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
5.2k
USC00042863
2,019
annual_avg_temp
66.89
USC00042863
2,019
total_precip_in
13.03
USC00042863
2,019
hottest_day_temp
103.89
USC00042863
2,019
cdd
1,671.6
USC00042863
2,020
annual_avg_temp
66.95
USC00042863
2,020
total_precip_in
13.03
USC00042863
2,020
hottest_day_temp
103.89
USC00042863
2,020
cdd
1,683.3
USC00042863
2,021
annual_avg_temp
67
USC00042863
2,021
total_precip_in
13.03
USC00042863
2,021
hottest_day_temp
103.89
USC00042863
2,021
cdd
1,694.9
USC00042863
2,022
annual_avg_temp
67.06
USC00042863
2,022
total_precip_in
13.03
USC00042863
2,022
hottest_day_temp
103.89
USC00042863
2,022
cdd
1,706.6
USC00042863
2,023
annual_avg_temp
67.12
USC00042863
2,023
total_precip_in
13.03
USC00042863
2,023
hottest_day_temp
103.89
USC00042863
2,023
cdd
1,718.2
USC00042863
2,024
annual_avg_temp
67.17
USC00042863
2,024
total_precip_in
13.03
USC00042863
2,024
hottest_day_temp
103.89
USC00042863
2,024
cdd
1,729.9
USC00042863
2,025
annual_avg_temp
67.23
USC00042863
2,025
total_precip_in
13.03
USC00042863
2,025
hottest_day_temp
103.89
USC00042863
2,025
cdd
1,741.6
USC00166584
2,019
annual_avg_temp
67.61
USC00166584
2,019
total_precip_in
56.78
USC00166584
2,019
hottest_day_temp
100.92
USC00166584
2,019
cdd
2,792.5
USC00166584
2,020
annual_avg_temp
67.68
USC00166584
2,020
total_precip_in
56.78
USC00166584
2,020
hottest_day_temp
100.92
USC00166584
2,020
cdd
2,804.2
USC00166584
2,021
annual_avg_temp
67.74
USC00166584
2,021
total_precip_in
56.78
USC00166584
2,021
hottest_day_temp
100.92
USC00166584
2,021
cdd
2,816
USC00166584
2,022
annual_avg_temp
67.81
USC00166584
2,022
total_precip_in
56.78
USC00166584
2,022
hottest_day_temp
100.92
USC00166584
2,022
cdd
2,827.7
USC00166584
2,023
annual_avg_temp
67.87
USC00166584
2,023
total_precip_in
56.78
USC00166584
2,023
hottest_day_temp
100.92
USC00166584
2,023
cdd
2,839.5
USC00166584
2,024
annual_avg_temp
67.93
USC00166584
2,024
total_precip_in
56.78
USC00166584
2,024
hottest_day_temp
100.92
USC00166584
2,024
cdd
2,851.2
USC00166584
2,025
annual_avg_temp
68
USC00166584
2,025
total_precip_in
56.78
USC00166584
2,025
hottest_day_temp
100.92
USC00166584
2,025
cdd
2,862.9
USC00280734
2,019
annual_avg_temp
51.58
USC00280734
2,019
total_precip_in
44.92
USC00280734
2,019
hottest_day_temp
93.82
USC00280734
2,019
cdd
741.9
USC00280734
2,020
annual_avg_temp
51.63
USC00280734
2,020
total_precip_in
44.92
USC00280734
2,020
hottest_day_temp
93.82
USC00280734
2,020
cdd
747.9
USC00280734
2,021
annual_avg_temp
51.69
USC00280734
2,021
total_precip_in
44.92
USC00280734
2,021
hottest_day_temp
93.82
USC00280734
2,021
cdd
753.9
USC00280734
2,022
annual_avg_temp
51.75
USC00280734
2,022
total_precip_in
44.92
USC00280734
2,022
hottest_day_temp
93.82
USC00280734
2,022
cdd
759.9
USC00280734
2,023
annual_avg_temp
51.8
USC00280734
2,023
total_precip_in
44.92
USC00280734
2,023
hottest_day_temp
93.82
USC00280734
2,023
cdd
766
USC00280734
2,024
annual_avg_temp
51.86
USC00280734
2,024
total_precip_in
44.92
USC00280734
2,024
hottest_day_temp
93.82
USC00280734
2,024
cdd
772
USC00280734
2,025
annual_avg_temp
51.91
USC00280734
2,025
total_precip_in
44.92
USC00280734
2,025
hottest_day_temp
93.82
USC00280734
2,025
cdd
778
USC00286055
2,019
annual_avg_temp
54.71
USC00286055
2,019
total_precip_in
49.5
USC00286055
2,019
hottest_day_temp
97.5
USC00286055
2,019
cdd
1,109.7
USC00286055
2,020
annual_avg_temp
54.78
USC00286055
2,020
total_precip_in
49.5
USC00286055
2,020
hottest_day_temp
97.5
USC00286055
2,020
cdd
1,117.1
USC00286055
2,021
annual_avg_temp
54.85
USC00286055
2,021
total_precip_in
49.5
USC00286055
2,021
hottest_day_temp
97.5
USC00286055
2,021
cdd
1,124.5
USC00286055
2,022
annual_avg_temp
54.92
USC00286055
2,022
total_precip_in
49.5
USC00286055
2,022
hottest_day_temp
97.5
USC00286055
2,022
cdd
1,132
End of preview. Expand in Data Studio

