2019–2025 with 1976–2005 and puts the median city at +21% cooling-required days (high ≥ 95 °F), +48% hot nights (low ≥ 75 °F) and +19% cooling degree days. The rise is concentrated in Texas, the Gulf Coast and the Southwest. Each 1 °F on a summer day adds about 2% to peak demand (EIA-930, 2015–2018). By 2025, warming uses about a fifth of the reserve cushion NERC's areas planned above their reference level. ERCOT, already below its reference, falls to −1.3% (short of capacity) in a one-in-ten hot summer. Bills were a problem before any warming. A low-income single person was over an affordable bar in 45 of 46 states, and the forecast heat adds only $17–33 a year. (The bar is 6% of a poverty-line income times electricity's share of home energy spending.) Livability fails in two ways. In the North and West, many homes have no AC (33% in Washington). In the South, AC costs too much to run or the power goes out, and hot nights climb (Florida: 69 → 96 a year).
- What would we do about it, and who would do it? (a) NERC, regional grid operators and state utility commissions add a warming adjustment to the 50/50 peak forecast. They start where margins are at or below the reference level (ERCOT, MISO). (b) State commissions and utilities offer income-based rates, and LIHEAP funds summer cooling, not only winter heating. (c) Cities and states bring cooling to homes without it (heat pumps in rentals, an indoor limit of about 82 °F) and bar utility shut-offs on cooling-required days. The lever that serves all three is efficient cooling for low-income homes: heat pumps in place of old AC and electric-resistance heat, weatherization, and demand response at the peak.
- Why is it sustainable? It lowers peak demand instead of adding peaker plants. Heat pumps replace resistance and oil heat, so winter savings help pay for summer cooling. It runs through institutions that already exist (NERC assessments, utility efficiency programs, LIHEAP), and it plans for the trend our forecast caught, not for one heat wave. Limits: there is no humidity data, we assume 50/50 forecasts exclude warming, and our forecast runs low on heat days (deck slide 6).
Evidence: outputs/grid_heat_days/, outputs/grid_national/, outputs/grid_prices_national/, outputs/grid_projection_national/ (code: code/policy_grid_*.py). Every weather-driven input is dated 2018 or earlier; 2019–2025 values come from our forecast. Earlier versions of this idea were small mixed-use buildings in Pittsburgh and cooling-ready rentals for Seattle. We dropped Pittsburgh because it has few homes without AC and no rise in heat days (outputs/policy_check/thesis_check.md). Seattle is now part of lever (c) (outputs/cooling_gap/summary.md).
Feedback to the Ecosystem
These are the barriers we hit while reusing the ProtoStar climate artifact and linking it to the energy, housing and policy data a heat-risk study needs. Each item gives the evidence (with a way to check it), what it means for climate-risk research beyond this challenge, and what would help the next team.
The artifact's missing-data rules live in people's heads, not in its metadata.
- Evidence: in
annual_indices, 1,184 rows ofMax Tmaxhold-inffor years with no valid days (e.g.,USC00380072, 1895–1899). That breaks standard tools: scikit-learn fails with "Input y contains infinity". No column says whether a yearly value is complete enough to use. We had to derive a rule for each metric fromcoverage_and_stations:Miss Prcp;Miss Tmax; the larger ofMiss TmaxandMiss Tmin; and for cdd, days in the year minusDays Both. A one-line index bug in that derivation silently flipped about a quarter of the cdd validity flags at origin 1998 (see the Method section). Verify:np.isinf(annual['Max Tmax']).sum();code/shared.py,valid_flags. - What it means: every reuser of a derived climate index rebuilds its completeness rule. Two studies of the same 210 cities can then use different city-years and get different trends without either one being wrong. The rule matters most where the data is thinnest, in the early years and at closed stations.
- What would help: store NaN for undefined values. Publish a per-metric missing-day count or validity flag next to each yearly value, plus the code that computed it. Add a machine-readable data dictionary (for example, a Croissant or Frictionless schema) giving units, sentinels and the completeness threshold.
- Evidence: in
A city's record joins several stations, but nothing marks where the joins could shift the data.
- Evidence:
coverage_and_stationshas aStation Segmentcolumn, so the joins are visible, but there is no flag for breaks in the series or for adjustments. We tested a level that uses only the years at each city's latest station (Experiment 5). We also tested averaging nearby GHCN stations (Experiment 6), which pulled skill down (temperature +0.162 vs +0.177). Our reading is that part of each station's recent shift is local (site moves, growth around airports) and averaging dilutes it. We could not tell the two apart with the published metadata. - What it means: trend studies built on city records mix climate change with station history. Researchers who connect these trends to heat-health or energy outcomes inherit the mix without knowing it.
- What would help: publish break dates for each segment with any homogeneity test results. Link each segment to its GHCN-Daily ID and NOAA station history (HOMR) records, so a reuser can test a site-change effect instead of guessing.
- Evidence:
Results reproduce only if the data version is pinned, and evaluations cannot be checked from outside.
