We priced the first 40 Premier League matches of the season two different ways, then checked every price against what Bet365 was actually offering. One board leaked on 35 of 40 games. Here are the numbers, including the ones that embarrass us.
Every bookmaker has a number it never shows you: how often its own prices are wrong. Not wrong in the sense of "the favourite lost" — wrong in the only sense that matters to the house, which is that somebody with a sharper price than yours could bet into you at a profit. We built a page that measures exactly that, and this is its first public reading.
A quote leaks when backing it at our price has positive expected value under the market's own probabilities. To get those probabilities we take Bet365's odds, strip out their margin (the Shin method — it corrects the way books shade longshots), and ask: at our price, would a bettor who simply trusted the market's fair probability be getting paid more than fair? If yes, that line is a gift, and sharps collect gifts for a living.
We ran it on two boards. The first prices every match from our own model — an Elo rating system trained on 21 years of results. The second anchors each price to the cross-bookmaker consensus and lets the model nudge it by a small amount. Both carry the same 5% margin. Both were graded against the same 40 played matches, with a sharp staking ₦100 on every leaking line to see what they'd actually have won.
An 88% leak rate isn't a bad model. It's a decent one, and that's the lesson. Elo predicts Premier League results about as well as any public method — we measured it at a Brier score of 0.582 against the market's 0.570 — and that gap, small as it looks, is the entire problem. A model that is slightly worse than the market produces prices that are slightly wrong on almost every game, and a sharp doesn't need them to be very wrong. He needs them to be wrong by more than the margin. Ours were, 35 times in 40.
The worst one is instructive. Brentford at home to Tottenham, 22 August:
| Outcome | Our price | Bet365 | Sharp's edge | |
|---|---|---|---|---|
| Brentford | 1.63 | 2.40 | We had Brentford as a strong favourite. The market didn't. | |
| Draw | 3.93 | 3.70 | ||
| Tottenham | 5.50 | 2.75 | +90.6% EV | Twice the market's price on a near-coin-flip. Spurs lost — the sharp dropped ₦100 — but he'd take that bet every week. |
The model had Brentford at 1.63 because their Elo rating was high after a strong finish last season; the market, which knows about summer transfers, injuries and a hundred things a results-only model can't see, had them at 2.40. Betting Tottenham at our 5.50 when the true price was about 2.90 is a +90% edge. It lost. It was still the best bet on the board. That is the difference between a result and a price, and it's the whole reason we grade by closing-line value rather than by who won.
Not a better model. We tried — feeding the model expected-goals data made it genuinely more accurate and changed the leak rate by almost nothing, because the market already knows everything xG knows. What fixed it was humility: anchoring every price to the market consensus and letting the model move it by 5% at most. The leak rate fell from 88% to 5%, and the sharp's take went from +₦869 to −₦200.
This is how every serious bookmaker actually operates. The odds you see at Bet9ja or Bet365 are not that company's opinion about the match; they are the market's opinion, shaded a little for the book's own liabilities. The "model" that people imagine sitting in a bookmaker's basement mostly exists to decide how far from consensus the book dares to stand — and the honest answer, on our numbers, is "barely".
| Match | Our away price | Bet365 | Sharp's edge | |
|---|---|---|---|---|
| Arsenal v Coventry | 17.39 | 13.00 | +9.6% EV | Lost |
| Man City v Coventry | 16.77 | 12.00 | +15.0% EV | Lost |
Both are Coventry, and both are the away longshot. Coventry were promoted this summer; they have never played a Premier League match in our 21 years of data, so the model prices them off a default rating for a newly promoted side and thinks they are a 22-to-1 shot at the Emirates. Even at 5% influence, that opinion drags the anchored price out past the market. Everywhere the model has real information, the anchor holds. Where it's guessing, the shade shows. We could set the model's influence to zero on longshots and close both gaps; for now we'd rather leave them visible and let the season decide.
No bookmaker will show you its leak rate, for the same reason no shop labels which shelves are overpriced. We show ours because the entire point of this site is the number the others hide: whether the price you took beat the price the market closed at. We hold ourselves to it in public so that when we tell you your closing-line value on the bets you place, you have a reason to believe the grader.
The page updates every day at book.rolescout.uk/admin/sharp, and this article will be wrong within a week — which is fine. Being checkable is the product.