Every call we make. Hits and misses both.

How the algorithm is graded

Every fixture runs through a 10,000-trial match simulation that produces a projected home-goal lambda, an away-goal lambda, a both-teams-to-score (BTTS) probability, and an over 2.5 goals probability. We commit those projections to the database before kickoff. Once a match finishes, our grader compares the algorithm's call against the actual score and records a hit or a miss. Coverage and bias on total goals show whether the model tends to project high or low — the closer to zero, the better. Goals coverage measures how much of the actual goal total the algorithm captured: 84% over this window, with a mean bias of -0.43 goals per fixture.

We don't hide the misses. We don't back-test cherry-picked windows. The table below is the most recent 14 days of finished matches that had a published projection — that's it. When the projection improves, you'll see the hit rate rise here. When the algorithm gets recalibrated, you'll see it in real time.