This is the page we would delete if this were a tipping service. It is the reason to trust anything else here.
After testing across two seasons, we found something that looked like an edge. Take the model's strongest disagreements with the market, throw out home wins — a category the model was demonstrably bad at — and the resulting selections beat the closing price by about two percentage points across both seasons.
Two independent seasons agreeing is not nothing. It looked real.
We froze the filter — wrote the rules down, changed nothing afterwards — and ran it against a third season the model had never touched.
It died.
Against the closing price it came out dead flat, at −0.03%, having managed +1.4% and +2.5% on the two seasons it had been built from. On money it lost heavily. And both of the things it leaned on failed at once: concentrating on the strongest opinions stopped helping, and the "never back home teams" rule did not just stop working, it reversed. In the unseen season, home teams were the least bad selection available.
It means the pattern was fitted to the past rather than describing something true about football. Two seasons of agreement felt like strong evidence and turned out to be a coincidence with a good story attached.
This is the single most common way people fool themselves with data, and knowing about it in the abstract offers essentially no protection. We knew about it. We did it anyway. The only thing that caught it was committing to the test before seeing the result, and then not negotiating with the answer.
A later re-examination softened the reading slightly — the part that broke was mostly the bolted-on home-team rule, while the core "back our strongest disagreements" signal did hold up out of sample, at a small but consistent margin. That is a genuine finding, and it is still far too small to bet.
Nothing, which is the entire point.
Because the discipline was to prove it before staking it, a curve-fit that would have lost real money over a season instead cost a few days of computer time and some bruised confidence. The ledger still reads £0.
That is what the process is for. Not to produce a winner — to catch you when you are wrong, before it matters.
The fourth exam runs live. Every high-conviction call the model makes this season is published before kick-off and settled in public afterwards, judged on beating the close and on profit at the price taken. The bars were written down in advance.
If it clears them, we will say so. If it does not, that will be on this site too, in the same size type.