Every prediction is saved before the match and compared with the result.
In short
Over 18 matches the model got the outcome (1, X or 2) right 56% of the time. Always picking an away win would be right about 43% of the time. For now its percentages do no better than league averages alone.
With fewer than 30 matches this is only a first look: the numbers can change a lot.
How the last round went
22 October · outcome right 10 out of 18 (56%) · exact score right 1 time
The surprises
The best predictions
10 correct predictions out of 18. Last 12: 5 correct. More matches are needed before judging form.
Real results are needed to calculate a reliable score. No correction is applied.
Lab · look back
18 predictionsComparison on predictions actually saved before the result. No prediction is rebuilt after the fact.
Model quality
For now the model does no better than the simple method (1% more error). With 18 matches the figure is still shaky: it needs at least fifty matches, ideally a hundred.
When I say 60%, does it happen?
Each point is a band of predictions (the number next to it says how many matches). The closer the points are to the dashed line, the more reliable the percentages.
If "Stated" and "Actual" are close, the percentages are reliable.
New model or old?
Over 18 matches, statistics only: error 0.36 for the current model, 0.38 for the first version. Too early to say.
Reliability by market
A 0–100 indicator built from calibration and Brier score. It is descriptive, not a guarantee of results. The model's picks use the history only when there are at least 10 observations per category.
Sample still too small: the system keeps collecting data without changing the model's probabilities.