The Method

Every match updates an overall rating plus separate hard, clay and grass ratings — results transfer partially across surfaces, so clay form informs hard-court form — with a dynamic K-factor (new players move fast, veterans settle) and margin-of-victory weighting from games won. At prediction time the overall and surface ratings are blended with a per-tour tuned weight, currently about 60/40 toward the surface rating. ATP ratings also learn from Challenger and qualifying matches, though predictions are trained and scored on tour-level main draws only.

Each player carries a time-decayed serve and return skill, computed per surface and adjusted for the strength of the opponents faced. These feed a hierarchical Markov model (point → game → tiebreak → set → match) that yields a full set-score distribution and handles best-of-3 vs best-of-5 correctly.

From the Match Charting Project, each charted player gets an eight-dimension style profile — serve dominance, placement variety, net play, serve-and-volley rate, aggression, forehand/backhand balance, return depth, break-point clutch — attached as features where available.

An ensemble of seed-bagged XGBoost classifiers fuses Elo, the point model, ranking, rest, fatigue, recent form, head-to-head, home advantage, player profile, style and match context (surface, format, tier, round) into one win probability, then Platt-calibrated so the numbers mean what they say. Elo remains the single most important input.

Pairwise probabilities drive thousands of simulated single-elimination draws for per-round and title odds. Live tournaments are simulated on the real remaining bracket — full draws parsed from Wikipedia, eliminations from ESPN — not a hypothetical re-seed.

Live scores and results are re-pulled hourly and every figure on this site regenerated; once a day the full pipeline re-downloads all data, re-walks the ratings and retrains the combiner — all walk-forward and leakage-free (betting odds are used only to benchmark, never as inputs).