Every prediction is stored before the match and settled against the official result — wins and losses alike. When the model changes, the counter resets: what you see below is the current model only, never a blend with retired versions.
When the model says 70%, it should win ~70% of those matches. Predicted vs actual, by confidence bucket.
Cumulative win rate as picks settle, against the market-favorite baseline.
| Date | Match | Pick | Odds | EV | Result |
|---|---|---|---|---|---|
| 07-28 | Ashlyn Krueger vs Katie BoulterWTA · WTA 500 Washington (Hard) | Krueger 51.6% | 1.67 | -13.9% | HIT · +0.67u |
| 07-28 | Madison Keys vs Liudmila SamsonovaWTA · WTA 500 Washington (Hard) | Keys 79.6% | 1.31 | +4.3% | MISS · −1.00u |
| 07-28 | Elvina Kalieva vs Katherine SebovWTA · WTA Memphis (Hard) | Kalieva 53.6% | — | — | HIT · Kalieva |
| 07-28 | Tristan Boyer vs Gabriel DialloATP · ATP Los Cabos (Hard) | Diallo 51.2% | — | — | MISS · Boyer |
| 07-28 | Tatiana Prozorova vs Solana SierraWTA · WTA Memphis (Hard) | Prozorova 69.3% | — | — | HIT · Prozorova |
| 07-28 | Terence Atmane vs Frances TiafoeATP · ATP 500 Washington (Hard) | Tiafoe 74.5% | 1.31 | -2.4% | MISS · −1.00u |
| 07-28 | Camila Osorio vs Storm HunterWTA · WTA Memphis (Hard) | Osorio 56.5% | — | — | HIT · Osorio |
| 07-28 | Talia Gibson vs Tatjana MariaWTA · WTA Memphis (Hard) | Gibson 75.3% | — | — | MISS · Maria |
| 07-28 | Aleksandar Kovacevic vs Cameron NorrieATP · ATP Los Cabos (Hard) | Norrie 54.1% | — | — | HIT · Norrie |
Flat 1-unit stake on the 3 settled picks with recorded odds: -1.33u (-44.3% ROI). Odds are the market average captured before the match, not closing odds.
Latest 57 predictions of the current model. Each row is written before the match starts and never edited after settlement.
| Segment | Settled | Correct | Win rate |
|---|---|---|---|
| ATP | 3 | 1 | 33.3% |
| WTA | 6 | 4 | 66.7% |
| Hard | 9 | 5 | 55.6% |
The model (v5.2) combines surface-specific Elo ratings — recalibrated against a walk-forward backtest of 8,930 main-tour matches — with bounded situational adjustments. Match probabilities are published for every modeled match; picks are selective: matches where the model lacks reliable data are shown as probabilities only, never as picks.
Odds and EV. Each pick carries the market average odds captured before the match (not closing odds). EV is the expected return per unit staked at those odds: model probability × odds − 1. A pick can be very likely to win and still show negative EV — at 1.18 the market already demands 84.7% accuracy just to break even, so a 78% pick loses money over time. The tooltip also reports our disagreement with the market after removing the bookmaker margin (≈4.6% on current data), which is the honest measure of whether the model sees something the market does not.
Win rate and ROI are both headline numbers, on purpose. They answer different questions and regularly disagree. Win rate asks was the pick right?; ROI asks did backing it pay? — a flat 1-unit stake on every settled pick that had a recorded price. Because a favourite at 1.22 needs 82% accuracy just to break even, a perfectly calibrated 70% model wins most of those picks and still loses units. Reporting only the win rate would flatter the model; reporting both is the honest version. ROI covers the 3 priced picks only, never extrapolated to picks we could not price.
Settlement is automatic and immutable. A prediction is recorded before the match and settled against the official result. Retired matches and walkovers never settle game-total or set-score markets. Set-score and over/under simulations are published as context, not as picks.
Previous model v4.5 (2026-06-13 → 2026-07-27): 55.3% over 217 settled predictions. It was retired on 2026-07-28 after an internal audit found calibration failures; the current model was rebuilt from that audit. We keep the number here because a track record you can trust includes the versions that didn't work.
