Model Performance

How to read this page

A forecast here is one decision — one match, one stage, one title race — not one line of probability. Picking the winner of a Tour stage means scoring 160 riders, and almost all of them are easy calls; counting those as 160 forecasts would let cycling drown out football.

Raw Brier scores cannot be compared across markets either, because the difficulty differs. Guessing blindly scores about 0.25 on a tennis match and about 0.006 on a 160-rider stage. The skill score removes that: it measures how much better the model does than knowing only how often this kind of call comes in. Zero means it adds nothing, one is perfect, and a negative number means it is worse than guessing. 4 forecasts sit in market and field-size segments too small to set a reliable baseline, so they show a raw score only.

Calibration

When the model says 70%, does it happen 70% of the time? The closer to the diagonal, the better calibrated. Head-to-head markets and large-field markets are drawn separately: field markets are almost entirely low-probability lines and would otherwise hide everything else.

Skill over time

Monthly skill score by sport. Higher is better. Months with fewer than 5 scored forecasts are excluded.

Model versions

Each model change is tracked separately so improvements can be verified.