Predictor — learning loop
Aratea is a weather-factor discovery engine. Every named feature here is a hypothesis; every training run measures whether it carries signal. The bench is the same row-set kalshi_mid Brier — beat the market, on its own ground.
Everything the manifest carries: named factors with their leave-one-out delta, paper-trade ledger, training runs and Brier trajectory. This is the meteorologist / actuary view — no rounding, no sugar-coating.
Features tracked
17
Active
3
Experimental
6
Dropped
0
Paper bets (open / resolved)
15 / 650
Phase 1: 665/50
Manifest generated at 2026-08-10T19:29:07Z (schema v3).
Hybrid effective sample (N_eff)
α = 0.3730.4= 650 + 0.3 × 268 = 730.4
N_live (real paper trades)
650
N_backtest_strict (replay, point-in-time)
268
NAIVE-excluded (informational)
0
Phase 1 gate reached (live only) ✓
N_eff drives secondary decisions only — feature-set selection, reliability plots, complementary promotion check. The Phase 1 go/no-go gate stays strictly on N_live; backtest volume never substitutes for live trades there.
Read CONVENTION §6.bisSeries
Status
A. Live runs (Kalshi paper trades)
Each row is a real paper trade on Kalshi. The champion takes the position (real ledger row, real P&L); challengers and baselines run in shadow mode for Brier comparison. ★ marks the best Brier on a given run. The promotion rule (champion swap) needs a rolling-mean Brier dominance over N≥10 resolved trades — single-run wins are anecdotal.
| Run | When | Event / Bin | Side | Champion p | Challenger p | Baseline p | kalshi_mid | Outcome | P&L paper |
|---|---|---|---|---|---|---|---|---|---|
| 666 | 2026-08-09 | LOWTPHX 10/8B85.5 | YES | 17.2% | 16.6% | 6.5% | 6.5% | PENDING | — |
| 665 | 2026-08-09 | LOWTPHX 10/8B89.5 | NO | 19.2% | 38.6% | 47.5% | 47.5% | PENDING | — |
| 664 | 2026-08-09 | LOWTMIA 10/8B80.5 | NO | 18.4% | 12.9% | 38.5% | 38.5% | PENDING | — |
| 663 | 2026-08-09 | LOWTBOS 10/8B66.5 | NO | 3.4% | 5.8% | 9.5% | 9.5% | PENDING | — |
| 662 | 2026-08-09 | LOWTDC 10/8B72.5 | NO | 2.7% | 6.2% | 9.0% | 9.0% | PENDING | — |
| 661 | 2026-08-09 | LOWTDC 10/8B76.5 | YES | 54.5% | 90.3% | 45.5% | 45.5% | PENDING | — |
| 660 | 2026-08-09 | LOWTCHI 10/8B72.5 | NO | 17.9% | 25.0% | 36.5% | 36.5% | PENDING | — |
| 659 | 2026-08-09 | LOWTCHI 10/8B70.5 | YES | 60.9% | 93.3% | 36.5% | 36.5% | PENDING | — |
