Transparent backtesting: the numbers to publish

A percentage without its denominator proves little. A public record should show the date, market, recorded line, outcome and grading rule. Losses must not be removed, and several checkpoints of the same decision must not be counted as independent wins.

One forecast, one outcome

Exact-bet deduplication combines repeat records of the same selection. It is also useful to show one best signal per match, so that a dozen correlated selections do not create a misleading impression of a large sample.

The test period and a frozen version

After tuning, the model is frozen and checked on new games. If thresholds were changed after reviewing losses, the earlier percentage is exploratory rather than independent evidence.

What the model checks

• The time period and N
• Wins, losses, pushes and voids
• Duplicate removal
• The model version and a blind sample

Analysis describes probability and risk; it does not guarantee outcomes. This material is intended for readers aged 21+.