Step by step
How a number is built
- 1Official ADM history collection: complete, cleaned and versioned draws.
- 2Feature engineering: delays, frequencies, cyclometric relationships and other descriptive indicators.
- 3Calibrated ML model: trained on the past and tuned so the stated confidence matches the real frequency.
- 4Confidence + backtest: every suggestion ships with a calibrated score, validated walk-forward with no lookahead.
Calibration
What 'calibrated' means
A model is calibrated when the confidence it states matches how often it's right: if it assigns 0.7 confidence to a group of predictions, about 70% of them should turn out correct. We measure this gap with ECE (Expected Calibration Error): the lower it is, the more trustworthy the confidence.
We also use conformal prediction (MAPIE) to build intervals with a declared coverage: at 90% coverage the intervals contain the real outcome in 90% of cases on the history. It's an honest way to say how much to trust, without promising any single result.
Illustrative reliability diagram: the closer the points to the diagonal, the better calibrated the model.
ECE and conformal prediction, in brief
ECE (Expected Calibration Error) groups predictions by confidence level and, in each group, compares the average stated confidence with the observed success frequency. The weighted mean of these gaps is the ECE: the closer to zero, the more trustworthy the confidence.
Conformal prediction (MAPIE) does not promise a single number: it builds intervals with a guaranteed coverage on history (e.g. 90% coverage = the real outcome falls in the interval 90% of the time). It is an honest way to quantify uncertainty instead of hiding it.
Three levels
High, Medium, Low — readable at a glance
- High: pattern consistent over the long run
- Medium: signal detected but with variance
- Low: weak hint, to weigh with caution
- Never presented as a probability of winning
What the score is NOT
It is a measure of consistency of the detected pattern, not a probability of winning. High confidence does not guarantee the outcome: it only signals that the pattern is solid against history. The future stays uncertain and the game's expected value (EV) negative.