| 658 | 2026-08-09 | LOWTSFO 10/8B55.5 | NO | 5.4% | 1.9% | 13.0% | 13.0% | PENDING | — |
| 657 | 2026-08-09 | LOWTSFO 10/8B57.5 | NO | 40.1% | 25.7% | 61.5% | 61.5% | PENDING | — |
| 656 | 2026-08-09 | LOWTSFO 10/8B59.5 | YES | 49.8% | 80.7% | 20.5% | 20.5% | PENDING | — |
| 655 | 2026-08-09 | LOWTLAX 10/8B67.5 | NO | 4.5% | 6.2% | 48.5% | 48.5% | PENDING | — |
| 654 | 2026-08-09 | LOWTNYC 10/8B69.5 | NO | 2.6% | 5.8% | 14.0% | 14.0% | PENDING | — |
| 653 | 2026-08-09 | LOWTNYC 10/8B71.5 | NO | 12.3% | 12.5% | 28.5% | 28.5% | PENDING | — |
| 652 | 2026-08-09 | LOWTNYC 10/8B75.5 | YES | 31.8% | 61.4% | 13.5% | 13.5% | PENDING | — |
| 651 | 2026-08-07 | LOWTDEN 8/8B64.5 | NO | 2.9%B=0.0008 ★ | 5.0%B=0.0025 | 26.5%B=0.0702 | 26.5% | WIN (NO) | +$22.00 |
| 650 | 2026-08-07 | LOWTDEN 8/8B66.5 | NO | 5.3%B=0.8968 | 7.7%B=0.8511 | 33.5%B=0.4422 ★ | 33.5% | LOSS (YES) | −$61.18 |
| 649 | 2026-08-07 | LOWTPHX 8/8B88.5 | NO | 7.4%B=0.0055 ★ | 8.6%B=0.0074 | 15.5%B=0.0240 | 15.5% | WIN (NO) | +$11.16 |
| 648 | 2026-08-07 | LOWTPHX 8/8B90.5 | NO | 20.0%B=0.0400 ★ | 30.1%B=0.0905 | 48.5%B=0.2352 | 48.5% | WIN (NO) | +$57.71 |
| 647 | 2026-08-07 | LOWTMIA 8/8B81.5 | YES | 42.1%B=0.3348 | 66.1%B=0.1149 ★ | 35.0%B=0.4225 | 35.0% | WIN (YES) | +$113.75 |
| 646 | 2026-08-07 | LOWTMIA 8/8B79.5 | NO | 31.3%B=0.0979 | 23.4%B=0.0549 ★ | 39.5%B=0.1560 | 39.5% | WIN (NO) | +$39.90 |
| 645 | 2026-08-07 | LOWTBOS 8/8B72.5 | YES | 38.2%B=0.1459 | 62.3%B=0.3877 | 19.5%B=0.0380 ★ | 19.5% | LOSS (NO) | −$61.23 |
| 644 | 2026-08-07 | LOWTDC 8/8B75.5 | NO | 30.1%B=0.0905 ★ | 44.5%B=0.1983 | 39.0%B=0.1521 | 39.0% | WIN (NO) | +$39.00 |
| 643 | 2026-08-07 | LOWTCHI 8/8B67.5 | NO | 2.3%B=0.0005 ★ | 6.8%B=0.0046 | 14.0%B=0.0196 | 14.0% | WIN (NO) | +$9.94 |
| 642 | 2026-08-07 | LOWTCHI 8/8B71.5 | YES | 54.2%B=0.2942 | 88.9%B=0.7900 | 38.0%B=0.1444 ★ | 38.0% | LOSS (NO) | −$61.18 |
| 641 | 2026-08-07 | LOWTSFO 8/8B59.5 | YES | 44.0%B=0.1940 | 73.9%B=0.5468 | 24.5%B=0.0600 ★ | 24.5% | LOSS (NO) | −$61.25 |
| 640 | 2026-08-07 | LOWTSFO 8/8B57.5 | NO | 33.2%B=0.4459 | 23.0%B=0.5936 | 60.5%B=0.1560 ★ | 60.5% | LOSS (YES) | −$61.23 |
| 639 | 2026-08-07 | LOWTLAX 8/8B65.5 | NO | 2.9%B=0.0008 ★ | 3.3%B=0.0011 | 13.5%B=0.0182 | 13.5% | WIN (NO) | +$9.45 |
