Methodology
Know what is behind the number.
A useful record explains where the information came from, what was known at the time, and how the outcome was counted.
01
Observed data
Scores, clocks and bookmaker observations retain their source timestamps. Missing coverage stays unknown. A live score does not prove that a live forecasting model is available.
02
Separate models and scopes
Model support is specific to a sport, competition, format and phase. A model that passes a held-out evaluation before release is labelled validated, and a forecast from a model without a reviewed release is labelled as such. MLB pregame forecasts are the de-vigged consensus of the betting market rather than a model.
03
Independent checkpoints
Early and main forecasts are computed independently before the event. Golf adds separate later-round checkpoints. Once published, original probabilities and selections stay unchanged.
04
Prices on the record
Each published forecast is recorded with the best sportsbook price for its side, taken at publication or recovered afterwards from the odds history as of that moment. A price that was never observed is not estimated; that forecast counts for accuracy and carries no units.
05
Results and corrections
Sport-specific settlement uses authoritative outcomes and the market contract. Uncertain or contradictory evidence remains pending or under review. Corrections are audited without rewriting the original forecast.
06
The complete cohort
Reporting includes every graded forecast, losses included, and preserves legacy labels. Accuracy, calibration and units at the recorded price each measure the same forecasts differently. We do not blend them into a promotional success rate.
The record is the place to check.
Inspect real publication times, validation labels, sample sizes and outcomes. A small sample is not proof of future performance.