Datasets:
city_id stringclasses 210
values | year int64 2.02k 2.03k | metric stringclasses 4
values | predicted_value float64 2.76 5.26k |
|---|---|---|---|
USC00042863 | 2,019 | annual_avg_temp | 66.975981 |
USC00042863 | 2,019 | total_precip_in | 12.808132 |
USC00042863 | 2,019 | hottest_day_temp | 103.927659 |
USC00042863 | 2,019 | cdd | 1,727.960107 |
USC00042863 | 2,020 | annual_avg_temp | 67.033058 |
USC00042863 | 2,020 | total_precip_in | 12.808132 |
USC00042863 | 2,020 | hottest_day_temp | 103.927659 |
USC00042863 | 2,020 | cdd | 1,739.666423 |
USC00042863 | 2,021 | annual_avg_temp | 67.090134 |
USC00042863 | 2,021 | total_precip_in | 12.808132 |
USC00042863 | 2,021 | hottest_day_temp | 103.927659 |
USC00042863 | 2,021 | cdd | 1,751.410961 |
USC00042863 | 2,022 | annual_avg_temp | 67.147211 |
USC00042863 | 2,022 | total_precip_in | 12.808132 |
USC00042863 | 2,022 | hottest_day_temp | 103.927659 |
USC00042863 | 2,022 | cdd | 1,763.191349 |
USC00042863 | 2,023 | annual_avg_temp | 67.204287 |
USC00042863 | 2,023 | total_precip_in | 12.808132 |
USC00042863 | 2,023 | hottest_day_temp | 103.927659 |
USC00042863 | 2,023 | cdd | 1,774.995143 |
USC00042863 | 2,024 | annual_avg_temp | 67.261364 |
USC00042863 | 2,024 | total_precip_in | 12.808132 |
USC00042863 | 2,024 | hottest_day_temp | 103.927659 |
USC00042863 | 2,024 | cdd | 1,786.863655 |
USC00042863 | 2,025 | annual_avg_temp | 67.31844 |
USC00042863 | 2,025 | total_precip_in | 12.808132 |
USC00042863 | 2,025 | hottest_day_temp | 103.927659 |
USC00042863 | 2,025 | cdd | 1,798.788286 |
USC00166584 | 2,019 | annual_avg_temp | 67.130221 |
USC00166584 | 2,019 | total_precip_in | 55.744821 |
USC00166584 | 2,019 | hottest_day_temp | 100.675662 |
USC00166584 | 2,019 | cdd | 2,726.147145 |
USC00166584 | 2,020 | annual_avg_temp | 67.187298 |
USC00166584 | 2,020 | total_precip_in | 55.744821 |
USC00166584 | 2,020 | hottest_day_temp | 100.675662 |
USC00166584 | 2,020 | cdd | 2,738.183912 |
USC00166584 | 2,021 | annual_avg_temp | 67.244374 |
USC00166584 | 2,021 | total_precip_in | 55.744821 |
USC00166584 | 2,021 | hottest_day_temp | 100.675662 |
USC00166584 | 2,021 | cdd | 2,750.234003 |
USC00166584 | 2,022 | annual_avg_temp | 67.301451 |
USC00166584 | 2,022 | total_precip_in | 55.744821 |
USC00166584 | 2,022 | hottest_day_temp | 100.675662 |
USC00166584 | 2,022 | cdd | 2,762.306397 |
USC00166584 | 2,023 | annual_avg_temp | 67.358528 |
USC00166584 | 2,023 | total_precip_in | 55.744821 |
USC00166584 | 2,023 | hottest_day_temp | 100.675662 |
USC00166584 | 2,023 | cdd | 2,774.406844 |
USC00166584 | 2,024 | annual_avg_temp | 67.415604 |
USC00166584 | 2,024 | total_precip_in | 55.744821 |
USC00166584 | 2,024 | hottest_day_temp | 100.675662 |
USC00166584 | 2,024 | cdd | 2,786.547389 |
USC00166584 | 2,025 | annual_avg_temp | 67.472681 |
USC00166584 | 2,025 | total_precip_in | 55.744821 |
USC00166584 | 2,025 | hottest_day_temp | 100.675662 |
USC00166584 | 2,025 | cdd | 2,798.718614 |
USC00280734 | 2,019 | annual_avg_temp | 52.063883 |
USC00280734 | 2,019 | total_precip_in | 44.738749 |
USC00280734 | 2,019 | hottest_day_temp | 94.733333 |
USC00280734 | 2,019 | cdd | 816.800756 |
USC00280734 | 2,020 | annual_avg_temp | 52.120959 |
USC00280734 | 2,020 | total_precip_in | 44.738749 |
USC00280734 | 2,020 | hottest_day_temp | 94.733333 |
USC00280734 | 2,020 | cdd | 823.09309 |
USC00280734 | 2,021 | annual_avg_temp | 52.178036 |
USC00280734 | 2,021 | total_precip_in | 44.738749 |
USC00280734 | 2,021 | hottest_day_temp | 94.733333 |
USC00280734 | 2,021 | cdd | 829.425514 |
USC00280734 | 2,022 | annual_avg_temp | 52.235112 |
USC00280734 | 2,022 | total_precip_in | 44.738749 |
USC00280734 | 2,022 | hottest_day_temp | 94.733333 |
USC00280734 | 2,022 | cdd | 835.788807 |
USC00280734 | 2,023 | annual_avg_temp | 52.292189 |
USC00280734 | 2,023 | total_precip_in | 44.738749 |
USC00280734 | 2,023 | hottest_day_temp | 94.733333 |
USC00280734 | 2,023 | cdd | 842.19123 |
USC00280734 | 2,024 | annual_avg_temp | 52.349266 |
USC00280734 | 2,024 | total_precip_in | 44.738749 |
USC00280734 | 2,024 | hottest_day_temp | 94.733333 |
USC00280734 | 2,024 | cdd | 848.614393 |
USC00280734 | 2,025 | annual_avg_temp | 52.406342 |
USC00280734 | 2,025 | total_precip_in | 44.738749 |
USC00280734 | 2,025 | hottest_day_temp | 94.733333 |
USC00280734 | 2,025 | cdd | 855.061999 |
USC00286055 | 2,019 | annual_avg_temp | 54.557069 |
USC00286055 | 2,019 | total_precip_in | 49.055133 |
USC00286055 | 2,019 | hottest_day_temp | 97.6934 |
USC00286055 | 2,019 | cdd | 1,128.988935 |
USC00286055 | 2,020 | annual_avg_temp | 54.614146 |
USC00286055 | 2,020 | total_precip_in | 49.055133 |
USC00286055 | 2,020 | hottest_day_temp | 97.6934 |
USC00286055 | 2,020 | cdd | 1,136.067179 |
USC00286055 | 2,021 | annual_avg_temp | 54.671222 |
USC00286055 | 2,021 | total_precip_in | 49.055133 |
USC00286055 | 2,021 | hottest_day_temp | 97.6934 |
USC00286055 | 2,021 | cdd | 1,143.150949 |
USC00286055 | 2,022 | annual_avg_temp | 54.728299 |
USC00286055 | 2,022 | total_precip_in | 49.055133 |
