Helping you make the right call.The football model that shows its working.

Every club in five leagues carries an Elo rating that moves after each match by how surprising the result was. Separately, every club carries two card tendencies — how many it collects, and how many it provokes in opponents. Both are shrunk toward the league average so a club with three matches played doesn't produce a wild number.

A fixture's expected card count comes from multiplying those tendencies against the league baseline. Cards fit a Poisson distribution unusually well: across 5,256 matches the mean is 3.85 and the variance 4.07, and Poisson requires those to be equal. That's why the card markets are the ones we publish.

What we have tested

• Cards model calibration. Over 1,602 walk-forward matches, when the model said 55% it happened 57% of the time; 64% → 65%; 74% → 75%. It is honestly calibrated and slightly underconfident.
• Match results against real closing odds. Across 1,552 matches in five leagues, backing the model's high-confidence picks returned between −5.8% and +3.3% depending on threshold, with every Z-score inside ±2. That is indistinguishable from zero.
• Our model versus the market. RPS 0.2054 for us, 0.1982 for the closing line. The market is about 3.6% sharper than we are at calling results.

What that means for the card

Because no source we can reach publishes card prices, we cannot check our card numbers against a market the way we can with results. The odds shown on each slip are our own estimates with an assumed margin, not quotes. Take the selection to your own bookmaker and use their price.

So the card is a most-likely-outcomes card, not a proven value card. A calibrated probability is necessary for an edge but not sufficient — you also need a price that is wrong, and we cannot yet demonstrate that we find those.

Things we tried that didn't work

• Referee effects. Card rates by referee spread from 3.54 to 4.52, which looks strong, but the variance between referees is smaller than sampling noise alone would produce. No signal.
• Corners. Modelled, tested, and dropped — the high-confidence tail was badly miscalibrated and it lost money at every threshold.
• Reacting faster early in the season, and pulling clubs toward their long-run level. Both made forecasts slightly worse on out-of-sample data, so neither is used.

Why the Red tier is priced differently

Red combines a draw with over 4.5 cards in the same match. Legs from one match aren't independent — drawn games carry more cards than decisive ones (4.42 against 4.19 across 5,256 matches) — so multiplying the two probabilities understates the true chance by about 9%. Red is priced with that correction applied. The other tiers only ever combine legs from different matches, where multiplying is legitimate.