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, for the Protostar Challenge: Climate Risk Intelligence. Only data dated on or before 2018-12-31 was used as a forecast input (the code asserts it), and every model choice was made on backtests inside those years: no leaderboard score and no 2019–2025 observation was used to choose or tune anything.

Team

Name Hugging Face username University, Program
Alec Alameddine @AlecAlameddine Carnegie Mellon University, M.S. AI and Innovation

Method

  • Model: one small, per-metric model chosen by a backtest ladder (below).
    • annual_avg_temp: a straight line through the city's 1989–2018 mean whose slope is half the city's own least-squares slope and half the mean slope of its 32 nearest cities (great-circle distance). Pooling removes most of the noise that makes per-city trend lines overshoot.
    • cdd: a per-city regression of cooling degree days on annual average temperature (1989–2018), applied to the temperature forecast above. More stable across backtests than a trend fitted to CDD itself.
    • total_precip_in, hottest_day_temp: the 30-year mean (1989–2018). For precipitation every alternative was worse or within noise; for the hottest day a recent-weighted blend gained on the two selection folds but lost on both earlier folds and was vetoed. We ship the baseline rather than a model that only looks better on two folds.
  • Features: each city's own valid years (the organizers' rule: at most 10 missing days), the station latitude/longitude from the dataset (to find neighbors), nothing else.
  • How we checked it: rolling-origin backtests with the real 7-year horizon and the challenge's own scoring (ported from the Space's scoring.py): train ≤ 2011 → score 2012–2018 and train ≤ 2004 → score 2005–2011 rank the candidates; two earlier folds can only veto. The harness reproduces the organizers' published practice-run numbers: CDD MAE 211.7 (30-year average) and 174.6 (linear trend) under the official rule, and both columns of the example notebook's table to three decimals (its 10-year-mean column when scored on the notebook's own rows).
  • Selection rule (mechanical, in code): per metric, candidates climb a ladder: 10-year mean → trend (pooled over k ∈ {8, 16, 32} nearest cities, own-slope weight ∈ {0, ¼, ½, ¾, 1}; for CDD the regression on forecast temperature competes on the same rung) → convex blend of 10-year mean, 30-year mean and trend (weights in steps of 0.1, shared by all cities). The best candidate of a rung replaces the simpler model only if it gains on both selection folds, by more than two bootstrap standard errors, and does not lose to the 30-year average on the earlier folds. The 30-year window and the grids were set before the first backtest. The rule itself was tightened once after a first run on these same folds: the veto by the earlier folds was added (the first version had picked a 20-year mean for the hottest day that the earlier decades contradict) and the two CDD trend families were put on one rung. So the earlier folds are a check we designed after looking, not a pristine holdout; the history is in NOTES.md §3.
  • Leak checks: loaders and predict() assert that no input is dated after 2018; a test corrupts every raw value after each backtest origin, rebuilds the panels and checks that no forecast changes (it fails when a leak is injected); the challenge's discussion threads were screened so that no score or outcome reached the modeling; the leaderboard was not opened until the model was frozen (commit in NOTES.md); external weather downloads carry an explicit 2018-12-31 end date.

