Methodology
How a forecast is made, recorded and judged, and where the method falls short.
1
History
About 360,000 tour-level matches. Every input is built in date order from earlier matches only.
2
Ratings
Elo (start 1500, K 32) and surface Elo (K 48, blended half and half with overall).
3
Inputs
Ranking, days since last match, form against Elo, and opponent-adjusted serve and return.
4
Model
A logistic regression per tour, fitted 2010 to May 2026 and frozen. Inputs add up to the log-odds.
5
Record
Issued at least 30 minutes before play, stored unchangeably, hash-chained and scored afterwards.
Two models
Tournament start uses information from before the tournament began.
Round-aware adds this tournament's completed earlier rounds. It applies to knockout matches after the first round. When it cannot apply, the site shows why and never borrows the other model's number.
How forecasts are judged
Log loss rewards honest probabilities; calibration checks that 70% forecasts win about 70% of the time.
Historical tests were fixed before their results were seen. The live record began on 9 November 2026 and is judged once, after 30 June 2027. Model Lab
What works, and what does not
Ranking, inactivity and form improve on Elo alone, and the round-aware update helps later rounds. Official rankings were the benchmark the earlier models had to beat; in post-hoc checks, the model's probabilities did, though its edge in picking winners narrowed over time.
Serve-and-return gains were not confirmed, matchup and fatigue features added nothing, and the betting market forecasts better.
Limits
- Live data has no serve statistics; those estimates refresh with new data files.
- A player without a ranking on record is treated as unranked.
- Unconfirmed identities are not forecast until checked.
- Every result is an association, not a cause. Nothing here is betting advice.
Data sources
Jeff Sackmann's ATP and WTA data (CC BY-NC-SA 4.0), TennisIsMyLife and TennisData.app files, ESPN fixtures and results, and Valuebetennis prices for research only (CC BY 4.0). Because of the non-commercial licence, this site is non-commercial.