Three full seasons, replayed match by match with every price set before the result was known. No selections removed afterwards. Here is all of it.
CLV is closing line value — how much better than the closing price the model's picks were, which is the honest measure of whether it saw something real. Simulated ROI is what fixed reference stakes would have returned in the replay — no money was staked, and the number depends entirely on the staking rules chosen. We treat it as reference data from the collector, not gospel.
| Season | CLV | Simulated ROI | Note |
|---|---|---|---|
| 2023/24 | -0.03% | -18.8% | the unseen exam — the filter died here |
| 2024/25 | +1.36% | -5.6% | tuned season |
| 2025/26 | +2.46% | -11.2% | tuned season |
The same machinery run on scrambled data, subtracted from the real thing. Anything that survives is signal rather than the testing procedure fooling itself. This is the part that held up every time, including on the season the model had never seen.
| Season | Skill vs placebo |
|---|---|
| 2023/24 | +0.69 |
| 2024/25 | +0.53 |
| 2025/26 | +0.70 |
Real, replicated, and far too small to bet.
We found what looked like an edge across two seasons — the model's strongest disagreements, excluding home wins, beating the close by about two points. We froze the rules and tested them on a third season the model had never touched.
It came out flat against the closing price and lost heavily on money. Both pillars broke: conviction stopped concentrating the edge, and the "never back home teams" rule reversed outright. That is what a pattern fitted to the past looks like when it finally meets fresh data.
The fourth exam runs in public. Every high-conviction call is published before kick-off and settled afterwards, judged on beating the closing price and on profit at the price taken. The bars were written down before the season started and cannot be moved.