USC00286055 | 2,022 | hottest_day_temp | 97.6934 |
USC00286055 | 2,022 | cdd | 1,150.248351 |
- What changed from v1, and why (honest disclosure)
- Method
- Validation (pre-2019 only)
- v9: first-difference-correlation neighbour weights (pre-registered)
- v7: second round of ideas from teams ranked above us (pre-registered)
- v6: ideas from teams ranked above us (pre-registered)
- v4: five further fixed-parameter ideas (pre-registered)
- v3: pre-registered ablations of the "SHEBE" component ideas (none adopted)
- Protocol deviation (CDD, inherited from v2b), disclosed
- Policy Idea
- Feedback to the Ecosystem
- Files and reproduction
- Data
Heat Seekers β v9
Negative result: no
Release: v9: v7 with one change that passed a pre-registered rule (research/V9_PROTOCOL.md): average temperature uses the same region + departure view as v7, but the 32 neighbouring cities in the regional anomaly are weighted by the correlation of their year-to-year temperature changes (first differences) instead of equally. The idea comes from a Gemini brainstorm; we pre-registered the exact rule and tested it on pre-2019 data only. Precipitation, hottest day and CDD are unchanged from v7. See "v9" below.
Previous release, v7: v6 with one change that passed a pre-registered rule (research/V7_PROTOCOL.md): cooling degree days now come from a temperature-to-CDD map driven by our v6 temperature forecast. The idea comes from team PstarAlu's public README; temperature, precipitation and hottest day are unchanged from v6. See "v7" below.
Previous release, v6: v4 plus three small changes that passed a pre-registered rule (research/V6_PROTOCOL.md): a "region + departure" temperature view, and a neighbour-departure nudge for hottest day and precipitation. CDD is unchanged (v2b daily shift). The ideas were taken from Team Tatakae's public README and tested on pre-2019 data only; see "v6" below. The v3 (42 candidates), v4 and v6 ablation studies are documented below.
Status: local release candidate, not yet published or submitted. Team: Heat Seekers (contact Tejas Chaudhari, tchaudha; teammate Parth Dave, BossCoder2025 teammate Manas dhawkar`).
What changed from v1, and why (honest disclosure)
Our v1 entry (city exponentially weighted mean for temperature/CDD, 30-year mean for precipitation and hottest day) was selected on four development origins (1983β2004). After submission we saw it score 3.9 (rank 19) on the live leaderboard (2019β2021), while our teammate's earlier trend-based entry had scored 12.7. This redesign was started because of that observation, and we cannot un-see it. To keep the forecast compliant with the data rule:
- no 2019β2025 observation was downloaded or used, and no window, weight or candidate was chosen by comparing to live-leaderboard numbers;
- the full candidate list and selection rules were written and hashed before any v2 candidate was backtested (
research/V2_PROTOCOL.md, SHA-256 inresults/v2_protocol_hash.txt;code/backtest_v2.pyrefuses to run if the file changed); - selection uses only pre-2019 rolling origins. Independent pre-2019 evidence for trend-aware levels already existed: the organizers' practice run (CDD 212 β 175 with a linear trend) and our teammate's pre-2019 backtests. Our v1 design simply never tested origins after 2004, where the 30-year windows no longer include mid-century cooling.
Method
Per city, from data through 2018 only (code/v2lib.py):
- Average temperature β level + national warming slope (
LT_L15_a0_national). Level = mean of the last 15 valid years in the current station segment (the last 15 valid years if fewer than 8 exist there). Slope = the mean, over all cities with β₯10 valid years, of each city's 1989β2018 OLS slope fitted with one intercept per station segment (+0.057 Β°F/yr at 2018). Forecast(year) = level + slope Γ (year β centre of level years). A single pooled slope beats each city's own noisy 30-year slope: own-slope-only variants rank last among the 21 trend configurations on the development origins. - Cooling degree days β daily shift (
SHmean_L15_a0_national, used v2bβv6; replaced in v7, see below): for each of the city's last 15 valid years y, every day's (Tmax+Tmin)/2 is shifted by slope Γ (target year β y), using the same national slope as temperature. CDD is recomputed from the shifted daily pairs and added as an increment to the official annual CDD. The forecast is the mean over those years. CDD therefore stays consistent with the temperature forecast and rises faster in cities with many days near 65 Β°F. This choice deviates from the frozen protocol; see below. - v4 changes: the temperature level is now the trimmed mean (drop the single max and min) of each year's value detrended to the centre with the same national slope, over calendar years [2004, 2018] in the current segment (falling back to the v2 level years if fewer than 5); hottest day = 0.9 Γ 30-year mean + 0.1 Γ mean of the valid values in calendar years [2004, 2018] in the current station segment (30-year mean alone if fewer than 5). See "v4" below.