| 638 | 2026-08-07 | LOWTLAX 8/8B67.5 | NO | 13.1%B=0.7554 | 11.7%B=0.7795 | 55.0%B=0.2025 ★ | 55.0% | LOSS (YES) | −$61.20 |
| 637 | 2026-08-07 | LOWTNYC 8/8B75.5 | YES | 58.3%B=0.3399 | 57.2%B=0.3267 | 40.5%B=0.1640 ★ | 40.5% | LOSS (NO) | −$61.16 |
| 636 | 2026-08-04 | LOWTDEN 5/8B53.5 | NO | 13.5%B=0.0183 ★ | 18.6%B=0.0347 | 29.5%B=0.0870 | 29.5% | WIN (NO) | ±$0.00 |
| 635 | 2026-08-04 | LOWTDEN 5/8B57.5 | YES | 23.2%B=0.0537 | 39.5%B=0.1563 | 4.5%B=0.0020 ★ | 4.5% | LOSS (NO) | −$72.36 |
| 634 | 2026-08-04 | LOWTPHX 5/8B85.5 | YES | 17.5%B=0.0306 | 15.0%B=0.0225 | 4.0%B=0.0016 ★ | 4.0% | LOSS (NO) | −$72.32 |
| 633 | 2026-08-04 | LOWTPHX 5/8B89.5 | NO | 19.8%B=0.0394 ★ | 22.6%B=0.0509 | 45.5%B=0.2070 | 45.5% | WIN (NO) | +$60.06 |
| 632 | 2026-08-04 | LOWTMIA 5/8B78.5 | NO | 31.8%B=0.4652 | 16.6%B=0.6957 | 39.0%B=0.3721 ★ | 39.0% | LOSS (YES) | −$71.98 |
| 631 | 2026-08-04 | LOWTMIA 5/8B76.5 | NO | 18.7%B=0.0351 | 6.4%B=0.0040 ★ | 31.5%B=0.0992 | 31.5% | WIN (NO) | +$33.07 |
| 630 | 2026-08-04 | LOWTMIA 5/8B80.5 | YES | 26.4%B=0.0696 | 47.9%B=0.2296 | 13.5%B=0.0182 ★ | 13.5% | LOSS (NO) | −$72.23 |
| 629 | 2026-08-04 | LOWTDC 5/8B73.5 | YES | 37.1%B=0.1374 | 53.2%B=0.2825 | 30.5%B=0.0930 ★ | 30.5% | LOSS (NO) | −$72.28 |
| 628 | 2026-08-04 | LOWTDC 5/8B71.5 | NO | 3.5%B=0.0012 ★ | 6.3%B=0.0040 | 11.5%B=0.0132 | 11.5% | WIN (NO) | +$9.31 |
| 627 | 2026-08-04 | LOWTCHI 5/8B65.5 | YES | 15.1%B=0.0228 | 3.9%B=0.0015 ★ | 4.5%B=0.0020 | 4.5% | LOSS (NO) | −$72.36 |
| 626 | 2026-08-04 | LOWTCHI 5/8B67.5 | YES | 31.5%B=0.0991 | 14.8%B=0.0220 ★ | 20.5%B=0.0420 | 20.5% | LOSS (NO) | −$72.16 |
| 625 | 2026-08-04 | LOWTCHI 5/8B69.5 | NO | 30.5%B=0.4833 | 14.3%B=0.7352 | 43.0%B=0.3249 ★ | 43.0% | LOSS (YES) | −$71.82 |
| 624 | 2026-08-04 | LOWTSFO 5/8B56.5 | NO | 11.0%B=0.0121 | 1.6%B=0.0002 ★ | 22.5%B=0.0506 | 22.5% | WIN (NO) | +$20.92 |
| 623 | 2026-08-04 | LOWTSFO 5/8B58.5 | NO | 50.7%B=0.2428 | 70.1%B=0.0895 ★ | 62.5%B=0.1406 | 62.5% | LOSS (YES) | −$72.00 |
| 622 | 2026-08-04 | LOWTLAX 5/8B67.5 | NO | 44.9%B=0.3038 | 78.0%B=0.0484 ★ | 56.5%B=0.1892 | 56.5% | LOSS (YES) | −$72.21 |
| 621 | 2026-08-04 | LOWTNYC 5/8B67.5 | NO | 22.7%B=0.0514 ★ | 24.9%B=0.0618 | 39.5%B=0.1560 | 39.5% | WIN (NO) | +$47.00 |
| 620 | 2026-08-02 | LOWTDEN 3/8B68.5 | NO | 8.0%B=0.0064 | 2.6%B=0.0007 ★ | 15.0%B=0.0225 | 15.0% | WIN (NO) | ±$0.00 |
