Backing every side quoted at 5.00 or longer - about +400 in American odds - returned about -26% per bet. Those sides were priced as though they win 13.5% of the time and won 10.4%, a 3.1 point gap at 7.7 sigma. Both the favorite and the underdog cut came out negative too, and all three cuts were negative in every one of the 19 seasons.
Realized EV per bet
| Cut | Figure | Standard error |
|---|---|---|
| Back every favorite | -4.86% | ± 0.34 |
| Back every underdog | -12.15% | ± 0.86 |
| Back every side at 5.00 or longer | -26.29% | ± 2.78 |
| Mean two-sided hold in the quotes | 6.20% | ± 0.01 |
The two-sided hold is the book's margin baked into each quote, not the realized cost of a bet: those are separate numbers and the table keeps them separate. At 6.2% the ATP hold is well above the roughly 3.9% measured on an MLB game line, which is most of what the graded results reflect.
Price ladder, all 35,346 matches
| Price implies | 91.3% | 75.5% | 60.5% | 42.8% | 28.0% | 13.5% |
|---|---|---|---|---|---|---|
| Actually won | 88.8% | 72.8% | 56.0% | 40.7% | 24.2% | 10.4% |
The ladder is monotone: every bucket underperformed its own price and the miss grows with the price. That is the longshot bias, and the recent-seasons cut reproduces it - 2024 through mid-2026, 6,183 matches, favorites at -3.25% ± 0.86 and underdogs at -10.49% ± 2.04.
Method
Every completed ATP match with a recorded Bet365 closing price is graded against the real winner, which sits in the same source row as the price, so there is no join to get wrong. Both sides of each match are graded at the recorded decimal close and expected value is reported per bet; standard errors are match-clustered. A single book is quoted rather than a blend of books. The census covers only the ~60% of source rows that resolve to known player ids, and the unresolved rows were checked against the resolved ones over 2024-26 - all-sides EV of -6.71% ± 0.65 against -6.81% ± 0.70 - so the selection does not move the headline. Tool: atp_moneyline_realized_census.py.
Inside the noise floor
The favorite win rate is not quoted here. The player-id resolution tilts the tier mix enough to move it (67.3% unresolved against 69.5% resolved) while leaving expected value flat, so expected value is the figure the sample supports and win rate is not.
Top comments (0)