- v6 changes: temperature = 0.75 Γ a "region + departure" view + 0.25 Γ the v4 temperature forecast. The view is the city's 1989β2018 mean + the regional anomaly (mean of the 32 nearest cities' anomalies, each against its own 1989β2018 mean) carried forward from its 2004β2018 mean with the same national slope + the city's mean departure from that regional anomaly over its last 5 valid years. Hottest day = 30-year mean + 0.3 Γ the city's mean departure from its 31 nearest neighbours over 2009β2018 (each measured against its own 30-year mean); this replaces the v4 hottest-day blend. Precipitation = 30-year mean Γ (1 + 0.3 Γ the same departure measured as a ratio). Code:
code/v6lib.py. - v7 change: CDD = temperature-to-CDD map. For each city we take its last 30 and, separately, its last 10 valid years with at least 360 days of data (the last n valid years if fewer than 5 such years). For each base year we shift that year's daily (Tmax+Tmin)/2 by (v6 temperature forecast for the target year β that base year's observed average temperature), recompute CDD from the shifted days (same daily-pair increment table as v2b, shift clipped to β8β¦+12 Β°F), and average over base years. The forecast is the mean of the 30-year and 10-year maps, floored at 0. Code:
c_mapincode/v7lib.py. - v9 change (current release): temperature is still 0.75 Γ the region + departure view + 0.25 Γ the v4 forecast, but the regional anomaly is a weighted mean of the 32 nearest cities. The weight of neighbour j is max(0, correlation of first differences of annual temperature between the city and j over the last 30 years), using a correlation only when at least 10 overlapping year-pairs exist; if every weight is 0, the equal-weight v7 regional mean is used. The idea came from a Gemini brainstorm; the exact rule was pre-registered by us. Code:
t_fdwincode/v9lib.py. - Precipitation (before v6; and, before v4, hottest day): official 30-year mean (1989β2018). Median-30, 60-year mean, and hottest-day + Β½ or 1 Γ warming were pre-registered challengers; none passed (all failed the β₯12/20 win rule).
- Fallback rows: 9 city-metrics (63 rows) have no official baseline (<10 valid years in 1989β2018) and are unscored; they use the same method on the last valid years at any time (
results/v2_fallback_rows_v2b.csv). Huntsville temperature/CDD therefore extrapolate from a 1940sβ50s level and should not be read as a meaningful forecast.
The structure (recent level + pooled warming + CDD derived from temperature) is a combination of standard ideas, not a new estimator. It is similar to what several other teams describe independently (see research/OTHER_TEAMS_ANALYSIS.md); our teammate Parth Dave's public method (15-year level, regional slope, daily-shift CDD) is included in the grid as a reference and credited here.
Validation (pre-2019 only)
Harness: official formula (skill_m = 1 β MAE_m / MAE of the 30-year mean at the same origin; score = 100 Γ max(0, mean skill)); official 10-missing-day validity and β₯10-valid-year baseline eligibility; truth rows left-joined to forecasts with an assertion that every scored row has a finite forecast; aggregates average unclipped skills. It reproduces the organizers' practice run exactly (origin 2010, 2011β2018: CDD MAE 211.69 / 174.59 vs published 212 / 175) and 9 city-metrics / 63 unscored rows at 2018. The official score() and check_forecasts() are executed in the tests.
Development origins 1985β2004 (20 seven-year blocks, selection); late origins 2005β2011 (7 blocks, consistency gate only β origin 2011 was already used as v1 confirmation and is not untouched). Mean unclipped skill Γ 100 (score-like):
| Model | dev h1β3 | dev h4β7 | late h1β3 | late h4β7 | 2011β2012β18 all 7 yrs |
|---|---|---|---|---|---|
| 30-year mean (reference) | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 |
| Official linear trend | β1.44 | β4.31 | β0.56 | 2.24 | β |
| v1 submitted | 2.50 | 2.43 | 1.18 | 1.27 | 1.44 |
| Teammate-like (L15, ΒΌ own + ΒΎ 10-NN slope, median daily-shift CDD, median30) | 1.85 | 2.66 | 1.96 | 7.36 | β |
| v2 protocol-strict (CDD = v1) | 3.20 | 3.81 | 0.73 | 4.23 | 5.39 |
| v2b (CDD = daily shift) | 2.78 | 3.62 | 2.60 | 8.73 | 9.75 |
| v4 (v2b + detrended trimmed temp level + hottest-day blend) | 2.86 | 3.69 | 2.73 | 8.83 | 9.81 |
| v6 (v4 temp blended with region + departure view; neighbour-departure nudge for hottest day and precipitation; CDD v2b) | 3.52 | 4.08 | 3.19 | 9.03 | 9.95 |
| v7 (v6 with CDD from the temperature-to-CDD map) | 4.37 | 4.73 | 3.64 | 9.42 | 10.38 |
| v9 (v7 with first-difference-correlation neighbour weights on temperature) | 4.39 | 4.76 | 3.64 | 9.43 | 10.38 |
Per-metric dev h4β7 skill: temperature 0.124 (v1 0.069, linear trend 0.059); CDD 0.029 (exp10) vs 0.021 (daily shift); late h4β7 CDD 0.026 (exp10) vs 0.206 (daily shift). At h4β7 the temperature method beats the official linear trend at all 27 origins and beats clim30 at h4β7 in 16 of 20 development origins. Organizers' practice run (2010 β 2011β2018): v1 1.05, v2-strict 4.77, v2b 8.89, linear trend 5.03.
