← All explorations

Machine Learning in Football · 01

The shape of the football world

Treat every nation as a node and every match as an edge, then read 150 years of results as a single graph. Its structure explains why ranking national teams — and seeding a 48-team World Cup — is genuinely hard: the world barely plays itself, it has only grudgingly come together, and it hangs on a handful of teams.

Part I

Six blocs, barely touching

How the six blocs connect

Each arc is a confederation, sized by how much it plays outside itself. Ribbons are cross-continental match flow — the only games that link the world. Hover an arc to isolate its links.

One ribbon dwarfs the rest: CONMEBOL↔CONCACAF. South and North/Central America play each other so much they barely read as separate.

Where every game goes

The same data, exact. Each row is a confederation; cells show the share of its games played against each bloc. The diagonal — games kept at home — dominates every row.

Read along a row: the bright diagonal is insularity. The only warm off-diagonal cell is CONMEBOL → CONCACAF — the Americas link again.

How inward-looking each confederation is

Share of games a bloc plays within its own ranks. Big self-sufficient confederations almost never look outward.

CONMEBOL is the outlier — ten teams forced to travel for games. CAF and UEFA, with dozens of members, stay home ~90% of the time.

Part II

Is the world coming together?

A century of globalization — then a retreat

Of all internationals played each year, what share crossed continents? Intercontinental play climbed for a hundred years… and then went into reverse.

Cross-continental play peaked at 18% in the 1990s, then fell back toward 12% — as confederations built out their own packed calendars (continental Nations Leagues, more regional qualifiers), friendlies abroad got crowded out.

Part III

Two degrees of football

Everyone is two matches from everyone

For all the silos, the graph is a small world: short paths, tight clustering. Pick any two nations and count the matches between them.

Clustering against for a random graph of the same size — high clustering with short paths is the textbook small-world signature.

Connect any two nations

The shortest chain of real matches linking two teams. Even the most isolated sides are only a hop or two from the giants.

Part IV

Who really holds it together

Brokers aren't the same as travelers

Cross-share asks who plays abroad most. Betweenness asks who sits on the most shortest paths between blocs — the teams the world routes through. They disagree.

Mexico brokers and travels. But Qatar, Ghana and the Central American sides are high-leverage brokers despite modest travel — and the densely-connected core of world football, by eigenvector centrality, is the Americas, not Europe.

Part V

Structure at every scale

Zoom in and the blocs keep splitting

Turn up the community-detection resolution and the six confederations fracture into recognizable sub-regions — the football world is structured all the way down.

Inner ring: confederations. Outer ring: the sub-regions Louvain finds at high resolution — Western Europe, the Caribbean, the Gulf, two African halves. Hover a wedge for its members.

Part VI

Robust world, heavy bridges

Pull out the brokers and watch the bridges thin

Remove teams worst-first by betweenness and track how many cross-continental match-links survive — against removing the same number at random.

The top 10 brokers carry a wildly disproportionate load — removing them strips a quarter of all cross-continental links, twice as fast as random. And yet the graph never breaks: every confederation stays reachable, so world connectivity is resilient — there is no single point of failure.