When the Oracle says 70%, does it happen?
Mostly — and when it doesn't, it errs on the side of caution. Out of sample, the Oracle's strong favourites win more often than it claims: it is a modest forecaster, not a loud one.
Predicted vs actual
Each blue dot is a band of favourites from the out-of-sample backtest (bigger dot, more matches). Gold diamonds are live calls locked before the toss. Dots above the dashed line mean favourites won more often than the Oracle said.
By format
Same bands, split by format (out of sample).
| FORMAT | 50–60% | 60–70% | 70–80% | 80–90% | 90–100% |
|---|---|---|---|---|---|
| Tests n=99 | 56.5%n23 | 60.0%n20 | 68.2%n22 | 82.6%n23 | 72.7%n11 |
| ODIs n=435 | 55.8%n138 | 58.7%n126 | 72.5%n102 | 90.7%n54 | 93.3%n15 |
| T20s n=3,526 | 56.4%n1646 | 68.6%n993 | 82.3%n537 | 93.5%n292 | 93.1%n58 |
Live calls so far
The same bands over every prediction locked before the toss since the Oracle went live. Small samples swing — the backtest is the steadier read until hundreds of calls settle.
Calibration questions
What does "calibrated" mean for a cricket prediction?
A calibrated forecaster is right as often as it says it will be: of all the teams it gives 70%, about 70% should win. Accuracy alone can hide over-confidence; calibration cannot.
Is the CricMind Oracle over-confident?
No — if anything it is too modest. Out of sample, favourites it rated 75% won 80.3%, and favourites rated 84% won 92.4%.
Is this tested on matches the model had already seen?
No. The backtest fits the model on matches before 2024-01-01 and scores it only on the 4,060 decisive matches after that date. Live calls are locked before the toss and can never be edited.
Out-of-sample: weights fitted on matches before the holdout start, scored on matches after it. Methodology →