How Accurate Are Betting Models?
Per-game accuracy is mostly variance, so it is the wrong question. What separates a useful betting model from a tout is calibration over many games, measured against the sharp closing line, and published where you can check it.
Updated Jul 2026 · Part of the getting started series
Ask how accurate a betting model is, and most people mean how often it calls the game correctly. Per-game results are dominated by variance, the swings in outcome driven by chance rather than skill, covered in Variance in betting. That per-game question cannot separate a good model from a lucky one on any sample size a bettor will ever actually see.
Calibration
A checkable question replaces the per-game one. When the model gives a side a 60% chance to win, does that side actually win about 60% of the time across hundreds of similar calls? That check is calibration. Calibration accumulates over many games. A single game’s outcome does not. Per-game correctness stays noise no matter how many times you check it.
The sharp close
The closing line at a sharp book gives calibration a benchmark to compare against. By kickoff, the price has absorbed injury news, confirmed lineups, weather, and the money of bettors who have historically beaten the market. That combination makes it the best publicly available forecast for a game’s outcome. That is the idea Closing line value covers in full. Academic comparisons have tested well-known public models against that closing line across tens of thousands of games, and the market price won.
A public-data model that ties the close over a large sample is performing at the practical ceiling for what public information can produce. A claim to broadly beat it deserves skepticism. The one exception Fairline claims is documented in the MLB underdog edge, with its limits shown.
Three numbers
Three numbers measure any model, checkable without taking the builder’s word for anything. A calibration curve plots predicted probability against realized outcome by bucket, every pick made in the 60% range grouped together and checked against how often that bucket actually won, repeated bucket by bucket from long shots to near locks. A Brier score grades those same predictions against the base rate, the accuracy of just predicting the historical average outcome every time, and shows whether the model adds anything beyond that plain baseline. The CLV of the model’s picks shows whether its specific selections beat the market before a single result gets decided by variance.
A model seller who publishes none of these three is asking you to trust an adjective.