Full tables: results/v2_tables.md (per origin), results/v2_candidate_summary.csv (all 76 candidate Γ metric rows), results/v2_skills_by_origin.csv (per origin Γ metric Γ split Γ method MAE and skill), results/v2_selection.json.
Limits. 27 overlapping blocks are not 27 independent tests; regional weather and national anomalies correlate errors across cities, so these numbers are descriptive. The live score and final ranking may differ substantially.
v9: first-difference-correlation neighbour weights (pre-registered)
How the ideas were chosen (disclosure). On 2026-11-10 Tejas asked us to test further improvements over v7, including ideas from a Gemini brainstorm he pasted (part 1, and a later part 2). The live board was not used to set any parameter or to choose among candidates. No other team's forecasts or submission.csv were used as inputs; no observation after 2018 was used; no slope multipliers were allowed; no hottest-day trend damping factor was tuned (Ξ³ was unused). An intermediate v8 experiment (a teammate's land-cover / climate-character temperature slope, and shrinkage of hottest day and precipitation toward the 30-year mean) was pre-registered and tested earlier the same day; nothing was adopted, so v7 remained the reference. Before computing any v9 candidate we wrote and hashed research/V9_PROTOCOL.md (SHA-256 2e70520c86949493151525521b7913d7942cd33dbbc07b53ec48ae4672f5db23; code/backtest_v9.py asserts it). It fixes 12 candidates. The reference is v7. A candidate had to (1) beat v7's mean h4β7 skill on the dev origins (1985β2004) and be no worse on the late origins (2005β2011), and (2) remain significant after a Holm correction over all 12 candidates (one-sided moving-block bootstrap of the 20 dev-origin differences, block length 5, 20,000 resamples, Ξ± = 0.10). The harness reproduces v7 exactly (4.73 / 9.42), and a future-data poisoning test passed at origins 1995, 2011 and 2018. Full tables: results/v9_tables.md.
Credit. The adopted temperature change β weighting neighbouring cities by the correlation of first differences of annual temperature β comes from a Gemini brainstorm. We wrote the exact weights, neighbour set (32 cities), 10-pair minimum and equal-weight fallback into the pre-registered protocol and implemented the test ourselves. Any errors are ours. The other 11 candidates (EWMA departures, Huber/winsorised means, a 45-year empirical-Bayes slope, cube-root and long-history precipitation, HodgesβLehmann and top-10 hottest-day variants, and two CDD-map averaging rules) did not pass the rule.
Differences vs v7 in each metric's mean h4β7 skill (Γ100):
| Metric | Candidate (source) | dev | late | 2008β2011 | p (Holm) | Result |
|---|---|---|---|---|---|---|
| Temperature | t_ewma_d: EWMA (half-life 10 y) of cityβneighbour departure |
β0.37 | β0.18 | +0.15 | 1.00 | fails |
| Temperature | t_huber: 1.65-MAD winsorised mean and departure |
β0.01 | β0.07 | β0.07 | 1.00 | fails |
| Temperature | t_fdw: first-difference-correlation neighbour weights (Gemini brainstorm) |
+0.11 | +0.04 | +0.04 | <0.001 | passes β adopted |
| Temperature | t_eb45: 45-year EB regional slope + 19-year anchor |
β10.46 | β0.63 | β0.70 | 1.00 | fails |
| Precipitation | p_ewma_d: EWMA on the v7 ratio departure |
β0.03 | β0.17 | β0.04 | 1.00 | fails |
| Precipitation | p_cbrt: 30-year cube-root mean + 0.3 neighbour departure |
+0.04 | β0.57 | β1.43 | 1.00 | fails late |
| Precipitation | p_eb_ratio: long-history city/neighbour ratio, EB blend |
+0.58 | β1.11 | β1.73 | 1.00 | fails late |
| Hottest day | h_ewma_d: EWMA on the v7 departure |
β0.12 | β0.15 | β0.16 | 1.00 | fails |
| Hottest day | h_hl: HodgesβLehmann 30-year location + 0.3 nudge |
β0.12 | β0.26 | +0.50 | 1.00 | fails |
| Hottest day | h_q10: mean of top-10 daily Tmax + half the shrunk tail gap |
β8.70 | β11.97 | β7.11 | 1.00 | fails |
| CDD | c_pmean: sqrt power mean (30 y) + 20% trimmed mean (10 y) |
+0.20 | β0.47 | β0.63 | <0.001 | fails late (Holm-significant) |
| CDD | c_detrend: 30-year map centred at the window midpoint, then shifted |
β0.00 | +0.00 | +0.00 | 1.00 | fails dev |
v9 = v7 with temperature t_fdw. Mean over the four metrics at h4β7: dev 4.76 (v7 4.73), late 9.43 (v7 9.42), 2008β2011 9.93 (v7 9.92); h1β3: dev 4.39 (v7 4.37), late 3.64 (v7 3.64). v9 beats v7 at 25 of 27 origins at h4β7; the two losses are tiny (the worse is β0.007 at origin 2004). The gain is small but it passed the pre-registered rule, including the Holm check. These are backtests, not a promise of live or final scores. Compared with v7, the 2019β2025 temperature forecasts change by a mean of β0.003 Β°F (mean absolute 0.013 Β°F, range β0.181 to +0.075); precipitation, hottest-day and CDD forecasts are identical to v7.
