Replays 10212 finished matches since 2026-01-02 the way the live model learns (start 1500, K 32, wins only), then asks what it misses. Residual = how often the player won minus how often Elo said they would. A result must clear |z| ≥ 3.10 to hold, because 26 ideas were tested at once. Dataset 2026-11-12 02:29 UTC.
The live model is now duel-elo-v4 (promoted by predict-41's held-out gate). The lab's baseline is v1, and it reproduces the v1 backtest stored beside it exactly: 4570 scored matches, 2738 hits (59.9%), Brier 0.2333.
The first verdict is against v1 Elo (the lab's baseline). The second is against duel-elo-v4's own pre-match chance on the same 4570 seasoned matchups: an idea that holds against v1 but not against the live model is already inside it; one that still holds is what the live model misses.
| Idea | The player of interest | Against v1 Elo | Against the live model | |||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Matches | Won | Elo said | Residual (95%) | z | Verdict | Residual (95%) | z | Verdict | ||
| Previous match | Played 6+ more games than the opponent in the previous round (both came through it within 3 days) | 1116 | 44.4% | 48.4% | -4.0 pts -6.8 pts to -1.1 pts | -2.75 | Hint · worse than expected, not proven | -3.0 pts -5.8 pts to -0.3 pts | -2.19 | Hint · worse than the live model says, not proven |
| Previous match | Came off a three-setter; the opponent won the previous round in straight sets | 893 | 45.2% | 48.4% | -3.1 pts -6.3 pts to +0.1 pts | -1.93 | No effect beyond Elo | -2.5 pts -5.6 pts to +0.5 pts | -1.62 | No effect beyond the live model |
| Previous match | Played the day before; the opponent had a day or more off | 460 | 47.6% | 49.8% | -2.2 pts -6.6 pts to +2.2 pts | -0.98 | No effect beyond Elo | +1.1 pts -3.2 pts to +5.4 pts | 0.49 | No effect beyond the live model |
| Previous match | First match back after 30+ days away; the opponent played within two weeks | 154 | 46.8% | 51.6% | -4.8 pts -12.4 pts to +2.8 pts | -1.23 | No effect beyond Elo | -8.9 pts -16.2 pts to -1.7 pts | -2.42 | Hint · worse than the live model says, not proven |
| Previous match | First match after losing the previous one by retirement | 21 | 42.9% | 57.0% | -14.1 pts -34.8 pts to +6.6 pts | -1.33 | Too few matches | -15.6 pts -35.9 pts to +4.7 pts | -1.50 | Too few matches |
| Elo peak and drop | Sits 50+ Elo further below their own season peak than the opponent does | 658 | 46.4% | 43.0% | +3.4 pts -0.3 pts to +7.0 pts | 1.79 | No effect beyond Elo | -0.5 pts -4.1 pts to +3.1 pts | -0.26 | No effect beyond the live model |
| Elo peak and drop | Gained 30+ more Elo than the opponent over their last five matches | 2132 | 51.4% | 52.2% | -0.8 pts -2.9 pts to +1.2 pts | -0.80 | No effect beyond Elo | +1.8 pts -0.2 pts to +3.8 pts | 1.78 | No effect beyond the live model |
| Serve and power | The bigger server: a tiebreak share of sets 8+ points higher than the opponent | 837 | 49.3% | 48.7% | +0.6 pts -2.7 pts to +3.9 pts | 0.38 | No effect beyond Elo | -0.7 pts -3.9 pts to +2.4 pts | -0.46 | No effect beyond the live model |
| Serve and power | The favourite when both are big servers (18%+ of their sets go to a tiebreak) | 655 | 58.8% | 60.4% | -1.7 pts -5.4 pts to +2.0 pts | -0.88 | No effect beyond Elo | -0.4 pts -3.9 pts to +3.2 pts | -0.20 | No effect beyond the live model |
| Serve and power | Wins 3+ points more of all games played than the opponent (dominance the win-loss record hides) | 2215 | 65.9% | 63.7% | +2.3 pts +0.3 pts to +4.2 pts | 2.24 | Hint · better than expected, not proven | +0.5 pts -1.4 pts to +2.4 pts | 0.55 | No effect beyond the live model |
| Injury, slump and struggle | Retired hurt or gave a walkover in the last 60 days; the opponent did not | 102 | 42.2% | 53.4% | -11.3 pts -20.7 pts to -1.9 pts | -2.35 | Hint · worse than expected, not proven | -10.0 pts -19.2 pts to -0.8 pts | -2.12 | Hint · worse than the live model says, not proven |