| 619 | 2026-08-02 | LOWTPHX 3/8B92.5 | NO | 34.3%B=0.4320 | 62.4%B=0.1415 ★ | 46.5%B=0.2862 | 46.5% | LOSS (YES) | −$82.93 |
| 618 | 2026-08-02 | LOWTPHX 3/8B90.5 | YES | 23.6%B=0.0559 | 29.7%B=0.0880 | 10.5%B=0.0110 ★ | 10.5% | LOSS (NO) | −$82.95 |
| 617 | 2026-08-02 | LOWTMIA 3/8B77.5 | NO | 21.2%B=0.0450 | 4.6%B=0.0021 ★ | 32.5%B=0.1056 | 32.5% | WIN (NO) | +$39.65 |
| 616 | 2026-08-02 | LOWTMIA 3/8B79.5 | YES | 49.0%B=0.2400 | 38.4%B=0.1478 | 35.5%B=0.1260 ★ | 35.5% | LOSS (NO) | −$82.71 |
| 615 | 2026-08-02 | LOWTMIA 3/8B81.5 | YES | 31.4%B=0.0984 | 13.1%B=0.0170 | 9.5%B=0.0090 ★ | 9.5% | LOSS (NO) | −$82.94 |
| 614 | 2026-08-02 | LOWTBOS 3/8B68.5 | YES | 26.1%B=0.0682 | 29.7%B=0.0885 | 15.5%B=0.0240 ★ | 15.5% | LOSS (NO) | −$82.93 |
| 613 | 2026-08-02 | LOWTDC 3/8B74.5 | YES | 66.1%B=0.1149 | 73.9%B=0.0680 ★ | 45.5%B=0.2970 | 45.5% | WIN (YES) | +$99.19 |
| 612 | 2026-08-02 | LOWTCHI 3/8B62.5 | YES | 46.5%B=0.2867 ★ | 32.8%B=0.4517 | 32.5%B=0.4556 | 32.5% | WIN (YES) | +$172.12 |
| 611 | 2026-08-02 | LOWTCHI 3/8B64.5 | YES | 49.5%B=0.2450 | 41.1%B=0.1686 | 35.5%B=0.1260 ★ | 35.5% | LOSS (NO) | −$82.71 |
| 610 | 2026-08-02 | LOWTSFO 3/8B56.5 | NO | 55.4%B=0.1986 | 51.0%B=0.2399 | 67.0%B=0.1089 ★ | 67.0% | LOSS (YES) | −$82.83 |
| 609 | 2026-08-02 | LOWTSFO 3/8B54.5 | YES | 28.0%B=0.0786 | 8.0%B=0.0063 ★ | 13.5%B=0.0182 | 13.5% | LOSS (NO) | −$82.89 |
| 608 | 2026-08-02 | LOWTLAX 3/8B67.5 | NO | 40.0%B=0.3606 | 22.8%B=0.5953 | 61.0%B=0.1521 ★ | 61.0% | LOSS (YES) | −$82.68 |
| 607 | 2026-08-02 | LOWTNYC 3/8B70.5 | NO | 6.1%B=0.8817 | 1.4%B=0.9731 | 24.5%B=0.5700 ★ | 24.5% | LOSS (YES) | −$82.30 |
| 606 | 2026-08-02 | LOWTNYC 3/8B72.5 | NO | 24.3%B=0.0593 | 6.0%B=0.0036 ★ | 45.0%B=0.2025 | 45.0% | WIN (NO) | +$67.50 |
| 605 | 2026-08-02 | LOWTNYC 3/8B74.5 | YES | 65.3%B=0.4261 | 71.0%B=0.5045 | 11.5%B=0.0132 ★ | 11.5% | LOSS (NO) | −$82.92 |
| 604 | 2026-07-31 | LOWTDEN 1/8B64.5 | YES | 41.0%B=0.1682 | 55.3%B=0.3053 | 13.5%B=0.0182 ★ | 13.5% | LOSS (NO) | ±$0.00 |
| 603 | 2026-07-31 | LOWTMIA 1/8B79.5 | NO | 25.7%B=0.0658 | 11.1%B=0.0124 ★ | 40.5%B=0.1640 | 40.5% | WIN (NO) | +$53.87 |
| 602 | 2026-07-31 | LOWTMIA 1/8B77.5 | NO | 17.6%B=0.0311 | 3.2%B=0.0011 ★ | 35.5%B=0.1260 | 35.5% | WIN (NO) | +$43.31 |
| 601 | 2026-07-31 | LOWTDC 1/8B68.5 | NO | 4.4%B=0.0020 ★ | 7.5%B=0.0056 | 15.0%B=0.0225 | 15.0% | WIN (NO) | +$13.95 |