Score (0–100, the challenge's formula) by model and fold

Model <=2011 -> 2012-2018
(selection)
<=2004 -> 2005-2011
(selection)
<=1997 -> 1998-2004
(confirmation)
<=1990 -> 1991-1997
(confirmation)
30-year average 0.00 0.00 0.00 0.00
Linear trend 2.31 3.05 0.00 0.00
Last year repeated 0.00 0.00 0.00 0.00
1 climatology: 10-year mean 5.44 4.58 1.31 0.25
2 pooled trend k=8, own-slope weight 0.5 5.14 5.62 0.21 0.00
2 pooled trend k=16, own-slope weight 0.5 5.82 5.93 0.96 0.00
2 pooled trend k=32, own-slope weight 0.5 6.72 6.09 1.23 0.00
2 pooled trend, best (k, weight) per metric 7.25 6.26 1.57 0.00
3 as 2, but cdd ~ temp regression on forecast temp 6.46 8.51 1.87 0.00
4 best blend (recent, long, trend) per metric 8.43 8.93 3.68 0.58
FINAL (selected per metric) 8.22 8.66 4.13 0.76

MAE (skill vs the 30-year average) by metric

Fold Model avg temp (°F) precip (in) hottest day (°F) CDD (°F-days)
<=2011 -> 2012-2018 30-year average 1.710 (+0.0 %) 5.978 (+0.0 %) 2.415 (+0.0 %) 201.3 (+0.0 %)
<=2011 -> 2012-2018 Linear trend 1.488 (+13.0 %) 6.789 (-13.6 %) 2.606 (-7.9 %) 165.7 (+17.7 %)
<=2011 -> 2012-2018 10-year mean 1.563 (+8.6 %) 6.229 (-4.2 %) 2.393 (+0.9 %) 168.2 (+16.4 %)
<=2011 -> 2012-2018 FINAL 1.453 (+15.0 %) 5.978 (+0.0 %) 2.415 (+0.0 %) 165.4 (+17.8 %)
<=2004 -> 2005-2011 30-year average 1.207 (+0.0 %) 6.145 (+0.0 %) 2.550 (+0.0 %) 190.0 (+0.0 %)
<=2004 -> 2005-2011 Linear trend 1.027 (+14.9 %) 6.316 (-2.8 %) 2.693 (-5.6 %) 179.2 (+5.7 %)
<=2004 -> 2005-2011 10-year mean 1.040 (+13.9 %) 6.181 (-0.6 %) 2.565 (-0.6 %) 179.3 (+5.6 %)
<=2004 -> 2005-2011 FINAL 0.983 (+18.6 %) 6.145 (+0.0 %) 2.550 (+0.0 %) 159.5 (+16.1 %)
<=1997 -> 1998-2004 30-year average 1.261 (+0.0 %) 5.525 (+0.0 %) 2.727 (+0.0 %) 158.9 (+0.0 %)
<=1997 -> 1998-2004 Linear trend 1.153 (+8.5 %) 5.966 (-8.0 %) 3.004 (-10.2 %) 166.1 (-4.5 %)
<=1997 -> 1998-2004 10-year mean 1.125 (+10.8 %) 5.676 (-2.7 %) 2.877 (-5.5 %) 154.7 (+2.7 %)
<=1997 -> 1998-2004 FINAL 1.093 (+13.3 %) 5.525 (+0.0 %) 2.727 (+0.0 %) 153.8 (+3.2 %)
<=1990 -> 1991-1997 30-year average 1.079 (+0.0 %) 5.496 (+0.0 %) 2.622 (+0.0 %) 172.5 (+0.0 %)
<=1990 -> 1991-1997 Linear trend 1.108 (-2.7 %) 6.140 (-11.7 %) 3.046 (-16.2 %) 180.5 (-4.6 %)
<=1990 -> 1991-1997 10-year mean 1.006 (+6.8 %) 5.693 (-3.6 %) 2.755 (-5.1 %) 167.5 (+2.9 %)
<=1990 -> 1991-1997 FINAL 1.063 (+1.4 %) 5.496 (+0.0 %) 2.622 (+0.0 %) 169.7 (+1.6 %)

