Full transparency
We show our real accuracy, the model's calibration, and the backtests — the numbers exactly as they come out.
We're not trying to beat the market.
Our goal isn't to beat the house at all costs, but to give you the best probabilistic read of the match: who wins, with what probability, and why. Everything you see here comes from paper backtests, audited and leak-free — no dressing up and no promising future profits.
65.1%
ACCURACY (2025, out-of-sample)
0.216
BRIER SCORE
4
AUDITED SEASONS (+2026 in progress)
Live accuracy · published predictions6 Jul – 7 Jul · n=62
76%
LIVE ACCURACY
0.154
LIVE BRIER
62
SETTLED PREDICTIONS
50–60%38% actual · n=8
60–70%40% actual · n=5
70–80%80% actual · n=30
80–90%93% actual · n=15
90–100%100% actual · n=4
Challenger 81% (n=42)ITF 82% (n=11)ATP 80% (n=5)WTA 0% (n=4)
Every prediction published in Analyze, settled against official results — ATP with the model and Challenger/ITF by Elo, none excluded. Each band's vertical mark is the predicted probability; the bar, the actual frequency. Small sample = noise: the number settles with volume. Past ≠ future.
Accuracy by confidence level Actual accuracy Declared level
When the model says "maximum confidence," does it hit what it promises? This is what actually happened, match by match, out of sample — not what we'd like to have happened.
90–100%
93.3%n=283
80–90%
83.6%n=1,259
70–80%
73.4%n=2,419
60–70%
63.5%n=3,405
50–60%
53.9%n=4,343
Walk-forward out-of-sample validation (2022–2026, 11,709 matches). This isn't marketing: it's the same table we use internally to decide whether the model needs recalibrating.
Live pick track recordsettled vs official results
2–0
DAILY PICK · 1 pending
1–0
VALUE PICK · 1 pending
100%
DAILY HIT RATE (2 settled)
Two separate series, both settled against official results: the Daily Pick (highest-confidence) and the Value pick (positive-EV underdog basket). The value series is tracked apart and never inflates the Daily Pick hit rate.
Model calibration Model Perfect calibration
When the model says X%, it wins very close to X% of the time — it hugs the diagonal, with a minimal over-confidence in its favorite (~1pp) that isn't worth recalibrating. Data from the out-of-sample walk-forward validation (2022–2026, n≈11,700).
Maximum-EV picks — real track recordAccuracy 35% · 3/8 in the sample
Our boldest bets: where the model disagrees most with the market. They hit rarely but at high odds. We show hits and misses — no cherry-picking.
22 MARGoffin vs Alcaraz@12.53EV +211%✅ WON
25 MAYSchoolkate vs Fucsovics@8.03EV +248%❌ MISSED
29 MAYRocha vs Mensik@11.78EV +189%✅ WON
12 JUNEngel vs Michelsen@6.07EV +190%✅ WON
01 JULCollignon vs Cilic@7.66EV +260%❌ MISSED
02 SEPLehecka vs Alcaraz@10.11EV +264%❌ MISSED
20 OCTRodionov vs De Minaur@9.18EV +233%❌ MISSED
05 NOVEchargui vs Tien@7.66EV +243%❌ MISSED
Real sample of 8 picks from the 2025 backtest (of 1,090 in the EV≥15% band, 35% hit rate). On paper, at average odds. They lose more than they win, but when they hit they pay big. Past ≠ future.
How we measure it
Walk-forward validation: we train only on the past and measure on future matches the model never saw.
Leak-free: we audit and fix leaks that inflated false results (+32%, +294% turned out to be noise).
Calibration proven, not assumed: the model is already well calibrated (slight over-confidence, ~1pp) — we tested walk-forward recalibration (Platt) and it doesn't improve it, so we don't apply it.
Paper-only: all metrics are on historical data, not real money.
The «Live accuracy» block above is fed automatically by every published prediction, settled daily against official results — by tier (ATP/Challenger/ITF) and by probability band.