| Injury, slump and struggle | Lost 3+ of their last 4 matches; the opponent lost at most 1 of theirs | 476 | 37.8% | 38.4% | -0.6 pts -4.8 pts to +3.7 pts | -0.27 | No effect beyond Elo | -5.0 pts -9.2 pts to -0.9 pts | -2.37 | Hint · worse than the live model says, not proven |
| Injury, slump and struggle | Dropped 4+ more sets than the opponent over their last 5 matches | 775 | 38.8% | 39.3% | -0.5 pts -3.8 pts to +2.9 pts | -0.27 | No effect beyond Elo | -2.8 pts -6.0 pts to +0.4 pts | -1.70 | No effect beyond the live model |
| Injury, slump and struggle | Won 5+ points less of the games over their last 5 matches than earlier in the season (a shift from their power); the opponent did not | 1263 | 51.4% | 49.4% | +2.0 pts -0.6 pts to +4.7 pts | 1.49 | No effect beyond Elo | +0.2 pts -2.3 pts to +2.8 pts | 0.18 | No effect beyond the live model |
| Head to head since 2020 | Won more of all their earlier meetings since 2020 (walkovers left out) | 1911 | 59.7% | 54.3% | +5.4 pts +3.2 pts to +7.5 pts | 4.88 | Holds · Elo under-rates them | +1.6 pts -0.5 pts to +3.6 pts | 1.47 | No effect beyond the live model |
| Format | The favourite in a men's Grand Slam main-draw match (best of five, scored as best of one) | 330 | 63.0% | 62.3% | +0.7 pts -4.4 pts to +5.8 pts | 0.26 | No effect beyond Elo | -2.0 pts -6.9 pts to +2.8 pts | -0.82 | No effect beyond the live model |
| Format | The favourite in a qualifying match | 652 | 58.0% | 58.6% | -0.6 pts -4.3 pts to +3.2 pts | -0.31 | No effect beyond Elo | -1.5 pts -5.2 pts to +2.2 pts | -0.79 | No effect beyond the live model |
| Player profile (outside records) | Left-hander against a right-hander | 854 | 48.1% | 48.2% | -0.0 pts -3.3 pts to +3.2 pts | -0.02 | No effect beyond Elo | +0.3 pts -2.9 pts to +3.4 pts | 0.17 | No effect beyond the live model |
| Player profile (outside records) | 8+ cm taller than the opponent (height as serve power) | 1502 | 54.4% | 52.6% | +1.8 pts -0.7 pts to +4.2 pts | 1.41 | No effect beyond Elo | -2.0 pts -4.4 pts to +0.3 pts | -1.69 | No effect beyond the live model |
| Player profile (outside records) | Aged 21 or under against an opponent 25 or older (Elo lags a rising player) | 424 | 55.4% | 53.8% | +1.6 pts -3.0 pts to +6.2 pts | 0.70 | No effect beyond Elo | +3.2 pts -1.2 pts to +7.6 pts | 1.42 | No effect beyond the live model |
| Player profile (outside records) | Charted serve + return points won 4+ points higher (matches charted before 2026) | 1448 | 62.4% | 54.7% | +7.7 pts +5.3 pts to +10.2 pts | 6.13 | Holds · Elo under-rates them | +0.4 pts -1.9 pts to +2.7 pts | 0.34 | No effect beyond the live model |
| Player profile (outside records) | A charted Big server against a charted Returner (the server's result) | 184 | 54.9% | 51.5% | +3.4 pts -3.6 pts to +10.4 pts | 0.96 | No effect beyond Elo | -0.2 pts -6.8 pts to +6.4 pts | -0.05 | No effect beyond the live model |
| Player profile (outside records) | A charted Net rusher against a charted Baseliner | 87 | 54.0% | 51.0% | +3.0 pts -7.1 pts to +13.1 pts | 0.58 | Too few matches | +7.5 pts -2.4 pts to +17.4 pts | 1.49 | Too few matches |
| Player profile (outside records) | Ranked higher in the latest official ranking, yet the Elo underdog | 681 | 54.8% | 42.0% | +12.7 pts +9.1 pts to +16.4 pts | 6.79 | Holds · Elo under-rates them | +8.6 pts +3.6 pts to +13.7 pts | 3.36 | Still missed · the live model under-rates them |
| Previous match | Won more of their earlier meetings this season | 846 | 59.9% | 55.0% | +4.9 pts +1.7 pts to +8.2 pts | 2.96 | Hint · better than expected, not proven | +3.3 pts +0.1 pts to +6.5 pts | 2.04 | Hint · better than the live model says, not proven |
| Head to head since 2020 | Both in the official top 20: won more of their earlier meetings since 2020 | 106 | 59.4% | 51.4% | +8.0 pts -1.2 pts to +17.2 pts | 1.71 | No effect beyond Elo | +2.4 pts -6.5 pts to +11.4 pts | 0.53 | No effect beyond the live model |
Every study that marks either player today, from their record up to the last result. Only proven studies (holding against the live model) count toward the exposure; hints and no-effect studies are listed so you can see them, not because they move the chance.
