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
3.01
5.15k
USC00042863
2,019
annual_avg_temp
66.9
USC00042863
2,019
total_precip_in
13.03
USC00042863
2,019
hottest_day_temp
103.97
USC00042863
2,019
cdd
1,762.69
USC00042863
2,020
annual_avg_temp
66.96
USC00042863
2,020
total_precip_in
13.03
USC00042863
2,020
hottest_day_temp
103.98
USC00042863
2,020
cdd
1,774.25
USC00042863
2,021
annual_avg_temp
67.01
USC00042863
2,021
total_precip_in
13.03
USC00042863
2,021
hottest_day_temp
103.99
USC00042863
2,021
cdd
1,785.81
USC00042863
2,022
annual_avg_temp
67.07
USC00042863
2,022
total_precip_in
13.03
USC00042863
2,022
hottest_day_temp
103.99
USC00042863
2,022
cdd
1,797.38
USC00042863
2,023
annual_avg_temp
67.12
USC00042863
2,023
total_precip_in
13.03
USC00042863
2,023
hottest_day_temp
104
USC00042863
2,023
cdd
1,808.94
USC00042863
2,024
annual_avg_temp
67.18
USC00042863
2,024
total_precip_in
13.03
USC00042863
2,024
hottest_day_temp
104.01
USC00042863
2,024
cdd
1,820.5
USC00042863
2,025
annual_avg_temp
67.24
USC00042863
2,025
total_precip_in
13.03
USC00042863
2,025
hottest_day_temp
104.02
USC00042863
2,025
cdd
1,832.39
USC00166584
2,019
annual_avg_temp
67.56
USC00166584
2,019
total_precip_in
56.78
USC00166584
2,019
hottest_day_temp
101.15
USC00166584
2,019
cdd
2,814.67
USC00166584
2,020
annual_avg_temp
67.6
USC00166584
2,020
total_precip_in
56.78
USC00166584
2,020
hottest_day_temp
101.16
USC00166584
2,020
cdd
2,824.95
USC00166584
2,021
annual_avg_temp
67.65
USC00166584
2,021
total_precip_in
56.78
USC00166584
2,021
hottest_day_temp
101.18
USC00166584
2,021
cdd
2,835.22
USC00166584
2,022
annual_avg_temp
67.7
USC00166584
2,022
total_precip_in
56.78
USC00166584
2,022
hottest_day_temp
101.19
USC00166584
2,022
cdd
2,845.5
USC00166584
2,023
annual_avg_temp
67.75
USC00166584
2,023
total_precip_in
56.78
USC00166584
2,023
hottest_day_temp
101.2
USC00166584
2,023
cdd
2,855.78
USC00166584
2,024
annual_avg_temp
67.8
USC00166584
2,024
total_precip_in
56.78
USC00166584
2,024
hottest_day_temp
101.22
USC00166584
2,024
cdd
2,866.05
USC00166584
2,025
annual_avg_temp
67.85
USC00166584
2,025
total_precip_in
56.78
USC00166584
2,025
hottest_day_temp
101.23
USC00166584
2,025
cdd
2,876.51
USC00280734
2,019
annual_avg_temp
51.68
USC00280734
2,019
total_precip_in
44.92
USC00280734
2,019
hottest_day_temp
94.87
USC00280734
2,019
cdd
800.75
USC00280734
2,020
annual_avg_temp
51.73
USC00280734
2,020
total_precip_in
44.92
USC00280734
2,020
hottest_day_temp
94.88
USC00280734
2,020
cdd
805.95
USC00280734
2,021
annual_avg_temp
51.78
USC00280734
2,021
total_precip_in
44.92
USC00280734
2,021
hottest_day_temp
94.89
USC00280734
2,021
cdd
811.16
USC00280734
2,022
annual_avg_temp
51.82
USC00280734
2,022
total_precip_in
44.92
USC00280734
2,022
hottest_day_temp
94.9
USC00280734
2,022
cdd
816.36
USC00280734
2,023
annual_avg_temp
51.87
USC00280734
2,023
total_precip_in
44.92
USC00280734
2,023
hottest_day_temp
94.91
USC00280734
2,023
cdd
821.57
USC00280734
2,024
annual_avg_temp
51.92
USC00280734
2,024
total_precip_in
44.92
USC00280734
2,024
hottest_day_temp
94.92
USC00280734
2,024
cdd
826.77
USC00280734
2,025
annual_avg_temp
51.97
USC00280734
2,025
total_precip_in
44.92
USC00280734
2,025
hottest_day_temp
94.93
USC00280734
2,025
cdd
831.98
USC00286055
2,019
annual_avg_temp
54.67
USC00286055
2,019
total_precip_in
49.5
USC00286055
2,019
hottest_day_temp
97.65
USC00286055
2,019
cdd
1,122.12
USC00286055
2,020
annual_avg_temp
54.72
USC00286055
2,020
total_precip_in
49.5
USC00286055
2,020
hottest_day_temp
97.66
USC00286055
2,020
cdd
1,127.96
USC00286055
2,021
annual_avg_temp
54.77
USC00286055
2,021
total_precip_in
49.5
USC00286055
2,021
hottest_day_temp
97.67
USC00286055
2,021
cdd
1,133.81
USC00286055
2,022
annual_avg_temp
54.81
USC00286055
2,022
total_precip_in
49.5
USC00286055
2,022
hottest_day_temp
97.68
USC00286055
2,022
cdd
1,139.65
End of preview. Expand in Data Studio