v7: second round of ideas from teams ranked above us (pre-registered)
What we read (disclosure). On 2026-11-09 v6 ranked 8th on the 2019β2021 live leaderboard, and Tejas asked us to study the seven teams above us more thoroughly (PstarAlu, GuaGuaBear, Team Hurricane, TROUBLESHOOTERS, Ceviche, Ashley boys, Klepon Nusantara): their READMEs at the scored commit, their public code and their metadata. So the choice of which teams to read again used the live board. One base repository we read (Ceviche's) reports post-hoc 2019β2025 results; we did not use them. Some ideas we tested were tuned by their authors on origins that overlap our late set (2005β2011), so late results for those ideas are not fully independent. No live-board number entered any parameter or rule; no other team's forecasts, submission.csv or outputs were downloaded or used; no observation after 2018 was used; no slope multipliers were allowed.
Protocol. Before computing any candidate we wrote and hashed research/V7_PROTOCOL.md (SHA-256 2f72f9b81dbc536da005636dbe1340810c01b9aa282ec9818c20e817e77cf55d; code/backtest_v7.py asserts it). It fixes 15 candidates and their parameters. The reference is v6. A candidate had to (1) beat v6's mean h4β7 skill on the dev origins (1985β2004) and be no worse on the late origins (2005β2011), and (2) remain significant after a Holm correction over all 15 candidates (one-sided moving-block bootstrap of the 20 dev-origin differences, block length 5, 20,000 resamples, Ξ± = 0.10). The harness reproduces v6 exactly (4.08 / 9.03), and a future-data poisoning test passed at origins 1995, 2011 and 2018. Full tables: results/v7_tables.md.
Credit. The adopted CDD map follows the CDD layer described in team PstarAlu's public README (nbiswal-star/protostar-climate-submission @8c8dc3f): the mean of 30-year and 10-year maps built from complete (β₯ 360-day) base years. Ours is driven by our own v6 temperature forecast, uses our own daily-shift table, and is our own transcription; any errors are ours. Rejected ideas came from PstarAlu (two-scale fading departure for temperature), GuaGuaBear (ExtraTrees CDD), TROUBLESHOOTERS (wet-day Γ intensity precipitation), Ceviche (pooled-trend AR temperature and ridge residual corrections), Team Hurricane (precipitation cycle, which that team disclosed choosing after seeing their live score) and our own v6 round (hottest-day mix).
Differences vs v6 in each metric's mean h4β7 skill (Γ100):
| Metric | Candidate (source) | dev | late | 2008β2011 | p (Holm) | Result |
|---|---|---|---|---|---|---|
| Temperature | t_twoscale: two-scale fading departure (PstarAlu) |
β2.49 | β0.97 | +0.46 | 1.00 | fails |
| Temperature | t_twoscale50: Β½ two-scale + Β½ v6 |
+0.17 | β0.04 | +0.70 | 1.00 | fails late |
| Temperature | t_arma25: ΒΎ v6 + ΒΌ pooled-trend AR(1) (Ceviche) |
+0.39 | +0.72 | +1.50 | 0.91 | passes step 1, not significant |
| Temperature | r25_temp: 0.25 Γ ridge residual correction (Ceviche) |
β0.33 | +0.27 | +0.64 | 1.00 | fails dev |
| CDD | c_map: temperature-to-CDD map (PstarAlu) |
+2.63 | +1.58 | +1.03 | <0.001 | passes β adopted |
| CDD | c_map_blend: Β½ map + Β½ v2b |
+1.89 | +1.15 | +0.95 | <0.001 | passes, lower dev |
| CDD | c_et50: Β½ ExtraTrees + Β½ v2b (GuaGuaBear) |
+1.59 | β0.91 | β0.37 | 0.012 | fails late |
| CDD | r25_cdd: ridge correction |
β0.82 | +0.23 | +0.26 | 1.00 | fails dev |
| Precipitation | p_wet: wet days Γ intensity, F = 1.01 (TROUBLESHOOTERS) |
β2.37 | +0.09 | β0.45 | 1.00 | fails dev |
| Precipitation | p_wet_f1: same, F = 1.00 |
β1.76 | β0.78 | β1.84 | 1.00 | fails |
| Precipitation | p_nudge_wet: v6 Γ wet-day factor |
β2.55 | β1.28 | β2.47 | 1.00 | fails |
| Precipitation | p_nudge_cyc: v6 + Β½ Γ 8β30-yr regional cycle (Team Hurricane spec) |
β0.15 | +1.03 | +0.80 | 1.00 | fails dev |
| Precipitation | r25_precip: ridge correction |
β0.18 | +0.20 | +0.52 | 1.00 | fails dev |
| Hottest day | h_mix_os2: Β½ v6 + Β½ (second-highest day + nudge) |
+0.72 | +0.25 | +0.49 | 0.63 | passes step 1, not significant |
| Hottest day | r25_hot: ridge correction |
β0.01 | β0.02 | +0.11 | 1.00 | fails |
v7 = v6 with CDD c_map. CDD h4β7 skill Γ100: dev 4.73 (v6 2.10), late 22.17 (v6 20.59), origins 2008β2011 21.99 (v6 20.96); the map beats v6's CDD at 25 of 27 origins (losses at 1996 and 1998). Mean over the four metrics at h4β7: dev 4.73 (v6 4.08), late 9.42 (v6 9.03), 2008β2011 9.92 (v6 9.66); h1β3: dev 4.37 (v6 3.52), late 3.64 (v6 3.19). These are backtests, not a promise of live or final scores. Compared with v6, the 2019β2025 CDD forecasts change by a mean of +1.6 (mean absolute 28.6, range β195 to +125); temperature, precipitation and hottest-day forecasts are identical to v6.
