Provider Models Compared in the football data vault

Two companies dominate the public supply of advanced football data, and their models do not produce identical numbers. Understanding where they differ is the only way to read a figure that arrived without its definition.

Two pipelines, one match

The two principal providers of advanced football data both watch the same match and both publish event-level records and a shot-value model. Their figures for a given match are close, usually within a few tenths of a goal, and they are not identical, because each has made a long series of definitions that the other has made differently.

The difference is not a question of competence. Each provider has built its model to serve particular clients and to answer particular questions, and those choices propagate into the numbers. Reading either figure without knowing which pipeline produced it is reading a measurement whose instrument is unknown.

Where the definitions diverge

The recurring divergences are consistent. Providers differ on whether a blocked shot counts as an attempt, on how a deflection is attributed, on whether a cross that drifts goalwards becomes a shot, and on where a shot taken from the edge of the area is recorded as having been taken. Each of those choices moves a small number of events, and small numbers of events move a total.

The divergences compound over a season because they are systematic rather than random. A provider that counts blocked shots as attempts will report higher shot volumes for every side in the league, and the bias does not cancel out. That is why mixing providers inside one table is a genuine error rather than a tolerable approximation.

The shot-value models

Both providers publish a shot-value model, and both build it from the same broad features: location, angle, body part and the characteristics of the preceding pass. The difference is in the richness of the features and in the underlying data that feeds them.

One pipeline is built on a long historical event archive with a very large training set, which favours stability and comparability over decades. The other is built around a finer-grained event taxonomy and positional data, which favours detail about pressure and defensive structure. Neither is plainly superior; they are optimised for different questions.

The event taxonomy difference

The two providers do not use the same event vocabulary. One distinguishes a larger number of pass types, including the height and the intent of the delivery, and records defensive pressure as an event in its own right. The other uses a more compact taxonomy with a very long historical record behind it.

That difference has a direct analytical consequence. Metrics that need fine distinctions about service type or pressure are computable on the richer taxonomy and have to be approximated on the more compact one. Comparing a metric across providers therefore requires checking whether both were capable of computing it as defined, which is a step that is almost always skipped.

Which to use for what

Historical comparison over many decades is best served by the provider with the longest consistent archive, because the value of the number lies in its comparability with the past. Analysis of defensive pressure, positioning and the geometry of defensive lines is best served by the pipeline built on positional data.

The practical rule for a reference site is to pick one provider for a given table and to state it. Where two providers are compared deliberately, the comparison should be of the two models rather than of the two numbers, because the numbers differ for reasons that only the model descriptions explain.

Reporting provider differences

This site names the providers when explaining how a model works and never links to them. Where a figure is presented, the caption says which family of model produced it and, where relevant, notes that the other provider reports a slightly different value for the same events.

That practice is more informative than choosing one number and presenting it as settled. Two credible totals for the same match are not a scandal; they are the expected consequence of two careful organisations defining a complex game differently, and stating that is part of describing the measurement honestly.

Key reference points

• Two dominant providers both publish event data and shot-value models.
• Their figures agree closely and never exactly, because their definitions differ.
• One archive favours historical comparability; the other favours positional detail.
• Divergences are systematic, so mixing providers inside a table introduces a bias.
• Check whether a metric can even be computed under a given taxonomy before comparing.
• Pick one provider per table, state it, and compare models rather than numbers.

A provider is an instrument, and a measurement quoted without its instrument is a number whose properties are unknown, however plausible it looks.