METHODS & DATA

Behind the probabilities.

Historical evaluation, training data and model details. Current live results are on the Results page.

The deployed model was trained through 2022. Live history refreshes update its inputs; experimental retraining is evaluated separately before any promotion. No market odds enter model features.

Training, historical evaluation & feature importance

Training data

Matches trained on

1,741,054

1967-12-25 → 2026-06-02

Main tour

359,005

ATP + WTA

Challenger

239,530

pre-main-tour history

ITF / Futures

1,142,519

cold-start warm-up

Last retrained

2026-09-19

13:15 UTC

How predictions score

These matches happened after training ended, so they were not used to fit the model. Accuracy is the share of matches where the more likely player won; log loss and Brier score measure probability quality (lower is better). Closing market prices are shown only as an independent comparison.

Prediction source	Accuracy · right picks	Log loss	Brier	Matches
Elo only	65.6%	0.6101	0.2118	14,150
Our model	66.9%	0.5963	0.2061	14,150
Real market (Pinnacle, de-vigged)	68.3%	0.5885	0.2023	11,052

Matched 11,052 / 14,150 main-tour 2024+ matches to closing odds (78% coverage). We pick the same favorite as the market 90% of the time.

Calibration: our model vs. the market

When a predictor says "70% to win," does that side actually win about 70% of the time? The closer a line sits to the diagonal, the more trustworthy its numbers are.

What the model weighs most

elo_pre_diff

surface_elo_pre_diff

p1_career_matches_before

p2_career_matches_before

p2_dominance_ratio_before

p1_dominance_ratio_before

p2_days_since_last_match

p2_surface_matches_before

p2_avg_sets_played_before

p2_career_win_pct_before

p1_days_since_last_match

p1_surface_matches_before

p1_avg_sets_played_before

p1_career_win_pct_before

p1_venue_win_pct_before

Newer model experiment · prospective trial

Candidate training ends 2025-09-30; calibration uses October–December 2025. Historical evaluation covers 2026-01-04 to 2026-06-02. The active model has not been replaced.

Variant	Accuracy	Brier	Log loss	Matches
recent base	69.76%	0.19458	0.57049	23770
recent base calibrated	69.76%	0.19449	0.57030	23770
recent context	69.97%	0.19429	0.56983	23770
recent context calibrated	69.97%	0.19420	0.56960	23770
deployed model replayed	69.64%	0.19510	0.57435	23770
surface elo	67.83%	0.20458	0.59563	23770
context elo blend	69.97%	0.19429	0.56983	23770

Exploratory historical evaluation: these dates were inspected in earlier experiments. Tournament dates make historical workload approximate. Live paired predictions are collected before outcomes; a future promotion needs prospective evidence.

Got an idea to make this better?

Feature ideas, a stat you'd want tracked, a bug in the numbers — all welcome.