How likely it is. Not what to bet.
Every pick is written down before kick-off and graded against the result afterwards, in public, whichever way it goes.
Does an 85% pick win 85% of the time?
Every graded selection, bucketed by what it claimed and plotted against what then happened. A dot on the dashed line is a band that came in exactly where it said it would; the shaded strip is one percentage point either side of it.
All 14 bands land inside that strip. The worst is off by 0.65 points. Dot size is the number of bets behind each one.
The same thing market by market, and where each one stops being checked.
Four steps, and the third is the one that matters.
First, the market price
Every number starts from the bookmakers' own. Raw 1/odds sums to about 1.07, and counting that margin as information is the easiest way to fool yourself — so it is removed first. The card is priced on the current line; everything it was checked against, on the closing one.
Then a model, walked forward
Team strengths are estimated from goals and shots on target with a low-score correction, refitted as the season moves, and every match is predicted using only what was known before it kicked off.
Then the calibration
The raw number is fitted to what actually happened, on folds it never saw. That is the step which makes “62%” mean 62% instead of meaning “confident”.
Finally, written down and graded
One pick per price band per match day goes into a ledger before kick-off and is settled against the result afterwards. The record on this page is that file, not a backtest.
The limits are the product too.
- It will not tell you a bet is good value. A calibrated probability says how often something happens. Whether the price on offer is worth taking is a different question, and it needs a price this does not always have.
- It is mostly the market. The bookmakers' line carries most of the forecast by design; the model's job is to add to it without damaging it. Measured out of sample over every season on file, it does not beat the closing line — it keeps up with it.
- The record is young. A handful of settled picks is not evidence of anything. The reliability tables run on tens of thousands of matches; the daily ledger does not, and is labelled with how few it has.
- It is research, not advice. Nothing here accounts for your stake, your book, or your ability to get the price shown.
The card is rebuilt every morning
Results, model, calibration and prices refresh on a schedule, and the day's pick goes out before the first kick-off.