Leagues

Champions League prediction model explained: why it isn't the league model

By · Published · Updated 19 Oct 2026 · 4 min read

In short: a domestic goal model learns each team’s strength from a full season against the same opponents. The Champions League gives it thirty-six clubs from about fifteen leagues, eight matches each, and almost no shared opponents, so the real problem is comparing teams across leagues, not fitting a small sample. Any champions league prediction model has to answer that question first, and the only test of its answer is a public record of its calls.

What a domestic model relies on What the Champions League offers
38 matches per team, one pool 8 league-phase matches per team
Every pair meets twice Most pairs never meet
One scoring rate for the league Fifteen leagues, fifteen rates

Building a champions league prediction model looks, at first glance, like reusing a domestic one with a different badge on the crest. It is not. A domestic model earns its attack and defence ratings from dozens of matches inside one league, played between the same pool of teams every season. The Champions League gives a model none of that: thirty-six clubs from roughly fifteen different leagues, each playing only eight league-phase matches against eight of the other thirty-five. There is not enough within-competition data to estimate a Bundesliga side’s attack against a Ligue 1 side’s defence the way a domestic model estimates two Premier League sides against each other, and the two leagues are not directly comparable in the first place. Whatever replaces the domestic approach has to solve a cross-league comparison problem, not just a small-sample one.

Why the domestic method doesn’t carry over

A domestic goal model such as Dixon-Coles fits an attack strength and a defence strength per team from a full season of matches against a shared pool of opponents. That’s what makes the ratings comparable: every team’s numbers were estimated against roughly the same set of rivals, home and away, many times over.

None of that holds in a cup competition assembled from several leagues at once. Eight matches per team is a thin sample to fit two parameters from even within a single league; across two teams that have never met and whose domestic form comes from different competitions entirely, it’s not really an estimation problem any more; it’s a comparability problem. A club’s Bundesliga attack rating and a club’s Eredivisie attack rating aren’t on the same scale to start with, because the two leagues score goals at different rates and defend at different levels. Feeding both into one model as if they were interchangeable would quietly bake that mismatch into every prediction.

Three ways a model can bridge leagues

There is no single textbook answer, but the public approaches fall into three families, and they can be combined.

A shared-scale rating. Public club rankings such as Club Elo rate every European club on one scale: a rating moves match by match, up for a win and down for a defeat, by more when the opponent was stronger. Because the rating is carried through every fixture a club plays, domestic and continental, two clubs that have never met can still be compared. The weakness is the starting point: a club’s first European match cannot wait for evidence, so the rating it walks in with has to come from somewhere, usually from how its league has fared before.

A league-strength adjustment. Instead of one scale, keep the domestic ratings and correct them by a measured gap between leagues, estimated from the matches where leagues do meet: European ties, pre-season and the clubs that move between them. This keeps the precision of the domestic fit and adds one cross-league term; it works only as well as the gap is measured, and the gap moves.

The market as a prior. Bookmakers price every Champions League match, and their closing prices are the strongest public predictor of football there is. A model can start from them and look for where its own evidence disagrees. The cost is independence: a model that leans on the market cannot be evaluated against it, and beating the closing price is the test that matters.

Whichever route a model takes, the output is the same thing a domestic model produces: an expected scoring rate for each side, which a scoreline model turns into prices for the result, the goals and both teams to score.

What this buys, and what it doesn’t

The payoff of a cross-league method is that it can put a club from any of the fifteen-plus leagues in the competition on one footing. What it doesn’t do is inherit the precision of a model fitted on a full domestic season of the same opponents played over and over. Eight matches a team, however well the leagues are bridged, is still a small sample, and any site pricing the competition should say how it handles that.

FootInsights publishes a probability for every Champions League fixture it covers and settles the competition on its own line of the track record, so a prediction for a league phase match is never quietly averaged in with domestic results. How the model bridges leagues is part of the method, and the method stays private; what it produces does not.

That separate settlement matters for honesty as much as engineering. A model’s design is a description of what it’s supposed to do; whether it actually does it is a separate, measurable question, and the only place that question gets answered is the public track record, scored match by match rather than argued for in a blog post.