Here's an experiment that produces uncomfortable results. Take a single match, take the event data from two different providers, and compare the totals. Passes, tackles, duels, take-ons. The numbers won't match. Depending on the metric, they can differ by 10 or 20%.

Both providers are competent. Neither made mistakes in any ordinary sense. The disagreement is definitional, and it goes much deeper into football analytics than most people using the data realise.

What's a tackle

Start with the simplest case. A defender makes contact with the ball while an opponent is in possession, and the ball goes out of play off the attacker.

Is that a tackle? Some definitions require the tackler's team to gain possession. Others count any dispossession. Others require the defender to have made a deliberate attempt rather than a block. A ball that ricochets off a standing leg might be a tackle, a clearance, an interception, or nothing at all, depending on whose rulebook you're reading.

And this isn't an edge case — it's a large fraction of defensive actions. Football is mostly ambiguous contact.

What's a pass

Seems obvious. A player deliberately plays the ball to a teammate.

Now: a defender under pressure hooks the ball forward. It reaches a teammate. Is that a pass or a clearance? The distinction is intent, and intent isn't observable. Different providers resolve this differently and it materially affects pass completion percentages, particularly for defenders in teams that play direct.

Or: a player is fouled while playing the ball, and the ball reaches a teammate. Pass, or not, since play was stopped? A player intercepts and immediately plays it to a teammate in one motion — is that one event or two?

Each of these is a policy decision made by a data provider, documented somewhere in a specification document that essentially nobody reads.

Why this matters more than it seems

The obvious consequence is that you can't mix data sources. Fine — analysts know that.

The less obvious consequence is that any model built on event data inherits its definitions, and therefore its judgements. An expected-goals model trained on one provider's shot data encodes that provider's definition of a shot, including whether blocked efforts count and how they classify deflected balls.

Move the model to another provider's data and it will produce different outputs, not because the football changed but because the ontology did. This is a genuine problem in the literature — results that don't replicate across data sources, sometimes without the authors realising why.

It also affects the comparisons that get published. A player's ranking on a defensive metric depends on how his league's data was collected, which affects cross-league comparisons in ways that are hard to correct for.

The context collapse

A deeper issue. Event data records what happened at a point in space and time. It doesn't record why, or what was available, or what the alternatives were.

A sideways pass in midfield is recorded identically whether it was the only option available under heavy pressure or a lazy choice with three forward passes on. Those are completely different actions and the data can't distinguish them.

This is the fundamental limitation and it's why tracking data — the positions of all 22 players continuously — is such a significant advance. With tracking, you can reconstruct what options existed, which turns "he played sideways" into "he played sideways when a line-breaking pass was available to a free teammate."

That's the difference between recording football and understanding it.

Attempts at fixing it

There have been standardisation efforts, with mixed success. The obstacles are partly technical and mostly commercial.

Providers compete on their data. A proprietary event taxonomy with proprietary derived metrics is a product differentiator and a lock-in mechanism. There's no strong incentive to converge on a shared standard that would make customers interchangeable.

There's also a genuine argument that different use cases want different definitions. A broadcaster wants tackle counts that match what a viewer thinks they saw. An analytics department wants a definition that's consistent and predictive, even if it feels wrong. Those aren't reconcilable into one number.

The practical advice

For anybody working with football data, a few habits that save trouble.

Read the specification document for your provider. Actually read it. Most of the surprises in an analysis turn out to be documented behaviour that nobody checked.

Never mix sources within an analysis, and be explicit about which source produced any number you publish.

Be sceptical of any metric whose definition you can't state precisely. If you can't say what counts as a progressive pass in the dataset you're using, you don't know what you're measuring.

And treat cross-league or cross-era comparisons as approximate at best. The data collection changed, the definitions changed, and the differences are large enough to swamp the effects people usually claim to find.

None of which is a counsel of despair. Event data is enormously useful and football analytics has produced real insight. It's just worth knowing that the foundation is a set of human judgements about an ambiguous sport, rather than a measurement of physical reality.