v6: ideas from teams ranked above us (pre-registered)
How the ideas were chosen (disclosure). On 2026-11-09 v4 ranked 10th on the 2019β2021 live leaderboard. Tejas asked us to read the public READMEs of the nine teams ranked above us and test their most promising ideas against v4. So the choice of which teams to read used the live board. To limit that influence: no live-board number entered any parameter or rule; every parameter was copied from the source README or fixed in the protocol before any run; no other team's forecasts, submission.csv or outputs were downloaded or used; and no 2019β2025 observation was used. Before computing any candidate we wrote and hashed research/V6_PROTOCOL.md (SHA-256 87bdcf7907166088c13f6a1fe78749bd4bc3524b7d2c209ba71637a33839c30a; code/backtest_v6.py asserts it). The rule was unchanged from v3/v4: higher dev (1985β2004) mean h4β7 skill than v4 and no worse on the late origins (2005β2011), with no slope multipliers. A future-data poisoning test passed at origins 1995, 2011 and 2018. Full tables: results/v6_tables.md.
Credit. All three adopted changes come from Team Tatakae's public README (ironMan334/protostar-climate-team-tatakae): the "region + departure" view with a 75 % weight, and the 30 %, 10-year, 31-neighbour "station-move nudge" for hottest day and precipitation. We changed one thing: we carry the regional anomaly forward with our national slope rather than Tatakae's regional 30-year trend, because the version with the regional trend failed our late check. Both versions were listed in the protocol before running. The rejected ideas came from Tatakae, Ashley boys, MKDir and PstarAlu (CDD-per-Β°F blend) and from PstarAlu and TROUBLESHOOTERS (second-highest-day anchor, 45-year precipitation half).
| Metric | Candidate (source) | dev h4β7 | late h4β7 | v4 dev / late | Result |
|---|---|---|---|---|---|
| Temperature | t_regdep75: region + departure, regional 30-y trend (Tatakae) |
0.1267 | 0.1413 | 0.1262 / 0.1446 | fails late |
| Temperature | t_regdep75_nat: region + departure, national slope |
0.1362 | 0.1477 | 0.1262 / 0.1446 | passes β adopted |
| CDD | c_sens50: Β½ daily shift + Β½ CDD-per-Β°F Γ slope |
0.0323 | 0.1962 | 0.0210 / 0.2059 | fails late |
| Hottest day | h_nudge: 30-y mean + 0.3 Γ neighbour departure (Tatakae) |
0.0042 | 0.0043 | 0.0004 / 0.0028 | passes, best dev β adopted |
| Hottest day | h_blend15_nudge: v4 blend + same nudge |
0.0037 | 0.0060 | 0.0004 / 0.0028 | passes, lower dev |
| Hottest day | h_os2: second-highest daily max + pooled spacing |
0.0101 | β0.0035 | 0.0004 / 0.0028 | fails late |
| Precipitation | p_nudge: 30-y mean Γ (1 + 0.3 Γ ratio departure) (Tatakae) |
0.0016 | 0.0031 | 0 / 0 | passes β adopted |
| Precipitation | p_mm45: Β½ median30 + Β½ mean30/45 |
0.0021 | β0.0023 | 0 / 0 | fails late |
v6 = temperature t_regdep75_nat + CDD v2b + precipitation p_nudge + hottest day h_nudge. Mean h4β7 score Γ100: dev 4.08 (v4 3.69), late 9.03 (v4 8.83); h1β3: dev 3.52 (v4 2.86), late 3.19 (v4 2.73). v6 beats v4 at 21 of 27 origins at h4β7; its worst origin is β0.34 (1996). The gains are small. The late gain comes mostly from origins 2005β2006; at origins 2008β2011 v6 and v4 are within Β±0.2. Eight candidates were tested, so part of each margin may be luck. Compared with v4, the 2019β2025 forecasts change by a mean absolute 0.19 Β°F for temperature (range β0.96 to +0.77), 0.12 Β°F for hottest day and 0.35 in for precipitation; CDD is identical.
v4: five further fixed-parameter ideas (pre-registered)
After v3, an outside AI assistant (pasted by Tejas) suggested five ideas. They came after we had seen the v2/v3 backtest results, so they are not independent of the data. Before computing anything we wrote and hashed research/V4_PROTOCOL.md (SHA-256 60b33c4dadf9b06e62ca999156d6edf611a2b9aff9bdf5923e48a9adec65e597; code/backtest_v4.py asserts it). All parameters were fixed in the protocol and none were tuned. The binding rule was unchanged from v3: a candidate replaces the reference only if it has a higher mean h4β7 skill on the development origins (1985β2004) and is no worse on the late origins (2005β2011). We also report a pooled ratio-of-sums skill (1985β2011) and three regime blocks, but these are secondary and were not used for selection. Full tables are in results/v4_tables.md.