Selected per metric: annual_avg_temp → pooled trend (k=32, own-slope weight 0.5); total_precip_in → 30-year mean; hottest_day_temp → 30-year mean; cdd → cdd ~ temp regression on [pooled trend (k=32, own-slope weight 0.5)].

Reading the table: the shipped system beats the 30-year average, the linear trend and the organizers' 10-year-mean example on both selection folds. The earlier folds are the honest caveat. The veto only asks a model not to lose to the 30-year average there; against the plain 10-year mean the shipped temperature and CDD models are behind on those two folds (mean skill 7.4 % vs 8.8 %, and 2.4 % vs 2.8 %), and ahead on the two recent ones. Trend extrapolation pays off only while the warming of recent decades continues; averaged over all four folds the shipped system scores 5.4, against 4.2 with the 10-year mean in their place. Because the veto decided the hottest-day model, the earlier folds are a clean holdout for the other three metrics only. The bootstrap resamples cities, which are spatially correlated, so its standard errors are lower bounds. Full tables, the ladder with standard errors, every candidate and a sensitivity table: reports/backtest/BACKTEST.md.

Tried and not shipped (numbers in reports/backtest/BACKTEST.md and reports/experiments/EXPERIMENTS.md):

  • worse than the simpler model: blends of 10-year mean, 30-year mean and trend for temperature and CDD; the 10-year mean and trends for precipitation;
  • vetoed by the earlier folds: a direct pooled CDD trend (the regression on temperature outranks it, and it loses to the 30-year average on the earlier folds); the recent-weighted hottest-day blend;
  • within noise: correcting the documented station changes with shifts estimated against neighboring cities (at best +0.3 points for CDD, below two standard errors); shrinking the CDD–temperature coefficient toward neighbors; filling invalid years from NOAA;
  • not candidates, shown as sensitivity only: 40- and 50-year trend windows, pooling over 64 or all cities.

No neural nets or gradient boosting: with 7 test years per city they fit noise. A forecast that knew every city's true 7-year mean would reach only about 28–30 % skill for temperature and 24–33 % for CDD on the selection folds, so little headroom is left above 15–19 %.

Rows without a baseline. Nine city–metrics (63 rows: Huntsville all four, Belvidere's three temperature metrics, Grand Island and Elko precipitation) have fewer than 10 valid years in 1989–2018 and are not scored. They get the city's last valid years plus the change its 8 nearest fitted cities forecast relative to those same years. NOAA's own records for six of them put this fallback within 0.4 °F, 0.3 in and 15 °F-days of a model fitted on the repaired record (Elko precipitation is the exception, 1.8 in low).

