Every match prediction the model serves, logged before the game and graded after it. Research project — paper predictions, no real money.
Whenever the model rates one side at 51% or higher, that side counts as its pick — no odds, no edge filter, no cherry-picking. Picks are logged before the match starts and graded automatically from results; forfeits and walkovers are void. Sub-51% calls are declared coin flips and sit out of the winrate.
This is a research scoreboard for a prediction model — not betting advice, and no money rides on it.
The model's pick won 1233 of 1904 (64.8%) since 30 Jun 2026.
65 coin-flip call(s) (under 51% either way) sit out · 126 void (forfeit/walkover) · 328 awaiting a result
Head-to-head with the market, on the 871 graded pick(s) that had a captured market price: model pick won 67.7% vs 69.0% for simply backing the market favorite (1033 pick(s) had no market and sit outside this comparison).
A raw winrate flatters any model that mostly agrees with the market — the favorite-vs-model bar above, scored on the same matches, is the honest comparison.
Grouping picks by the model's stated confidence: a well-calibrated model's 70–80% bucket should win about 70–80% of the time. That, not the headline winrate, is the number to watch.
| confidence | picks | won | win rate | avg stated |
|---|---|---|---|---|
| 51-60% | 663 | 367 | 55.4% | 55.3% |
| 60-70% | 583 | 375 | 64.3% | 64.8% |
| 70-80% | 418 | 296 | 70.8% | 74.5% |
| 80%+ | 240 | 195 | 81.2% | 85.9% |
Each model promotion starts a new era; every graded pick lands as a tick on its era's band.
| era | what changed | live | graded | W–L | win rate (95% CI) | avg conf |
|---|---|---|---|---|---|---|
| v5-era | pre-refit era | 21 Jun – 23 Jun | 0 | — | — | — |
| v6 | v6 baseline | 23 Jun – 2 Jul | 18 | 11–7 | 61.1% CI 38.6–79.7% — n too small | 64.3% |
| v6-refit | refit-all baseline | 2 Jul – 13 Jul | 136 | 95–41 | 69.8% CI 61.7–76.9% | 65.3% |
| v9 | post-Phase-A data | 13 Jul – 15 Jul | 18 | 15–3 | 83.3% CI 60.8–94.2% — n too small | 69.4% |
| v10 | lineup-aware h2h | 15 Jul | 16 | 11–5 | 68.8% CI 44.4–85.8% — n too small | 62.7% |
| v11 | Valve rankings | 16 Jul | 18 | 9–9 | 50.0% CI 29.0–71.0% — n too small | 62.9% |
| v12 | schedule strength | 17 Jul – 26 Jul | 250 | 152–98 | 60.8% CI 54.6–66.6% | 65.6% |
| v13 | Elo decay-to-rank | 28 Jul – 11 Aug | 378 | 250–128 | 66.1% CI 61.2–70.7% | 66.5% |
| v14 | per-format cal + data refresh | 12 Aug – 23 Aug | 322 | 196–126 | 60.9% CI 55.4–66.0% | 64.6% |
| v15 | as-of RAPM player ratings | 24 Aug – 14 Sep | 813 | 527–286 | 64.8% CI 61.5–68.0% | 66.1% |
Eras face different slates — compare each model's calibration, not raw win rates. No per-era verdict before 100 graded picks.