| Match | Pick | Conv. | Proven exposure | What the studies see |
|---|---|---|---|---|
| Alejandro Tabilo v Alex Michelsen 11-12 12:00 UTC · Bybit Stockholm Open | Alex Michelsen | 61.5% | none |
|
| Andrea Lazaro Garcia v Elizabeth Mandlik 11-12 15:00 UTC · DIRECTV Open Presentado por Compesar | Andrea Lazaro Garcia | 69.7% | none |
|
| Irene Burillo v Laura Samson 11-12 15:00 UTC · DIRECTV Open Presentado por Compesar | Laura Samson | 69.3% | none |
|
| Marina Bassols Ribera v Bianca Andreescu 11-12 16:30 UTC · DIRECTV Open Presentado por Compesar | Bianca Andreescu | 70.3% | none |
|
| Polona Hercog v Alice Tubello 11-12 16:30 UTC · DIRECTV Open Presentado por Compesar | Alice Tubello | 68.1% | none |
|
| Holger Rune v Taylor Fritz 11-12 17:30 UTC · Bybit Stockholm Open | Taylor Fritz | 67.2% | none |
|
| Ugo Humbert v Lorenzo Musetti 11-12 19:00 UTC · Bybit Stockholm Open | Lorenzo Musetti | 69.2% | none |
|
Every golden (60%+) tennis pick the ledger graded (371 of 381 found in the replay), with what each study saw before the match: once when it marked our player, once when it marked the opponent. A proven weakness (a study that holds against the live model, pointing against our pick) was there before 6 of 107 losses and 8 of 264 wins. Small groups prove nothing on their own; the verdict column is the full-season test.
| Before the match | Our pick had it: picks · W-L · hit · units | The opponent had it | Season test |
|---|---|---|---|
| Played 6+ more games than the opponent in the previous round (both came through it within 3 days) | 30 · 18-12 · 60.0% (said 68.0%) · -3.55u | 41 · 33-8 · 80.5% (said 68.9%) · +7.04u | hint |
| Came off a three-setter; the opponent won the previous round in straight sets | 26 · 17-9 · 65.4% (said 68.0%) · -0.85u | 32 · 27-5 · 84.4% (said 69.3%) · +7.24u | no effect |
| Played the day before; the opponent had a day or more off | 18 · 12-6 · 66.7% (said 67.4%) · -0.30u | 23 · 16-7 · 69.6% (said 69.7%) · +0.22u | no effect |
| First match back after 30+ days away; the opponent played within two weeks | 7 · 2-5 · 28.6% (said 63.7%) · -3.90u | 26 · 22-4 · 84.6% (said 66.8%) · +6.87u | hint |
| First match after losing the previous one by retirement | 7 · 4-3 · 57.1% (said 69.4%) · -1.39u | 1 · 1-0 · 100.0% (said 65.4%) · +0.53u | too few |
| Sits 50+ Elo further below their own season peak than the opponent does | 10 · 4-6 · 40.0% (said 66.8%) · -4.30u | 49 · 32-17 · 65.3% (said 67.3%) · -1.37u | no effect |
| Gained 30+ more Elo than the opponent over their last five matches | 109 · 88-21 · 80.7% (said 68.6%) · +20.00u | 49 · 29-20 · 59.2% (said 68.6%) · -7.05u | no effect |
| The bigger server: a tiebreak share of sets 8+ points higher than the opponent | 15 · 13-2 · 86.7% (said 69.4%) · +3.91u | 30 · 20-10 · 66.7% (said 67.9%) · -0.53u | no effect |
| The favourite when both are big servers (18%+ of their sets go to a tiebreak) | 56 · 40-16 · 71.4% (said 70.0%) · +0.57u | 1 · 1-0 · 100.0% (said 66.6%) · +0.50u | no effect |
| Wins 3+ points more of all games played than the opponent (dominance the win-loss record hides) | 181 · 135-46 · 74.6% (said 69.7%) · +13.02u | 1 · 1-0 · 100.0% (said 60.8%) · +0.64u | no effect |