Carbon Copies: 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.

In one line: the city's recent level plus a warming trend pooled across neighboring cities. Cooling degree days come from replaying each city's last 30 years of daily weather, warmed to that forecast. In backtests it beats the 30-year average by 11–17% on temperature and 8–17% on CDD. On precipitation, honestly, nothing we tried beat the 30-year average. A look-ahead test (code/check_leakage.py) proves the model uses no data after its forecast date.

flowchart LR
    A["Challenge data, to 2018-12-31<br/>annual indices + daily Tmax/Tmin"] --> B["Validity: ≀ 10 missing days<br/>(the scoring rule)"]
    B --> S["Station-change correction<br/>(32 moves in 1989–2018)"]
    S --> T["Avg temperature<br/>15-yr level + trend pooled<br/>over cities within 1,000 km"]
    A --> D["Daily β†’ CDD-if-warmer curves<br/>(data/cdd_shift_curves.parquet)"]
    T --> C["CDD: delta method<br/>last 30 yrs of daily weather,<br/>warmed to the forecast"]
    D --> C
    S --> H["Hottest day: 30-yr mean<br/>+ Β½ pooled trend"]
    B --> P["Precipitation: 30-yr mean"]
    T --> O["submission.csv"]
    C --> O
    H --> O
    P --> O
    O --> X["outputs/: GeoJSON, wide CSV,<br/>cooling-gap cities (policy)"]
    L["check_leakage.py: future scrambled<br/>β†’ identical forecasts"] -.-> O

Team

Name Hugging Face username University, Program
Naina Dhingra @ndhingraa Carnegie Mellon University
Other members: to be added

Method

Model (per metric, code/forecast.py):

Metric Forecast
annual_avg_temp Mean of the last 15 years + regional trend Γ— years since the middle of that window
cdd Delta method: take every day of the city's last 30 valid years and warm it by (temperature forecast βˆ’ mean temperature of those years). Recompute each year's CDD and average
hottest_day_temp 30-year mean + half the regional trend
total_precip_in 30-year mean (nothing beat it; see below)
  • Regional trend: the slope of the 50-year anomaly series averaged over all cities within 1,000 km. One station's 30-year trend is mostly noise. Pooling neighbors keeps the regional warming signal and averages away station quirks.
  • Station changes: 32 station moves fall in 1989–2018. Years recorded at an earlier station are shifted to the current station's level. The step is estimated from 10 years each side of the move, net of the national anomaly, so warming is not mistaken for a station jump.
  • Validity: we drop years with more than 10 missing days, the same rule the challenge scores with.
  • Why the delta method for CDD: CDD only counts warmth above 65 Β°F. A city with many days just under 65 Β°F gains a lot of CDD per degree of warming, while a city that is rarely warm gains little. A yearly trend line cannot see that; the daily record can. code/prepare_data.py condenses the 10-million-row daily table into a 2 MB table: each city-year's CDD if every day were s Β°F warmer, s = βˆ’3 … +6 in 0.25 Β°F steps. At s = 0 it matches the official cdd column exactly.
  • Learned models (code/learned.py, tested, not in the final model): ridge and gradient boosting, trained across cities on past forecast origins. The target is the standardized departure from the 30-year mean. Features are recent anomalies of the city, its region and the nation, plus trends Γ— lead time. Training uses only origins whose targets were fully observed (origin + 7 ≀ forecast origin).

