Why Prediction Accuracy Fluctuates

Across a weekend of ten matches, a model issuing 60% selections will land somewhere between three and nine of them purely by chance. Both extremes are entirely consistent with the model being correct about the underlying probabilities.

Judging any prediction service on a week of results tells you almost nothing. Hundreds of predictions are needed before a hit rate carries a meaningful signal.

A single deflection, red card or offside call swings matches that were otherwise going one way. Because scorelines are low, individual events carry disproportionate weight compared with higher-scoring sports.

That structural variance places a hard ceiling on 1X2 accuracy. No model reaches the high accuracy numbers some sites advertise, and claims above roughly 60% on 1X2 across a large sample should be treated with scepticism.

The useful question is not how often the model was right, but whether its stated probabilities matched reality. Group every 60% selection together: if roughly 60% of them landed, the model is calibrated.

Our accuracy page reports hit rates alongside a Brier score and sample size, so performance can be read in context rather than as a marketing number.