| Idea | Candidate | dev h4β7 | late h4β7 | vs reference | Result |
|---|---|---|---|---|---|
| 1 Hinge (post-1975) national slope | t_hinge (temperature) | 0.1175 | 0.1502 | v2 0.1239 / 0.1432 | fails dev |
| 1 | c_v2b_hinge (CDD) | 0.0034 | 0.1987 | v2b 0.0210 / 0.2059 | fails |
| 2 Detrended level, calendar window [Tβ14,T] | t_det_v2 | 0.1254 | 0.1457 | v2 | passes |
| 2+4 Detrended, trimmed (drop max and min) | t_dettrim_v2 | 0.1262 | 0.1446 | v2 | passes, best dev β adopted |
| 1+2(+4) | t_det_hinge / t_dettrim_hinge | 0.1194 / 0.1202 | 0.1522 / 0.1507 | v2 | fail dev |
| 3 Per-year shifted CDD, calendar window, trimmed | c_trim_v2 | 0.0257 | 0.2029 | v2b | fails late |
| 3 with hinge slope | c_trim_hinge | 0.0087 | 0.1953 | v2b | fails |
| 4 Trimmed 30-yr mean precipitation | p_trim | 0.0001 | β0.0001 | clim30 0 / 0 | fails late |
| 5 Hottest day: clim30 + warming tether | h_tether_v2 / h_tether_hinge | β0.0010 / β0.0016 | 0.0014 / β0.0005 | clim30 | fail dev |
| 5 Hottest day: 0.9 Γ clim30 + 0.1 Γ 15-yr current-segment mean | h_blend15 | 0.0004 | 0.0028 | clim30 | passes β adopted |
The v2b CDD method already applies a per-year lag (each level year y is shifted by slope Γ (target β y)). Idea 3 therefore differs from v2b only in its calendar window and trimming. That change helped the dev origins but hurt the late origins, so v2b's CDD is kept.
v4 = temperature t_dettrim_v2 + CDD v2b + precipitation clim30 + hottest day h_blend15. Mean h4β7 score Γ100: dev 3.69 (v2b 3.62), late 8.83 (v2b 8.73); h1β3: dev 2.86 (v2b 2.78), late 2.73 (v2b 2.60). For the 2011 origin over all 7 years: 9.81 (v2b 9.75). These gains are small, probably within noise: +0.0023 dev / +0.0014 late temperature skill and +0.0004 / +0.0028 hottest-day skill. Compared with v2b, v4 changes 2019β2025 temperature forecasts by a median of Β±0.04 Β°F (max 0.28 Β°F) and raises hottest-day forecasts by 0.02 Β°F on average (range β0.10 to +0.25 Β°F). CDD and precipitation forecasts are identical to v2b.
v3: pre-registered ablations of the "SHEBE" component ideas (none adopted)
Tejas pasted an outside AI assistant's proposal ("SHEBE"). It suggested empirical-Bayes regional slopes, a regional-anomaly level, seasonal CDD shifts, a Jacobian CDD blend, robust precipitation and hottest-day levels, and station-move homogenisation. The backtest numbers it claimed were not used, and nothing we ran comes close to them. We wrote research/V3_PROTOCOL.md and hashed it (results/v3_protocol_hash.txt) before computing any candidate. It defines 42 candidates and one rule: a component is adopted only if its mean h4β7 skill on dev origins 1985β2004 beats the reference (v2 temperature, v2b CDD, 30-year mean for precipitation and hottest day) and its mean on the locked late origins 2005β2011 is no worse. No slope multipliers were allowed. Code: code/v3lib.py, code/backtest_v3.py. Full tables: results/v3_tables.md.
No candidate passed, so the v3 forecast is v2b, byte for byte. Mean h4β7 skill (dev / late):
| Idea | Best variant | dev | late | Reference dev / late | Outcome |
|---|---|---|---|---|---|
| EB-shrunk city slopes (OLS/Huber/TheilβSen, 30/40/50 y, K=12 IDW prior) | eb_huber_30 |
0.096 | 0.126 | 0.124 / 0.143 | worse on both |
| Regional (IDW) slope | reg30 |
0.093 | 0.120 | 0.124 / 0.143 | worse on both |
| Longer national slope window | nat40 |
0.120 | 0.154 | 0.124 / 0.143 | worse on dev |
| Regional-anomaly level (12/15 y) | ra15 |
0.112 | 0.131 | 0.124 / 0.143 | worse on both |
| Station-move homogenisation | homog |
0.1251 | 0.1420 | 0.1239 / 0.1432 | +0.001 dev, β0.001 late: fails |
| Seasonal (MayβOct / NovβApr) CDD shift, national slopes | season_nat_L15 |
0.038 | 0.177 | 0.021 / 0.206 | better dev, worse late: fails |
| Seasonal CDD shift, EB slopes | season_eb_L15 |
0.017 | 0.153 | 0.021 / 0.206 | worse on both |
| 0.65 shift + 0.35 Jacobian blend | jacw |
0.031 | 0.196 | 0.021 / 0.206 | fails late |
| Leap-year / coverage normalisation | norm_leap |
0.019 | 0.208 | 0.021 / 0.206 | fails dev |
| Precip: 30/45-y blend, SHEBE blend, Huber, regional ratio | blend3045 |
+0.003 | β0.001 | 0 / 0 | fails late |
| Hottest: Huber, top-3 heatwave anchor, damped JJA slope (0.25β0.5) | huber30 |
β0.001 | β0.001 | 0 / 0 | none beats 30-y mean on dev |
Findings. (1) Shrinking city slopes toward regional ones loses to a single national slope by about 0.03 skill in both blocks. At 4β7 years ahead, any city-specific slope carries more noise than signal. (2) Station homogenisation is real: city_info lists 461 station records with start and end dates, which match the segments in coverage_and_stations. Its effect on scores is negligible because v2 already fits one intercept per station segment. (3) Seasonal CDD slopes are the only idea that beats v1's exp10 on dev origins (0.038 vs 0.029), which would have removed the need for v2b's protocol deviation. However, they lose 0.03 on the late block, so the rule rejects them. (4) For precipitation and hottest day, every robust-level or trend idea stays within Β±0.01 of the 30-year mean, as the organizers' guidance predicts.