| Retired hurt or gave a walkover in the last 60 days; the opponent did not | 28 · 19-9 · 67.9% (said 66.5%) · +0.74u | 14 · 10-4 · 71.4% (said 66.8%) · +0.69u | hint |
| Lost 3+ of their last 4 matches; the opponent lost at most 1 of theirs | 2 · 1-1 · 50.0% (said 67.1%) · -0.59u | 42 · 33-9 · 78.6% (said 69.6%) · +5.93u | hint |
| Dropped 4+ more sets than the opponent over their last 5 matches | 3 · 1-2 · 33.3% (said 63.5%) · -1.50u | 66 · 52-14 · 78.8% (said 69.8%) · +8.61u | no effect |
| Won 5+ points less of the games over their last 5 matches than earlier in the season (a shift from their power); the opponent did not | 17 · 12-5 · 70.6% (said 70.7%) · -0.19u | 34 · 27-7 · 79.4% (said 69.1%) · +5.29u | no effect |
| Won more of all their earlier meetings since 2020 (walkovers left out) | 90 · 65-25 · 72.2% (said 69.2%) · +3.84u | 33 · 18-15 · 54.5% (said 65.5%) · -5.76u | no effect |
| The favourite in a men's Grand Slam main-draw match (best of five, scored as best of one) | 18 · 13-5 · 72.2% (said 68.4%) · +1.21u | — | no effect |
| The favourite in a qualifying match | 115 · 82-33 · 71.3% (said 64.9%) · +10.70u | 5 · 4-1 · 80.0% (said 67.5%) · +0.81u | no effect |
| Left-hander against a right-hander | 35 · 29-6 · 82.9% (said 67.7%) · +8.26u | 35 · 30-5 · 85.7% (said 68.1%) · +9.25u | no effect |
| 8+ cm taller than the opponent (height as serve power) | 47 · 35-12 · 74.5% (said 69.5%) · +3.21u | 43 · 33-10 · 76.7% (said 68.9%) · +5.07u | no effect |
| Aged 21 or under against an opponent 25 or older (Elo lags a rising player) | 14 · 9-5 · 64.3% (said 68.3%) · -0.83u | 8 · 6-2 · 75.0% (said 70.2%) · +0.57u | no effect |
| Charted serve + return points won 4+ points higher (matches charted before 2026) | 57 · 42-15 · 73.7% (said 71.1%) · +1.49u | 30 · 20-10 · 66.7% (said 66.8%) · +0.17u | no effect |
| A charted Big server against a charted Returner (the server's result) | 5 · 3-2 · 60.0% (said 65.0%) · -0.30u | 8 · 7-1 · 87.5% (said 74.6%) · +1.35u | no effect |
| A charted Net rusher against a charted Baseliner | 1 · 0-1 · 0.0% (said 62.1%) · -1.00u | — | too few |
| Ranked higher in the latest official ranking, yet the Elo underdog | — | 14 · 8-6 · 57.1% (said 63.8%) · -1.12u | proven |
| Won more of their earlier meetings this season | 53 · 36-17 · 67.9% (said 68.1%) · -0.38u | 15 · 10-5 · 66.7% (said 65.3%) · +0.14u | hint |
| Both in the official top 20: won more of their earlier meetings since 2020 | 2 · 0-2 · 0.0% (said 65.5%) · -2.00u | — | no effect |
| Before the match | Picks · W-L · hit · units |
|---|---|
| Our pick led the head to head | 90 · 65-25 · 72.2% (said 69.2%) · +3.84u |
| Level (met, same number of wins) | 19 · 10-9 · 52.6% (said 69.4%) · -4.38u |
| Our pick trailed the head to head | 33 · 18-15 · 54.5% (said 65.5%) · -5.76u |
| Never met before | 229 · 171-58 · 74.7% (said 67.1%) · +25.83u |
Does knowing every earlier meeting (46,936 matches before this season, plus this season's) make duel-elo-v4's own chance better? The live chance is moved by beta × (wins − losses) / (meetings + 2), beta fitted on the first 7148 matches of the season, judged once on the last 3064. Two rules tried, so a gain needs |z| ≥ 2.24.
