NINETYSTATS

Accuracy, measured. Not promised.

Everything below comes from matches the model predicted before knowing the result. The headline numbers are measured on three seasons it was never tuned on, so they cannot be flattered — and any market that fails this page’s standard is hidden from the product until it passes.

Distance behind the market
+2.58%
The betting market’s consensus is the world’s sharpest forecast, priced with team news we cannot see. This is how close we get without any of that — smaller is better, and we publish it either way.
Average drift from reality
1.5pp
When we state a percentage, this is how far it sits from what actually happens, on average — about one and a half points. Home, draw and away are each checked separately too.
Most likely outcome correct
52.7%
Across 2,660 matches. The market manages 54.6% on the same games — and draws, about a quarter of all matches, are almost never anyone’s most likely outcome.

The percentages are true

The heart of the product. Take every probability we have ever stated, group them by size, and check how often each group actually happened:

When we say 30–40%happened 34.7% of 354
When we say 40–50%happened 43.5% of 947
When we say 50–60%happened 57.6% of 668
When we say 60–70%happened 65.1% of 424
When we say 70–80%happened 73.2% of 198
When we say 80%+happened 88.4% of 69

Green bar: how often that outcome actually happened. White tick: the average probability we had stated. Premier League backtest, 2,660 matches, 2019–2026.

The live record starts at matchweek one

From the first round of the 2026–27 season, every published probability is frozen 24 hours before kick-off and graded after full time, hit or miss, alongside the market’s number for the same match. This page will show that record as it accumulates; until then, the figures above are from the historical backtest and labelled as such.

Season starts August 2026

How to read these numbers honestly

  • A 60% probability failing is not a wrong prediction — it is expected 4 times in 10. Judge the percentages over many matches, not one Saturday.
  • We sit behind the betting market and say so. Its edge is information (line-ups, late team news), not better judgement; where no special information exists, we measure at parity with it.
  • The model is tuned on one era of matches and judged on a different one, so the headline cannot flatter itself.