Protocol deviation (CDD, inherited from v2b), disclosed
The frozen rule (select on origins 1985β2004) kept v1's exp10 for CDD: the daily-shift challenger scored J 0.024 vs 0.030 there. We adopted the daily shift anyway, for three reasons, all based on pre-2019 data:
- Consistency: a flat CDD forecast contradicts the rising temperature forecast that the protocol selected. The shift derives CDD from that same slope.
- Regime evidence: the shift beats
exp10at h4β7 in every origin from 2000 to 2011 (12 of 12; late-block CDD skill 0.206 vs 0.026). It loses mainly at origins 1988β1994, whose 30-year windows include mid-century cooling. The 2018 forecast window (1989β2018) lies entirely in the post-1980 warming period. - Official practice run (2010 β 2011β2018): CDD MAE 169.7 vs 204.7 (exp10) and 174.6 (linear trend).
Honest caveat: we made this decision after we saw that v1 scored poorly on the live leaderboard, and after we saw other teams' daily-shift methods. No live number was used to set any parameter: the configuration is the protocol's selected temperature configuration, with no new tuning. Still, the choice of method may have been influenced by that exposure. The protocol-strict forecast is kept at results/submission_v2_strict.csv (SHA-256 ca058bcab452c0c814c75f47dda1d2aec62470ee38c78d0d45abbc0ec5c1d49c). Another team (Evanguo12) disclosed the same kind of CDD deviation.
Policy Idea
- Risk: cooling degree days rise faster than average temperature, because warming pushes more days above 65 Β°F. Using a 30-year normal, as many plans do, understates the cooling load a city should prepare for. Our forecast puts 2022β2025 CDD above the 1989β2018 normal in 208 of 210 cities.
- Action: municipal sustainability offices and utility planners publish an annual airport-station cooling-demand screening report that shows the trend-aware CDD forecast next to the 30-year normal, its historical error, missing-data and station-change flags. They use it to choose which public buildings (schools, libraries, senior centres) get building-level cooling and efficiency audits first. They should not use it to rank funding on its own.
- Why sustainable: it relies on free public station data and reruns each year in minutes. It steers audits toward efficiency and passive cooling before new peak capacity is added. CDD is not electricity demand or health risk. Building, occupancy and vulnerability data are needed before investment, and no energy saving or health benefit is claimed.
Feedback to the Ecosystem
- Origins matter more than models: backtests whose 30-year windows include mid-century cooling (origins before ~1995) favour climatology. Several teams, including our v1, were misled by this. A recommended rolling-origin range in the starter kit would help.
- Publish a machine-readable scored-mask spec: CDD sums complete Tmax/Tmin pairs while eligibility checks each separately (18 city-years differ in 1989β2018).
- Precipitation 0.00 for all-missing years should be stored as missing (teammate found 935 such city-years).
- Ship an offline validator that takes the pinned city list; the template checker reads it online.
- Clarify, in one place, whether choosing between one's own submitted versions by live score is allowed: team READMEs interpret it differently.
Files and reproduction
submission.csvβ v9, 5,880 rows, template order, official validator passes. SHA-256c17072c32a5fab8547c277ac6bafa18d03519d76f6a67f8c906e51401bcfd5b6(identical toresults/submission_v9.csv). The previous v7 forecast is kept atresults/submission_v7.csv(SHA-256ea5369650c47fb4b1b5052b7a415ca96b0b4a013f325f4e64e06dc3e7eb01ec7); the v6 forecast is kept atresults/submission_v6.csv(SHA-25692b2155f43e6f997adcfa5bdcddf8dd7769bd68256935a2ce01b9d10dbe8deed); the v4 forecast is kept atresults/submission_v4.csv(SHA-256d4ae4e81072cb5e67c4543ee8b3b69d67d1434b20619bd3759cf54fbed031442).code/βv2lib.py(data, forecasts, scoring),backtest_v2.py,report_v2.py,make_submission_v2.py,practice_run.py,test_v2.py,reproduce.py; v3v3lib.py,backtest_v3.py,report_v3.py,make_submission_v3.py,test_v3.py; v4v4lib.py,backtest_v4.py,report_v4.py,make_submission_v4.py,test_v4.py,reproduce_v4.py; v6v6lib.py,backtest_v6.py,make_submission_v6.py,reproduce_v6.py; v7v7lib.py,backtest_v7.py,make_submission_v7.py,reproduce_v7.py; v8v8lib.py(library used by v9); v9v9lib.py,backtest_v9.py,make_submission_v9.py,reproduce_v9.py; pinnedrequirements.txt.data/β byte-identical pinned copies ofannual_indices,coverage_and_stations,city_info,daily(needed for CDD shifts) fromCMU-OSPO/protostar-climate-train-to-2018@bbbdebda, with hashes indata/manifest.json.research/β frozen protocol, official rule/scorer snapshots, other-teams survey.results/β all backtest outputs.
python3.12 -m venv .venv && . .venv/bin/activate
pip install -r code/requirements.txt
python code/reproduce_v9.py # offline; reruns the v9 backtest and forecast and checks submission.csv byte-for-byte
# code/reproduce_v7.py, code/reproduce_v6.py and code/reproduce_v4.py reproduce earlier studies; their final checks compare submission.csv with v7/v6/v4 and will now fail by design
Data
No external data in the forecasts. Training data: Lai & Dzombak (2019) records, via CMU-OSPO/protostar-climate-train-to-2018 (CC0-1.0); cite DOI 10.1184/R1/7890488.v6, 10.1184/R1/7961012.v6 and J. Climate 32(14) 4299β4320. The release is retrospective; historical publication vintages and station geolocation history are not available. See data/README.md.
- Downloads last month
- 358