Forecast changes behind the policy idea

Quantity (our forecasts vs each city's 1989–2018 mean) Value
Cities with a baseline 208 of 210
CDD, mean over cities: 1989–2018 → forecast 2019–2025 1,396 → 1,508 °F-days (+8.0 %)
CDD change, median city +8.5 %
CDD, last 10 valid years (observed, up to 2018) vs the 1989–2018 mean +7.0 %
Cities with CDD ≥ +10 % / ≥ +15 % 75 / 25
Cities with ≥ 500 CDD and CDD ≥ +10 % or ≥ +200 °F-days (lever b) 73
CDD bias in backtests, 30-year normal → our model (forecast minus actual, share of actual) 2012–2018: -11.1 % → -6.3 %; 2005–2011: -8.4 % → -2.7 %; 1998–2004: -4.3 % → +0.0 %; 1991–1997: -0.7 % → +4.2 %
Average temperature change, mean over cities +0.97 °F
Hottest day of the year no trend survived the backtests: forecast = the 1989–2018 mean (largest difference 0.01 °F, rounding)
Largest CDD increases, °F-days* Needles +347, Phoenix +338, ElPaso +298, Bakersfield +270, Reno +269, DallasFortWorth +266, Abilene +266, Austin +246
Largest relative CDD increases (cities with ≥ 500 CDD)* Reno +30 %, SanDiego +29 %, ColoradoSprings +25 %, Lansing +19 %, Muskegon +19 %, Clayton +18 %, RochesterNY +17 %, GrandRapids +17 %, Pueblo +16 %, Worcester +16 %

* Rankings use cities with at least 25 valid years in 1989–2018 and no documented station change in that window (173 of 208); all cities are in reports/forecast/policy_cdd_by_city.csv.

Reuse and Integration Checks

What we verified about the artifacts before trusting them, each reproducible from code/:

  • The yearly table is exactly what the daily table gives (28,052 city-years: missing-day counts, totals, extremes and CDD identical; average temperature within 0.01 °F, a rounding effect). Units pass (°F, inches). reports/audit/AUDIT.md
  • Combining with NOAA GHCN-Daily (NCEI Access Data Service, 35 stations, 1975–2018, every request capped at 2018-12-31 and re-checked on read): on days present in both sources 92 % of temperatures and 97 % of precipitation values are identical; the rest is 2018 rounding and 11 stations where the challenge's threaded record holds other observations than NOAA's file under the same ID (Akron 1999–2017: 36 % of Tmax equal). Filling invalid years from NOAA adds 83 temperature and 132 precipitation city-years and repairs 6 of the 9 city–metrics without a baseline, but moves backtest skill by −0.5 / +0.7 points (temperature) and −0.7 / +0.4 (CDD) on the two selection folds, opposite signs and one city. So NOAA data is not a forecast input; it validates our fallback. reports/ghcnd/GHCND.md
  • Station history: the 32 documented station changes in 1989–2018 shift average temperature by up to 3 °F (Klamath Falls −3.0, Vancouver +2.1); a neighbor scan flags 106 more city–metrics, only 5 near a documented change. Correcting the documented shifts did not improve 7-year forecasts beyond noise in the backtest, so the published series is used as is. reports/experiments/EXPERIMENTS.md

AI Tools and Agents

  • Tools and models: Claude Code (Anthropic), working under the team's direction.
  • What they did: implemented the team's plan: wrote the code (data checks, backtest harness, models, tests), ran the audit, the backtests and the NOAA cross-check, generated submission.csv with that code, and drafted this README, the data card and the policy idea for the team's revision; a second pass re-checked the code. No forecast value comes from a model's memory: every predicted_value is the output of code/ on data ≤ 2018 (tests.py checks the file equals the code's output), and make all reproduces submission.csv without any AI tool.
  • What a person reviewed: the team designed the approach and guided the agent throughout. A written plan, made before any code, fixed the targets, the model ladder (climatology, trends pooled over nearby cities, CDD from temperature, blends), the two 7-year backtests, the rule to choose models by backtest and stop when gains fall within noise, the ban on neural nets and per-city tuning, and the ≤ 2018 assertion. The team then steered the work at its checkpoints (data access, this disclosure, the policy idea, the submission) and reviewed the results and the policy idea before upload. Code correctness rests on the automated tests and the second review pass.
  • Agent instructions file: code/AGENTS.md.

Policy Idea

Plan cooling on a forecast normal, not a 30-year one.