| Rule | beta (95%) | All held-out | Had met before | Met 2+ times | Both top 20, had met | Verdict (had met) |
|---|---|---|---|---|---|---|
| Every meeting the same | 0.075 ± 0.246 | 64.7% → 64.7% Brier 0.21797 → 0.21785, z -2.04 n 3064 | 66.2% → 66.1% Brier 0.21400 → 0.21366, z -2.04 n 1033 | 66.5% → 66.9% Brier 0.21649 → 0.21615, z -1.27 n 468 | 65.5% → 65.5% Brier 0.23262 → 0.23159, z -1.45 n 55 | Better, not proven |
| Recent meetings weigh more (half each year) | 0.197 ± 0.397 | 64.7% → 64.8% Brier 0.21797 → 0.21778, z -1.96 n 3064 | 66.2% → 66.4% Brier 0.21400 → 0.21345, z -1.96 n 1033 | 66.5% → 67.3% Brier 0.21649 → 0.21590, z -1.33 n 468 | 65.5% → 67.3% Brier 0.23262 → 0.22987, z -2.05 n 55 | Better, not proven |
Calibration: slope 1.018 ± 0.119 over 4570 scored matches (ATP 1.01, WTA 1.02). Above 1 means the favourite wins more often than the rating says.
Each rule replays the whole season and is judged on the held-out last 30%, which none was tuned on: 1371 seasoned matchups (the ones the live backtest scores) and 815 thin or warming-up matchups (never backtested - where the ledger loses money). A fix must clear |z| ≥ 2.84 (11 tried). Every rule is judged against the v1 baseline, not against the live duel-elo-v4 (which already carries history, the seed, margin weighting and K by experience); a rule here goes live only through predict-41's own held-out gate. Nothing here changes the live model.
| Rule | Seasoned matchups (held-out) | Thin and warming-up (held-out) | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Acc | Brier | Δ vs v1 | z | Verdict | Acc | Brier | Δ vs v1 | z | Verdict | |
| Baseline v1 (the first live model): K 32, wins only, walkovers count, everything best-of-1 | 59.4% | 0.2352 | — | — | baseline | 55.2% | 0.2414 | — | — | baseline |
| Skip walkovers and retirements | 58.6% | 0.2361 | +0.00087 | 2.75 | No proven change | 54.8% | 0.2415 | +0.00013 | 0.52 | No proven change |
| Margin-weighted (WElo: the step scales with the winner's share of games) (yardstick 63.2% of games, from training matches) | 59.8% | 0.2334 | -0.00181 | -4.16 | Better | 57.5% | 0.2396 | -0.00178 | -3.85 | Better |
| Variable K (538-style: fast for new players, slower as they play more) | 60.8% | 0.2338 | -0.00145 | -0.93 | No proven change | 58.4% | 0.2364 | -0.00491 | -1.56 | No proven change |
| Best-of-5 math for men's Grand Slam main draws | 59.2% | 0.2351 | -0.00014 | -1.46 | No proven change | 55.1% | 0.2413 | -0.00007 | -2.39 | No proven change |
| All four fixes together (yardstick 63.2% of games, from training matches) | 61.1% | 0.2327 | -0.00254 | -1.47 | No proven change | 60.2% | 0.2337 | -0.00766 | -2.25 | No proven change |
| Pre-season seed: new players start from their week-2 official ranking points, not 1500 (seed = 978 + 82.7·ln(points), fit on 254 ranked players, r² 0.34; unranked start 1476.8) | 61.2% | 0.2311 | -0.00412 | -4.78 | Better | 58.4% | 0.2330 | -0.00838 | -7.52 | Better |
| Pre-season seed + margin-weighted (yardstick 63.2% of games, from training matches) | 61.7% | 0.2294 | -0.00583 | -6.01 | Better | 60.0% | 0.2313 | -0.01004 | -8.34 | Better |
| Two more seasons of history: 2024-2025 replayed before 2026 | 61.9% | 0.2286 | -0.00664 | -3.80 | Better | 61.7% | 0.2295 | -0.01187 | -4.53 | Better |
| History + margin-weighted (yardstick 63.2% of games, from training matches) | 63.0% | 0.2269 | -0.00836 | -4.45 | Better | 62.0% | 0.2264 | -0.01500 | -5.41 | Better |
| History + pre-season seed for newcomers + margin-weighted (yardstick 63.2% of games, from training matches) | 62.7% | 0.2263 | -0.00892 | -3.83 | Better | 64.2% | 0.2177 | -0.02367 | -6.99 | Better |
| Four seasons of history (2022-2025) + pre-season seed + margin-weighted (yardstick 63.2% of games, from training matches) Against the same recipe on two seasons: Δ Brier +0.00014 (z 0.30) seasoned, -0.00661 (z -6.82) thin; 38368 history matches. | 63.2% | 0.2264 | -0.00878 | -3.27 | Better | 66.0% | 0.2111 | -0.03028 | -7.57 | Better |
Since 2026-11-04 08:31 UTC the seedMargin rule (seed 903 + 94.5·ln(points), margin-weighted) locks its own probability before every upcoming match, next to the live model's locked pick, and both are graded on the same final. It does not touch the live model or its ledger. 167 locked · 151 graded · 12 pending · 4 void (pairing changed).
