We graded 72 settled matches our model chose not to bet: flat-staking those passes would have returned -13.6%, and even the subset with the largest apparent edge went 13-11 and lost money. The first version also claimed the staked bets beat the close 72% of the time; against the vig-stripped close, that claim failed.
Update 2026-10-09: the first version of this article said the bets the model did make beat the closing line 72% of the time (average CLV +4.3 points) and that passing was where the edge came from. That figure was measured against the raw closing price, vig included. Against the vig-stripped close our CLV is negative: see our devigged CLV. The pass audit below stands; the CLV claim does not.
Most tipster services post a play on nearly every match. Our model passes on most days. To check whether those passes were right, we took every match it declined over a settled stretch and graded what would have happened had we bet them.
Across 72 settled matches the model passed, flat-staking every one would have returned −13.6%. The subset with the largest apparent closing-line edge went 13-11 and still lost money. Those matches were priced efficiently.
It shows the model was not leaving money on the table by passing. It does not show that the bets it made have an edge: that needs positive closing-line value against the fair close, and on that measure our number is negative. Our accuracy study also finds the model does not out-predict the closing market.
Our other studies explain why so many matches are passes: longshot "value" is mostly margin, and most "tired player" angles are already priced.
72 passes is a small sample. This is an audit of the process, not a statistical proof.
Based on the model's forward record and a leak-free audit of its declined matches. The graded record is at /bot.
See today's picks: published before the match, graded in public →