| 600 | 2026-07-31 | LOWTDC 1/8B72.5 | YES | 43.9%B=0.1929 | 66.8%B=0.4460 | 25.5%B=0.0650 ★ | 25.5% | LOSS (NO) | −$79.05 |
| 599 | 2026-07-31 | LOWTCHI 1/8B68.5 | YES | 44.9%B=0.2013 | 75.4%B=0.5679 | 35.5%B=0.1260 ★ | 35.5% | LOSS (NO) | −$79.16 |
| 598 | 2026-07-31 | LOWTCHI 1/8B64.5 | NO | 3.4%B=0.0012 ★ | 6.5%B=0.0043 | 13.5%B=0.0182 | 13.5% | WIN (NO) | +$12.28 |
| 597 | 2026-07-31 | LOWTCHI 1/8B66.5 | YES | 42.8%B=0.3270 | 67.3%B=0.1072 ★ | 27.5%B=0.5256 | 27.5% | WIN (YES) | +$208.07 |
| 596 | 2026-07-31 | LOWTSFO 1/8B54.5 | YES | 14.2%B=0.0202 | 2.5%B=0.0006 ★ | 5.5%B=0.0030 | 5.5% | LOSS (NO) | −$79.14 |
| 595 | 2026-07-31 | LOWTSFO 1/8B58.5 | YES | 40.1%B=0.1608 | 23.7%B=0.0562 ★ | 29.5%B=0.0870 | 29.5% | LOSS (NO) | −$79.06 |
| 594 | 2026-07-31 | LOWTSFO 1/8B56.5 | NO | 34.7%B=0.4269 | 14.4%B=0.7321 | 59.5%B=0.1640 ★ | 59.5% | LOSS (YES) | −$78.98 |
| 593 | 2026-07-31 | LOWTLAX 1/8B68.5 | NO | 37.3%B=0.1388 | 23.4%B=0.0550 ★ | 43.5%B=0.1892 | 43.5% | WIN (NO) | +$60.90 |
| 592 | 2026-07-31 | LOWTLAX 1/8B66.5 | NO | 25.7%B=0.5518 | 8.9%B=0.8302 | 32.5%B=0.4556 ★ | 32.5% | LOSS (YES) | −$78.98 |
| 591 | 2026-07-31 | LOWTNYC 1/8B71.5 | YES | 41.3%B=0.3440 ★ | 24.0%B=0.5782 | 29.5%B=0.4970 | 29.5% | WIN (YES) | +$188.94 |
| 590 | 2026-07-31 | LOWTNYC 1/8B67.5 | NO | 8.7%B=0.0075 | 1.6%B=0.0003 ★ | 26.5%B=0.0702 | 26.5% | WIN (NO) | +$28.36 |
| 589 | 2026-07-31 | LOWTNYC 1/8B69.5 | NO | 22.8%B=0.0520 | 5.3%B=0.0028 ★ | 41.5%B=0.1722 | 41.5% | WIN (NO) | +$56.03 |
| 588 | 2026-07-29 | LOWTDEN 30/7B65.5 | YES | 45.7%B=0.2086 | 33.5%B=0.1123 | 29.0%B=0.0841 ★ | 29.0% | LOSS (NO) | ±$0.00 |
| 587 | 2026-07-29 | LOWTPHX 30/7B90.5 | NO | 23.7%B=0.0563 ★ | 39.5%B=0.1564 | 37.5%B=0.1406 | 37.5% | WIN (NO) | +$49.12 |
| 586 | 2026-07-29 | LOWTMIA 30/7B80.5 | YES | 40.0%B=0.3600 ★ | 22.3%B=0.6037 | 31.0%B=0.4761 | 31.0% | WIN (YES) | +$182.85 |
| 585 | 2026-07-29 | LOWTMIA 30/7B78.5 | NO | 24.7%B=0.0610 | 6.2%B=0.0039 ★ | 35.5%B=0.1260 | 35.5% | WIN (NO) | +$45.08 |
| 584 | 2026-07-29 | LOWTBOS 30/7B61.5 | NO | 5.1%B=0.0026 ★ | 6.9%B=0.0048 | 11.0%B=0.0121 | 11.0% | WIN (NO) | +$10.12 |
| 583 | 2026-07-29 | LOWTDC 30/7B68.5 | NO | 12.8%B=0.0165 | 2.5%B=0.0006 ★ | 27.0%B=0.0729 | 27.0% | WIN (NO) | +$30.24 |
