Evaluation desk · recorded evidence
Research workspace
Model performance
Recorded probabilities, settled outcomes, and market snapshots—shown with the sample size needed to tell evidence from noise.
Evaluation window
180 days
Current production models only. Retired experiments excluded.
Our own record: the model and the blend
Record by source
How each source has done
Each source is graded on its own games, from the number recorded before the start, against the final result: how often the side it gave the highest probability won. For the home-field default, that is how often the home team won. Only the model and the blend are ours.
Probability quality
Calibration curve
Probability bands compared with their observed outcomes. The diagonal is perfect calibration.
This curve grades every published number together, whatever its source: soccer's blend and model with the bookmakers' or data provider's numbers, the NFL and college home-field default, and basketball and MLB numbers of unknown source. The record by source above gives each named source on its own.
Football parlay cards
Loading the record.
Reading the record
Every number is recorded before kickoff and graded against the final result, under the name of its source. Calibration matters more than a hot streak.