Same xG. Different Football.

Core Question

Same xG. Possibly very different football. Does that difference explain why some chances keep coming back and others don't?

Exploring the Question Through FSL

xG estimates the probability of a shot becoming a goal given its characteristics at the moment of contact. It is very good at what it does.

But it does not explain how the chance arrived.

A 0.30 xG shot produced through sustained positional control, stable spacing, and coordinated progression looks identical in the model to a 0.30 xG shot from a scrambled clearance, a fortunate deflection, and a recovering defence. Same event. Possibly very different behaviour.

FSL is interested in that difference — not to replace xG, but to describe the layer beneath it.

This suggests a different way of organising what football analysis measures, and it's worth stating explicitly as the ontology the rest of FSL is built on, not just a device for this piece:

Outcome — goal or miss.

Event — the shot, with its location, angle, and probability.

Behaviour — the coordinated sequence that produced the shot.

State — the underlying behavioural organisation governing that sequence.

FSL therefore studies football one layer earlier than event-based analytics.

This is not a minor distinction. Models predict. Ontologies organise. FSL is not attempting to produce a better predictive metric. It is attempting to define the behavioural entities — states, forces, transitions — that football actions arise from. Without agreed behavioural entities, researchers may measure different phenomena while using the same terminology. An ontology provides a shared vocabulary, allowing observations, coding, and models to refer to the same underlying concepts. Establishing what those entities are is a precondition for modelling them. Biology had taxonomies before predictive models. FSL is trying to do something similar: define the objects before measuring them.

Existing system metrics — EPV, VAEP, xThreat, packing, passing networks — already describe coordinated attacking behaviour through different methods. FSL attempts to classify organisational state itself.

The question FSL wants to ask is therefore not: did it go in?

It is: did the attacking system reliably recreate the conditions?

Pep Guardiola rarely says we need to finish better. He says we need to keep creating those situations. The implicit claim is that process repeats and outcomes fluctuate — and that controlling the process is therefore the more durable objective. FSL is an attempt to formalise what that process actually consists of.

Here, behavioural variance refers to variation in the organisation of possessions leading to comparable events, rather than variation in outcomes. Two possessions can produce shots of identical xG while differing substantially in spacing, option availability, decision pressure, and transition stability. If those differences are systematically related to FSL state classifications, the language is identifying something that event-based models currently treat as invisible.

This produces a testable hypothesis:

Possessions independently classified as High Agency × High Clarity will exhibit greater behavioural repeatability than possessions of equivalent xG generated from degraded or unstable behavioural states.

The coding unit matters here. FSL should not code shots. It should code possessions — tracking initial state, transitions, Forces observed, and terminal state, with the shot as one possible terminal event. That is consistent with the core FSL claim that finishing is an Outcome, not a State.

Alternative Explanations

FSL is simply redescribing possession quality. Structured build-up play produces lower variance by definition, and High Agency × High Clarity is just a new label for that.

Repeatability differences are better explained by team quality, opposition strength, or tactical setup — none of which require a behavioural ontology.

Existing system metrics may already capture much of what FSL describes. The language may be adding vocabulary without adding resolution.

The counterattack problem: a fast transition may reach the same terminal state as a longer structured move — equally high-agency, equally high-clarity, arrived at differently. If FSL cannot code both as such, it is measuring style rather than state. This is probably the hardest objection the language faces, and the coding manual has to answer it before the research begins.

What Evidence Would This Need?

Code possessions — not shots — from a single team across a full season. For each possession, record initial state, observed Forces, state transitions, and terminal event.

For all shot attempts above a minimum xG threshold, compare behavioural indicators across FSL state classifications: time between last defensive action and contact, number of viable options at the moment of the attempt, pressing intensity faced during the sequence.

Use two independent coders working without access to xG values. Establish inter-rater reliability before analysis.

Show that FSL state classification predicts behavioural variance over and above existing possession metrics. If it does not add explanatory power beyond territory, pressing intensity, or chance quality scores, the language is not yet earning its complexity.

Test whether the result holds for counterattacks as well as structured possession. If FSL only identifies high-agency states in slow build-up play, the coding scheme is capturing style preference, not behavioural organisation.

Open Question

Are Agency and Clarity real but inferred properties — like confidence or composure — or are they constructed directly from measurable indicators, like xG itself? The answer shapes how coding decisions are justified and how critics can challenge them. FSL does not need to resolve this before Phase 1. But it cannot avoid it indefinitely.

Potential Implications

If supported, the hypothesis would show that behavioural organisation and shot quality are separable dimensions — and that FSL is describing the former in a way that existing metrics do not.

For coaches: the question shifts from how do we finish better to how do we reliably recreate the conditions from which finishing naturally follows.

For analysts: a team can generate high-xG chances from low-agency states. If those chances are less reproducible, that matters for how performance is evaluated and how variance in results is interpreted.

For recruitment: a player who sustains High Agency × High Clarity possessions consistently may be more valuable than their goal or xG contribution suggests — because they are maintaining the process rather than simply benefiting from it.

FSL succeeds or fails not by predicting goals more accurately than existing models, but by demonstrating that behavioural organisation is a distinct and measurable layer of football that current event-based frameworks do not explicitly represent.