How accurate are football predictions? We marked all 38.
Gregor · Statteo ·
Anyone can post a football prediction. Almost nobody posts what happened next. So here is every match Statteo put a number on for Saturday 15 August — 38 of them across the Championship, League One, League Two and the Scottish divisions — checked against the results.
Start with the six we were most confident about. Cambridge United, we made them 71% to beat Wigan, and they won 3-2. East Kilbride at 68% beat Cove Rangers 3-0. Hamilton at 68% beat Airdrie 4-0. Bradford at 66% beat Peterborough 2-0. Bolton at 65% beat Preston 2-1. Grimsby at 62% beat Exeter 1-0. Six from six.
That is the pattern we most wanted to see. When the model is confident, it should be right more often — and on Saturday every single match it rated between 60% and 69% finished the way it said. Seven of its eight strongest calls came in.
Scottish League One was the standout. Four matches, and we called the winner in all four and the goals in all four. Eight calls, eight right. Alloa 3-0, East Kilbride 3-0, Hamilton 4-0, Peterhead 2-1.
It is worth being clear about what these percentages mean, because they are not tips. When the site says 'Bolton to win · 65%', it means we think that result happens about two times in every three. It is a measurement, not a shout. The useful thing you can do with it is hold it against whatever price you are being offered and decide for yourself whether that looks generous or mean. We do not publish odds and we are not a bookmaker.
Now the parts that went badly, because a results article that only lists the wins is an advert.
Both teams to score was our weakest call of the day: 17 correct from 38, which is worse than a coin toss. That market falls out of the model as a by-product rather than being something we built for directly, and Saturday made the cost of that obvious.
The second problem is subtler. Across the 38 matches the model averaged 25.5% on the draw, and six games finished level — 15.8%. We were consistently calling more draws than football actually produces, and those percentage points came out of away wins, which we had at 26% and which happened 34% of the time.
Here is the part we are genuinely pleased about, oddly enough. We had already found that draw bias earlier the same day, testing how hard one of the model's internal corrections was pushing. Saturday's results agreed with the arithmetic, independently. Two separate pieces of evidence pointing at the same fault is a much stronger reason to change something than either one alone.
Some context on the numbers, so nobody reads more into them than is there. Thirty-eight matches is one Saturday. It is enough to spot a systematic bias — a market landing 45% of the time when it should be near 50% is a finding — but nowhere near enough to judge a model overall. We will not be tuning anything on the strength of a single afternoon.
For anyone who wants the technical measure: across those 38 fixtures the model scored 0.580 on Brier, against 0.621 for someone who simply guessed football's long-run base rates every time. Brier grades how confident you were rather than just which box you ticked, and lower is better. Beating the base-rate benchmark on the first scored card is a reasonable place to start from.
Every match we publish now gets scored and kept, automatically, whether it made us look good or not. That record is what tells us what to fix, and it is the only honest basis for asking anyone to take our numbers seriously. Come back next Saturday and you can check that one too.
Statteo — tips that show their working
18+ · Patterns, not tips · begambleaware.org