The forecast, explained.
A probability estimate is useful only when you understand how it was made—and what it cannot tell you.
Completed results in. Probabilities out.
Our collector records source round IDs, outcomes, source timestamps and payouts. A central worker reads these records from InfluxDB, maintains a canonical history, and computes forecasts once per game update. Every viewer receives the same model snapshot.
Three views of the same history
Responsive uses recency-weighted counts from up to 100 rounds. Balanced uses smoothed frequencies from up to 500 rounds. Long history uses up to 3,000 rounds. These initial models are statistical baselines. They are not advertised as proof that random outcomes can be predicted.
Category accuracy and exact accuracy
Category accuracy measures whether the highest-probability group—number or bonus—matched the result. Exact accuracy measures whether the top single outcome matched. Groups have unequal frequencies, so a high category hit rate alone is not evidence of an advantage.
Historical evaluation versus live evaluation
Historical walk-forward evaluation uses earlier rounds to estimate a subsequent round without random train/test shuffling. The interface labels these results as historical. Sealed live evaluation uses forecasts actually saved before the target outcome’s source timestamp. Counts and evaluation types appear with each metric.
Payouts need context
Total paid is the source-reported aggregate monetary payout in the displayed currency. Multiplier is the source-reported maximum, not necessarily an individual player’s return. Payout figures are shown for context; the first baseline models do not yet use them as predictive features. Payout-aware candidates must pass separate evaluation before release.
Limits and freshness
Game outcomes may be independent. No model guarantees wins or positive returns. A delayed feed is explicitly marked; estimates must not be treated as current when source data is stale. Crazy Time and Crazy Time A use independent histories.