Accuracy record
Every published model summary is scored against what happened. The record is free and needs no account.
How a number gets onto this page
- A prediction is written down before the event starts. The database refuses one that arrives late.
- Once written it cannot be edited or deleted. A correction is a new row that points at the old one, so the original stays readable.
- Each prediction is linked to the one before it by a hash. Change any past row and every later row stops matching.
- After the event, the outcome is attached and the scores below are computed from it.
That chain is the point. It means the underlying record cannot quietly drop the predictions that went badly.
Scores by model and market
No scored predictions match that selection yet. That is the record saying so, not an error.
Calibration by confidence band
No calibration bands for that selection yet.
Check it yourself
Download the scored summaries and calibration bands shown on this page as one self-contained file:
The CSV contains its own field names, sample sizes, scores and calibration bands. It does not depend on a second website or repository link.
What this record does not prove
- A good score on one market says nothing about another. Read the row you care about, not the average.
- Rows marked low sample are noise. Ignore them until the count is real.
- Calibration is not profit. A well-calibrated model still loses to a market that is equally calibrated and charges no fee.
- Past accuracy is a measurement, not a promise. Rosters, rules and schedules change, and so does the model.