| Matches | n | v1 acc | v2 acc | v1 Brier | v2 Brier |
|---|---|---|---|---|---|
| All graded | 151 | 59.6% | 63.6% | 0.2399 | 0.2317 |
| seasoned | 119 | 58.0% | 62.2% | 0.2444 | 0.2403 |
| thin | 12 | 75.0% | 83.3% | 0.1972 | 0.1737 |
| warming-up | 20 | 60.0% | 60.0% | 0.2390 | 0.2148 |
Δ Brier v2 − v1 -0.00828, z -2.42 (needs |z| ≥ 1.96 and weeks of matches before it means anything). The two disagree on the winner in 16 of 151. Golden (60%+) paper units at each model's own fair price: v1 40-24 -3.52u, v2 52-26 -1.46u.
duel-elo-v2, duel-elo-v3, duel-elo-v4 is live in predict-41; its own locked chance sits beside every shadow pick it made, scored against v1 on the same finals. 103 graded · 8 pending. Accuracy live 64.1% v v1 57.3%; Brier live 0.2136 v v1 0.2373; Δ live − v1 -0.02372, z -2.50 (needs |z| ≥ 1.96 before it means anything).
1365 players in the tennis dataset: ESPN profile 391, Wikidata 968, charted (Match Charting Project, before 2026) 634, with a style 380; 96% of match appearances carry at least one record. 68 official ranking weeks.
| Suspect | Why | Measured |
|---|---|---|
| 16 walkovers and 64 retirements | are learned as full wins and losses; a walkover is no match at all. | Unproven: no clear change (Δ Brier +0.00087, z 2.75) |
| Wins only | A 6-0 6-0 and a 7-6 6-7 7-6 move the rating the same amount; the games in the score line are never used. | Real flaw: fixing it beats live (Δ Brier -0.00181, z -4.16) Power study: Hint · better than expected, not proven (+2.3 pts, z 2.24) |
| One K for everyone | A newcomer and a 60-match veteran move at the same speed; every player starts at 1500 from 2026-01-02 (no prior history), including 3192 qualifying matches. | Unproven: no clear change (Δ Brier -0.00145, z -0.93) |
| Every match is stored as best of one | 508 men's Grand Slam main-draw matches are best of five (492 of 504 were won with three sets), but the harvester hard-codes bestOf 1. ESPN does publish the format (format.regulation.periods). | Unproven: no clear change (Δ Brier -0.00014, z -1.46) Favourite study: No effect beyond Elo (+0.7 pts, z 0.26) |
| No rest or fatigue in the model | Rest days, the previous match and time away never enter the prediction. | Previous round 6+ games longer: Hint · worse than expected, not proven (-4.0 pts, z -2.75) Three-setter before: No effect beyond Elo (-3.1 pts, z -1.93) |
| 5642 matchups never backtested | The backtest scores only seasoned matchups (both players 12+ rated matches), but picks are published at every depth. | Walk-forward on them: 54.3%, Brier 0.2440 Ledger: 1378 warming-up picks, 50.3% hit, -33.00u |
| 12 repeat pairings within 3 days | The same two players listed twice in 3 days, mostly a qualifying final then round 1: a lucky loser, or a feed duplicate learned twice. e.g. Daniel Merida v Marco Trungelliti (Mutua Madrid Open: 176750 2026-04-21 Qualifying Final / 174071 2026-04-23 Round 1); Arthur Gea v Aleksandar Kovacevic (Bitpanda Hamburg Open: 178329 2026-05-17 Qualifying Final / 175459 2026-05-19 Round 1) | Not measured: needs a second source |
| No serve or power data | ESPN publishes no tennis match statistics (the statistics list is empty and the stats endpoint 404s), so serve is read from tiebreak share and power from games share. | Bigger server: No effect beyond Elo (+0.6 pts, z 0.38) |
| 0 pending picks with a TBD player | 1 pending picks are past their start time; a TBD slot can never be graded. | Owner's call |
| Depth | Graded | W-L | Hit | Conviction | Brier | Paper units |
|---|---|---|---|---|---|---|
| seasoned | 547 | 340-207 | 62.2% | 61.3% | 0.229 | +8.76u |
| thin | 108 | 69-39 | 63.9% | 59.8% | 0.215 | +5.76u |
| warming-up | 1378 | 693-685 | 50.3% | 51.5% | 0.248 | -33.00u |
Peak = the highest rating after their 12th match this season. Active = played in the 45 days before the last result.
