What Existing Football Analytics Already Explain (and What They Don't)

Core Question

Modern football analytics has developed serious tools for measuring how attacks function as systems. Before proposing anything new, FSL needs to be precise about what those tools already do — and where they stop.

What the Metrics Already Do Well

Football analytics has moved well beyond counting shots and goals. The most sophisticated current tools attempt to measure the quality, structure, and efficiency of attacking systems as coordinated wholes, not just collections of individual actions.

Expected Goals (xG) estimates the probability of a shot becoming a goal given its characteristics at the moment of contact — location, angle, assist type, game state. It is the most widely used system-level metric and, within its scope, one of the most validated. Its primary object is the shot.

Possession Value Models — EPV, VAEP, xPossession — extend the logic of xG backwards through the possession. Every action — pass, carry, dribble — is assigned a value based on how much it increases the probability of scoring. These models capture the flow of an attack rather than just its terminal event, showing which teams progress the ball into dangerous zones most effectively and which players contribute most to that progression.

Expected Threat (xThreat) measures how much danger a team creates by moving the ball into threatening areas, independent of whether a shot results — particularly useful for evaluating possession-heavy or positional systems that generate territorial pressure without always shooting.

Packing quantifies how many opponents a pass or carry eliminates from the play. High packing values indicate a system that breaks defensive lines effectively, and are especially useful for evaluating vertical attacks and transition-heavy teams.

Field Tilt measures a team's share of final-third passes relative to the opponent, capturing sustained territorial dominance — whether a team consistently operates in the attacking third even against organised defences.

Passing Networks map the structural connectivity of a team's play — which players combine most frequently, whether stable triangles and overloads form, how the team maintains width, depth, and central occupation — revealing something close to architectural properties of attacking organisation.

Taken together, these tools represent a serious and increasingly sophisticated attempt to describe not just what happened but how a team created it. Anyone proposing an additional layer of analysis needs to engage with them honestly, not treat them as a foil.

What They Share

Despite their differences, these metrics share a common underlying architecture: they are all action-aggregation models. Each assigns value to individual events — shots, passes, carries, ball movements — and builds a picture of system behaviour by accumulating those values. The system is inferred from the sum of its actions.

This is a powerful approach. It is also a specific one. It means the primary object of analysis is always the action, weighted by its relationship to an outcome. The question each metric is implicitly asking is: how much did this action contribute to the probability of scoring?

That is a well-formed and answerable question. But it is not the only question worth asking.

Where They Stop

The gap is easiest to see through a single sharp case: two possessions producing shots of identical xG can look comparable across every metric above — similar packing, similar accumulated xThreat — while the organisational conditions producing them were entirely different: one from sustained shared understanding, the other from momentary defensive disorder the team happened to exploit. "Same xG. Different Football." works through that case and its testable hypothesis in full; the point worth establishing here is the more general one it's a specific instance of.

None of the tools above are explicitly designed to answer: what was the organisational state of the team when those actions occurred? Not the value of the actions, not their relationship to outcomes, but the condition of the system itself — whether it was operating with coherence and intent, or producing similar-looking actions from a degraded or reactive state.

This distinction matters because process and outcome can come apart. A team can generate high-value actions from an unstable organisational state. If that state isn't reproducible — if it depended on the opponent's disorganisation rather than the team's own structure — the actions it produced may not recur reliably, even if their measured value was high. Existing metrics capture the value of the actions a system produces. They are not explicitly designed to classify the organisational condition of the system itself.

What This Means for FSL

FSL is not proposing to replace these tools. It is proposing to describe something they do not currently represent: the behavioural state of the attacking system as a primary object of analysis, independent of the value of the actions that state produces.

If that layer can be reliably identified and coded — if observers can independently classify whether a team is in a stable, high-initiative state or a degraded, reactive one — it becomes possible to ask questions existing metrics cannot yet answer:

Do high-agency, high-clarity states produce more repeatable attacking structures than equivalent-value states arising from disrupted conditions? (This is the specific hypothesis "Same xG. Different Football." sets out to test.)

Do teams that sustain stable organisational states across a season show lower variance in their attacking output than teams whose metrics are driven by opportunistic transitions?

Can the organisational state of the system at the moment of a chance predict something about that chance's repeatability that xG, xThreat, and possession value do not?

These are empirical questions. They may have uninteresting answers. Existing metrics may already capture the relevant variation, and FSL state classification may add nothing beyond what packing or possession value already explains. That would be a meaningful finding — it would mean the action-aggregation approach is sufficient, and that classifying organisational state independently adds no resolution.

But it is not yet known. Until it is tested, it's worth being precise: existing analytics describe coordinated attacking behaviour through the lens of action value. FSL is attempting to describe the organisational state that governs those actions. Those are different objects. Whether the difference is scientifically useful remains an open question.

Alternative Explanations

Organisational state is already implicit in existing metrics. Teams with high packing, high xThreat, and stable passing networks are, by definition, in high-agency states — in which case FSL is naming something the numbers already capture, and adds vocabulary without adding resolution. This is the account that most needs ruling out: if it holds, the entire project of a separate state-classification layer is redundant regardless of how conceptually clean the distinction is, because everything FSL would code is already recoverable from metrics that exist today.

The distinction between action value and organisational state is philosophically interesting but practically irrelevant. Coaches and analysts can make better decisions using existing tools without needing a separate state-classification layer at all — a claim not about whether the distinction is real, but about whether it's worth the operational cost of maintaining it.

Reliable coding of organisational state may prove impossible. If independent coders cannot agree on whether a team is in Command or Drift, the language cannot be operationalised regardless of its conceptual merits — a practical failure mode that would sink the project even if the underlying distinction is genuine.

What Evidence Would This Need?

Show that FSL state classifications correlate with but are not reducible to existing metrics. If High Agency × High Clarity possessions have systematically higher packing and xThreat values, that's expected — but FSL must show it explains variance that those metrics do not, which is precisely the test "Same xG. Different Football." lays out in detail.

Demonstrate inter-rater reliability in state coding above a threshold sufficient for scientific use. Without this, the language cannot be distinguished from subjective interpretation — this is the test that would settle the "reliable coding may prove impossible" objection above.

Identify at least one empirical question that FSL state classification answers and that existing metrics cannot, and test it directly.

Potential Implications

If existing metrics already capture organisational state implicitly, FSL's contribution is primarily terminological — a vocabulary for concepts the numbers already represent. That has some value but limited scientific novelty.

If FSL state classification adds explanatory power beyond existing metrics, it suggests that action value and organisational state are genuinely separable dimensions — and that football analytics currently has language for one but not the other.

For the field: the question is not whether to use xG, xThreat, or possession value models. They work. The question is whether there is a layer of football organisation those models do not yet explicitly represent — and whether naming and coding that layer produces knowledge the existing tools cannot. That is what FSL is trying to find out. Whether that layer exists is now an empirical question rather than a conceptual one.