Where bias correction wins big

Across our 18-league validation, five leagues stood out: J1 League, Saudi Pro League, MLS, La Liga, and Brasileirão. In each one, correcting for AI biases improved predictions by more than 10% on Brier score. The pattern across them says something important about where this methodology earns its keep.

The five-league shortlist

That's a meaningful gap. A 14-19% Brier improvement is the difference between a model that occasionally earns its keep and one that consistently does. It also held up across all five models we tested — not just the ones with the loudest documented biases.

What these five leagues have in common

Three of them — J1, Saudi Pro, MLS — are leagues where you would not bet heavily on AI training data being deep. There are far fewer match reports, tactical breakdowns, and English-language analyses written about a Sapporo vs Sanfrecce Hiroshima fixture than about Manchester City vs Arsenal. The models are working from a thinner base of priors.

That thin base of priors matters because it's exactly the situation where defaults — “Premier League players are better,” “tournament teams beat club teams,” “home advantage is worth +X” — fill the gap. Bias correction is most useful when the model is leaning hardest on those defaults instead of league-specific evidence.

La Liga is the odd one out. It's a league with mountains of training data. We'll cover why La Liga still benefited so much in a separate post — it's our cleanest validation of the methodology, and the explanation isn't the same as for J1 or Saudi Pro.

What it does not mean

A 19% Brier improvement does not mean we are picking 19% more match winners in J1. Brier penalises confidence, and most of the improvement comes from tempering over-confident predictions on outcomes the models had no real basis for. In Brier-score terms that's a big move; in “was the pick right” terms it's smaller and harder to read.

It also doesn't mean we should expect the same lift on every match in those leagues. The improvement is an average over many fixtures. On any single match the ensemble could land worse than the raw average — and in our data, did, sometimes. The methodology is about the long run.

What this implies for the World Cup

The World Cup pulls in teams from every confederation. By group stage we will see fixtures that look more like J1 or Saudi Pro than Premier League: matchups where the models have thin priors and reach for defaults. Those are precisely the matches where this study expects bias correction to add value.

Tomorrow: the four leagues where bias correction made things worse, and what we think that tells us about the limits of the approach.