| Player | Tour | Now | Peak | Peak date | Drop | Played |
|---|---|---|---|---|---|---|
| Anhelina Kalinina | wta | 1529.0 | 1719.9 | 2026-05-22 | −190.9 | 64 |
| Martin Landaluce | atp | 1423.1 | 1589.5 | 2026-05-28 | −166.4 | 42 |
| Victoria Jimenez Kasintseva | wta | 1411.3 | 1570.6 | 2026-03-04 | −159.3 | 37 |
| Moyuka Uchijima | wta | 1474.6 | 1626.1 | 2026-05-03 | −151.5 | 56 |
| Marcos Giron | atp | 1393.5 | 1537.5 | 2026-03-04 | −144.0 | 43 |
| Sebastian Baez | atp | 1439.2 | 1582.0 | 2026-03-08 | −142.8 | 51 |
| Tomas Martin Etcheverry | atp | 1495.4 | 1632.8 | 2026-04-26 | −137.4 | 58 |
| Tyra Caterina Grant | wta | 1491.5 | 1627.5 | 2026-06-30 | −136.0 | 24 |
| Lucia Bronzetti | wta | 1515.5 | 1650.2 | 2026-07-31 | −134.7 | 57 |
| Sho Shimabukuro | atp | 1453.5 | 1585.7 | 2026-06-11 | −132.2 | 38 |
| Veronika Erjavec | wta | 1451.6 | 1583.6 | 2026-05-02 | −132.0 | 45 |
| Mananchaya Sawangkaew | wta | 1539.4 | 1669.3 | 2026-07-30 | −129.9 | 48 |
| Linda Fruhvirtova | wta | 1421.4 | 1548.9 | 2026-03-17 | −127.5 | 47 |
| Laura Pigossi | wta | 1375.9 | 1503.1 | 2026-05-07 | −127.2 | 26 |
| Yuliia Starodubtseva | wta | 1489.6 | 1616.2 | 2026-05-27 | −126.6 | 57 |
| Player | Tour | Now | Peak | Peak date | Drop | Played |
|---|---|---|---|---|---|---|
| Carlos Alcaraz | atp | 1865.4 | 1865.4 | 2026-11-08 | 0.0 | 48 |
| Mirra Andreeva | wta | 1799.4 | 1799.4 | 2026-11-11 | 0.0 | 69 |
| Linda Noskova | wta | 1788.8 | 1788.8 | 2026-11-10 | 0.0 | 53 |
| Diana Shnaider | wta | 1762.0 | 1762.0 | 2026-11-01 | 0.0 | 61 |
| Zheng Qinwen | wta | 1737.4 | 1737.4 | 2026-10-11 | 0.0 | 39 |
| Alejandro Davidovich Fokina | atp | 1736.6 | 1736.6 | 2026-10-18 | 0.0 | 49 |
| Maja Chwalinska | wta | 1714.1 | 1714.1 | 2026-10-18 | 0.0 | 53 |
| Claire Liu | wta | 1671.9 | 1671.9 | 2026-11-08 | 0.0 | 48 |
| Clara Burel | wta | 1663.4 | 1663.4 | 2026-10-18 | 0.0 | 26 |
| Harriet Dart | wta | 1662.8 | 1662.8 | 2026-11-01 | 0.0 | 38 |