  1. What risk did we notice? Cooling demand is outrunning the "normal" used to plan for it. Our forecasts put cooling degree days (CDD) in 2019–2025 8 % above the 1989–2018 normal across 208 cities and at least 10 % above in 75. Largest increases: Needles +347, Phoenix +338, El Paso +298 °F-days; largest relative ones among cities with at least 500 CDD: Reno +30 %, San Diego +29 %, Colorado Springs +25 %. The hottest day showed no forecastable trend. Rate cases often still rest on 30-year normals, which ran 11 % and 8 % low in our two latest backtests (our forecast: 6 % and 3 %).
  2. What would we do about it, and who would do it? State utility commissions, which already rule on the weather normal in rate cases, and public-power boards adopt a forecast cooling normal: a trend-adjusted CDD normal, refitted and backtested yearly from public NOAA records with open code. It is used (a) to weather-normalize sales in rate cases and resource plans; (b) to target demand-side programs: in cities with at least 500 CDD whose forecast normal is 10 % or 200 °F-days above the 30-year one (73 of ours, Phoenix included), low-income households get cool-roof, shade and efficient heat-pump retrofits first; (c) by the state LIHEAP agency to set the cooling share of its plan. Cost: minutes of computation; program money is redirected, not added (more cooling aid means less heating aid, each state's call). Measured within three years: the forecast normal closer to realized CDD than the 30-year one; cooling kWh per CDD down 15 % in retrofitted homes that already had cooling; fewer summer shut-offs per 1,000 low-income customers.
  3. Why is it sustainable? It corrects a planning input instead of building something, and efficiency first lowers bills and emissions. It is also self-checking: the adjustment cut the normal's error by 16–18 % in our recent backtests but only 2–3 % in 1990s ones, so the yearly backtest tells a commission when to drop it.

Numbers: Forecast changes behind the policy idea (end of Method) and reports/forecast/POLICY_NUMBERS.md.

Feedback to the Ecosystem

Full log with evidence: OBSTACLES.md. The items that cost the most:

  • What got in the way
    1. Conflicting public facts. The blog and org card still say "November 2–13" and "2019 to 2026"; the Space closes on Nov 10 and scores 2019–2025. We planned from the blog and found the real deadline with hours left.
    2. Baselines are not reproducible from public material. The linear trend is one sentence; its forecasts are in a private repo. The fit, the valid-year rule and the 10-year minimum had to be inferred from a discussion thread and checked against practice-run numbers.
    3. No valid flag in the data. The 10-missing-days rule is re-implemented by every team from three places (About tab, dataset card, notebook); hottest_day_temp validity silently depends on Miss Tmin; Max Tmax uses -inf for empty years, which notna() misses.
    4. No dry run and no API. Testing a dataset link costs one of 10 daily submissions; registration and team state are invisible to scripts; the template checker needs the network.
    5. The dataset moved during the challenge (new subset and a dtype change on Nov 5) without tags, and its source_datasets return HTTP 401, so provenance stops at the card.
    6. No join keys. City names are run together (KeyWest, Zaneville) and there are no FIPS / CBSA codes, which blocks joins with census, energy or health data for the policy part.
    7. The same station ID is not the same record. The challenge names each city by its current GHCN ID, but at 11 of 35 stations we compared, whole runs of years hold other observations than NOAA's file under that ID (and at Milton NOAA's Tmax matches the challenge's next day). NOAA's service also returns values that failed its own quality checks unless flags are requested: without filtering them, Belvidere looked repaired (4 → 17 valid years, really 5).
    8. Two cities lose their precipitation baseline to a storage quirk: Grand Island and Elko miss precipitation on 25–72 days in almost every year before 2018 with complete temperature, and on none in 2018; NOAA has all 30 years.
  • What would fix it
    1. One machine-readable challenge.yaml (dates, schema, limits) rendered into every page.
    2. Publish the baseline code and its ≤ 2018 practice-run forecasts as test vectors.
    3. Add valid_<metric> and has_baseline columns; store missing values as nulls.
    4. check_submission.py --repo owner/name replaying the Space's checks; ship city_ids.txt.
    5. Tag dataset releases, keep a changelog in the card, and point source_datasets at the DOIs.
    6. Add county FIPS / CBSA codes and display names to city_info.
    7. Publish the GHCN ID actually used for each segment of each city's record, and say whether quality-flagged values were removed.
    8. Re-extract precipitation for Grand Island and Elko, or flag them in the dataset card.
  • What worked well (keep it): the Space's code is public, so the scoring could be ported and tested instead of guessed; coverage_and_stations made the station audit a 30-line job; the example notebook's numbers gave us an exact target to validate the harness against.
  • A negative finding worth sharing: correcting documented station changes did not improve 7-year forecasts beyond noise, and no model beat the 30-year mean for precipitation or the hottest day once earlier decades were used as a check. For the hottest day, two folds alone would have told us otherwise.

