How we test our model
Next to each game on our lines page you’ll see what the market says. For some markets you’ll also see what our model says. Here’s how deep the testing goes before any market gets our number.
21+ · 1-800-GAMBLERFive checks. All five, or no number.
We replay past games in time order. The model only sees data from before each game, so every test game is one it hadn’t seen. Every number is scored against the closing line with the margin removed.
Enough games
At least 500 games (250 for NFL), so a lucky streak can’t pass.
Accuracy
Brier score and log loss measure how close our percentages land to what happened. Ours must be no worse than the closing line’s on both.
Calibration
When we say 60%, it should happen about 60% of the time, give or take 3 points (ECE of 3 or less).
No weak spot
Split the games into ten groups by our percentage. No group can be off by more than 5 points.
Holds over time
It must do better than the close in at least 2 of 3 time periods, not just one hot stretch.
What a pass means: our numbers held up as well as the closing market across every check. It does not mean they beat it, and it makes no promises about any game. “Blended with market” means a fixed 50/50 average of our model and the no-margin closing line, set before testing.
Two numbers, kept apart
The chance the posted lines imply once the built-in margin is taken out. Shown for every game with a line.
Our own estimate, shown only for markets that passed all five checks. Where a market hasn’t passed, you see the market’s number only.
This badge marks a pick or leg where we show our own number from a market that passed every check. Blended with market marks numbers that average our model with the closing line.
Built on thousands of real games
Every market is replayed through at least 500 past games (250 for NFL) before it can show our number, and most were replayed through thousands. In total we’ve graded 98,784 past game-markets across six leagues (each game counted once per market): NFL, college football, NBA, NHL, MLB and college basketball.
Time order, like real life
Games are replayed in the order they happened. The model never sees a result before it has to make its number.
Checked against the close
Every number is graded against the closing line on accuracy and calibration, with the built-in margin removed.
Different time periods
The test history is split into three stretches, and a market has to hold up in at least two of them.
Updated as we go
This is a work in progress. We’ll keep testing and update this page as results change.
Factors we tested and dropped because the closing line already priced them in, or they made our numbers worse: bullpen workload and fatigue, rest and travel, ballpark roof and air density, line movement, pace, officials, and goalie quality on its own.
All numbers come from public game data. This is a calibration tool, not advice, and it makes no promises. Last updated October 26, 2026.
Our model numbers are a model estimate for entertainment, not a recommendation. Not gambling advice. GG Media does not accept wagers. Must be 21+. Gambling problem? Call 1-800-GAMBLER.