| 582 | 2026-07-29 | LOWTCHI 30/7B63.5 | YES | 42.5%B=0.3304 ★ | 27.2%B=0.5305 | 33.5%B=0.4422 | 33.5% | WIN (YES) | +$162.93 |
| 581 | 2026-07-29 | LOWTCHI 30/7B65.5 | YES | 33.3%B=0.1108 | 15.8%B=0.0249 ★ | 24.0%B=0.0576 | 24.0% | LOSS (NO) | −$82.08 |
| 580 | 2026-07-29 | LOWTSFO 30/7B58.5 | YES | 30.6%B=0.0935 | 11.7%B=0.0137 ★ | 17.0%B=0.0289 | 17.0% | LOSS (NO) | −$82.11 |
| 579 | 2026-07-29 | LOWTSFO 30/7B54.5 | YES | 22.7%B=0.0515 | 5.1%B=0.0026 ★ | 7.0%B=0.0049 | 7.0% | LOSS (NO) | −$82.11 |
| 578 | 2026-07-29 | LOWTSFO 30/7B56.5 | NO | 35.3%B=0.4190 | 15.0%B=0.7218 | 55.5%B=0.1980 ★ | 55.5% | LOSS (YES) | −$81.88 |
| 577 | 2026-07-29 | LOWTLAX 30/7B63.5 | YES | 12.0%B=0.0145 | 2.2%B=0.0005 ★ | 4.0%B=0.0016 | 4.0% | LOSS (NO) | −$82.16 |
| 576 | 2026-07-29 | LOWTLAX 30/7B67.5 | NO | 39.5%B=0.3661 | 20.6%B=0.6309 | 52.0%B=0.2304 ★ | 52.0% | LOSS (YES) | −$82.08 |
| 575 | 2026-07-29 | LOWTLAX 30/7B65.5 | YES | 28.7%B=0.0821 | 8.7%B=0.0075 | 5.0%B=0.0025 ★ | 5.0% | LOSS (NO) | −$82.15 |
| 574 | 2026-07-29 | LOWTNYC 30/7B64.5 | NO | 13.1%B=0.0171 | 2.4%B=0.0006 ★ | 24.5%B=0.0600 | 24.5% | WIN (NO) | +$26.46 |
| 573 | 2026-07-29 | LOWTNYC 30/7B68.5 | YES | 42.5%B=0.1810 | 24.9%B=0.0619 | 17.5%B=0.0306 ★ | 17.5% | LOSS (NO) | −$82.07 |
| 572 | 2026-07-27 | HIGHTSFO 28/7B71.5 | NO | 15.7%B=0.7102 | 17.3%B=0.6836 | 30.5%B=0.4830 ★ | 30.5% | LOSS (YES) | −$5.56 |
| 571 | 2026-07-27 | LOWTDEN 28/7B68.5 | YES | 53.7%B=0.2889 | 87.2%B=0.7609 | 12.0%B=0.0144 ★ | 12.0% | LOSS (NO) | −$79.44 |
| 570 | 2026-07-27 | LOWTPHX 28/7B87.5 | YES | 11.1%B=0.0123 | 11.7%B=0.0138 | 3.5%B=0.0012 ★ | 3.5% | LOSS (NO) | −$78.19 |
| 569 | 2026-07-27 | LOWTMIA 28/7B79.5 | YES | 41.2%B=0.1699 | 25.2%B=0.0636 | 19.5%B=0.0380 ★ | 19.5% | LOSS (NO) | −$79.37 |
| 568 | 2026-07-27 | LOWTMIA 28/7B77.5 | NO | 25.2%B=0.5599 | 3.8%B=0.9259 | 47.5%B=0.2756 ★ | 47.5% | LOSS (YES) | −$79.28 |
| 567 | 2026-07-27 | LOWTDC 28/7B67.5 | YES | 7.1%B=0.0050 | 12.1%B=0.0147 | 4.5%B=0.0020 ★ | 4.5% | LOSS (NO) | −$74.97 |
★ = best Brier this run · B = Brier score per model · P&L = champion only (challengers and baselines are shadow; no real exposure).
B. Named factors
Each row is a named hypothesis used by the learned predictor at training time. Brier Δ is the leave-one-out test delta from the most recent training run — sort by it to see what carried the model.