- Evidence: we pin the challenge dataset to revision
bbbdebd…(code/shared.py) and log it in everyrun_log.json, because a dataset revision on the Hub can change under the same name. The ground truth for scoring is private:hf_hub_download('CMU-OSPO/protostar-climate-ground-truth', …)fails without a token. So we rebuilt the 2019–2025 truth from GHCN-Daily. It reproduces the published reference Score (10.4) but not its MAE values exactly (precipitation 5.802 vs 5.782). - What it means: a forecast benchmark that cannot be rescored from outside cannot be built on. Later work cannot show that a new method beats an old one on the same truth.
- What would help: cite datasets by revision hash or DOI version, not by name alone. Once a scoring window closes, release the truth table (or its build script and a checksum) together with the scoring code.
- Evidence: we pin the challenge dataset to revision
Linking a weather station to the people and grids it represents means hand-built crosswalks every time.
- Evidence: our policy analysis had to place each station in a county (FCC Area API, cached in
outputs/policy_check/station_county.csv). It then joined counties to ZIP code areas using the 2010 ZCTA file, since the Heat & Health Index is published on 2010 ZCTAs. It assigned cities to EIA demand regions by state and mapped those regions to NERC assessment areas by hand, with special cases for El Paso and Amarillo (AREAS,CITY_AREAincode/policy_grid_projection_national.py). Cities near borders are assigned only approximately. - What it means: climate-to-impact research (health, energy burden, grid reliability) depends on these joins. Each team rebuilds them with different, undocumented choices, so results across studies are hard to compare, and the errors cluster at grid and state borders.
- What would help: publish a standard crosswalk with the climate artifact: station to county, ZCTA, EIA balancing authority and NERC assessment area, with the vintage of each geography.
- Evidence: our policy analysis had to place each station in a county (FCC Area API, cached in
The numbers that drive grid planning are locked in PDFs.
- Evidence: NERC's anticipated and reference reserve margins (2018 LTRA, Tables 3.1–3.2) are only in a PDF, and we extract them with
pypdf. ERCOT's capacity plan and EIA's RECS 2020 state energy spending table are also PDFs. - What it means: connecting a climate forecast to how much reserve the grid holds is the core of a reliability question. When the planning side is not machine-readable, few studies make that connection, and those that do cannot easily be rerun when the next assessment comes out.
- What would help: NERC, ISOs and EIA publish their assessment tables as CSV or an API with stable area IDs and stated assumptions. That includes whether a 50/50 peak forecast already includes a warming trend; we had to assume it does not.
- Evidence: NERC's anticipated and reference reserve margins (2018 LTRA, Tables 3.1–3.2) are only in a PDF, and we extract them with
There is no open, machine-readable record of heat policy or of the housing it covers.
- Evidence: we found no list of rental cooling rules (indoor maximum temperatures) or of building performance standard thresholds. We compiled 12 rows by hand from city codes and news reports (
data/policy_rules.csv, each row with its source). ASHRAE design conditions, the weather that buildings are sized for, are copyrighted, so we used 1976–2005 values from the challenge data instead. Building records do not line up either. Allegheny County records year built for only 2.4% of its 4,100 apartments-over-shops parcels, because they are classed commercial. Pittsburgh, New York (PLUTOlanduse 04) and Cook County (classes212/318) each code mixed use differently (data/policy_sources.md). - What it means: researchers cannot measure, at scale, which heat-exposed homes are covered by a cooling rule, or check whether design weather has drifted from current climate. Policy assessments stay case studies.
- What would help: an open registry of local heat and cooling rules (for example, kept by DOE or a city network), building records with a common use code, and design-condition values derived openly from NOAA data.
- Evidence: we found no list of rental cooling rules (indoor maximum temperatures) or of building performance standard thresholds. We compiled 12 rows by hand from city codes and news reports (
Key exposure data is missing, short or modelled.
- Evidence: the artifact has no humidity, so we used dry-bulb thresholds (95 °F days, 75 °F nights). These understate risk in humid places, and survival limits depend on wet-bulb temperature (Vecellio et al. 2022; Vanos et al. 2023). EIA-930 hourly demand starts in July 2015, so we had only 3.5 summers to estimate heat sensitivity in most regions. The only national data on homes without air conditioning is Census LACE 2023, an experimental, modelled estimate. The Census ACS API now requires a personal key; the keyless route is a 48 MB file per table.
- What it means: heat-health and energy-burden work leans on the variables that are easiest to get, not the ones that matter. Results are least reliable in the humid South, which is where our forecast puts the largest rise in hot nights.
- What would help: add dew point or wet-bulb temperature to the city records from the same stations, continue and validate an AC-prevalence measure, and keep keyless bulk downloads alongside APIs.
