Tennis — the record

Model validation · out-of-sample backtest

Validated on 103,520 ATP & WTA matches (2005–2026)

On the final 3,000 matches — data the model never trained on — its probabilities scored a log-loss of 0.577 against the sharpest bookmaker's closing price at 0.576: statistically level with the best market in the world. We publish that comparison because it is the honest ceiling — and the live record below is where we prove it holds in real time, starting 21 July 2026.

We also tested surface-specific ratings and world rankings: neither improved the model reliably, so we don't use them, and we say so rather than pretend otherwise. The model blends a 20-year Elo rating (men's and women's tours rated separately) with the market's opening price.

Is the model accurate?

On the same 3,000 held-out matches: two things prove accuracy — the model picks the winner about seven times in ten, and its probabilities are calibrated: when it says 78%, it happens 78% of the time.

Reliability — model probability predicted vs what actually happened

The bars line up on every band — that's calibration, shown in the open. Out-of-sample to 2026-07-19.

The live record

Lower Brier is better. Early sample — we let it build before drawing conclusions.

Daytime vs overnight

Split counts only picks logged before the first serve; picks logged after the start (disclosed on the corrections page) stay in the full record above.

Model favourite, its probability, and the result. Every prediction kept — wins and losses.

⬇ Download the full record (CSV — opens in Excel / Sheets) · showing the most recent 30 of 4413

105 voided match(es) are kept in the downloadable record and excluded from the accuracy figures (no result to score).