| Name | Hypothesis | Source | Added | Brier Δ ↓ | Status |
|---|---|---|---|---|---|
| p_ensemble | Mean of four vendor probabilities (ECMWF + GraphCast + GFS + JMA). Hypothesis: vendor disagreement washes out, the mean is the wisest single bet. (Bench 2026-05-11 N=138: ensemble Brier 0.1429 vs kalshi_mid 0.0845 — the average **lost** to the market, so we need to learn weights instead of averaging blindly.) | derived from `predictors/ensemble.py` | 2026-05-09 | ↑ +0.0041 | active |
| forecast_spread | Max − min of the per-vendor probabilities (proxy of model disagreement). Hypothesis: when vendors disagree, the prediction is less trustworthy and the market mid carries more weight than the model. | derived from `predictions.ensemble.inputs.individual_probs` | 2026-05-09 | ↑ +0.0033 | active |
| p_climatology | Historical base rate of (variable in [lower, upper]) over the same date-of-year window from the past 15 years. The dumb-but-honest prior every forecast must beat. | derived from `predictors/climatology.py` (Open-Meteo historical) | 2026-05-09 | ↑ +0.0015 | experimental |
| urban_density_5km | OSM `way["building"]` count within 5 km of the station. Hypothesis: urban heat island raises overnight lows above what a non-urban climatology predicts → biases low-temp markets in cities. Units: building count (not %-area; see README for why). | OSM Overpass API | 2026-05-11 | ↑ +0.0000 | dropped (v3, 2026-06-05 — noise as additive linear term) |
| elevation_m | USGS EPQS elevation at the station point. Hypothesis: thinner air at altitude amplifies the diurnal swing (Denver KDEN ~1638 m vs. Miami KMIA ~2 m at the extremes of our station set). | https://epqs.nationalmap.gov/v1/json | 2026-05-11 | ↑ +0.0000 | dropped (v3, 2026-06-05 — noise as additive linear term) |
| latitude | Station latitude (degrees, signed). Hypothesis: insolation, daylight length, and seasonal amplitude scale with `cos(latitude)` — explicit feature lets the learner discover the interaction with the date-of-year encoded in climatology. | NWS_STATIONS table | 2026-05-11 | ↑ +0.0000 | dropped (v3, 2026-06-05 — noise as additive linear term) |
| forest_pct_5km | OSM `natural=wood` + `landuse=forest` feature count within 5 km. Hypothesis: canopy cover lowers daytime highs (shade + evapotranspiration) and limits radiative night cooling (canopy traps). Units: feature count. | OSM Overpass API | 2026-05-11 | ↑ +0.0000 | dropped (v3, 2026-06-05 — noise as additive linear term) |
| days_ahead | Days between snapshot and target_date. Hypothesis: forecast skill decays with horizon, learned weights should interact non-linearly with this. | derived from `predictions.forecast_blend.inputs.days_ahead` | 2026-05-09 | · ±0.0000 | experimental |
| water_pct_10km | OSM `natural=water` + `waterway=*` feature count within 10 km. Hypothesis: large water bodies dampen diurnal swings via thermal inertia → tightens the [lower, upper] hit probability for both highs and lows. Units: feature count (kept the `_pct_` name from the spec for continuity). | OSM Overpass API | 2026-05-11 | ↓ −0.0000 | dropped (v3, 2026-06-05 — noise as additive linear term) |
| distance_to_coast_km | Haversine distance to the nearest Natural Earth 1:50m coastline vertex. Hypothesis: maritime regime (Boston, Miami, SFO) damps extremes; continental regime (Denver, Oklahoma City) amplifies them. | Natural Earth `ne_50m_coastline.geojson` | 2026-05-11 | ↓ −0.0000 | dropped (v3, 2026-06-05 — noise as additive linear term) |
| p_consensus | Mean of the three correlated probability views (`p_climatology` + `p_forecast_blend` + `p_ensemble`). Hypothesis: those three estimate the same P(YES) by different routes and are near-collinear; under L2 the learner splits one signal across three compensating coefficients (the +1.07 / -0.87 / -0.40 pattern measured on the v2 run). Collapsing them into their mean keeps the shared signal on one stable coefficient, with the orthogonal disagreement axis carried by `forecast_spread`. Standard mean+spread reparametrisation of a collinear block. | derived from `predictors/{climatology,forecast_blend,ensemble}.py` | 2026-06-05 | ↓ −0.0011 | experimental |
| p_forecast_blend | Open-Meteo deterministic forecast around target_date, blended with climatology by horizon. Hypothesis: state-of-art deterministic forecast carries calibrated short-horizon signal. | derived from `predictors/forecast_blend.py` | 2026-05-09 | ↓ −0.0021 | active |