What our results suggest for later research
- Nights and the hot tail behave differently. In the daily data up to 2018, mean highs and especially lows rose, but the hot tail of daily highs barely moved (shared slope −0.005 to +0.014 °F/yr,
outputs/exp05b/slopes.csv). Studies of heat risk should track night-time minimums, not only peak days. - Pooling beats going it alone. One warming slope shared by all cities beat each city's own trend for temperature and cdd, on average and in 5 of 6 metric-origin pairs (
outputs/exp03/heldout_results.csv). The data put only about 14% weight on a city's own slope (outputs/exp07/pooling.json). Single-station trends are too noisy to use alone for local planning. - Not every heat signal is a regional signal. Neighbor stations and gap filling did not improve the forecast (Experiments 6 and 9) even when the fill itself was accurate. This points to station history (item 2) as a source of error worth studying in its own right.
- Heat moved electricity use but not price. Across 46 states and 2002–2018, a hotter year did not change the real price of electricity (−0.002 ¢/kWh per 100 cdd, standard error 0.023,
outputs/grid_prices_national/). Cost risk from heat runs through usage and grid scarcity, which is where later work should look.
Files
| File | Content |
|---|---|
submission.csv |
Our forecasts: city_id,year,metric,predicted_value (Experiment 5B) |
code/ |
Every script; see code/README.md. The method is exp05b_daily_distribution.py, its helpers are shared.py, exp05_levels_slopes.py, exp03_shared_slope.py and exp02_line_forest.py. The other files are experiments not chosen, post-hoc scoring and the policy analysis. |
data/ |
The curated data our code uses; described in its data card, data/README.md |
outputs/ |
Results of every experiment and analysis, each with its decision, config, run log and summary |
docs/ |
The policy essay outline, the Experiment 10 rule, the team's experiment specs (specs/), and the challenge rules for reference (challenge/) |
index.html, presentation/ |
The policy slide deck (live page) |
obs_2018_2021.csv, obs_2021_2025.csv |
NOAA observations 2018–2025, analysis only (post-hoc scoring and policy checks, approved by a team member); never read by the forecast code |
image.png, image-1.png |
Screenshots of the deck |
Metric codes: annual_avg_temp (°F), total_precip_in (inches), hottest_day_temp (°F), cdd (°F-days).
City IDs: the station ID from the city_info table of the training dataset (e.g., Pittsburgh = USW00094823).
How to Reproduce
- Install the packages:
pip install -r code/requirements.txt(pinned versions, Python 3.14). - Run
python code/exp05b_daily_distribution.py. It downloads the challenge dataset at the pinned revisionbbbdebdab2a18e00f97ef87bd3f35a71e222180b(shared.REVISION) and writesoutputs/exp05b/submission_exp05b.csv, which is identical tosubmission.csv. It takes under a minute and has no randomness. - Check it:
python code/check_submission.py submission.csv. - Optional, the other experiments:
- Experiment 5:
python code/exp05_levels_slopes.py. - Experiment 3:
python code/exp03_shared_slope.py. - Experiment 6:
python code/exp06_fetch_neighbors.py(about 35 minutes of downloads; it can use the curateddata/neighbors/files instead), thenpython code/exp06_neighbors.py. - Experiment 9:
python code/exp09_daily_gapfill.py(needs the neighbor daily files fromexp06_fetch_neighbors.py). - Experiment 10:
python code/exp10_blend.py.
- Experiment 5:
- Optional, the policy analysis: run
code/policy_grid_*.py, thenpython code/export_deck_data.py, which rebuildspresentation/data/deck_data.js. Openindex.htmlin a browser (or runpython -m http.server 8765). - Optional, post-hoc scoring (analysis only):
python code/starter_score.py obs_2021_2025.csv 2022-2025.
Data
- External data:
- Experiments only, not chosen: NOAA NCEI GHCN-Daily (Menne et al. 2012), daily TMAX/TMIN/PRCP for 1,396 nearby stations, 1969–2018 only. This was an input to the candidates in Experiments 6 and 9; none was chosen, so the submitted forecast does not use it.
- Analysis and policy idea only, never read by the forecast code:
- NOAA GHCN-Daily observations 2019–2025 for the 210 stations.
- Census ACS 5-year 2019–2023 (B25024, B25034, B25003); FCC Area API.
- Allegheny County (WPRDC), NYC PLUTO and Cook County Assessor building data.
- Census LACE 2023; CDC/ATSDR Heat & Health Index 2024; Census 2010 ZCTA-to-county file.
- Electricity sources, read only up to 2018 where they respond to the weather: EIA-930 (Jul 2015–2018), ERCOT hourly load (2002–2018), EIA-861 (2002–2018), EIA RECS 2015 and 2020, BLS CPI-U, NERC 2018 LTRA, HHS 2018 poverty guidelines.
data/policy_rules.csv, compiled by hand.
- Curated data: see the data card in
data/README.md(each file, its source, license and preparation) anddata/policy_sources.md.
Training data (the only input to the submitted forecast): CMU-OSPO/protostar-climate-train-to-2018, tables annual_indices, coverage_and_stations and daily, pinned to revision bbbdebdab2a18e00f97ef87bd3f35a71e222180b, from:
- 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
- Downloads last month
- 128