How we checked it: a rolling-origin backtest on data up to 2018 only. From each origin year we forecast the next 7 years with only earlier data and score with the challenge formula, 100 Γ— mean of (1 βˆ’ MAE / MAE of the 30-year average). We chose settings on origins 1997–2004 ("tune") and confirmed them on 2005–2011 ("check"). We never looked at 2019–2025 observations, and we did not use the live leaderboard to choose between models. code/check_leakage.py reruns the model from 1998, 2004 and 2011 with every later value replaced by random noise and requires identical forecasts. It also confirms that a deliberately leaky model is caught.

Final model, backtest (code/backtest_results.csv). Each metric's number is the improvement over the 30-year average; 0 = no better.

Split Score Avg temp Precip Hottest day CDD MAE: temp Β°F / precip in / hottest Β°F / CDD Β°F-days
tune (1997–2004 origins) 6.14 +0.166 0.000 +0.003 +0.077 0.95 / 5.72 / 2.58 / 152
check (2005–2011 origins) 7.24 +0.114 0.000 +0.008 +0.169 1.32 / 6.00 / 2.68 / 165

What we tried (python code/backtest.py compare β†’ code/experiments.csv; same backtest):

Model Score tune / check Avg temp Precip Hottest day CDD
baseline: last year (score floored at 0) 0.00 / 0.00 -0.592 / -0.465 -0.297 / -0.323 -0.427 / -0.433 -0.527 / -0.478
baseline: city linear trend 30y 1.17 / 2.00 +0.135 / +0.070 -0.060 / -0.061 -0.091 / -0.061 +0.012 / +0.107
mean of last 10 years 2.59 / 3.46 +0.128 / +0.069 -0.024 / -0.012 -0.029 / -0.017 +0.029 / +0.099
mean of last 15 years 4.12 / 2.94 +0.143 / +0.069 -0.005 / -0.005 -0.017 / -0.002 +0.043 / +0.056
level15 + pooled trend (one setting) 4.50 / 6.04 +0.166 / +0.114 -0.022 / -0.001 -0.014 / +0.006 +0.049 / +0.123
level+trend, per-metric PARAMS 5.45 / 6.12 +0.166 / +0.114 +0.000 / +0.000 +0.003 / +0.008 +0.049 / +0.123
ridge, all origins since 1940 4.71 / 0.61 +0.182 / -0.008 -0.021 / -0.017 -0.019 / -0.032 +0.047 / +0.007
ridge, last 30 origins 4.24 / 2.24 +0.208 / +0.051 -0.033 / -0.021 -0.039 / -0.013 +0.034 / +0.073
GBM, all origins since 1940 4.72 / 2.90 +0.150 / +0.044 -0.007 / -0.030 -0.017 / -0.003 +0.063 / +0.106
GBM, last 30 origins 5.75 / 3.98 +0.199 / +0.084 -0.008 / -0.039 -0.025 / -0.021 +0.065 / +0.135
temp: trend fitted since 1975 – (temp only) +0.166 / +0.111 – – –
temp: trend x1.5 (amplified) – (temp only) +0.103 / +0.102 – – –
temp: Tmax and Tmin forecast separately – (temp only) +0.166 / +0.114 – – –
CDD from temperature forecast – (CDD only) – – – +0.059 / +0.139
CDD: from temperature + GBM (previous final) – (CDD only) – – – +0.070 / +0.143
CDD: delta method, 15 years – (CDD only) – – – +0.063 / +0.153
CDD: delta method, 30 years – (CDD only) – – – +0.077 / +0.169
FINAL (forecast.py) 6.14 / 7.24 +0.166 / +0.114 +0.000 / +0.000 +0.003 / +0.008 +0.077 / +0.169