| p_nws_ndfd | P(YES) computed from the NWS NDFD official forecast, gaussian around NDFD temp with sigma from climatology range. Hypothesis: the agency that *resolves* Kalshi weather markets (NWS Climatological Report Daily) should issue the highest-signal forecast available. | https://api.weather.gov | 2026-05-11 | TBD (forward-only — no historical coverage yet) | experimental |
| series_bias_prior | Known mean bias (p_consensus − y) per series_ticker over 61-date backfill. Hypothesis: each Kalshi weather series has a stable series-specific intercept (KXHIGHTSFO −0.090 to BOS/LAX ~0); this continuous prior generalises `is_hightemp` without per-series dummy variables. Expected coef ≈ −1. | backfill_dataset analysis (B24) | 2026-06-21 | TBD (v3b run, pending HOLDOUT > 20 dates) | experimental |
| forecast_revision | Change in p_consensus between earliest and latest capture of the same ticker. Hypothesis: drift velocity of the consensus toward YES/NO encodes atmospheric persistence; complementary to the level (p_consensus) and the horizon decay (days_ahead). | derived via dataset.annotate_revision_drift() across multi-day forward captures (B23) | 2026-06-21 | TBD (v4, pending multi-capture pipeline) | experimental |
| p_consensus_x_series_bias_fa | Interaction p_consensus × series_bias_fa. Hypothesis: bias correction should scale with confidence level — when p_consensus is high and series overestimates, the error is larger. Tested B38 2026-06-21: NO-GO (VALID p=0.912, 3/12 dates, Brier worse than incumbent). | derived from p_consensus × series_bias_fa | 2026-06-21 | +0.0002 (VALID, worse) | dropped (v3fb NO-GO, 2026-06-21) |
| days_ahead_x_series_bias_fa | Interaction days_ahead × series_bias_fa. Hypothesis: per-series calibration bias scales with forecast horizon — longer horizons may amplify series-specific miscalibration. Tested B38 2026-06-21: NO-GO (VALID p=0.633, 6/12 dates, tie). | derived from days_ahead × series_bias_fa | 2026-06-21 | 0.0000 (VALID, tie) | dropped (v3fb NO-GO, 2026-06-21) |
Click a row for the full hypothesis, source link, and per-run history. Brier Δ is the leave-one-out test-Brier delta from the latest run — negative (↓) = feature carried signal, positive (↑) = net noise on this split.
C. Latest training run
Snapshot of the most recent sklearn fit of the learned predictor on historical resolutions. This is not a paper trade — it's a cross-validation pass to see whether the current feature set has edge over kalshi_mid on past Kalshi events.
Latest training run
feature set v3 · 2026-06-05 12:34 UTC
n_train
144
n_test
84
Brier train
0.1368
Brier test
0.1359
Brier kalshi_mid
0.1098
gap (test − kalshi_mid)
+0.0261
log-loss test
0.4368
log-loss kalshi_mid
0.3611
D. Training run history
Every sklearn training pass, most recent first. This is not the paper-trade history — see section A for that. A training run with Brier test below Brier kalshi_mid on the same rows means the model has signal beyond the market mid in cross-validation.
| When (UTC) | Feature set | n_test | Brier test | Brier kalshi_mid | Gap | Verdict | Notes |
|---|---|---|---|---|---|---|---|
| 2026-06-05 12:34 UTC | v3 | 84 | 0.1359 | 0.1098 | +0.0261 | MARKET WINS | — |
| 2026-05-14 19:19 UTC | v2 | 84 | 0.1323 | 0.1305 | +0.0018 | MARKET WINS | Phase A.2 rerun under clean target_date split; supersedes invalidated 20260514T141925Z. Decision-gate run for the temporal-split fix. |
| 2026-05-12 13:45 UTC | v2 | 65 | 0.1301 | 0.0764 | +0.0537 | MARKET WINS | schema v2: add intercept + feature_means/stds for live inference |
| 2026-05-11 13:08 UTC | v2 | 42 | 0.1260 | 0.0282 | +0.0978 | MARKET WINS | First A.3 discovery run. V2 = V0 baseline + 6 static geographic features (urban_density_5km, water_pct_10km, forest_pct_5km, elevation_m, distance_to_coast_km, latitude). N=138 resolved, train=96 (older), test=42 (newer). Test slice currently collapses to a single capture date (20260510T171217Z) — limits temporal variance; geographic deltas mostly null on this split as expected. |
E. Training Brier trajectory
Learned model (test) vs kalshi_mid (same test rows) across all training runs. Dashed horizontal line is the most recent kalshi_mid Brier as all-time reference; vertical dashed markers flag a feature-set bump (v0 → v1 → v2 …).
F. Backtest replays
Replayed records produced by backtest.py against settled Kalshi events. Only strict point-in-time records count toward N_backtest_strict; NAIVE-mode rows are flagged and excluded from the hybrid sample. Filters above narrow both this table and the live runs section.
No backtest replays in the manifest. The aggregate count may still be non-zero — per-record detail is omitted by the manifest builder when the ledger exceeds the inline budget.