This is the one nobody selling picks will tell you, so it gets its own page.
We replayed the model across full past seasons — pricing every match before the result was known, betting it exactly as we would have at the time, with no hindsight and no quietly dropped losers.
It beat the closing price. And it still lost money.
Taking the disciplined version of the strategy through a season, starting with a £1,000 pot:
| Season | At the price a normal person gets | Shopping for the best price every time |
|---|---|---|
| 2024/25 | £1,000 → £769 | £1,000 → £1,024 |
| 2025/26 | £1,000 → £822 | £1,000 → £999 |
Scale those however you like — £100 divides by ten.
At the odds an ordinary person actually gets, betting at one bookmaker, you lose about a fifth of your money across a season. Chase the best available price on every single bet, which is a real job requiring accounts at a dozen books, and you get back to roughly breakeven.
And the ride is brutal either way. At the worst point in those seasons the pot was down between 30% and 60% before recovering. Very few people hold a strategy through a 50% drawdown. They quit at the bottom, which converts a paper breakeven into a real loss.
Two forces work against a small edge.
The margin. Every price on the board has the bookmaker's cut built into it. That cut applies to every bet you place, relentlessly, whether you are sharp or not.
Variance. Football results are close to coin flips with extra steps. A 3% edge is a whisper underneath swings that are an order of magnitude louder. Over enough bets the edge wins; a season is not enough bets.
The uncomfortable arithmetic is that an edge large enough to survive the margin and outrun the variance inside one season would have to be enormous — far larger than anything an xG model built on public data is going to find. If someone is claiming that, the claim is the product.
That a real edge does not print money over one season, and that ours is not big enough to bet.
So we have staked nothing. Not as a marketing line — the ledger shows £0 because the testing said do not, and the whole point of testing is to actually listen to it.