What we learned:

  • Warming is the signal; everything else is noise. All of the gain is in temperature and CDD. Precipitation and the hottest day vary so much from year to year that every attempt to follow recent years or trends did worse than the 30-year mean. We kept the 30-year mean there on purpose.
  • Pool trends, don't fit them per station. A per-city linear trend (a challenge baseline) is unstable. A trend pooled within 1,000 km is steadier and better in both splits.
  • Learned models trained on the whole record fail on recent decades. Trained on origins since 1940, ridge and gradient boosting learned the mid-century cooling and post-1980 warming as one relationship. They score about 4.7 in the tune split but only 0.6–2.9 in the check split. Restricting training to the last 30 origins helps: gradient boosting then has the best temperature number in the tune split. Alone, though, they still lose to the simple model in the check split, which is closer to the regime we forecast. A gradient-boosting + CDD-from-temperature blend was our previous final model; the delta method beat it in both splits with no learned parameters.
  • CDD is better forecast through temperature than on its own. CDD is a nonlinear function of temperature, and the temperature trend is estimated far more reliably. A per-city CDD-per-Β°F slope helped. Replaying the daily weather (delta method) helped more, and 30 years of days beat 15, because more summers sample the tail of hot days better.
  • Making the trend steeper did not help. Fitting the trend only since 1975 tied our 50-year window, and forecasting Tmax and Tmin separately gave identical results. Multiplying the trend by 1.5 lowered the backtest score in both splits, so we did not do it, even though 2019–2021 were warm years.
  • Caveat: we chose the 50-year trend window and the 30-year delta window partly by the check split, so the check score is slightly optimistic. The tune score is the cleaner estimate.

AI Tools and Agents

  • Tools and models: Claude Code (Claude Opus 5.5), an AI coding agent, in VS Code.
  • What they did: read the challenge pages, the template and other teams' public READMEs (for ideas; the delta method for CDD was used by other teams first, and we adopted it only after it won our own backtest). Downloaded the training data (2018 cutoff enforced in code), wrote the code in code/, ran the backtests, the experiment comparison and the leakage test, and drafted this README and the data card.
  • What a person reviewed: To be completed by the team before the deadline.
  • Agent instructions file: code/AGENTS.md (template rules plus our team notes). Guardrails: load_panel() stops if any input is dated after 2018. Every forecast value comes from code/forecast.py; the agent wrote no value itself.

Policy Idea

Cooling readiness for "newly hot" cities.

  1. What risk did we notice? Cooling demand rises almost everywhere: CDD goes up in 209 of 210 cities, by a median of 10%, in our 2019–2025 forecast versus 1989–2018. The biggest relative jumps are not in the South. They are in the 101 cities that historically needed little cooling (under 1,000 Β°F-days), where the median rise is 13.5%, against 8.6% elsewhere. The steepest: Sault Ste. Marie +33%, Elko +28%, San Diego +26%, Duluth +25%, Bluefield +23%, Helena +22%, Vancouver WA +21%, Rochester MN +21%, Colorado Springs +21%, Lansing +21%, Reno +20%, with Seattle (+20%), Burlington VT and Worcester (+19%) close behind (outputs/cdd_change_by_city.csv). Phoenix adds more Β°F-days, but it already has the air conditioning, building codes and heat plans to handle them. In mild-climate cities, many homes, schools and care facilities have no cooling, and residents and health systems have little heat experience. The same rise in heat does more harm there.
  2. What would we do about it, and who would do it? A "cooling-gap" trigger. Each year, a state climate office or university (e.g., a state climatologist) reruns this open pipeline on the newest records and flags cities whose cooling demand is rising fast from a low base. Flagged cities' health departments and electric utilities then (a) add cooling (heat pumps, which also replace inefficient heating) to existing low-income weatherization and efficiency programs, prioritizing renters, older adults and people living alone; (b) require that new and renovated multifamily housing and schools can be cooled; and (c) adopt a heat-response plan with cooling centers and outreach before the first bad summer, not after.
  3. Why is it sustainable? It uses money and institutions that already exist (utility efficiency programs, weatherization assistance, building-code cycles) instead of a new agency. Heat pumps pay back through lower heating bills, so they make sense even in years without a heat wave. The trigger costs almost nothing to run: the code, data and model are open, run in about a minute, and update as each year's records arrive.

