The record
Our record, in full
Every pick is written to an append-only log before kick-off and graded against the real result afterwards. Nothing is edited once a match starts. This page reports that log, including the parts that look bad.
Where it stands
| Measure | Value | What it means |
|---|---|---|
| Graded picks | 157 | Logged pre-kickoff, settled against a real scoreline |
| Tip accuracy | 43.3% | 68 of 157 model tips were the actual result |
| Value bets placed | 73 | Where the model's edge cleared the threshold |
| Value bets won | 19 | 26.0% strike rate at longer average prices |
| Profit and loss | −46.02 | Paper units, flat notional bankroll |
| ROI | −21.4% | Losing, and shown because it is losing |
The uncomfortable part
Always picking the home side wins about 47.1% of the time. Our live tip accuracy is 43.3%. On this sample, the model is currently doing worse than a rule a child could apply.
We are not going to dress that up. Three things are true at once, and all three matter:
- 157 picks is a small sample. At this size a run of draws moves the number by several points. It is not yet evidence of much in either direction.
- The sample has a hole in it. A four-week outage in July 2026 left the pipeline publishing an empty slate, and that gap sits inside this window.
- The backtest said +7.1pp, not −4pp. Held-out validation on an unseen 2021-22 season beat the baseline. Live results have not reproduced that yet. Either the sample is too small, or something about live conditions differs from the backtest. We do not know which, and saying so is more useful than picking the flattering answer.
Why the P&L is worse than the accuracy
The tip and the bet are different things. The tip is the most likely outcome; the bet is the one with the highest expected value, which is often a different result at a longer price. Across 73 bets, 31 were placed on an outcome other than the tip.
Longer prices mean a lower strike rate by construction — 26% here — and a losing run costs more before a winner arrives. The reasoning is set out here.
What this record does not cover
Only competitions with a results feed can be graded. That is 12 of the 43 currently on the board. Fixtures from the other 31 are published with a prediction but can never be settled, so they are not in the 157 and not in the accuracy figure.
Stated plainly: the record describes those 12 competitions, not the whole board. Closing that gap needs a results feed we do not have.
Why publish this at all
Prediction sites advertising 80-90% accuracy are not measuring 1X2 outcomes over a public, pre-committed log. The number that can be checked is worth more than the number that sounds good, even when the checkable one is negative.
If the model improves, this page will say so, and the log will show when the change happened. If it does not, this page will keep saying that too.
Figures from the forward log, current as of 27 August 2026. No wager is ever actually placed, so the P&L is paper arithmetic rather than a trading record.