Best bets

Every club gets an attack and a defence rating: what it creates and concedes, divided by what an average side would have managed against the same opponents. Those two rates go into a Poisson over every scoreline, and every market on this page - the result, the totals, both teams to score, each club’s own total - is read off that one grid, so they cannot contradict each other.

It is fitted on expected goals with a little weight left on the finishing (70/30), with a match halving in weight every 180 days. Both of those were chosen by the backtest rather than by preference: expected goals beat actual goals out of sample by a clear margin, on every setting tried.

Log loss, result

0.9717

Base rates would be

1.0669

Result called right

53.7%

Calibration error

1.7%

Measured over 1,610 EPL fixtures it had not seen, across 2021-22–2026-27. The ratings were refitted at the start of every round from the matches played before it and nothing after, which is the situation you are in on a Friday.

What the edge is, and what it is not

Take Crystal Palace to score at least once, at 1.53. The model makes it 73.3% and the books’ own prices make it 64.2%; blended, it is about 67%. That gives an edge of +2.8%, which means: stake one unit on bets like this over and over, and you finish 0.028 units ahead per unit staked, on average.

It is not 53%. A winning bet at 1.53 returns 1.53 — a profit of 0.53 — but it only wins about two times in three, and the third time the whole unit is gone. 0.67 × 0.53 − 0.33 × 1 is where the 2.8% comes from. The edge is the number that survives the losses; the price is the number that does not.

Why the biggest edge is not always the pick

A board is ranked on the stake, which is the Kelly fraction — the edge divided by what the bet stands to lose — taken at a quarter. At the same edge a shorter price is the better bet, because the same money is at risk for far less of the time, and an edge on its own cannot see that. On Aston Villa against Arsenal the outright at 7.00 carries almost three times the edge of Villa-or-the-draw at 2.60, and the double chance is still the better bet: 5.8% of a bankroll against 4.6%.

What confidence is made of

• Sample. How much football is behind the weaker of the two ratings. A promoted club in August has almost none, and the score says so.
• Corroboration. How many books quote the bet, and how far the best price sits above the rest. A standout and a stale quote look identical from here.
• Calibration. How the model behaved at this probability in the backtest - not on average. A model can be excellent overall and over-confident in one band, and a bettor bets the band, not the average.
• Agreement. How far the model has strayed from the market’s own devigged view. A model claiming a bet is twice as likely as the whole market thinks has almost certainly missed something - a suspension, a rested squad, a goalkeeper - and this is a brake, not an accelerator.

Why there are no player props here

Because they are measured against something else entirely, and putting the two in one ranked list would imply a comparison neither number supports. A team market has a complete set of outcomes, so the bookmakers’ own margin can be removed and the model checked against what the market really thinks. Australian books quote a player prop one way — there is no Under to devig against — so the only honest yardstick is the player’s own record, which is a fact about the past rather than a claim about the match.

They are on the value board, ranked on that record, in their own table for the same reason. Bringing them here needs one confidence scale that spans a model probability and a hit rate, and that does not exist yet.

What has not been proved

The model is measured against results, and it is good at them. It has never been measured against a closing price, and cannot be: this database has only ever stored the current round’s odds, so there are no historical lines to test against and there never will be for the seasons already gone. Until that record exists, the model is given at most equal weight with the market rather than more - the edges you see are already blended on that basis.

Every run now stores its probability beside the price it could have been struck at, so that record is being built from this round on.

Model epl-dixon-coles-1 · fitted 7 Sept 2026, 9:01 pm (Sydney) · nothing after that instant was read.