Feedback to the Ecosystem

  • What got in the way:
    • Validity rules lived in prose, not data. The "≀ 10 missing days" rule, and CDD's extra "days with both temperatures" condition, had to be rebuilt from the dataset card. Our count of unscorable city-metrics (10) differs from the challenge's (9), so one borderline rule is ambiguous.
    • Station changes are flagged but not adjusted. coverage_and_stations marks the 32 changes in 1989–2018, but there is no homogenized series, so every team must invent its own adjustment.
    • No uncertainty in the scoring. Point forecasts only. For noisy metrics like precipitation, the most useful output for planners is a range, and MAE rewards forecasting the median.
    • Inconsistent edge values. The extreme columns use Β±inf for years with no data, while other columns use NaN.
  • What would fix it:
    • Publish a boolean valid_<metric> column per city-year using the exact scoring rule, so training and scoring cannot drift apart.
    • Add a homogenized variant (e.g., NOAA's pairwise homogenization, or the offsets from Lai & Dzombak's method) alongside the raw threaded record.
    • Score a quantile or interval forecast (e.g., pinball loss) as an optional track.
    • Use NaN consistently, and add a machine-readable schema (units, missing-value conventions) to the dataset card.
    • Ship the daily data's derived products. The daily table is 10M rows, and most teams only need per-year summaries such as our CDD-if-warmer curves (2 MB). A small "derived" subset would make the daily record usable for teams without much compute.
    • For reuse: our climate_lib.py exposes one forecaster signature, a backtest scorer that reproduces the challenge formula, and check_leakage.leak_test(), which certifies any forecaster against look-ahead. Future editions could ship a shared harness like it, so teams compare models on the same backtest and organizers can verify the honor code automatically.

Files

File Content
submission.csv Our forecasts: city_id,year,metric,predicted_value
code/ Scripts (see code/README.md): forecast.py (final model), climate_lib.py and learned.py (reusable library), backtest.py (experiments), check_leakage.py (look-ahead test), export.py, policy_analysis.py, prepare_data.py, check_submission.py. Results: backtest_results.csv, experiments.csv
data/ The curated data our code uses; described in its data card, data/README.md
outputs/forecasts.geojson One point per city with all forecasts and the change versus 1989–2018. Opens in QGIS, ArcGIS, geojson.io, kepler.gl
outputs/forecasts_wide.csv One row per city and year, a column per metric, with name, latitude and longitude
outputs/cdd_change_by_city.csv Forecast change in cooling demand per city (Policy Idea)

Metric codes: annual_avg_temp (Β°F), total_precip_in (inches), hottest_day_temp (Β°F), cdd (Β°F-days). City IDs: the station ID from the city_info table of the training dataset (e.g., Pittsburgh = USW00094823).

How to Reproduce

  1. python -m venv .venv && source .venv/bin/activate && pip install -r code/requirements.txt (tested with Python 3.14.6)
  2. python code/forecast.py reads data/, prints the backtest, and writes submission.csv and code/backtest_results.csv (about a minute; deterministic, no random parts).
  3. python code/check_submission.py and python code/check_leakage.py
  4. python code/export.py and python code/policy_analysis.py write outputs/.
  5. Optional: python code/backtest.py compare (all experiments, ~15 min); python code/prepare_data.py re-downloads data/ and rebuilds the CDD curves from the daily table.

Data

  • External data: none. Only the challenge training data is used, for both forecasts and analysis.
  • Curated data: see the data card in data/README.md (each file, its source, license and preparation).

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

Downloads last month
103