Files

File Content
submission.csv Our forecasts: city_id,year,metric,predicted_value (5,880 rows)
code/protostar/ The pipeline: data.py (loading, valid-year rule, leak guard), scoring.py (the challenge's score), models.py, backtest.py (folds, ladder, selection), forecast.py (refit, submission.csv, plausibility), audit.py, experiments.py (what we tried and did not ship), ghcnd.py (NOAA cross-check), docs.py, tests.py
code/fetch_data.py, code/check_submission.py, code/requirements.txt, code/AGENTS.md Pinned-data check/download, the template's file check, exact package versions, agent rules
data/ Pinned copy of the training tables (data/raw/) and the NOAA extract used for the cross-check (data/external/ghcnd/); data card: data/README.md
reports/ Generated: audit/ (data audit), backtest/ (tables, ladder, every candidate), experiments/ (tried and not shipped), forecast/ (validation, policy numbers), ghcnd/ (NOAA cross-check)
NOTES.md, OBSTACLES.md Rules as read from the Space (verbatim), decisions, and the running log of obstacles
Makefile make all reproduces everything

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

  1. Clone this dataset repository (the three small training tables are included, pinned by sha256).
  2. make all (Python ≥ 3.11). It creates .venv, installs the exact versions in code/requirements.txt, verifies the data, and runs the audit, the backtests and model selection, the experiments, the forecast, the NOAA cross-check (from the committed extract), the file checks, the tests and the README tables. Network is needed only to install the packages and to fetch the 49 MB daily table used by the audit (skipped with a warning if unreachable).
  3. git status should show submission.csv unchanged: the run is deterministic (fixed seed, no randomness in the models).

Step by step: python code/fetch_data.py, then with PYTHONPATH=code: python -m protostar.audit, python -m protostar.backtest, python -m protostar.experiments, python -m protostar.forecast, python -m protostar.ghcnd --offline, python -m protostar.tests, python code/check_submission.py.

License

Data and text (this card, submission.csv, reports): CC BY 4.0. Code in code/: MIT (code/LICENSE). Training data: CC0 (see below).

Data

  • External data: NOAA GHCN-Daily through the NCEI Access Data Service (dataset=daily-summaries), 35 stations, 1975-01-01 to 2018-12-31. Analysis only (cross-check, gap-fill experiment, validation of the fallback); it is not read by the forecasting code. The forecasts use the challenge dataset alone.
  • Curated data: see the data card in data/README.md (each file, its source, license and preparation).
  • Sources behind the policy idea (text, not data; no weather observations): Won, Wang & Warren (2016), Climate normals and weather normalization for utility regulation, Energy Economics 54, 405–416 ("Conventionally, normal weather values are calculated using the Standard Climate Normal … the 30-year average"; it "can inadvertently introduce biases in the weather normalization adjustment"); NOAA NCEI (2021), NOAA Delivers New U.S. Climate Normals (15-year supplemental normals "for users who require information for periods closer to the present"): the precedent that a shorter normal is already an accepted product.

Training data: CMU-OSPO/protostar-climate-train-to-2018 at commit bbbdebdab2a18e00f97ef87bd3f35a71e222180b, from:

Downloads last month
102