The Ingredients of a Shot Value: definition and limits

A single expected-goals figure is the product of a short list of ingredients, each of which can be inspected. Learning the list turns an opaque decimal into a statement about the chance that can be argued with.

Geometry first

Two coordinates describe most of a shot: how far from goal it was struck and where the shooter stood in relation to the posts. Distance acts almost exponentially, so the difference between the six-yard box and the edge of the area is enormous, while the difference between two points a metre apart near the halfway line is nil. Angle then modulates that distance by how much of the goal the shooter can actually see.

The combination explains why the penalty spot is worth roughly three-quarters of a goal before anything else is considered, why a shot from the corner of the box is worth less than a tenth of that, and why the phrase a good position is meaningless in analytics without both numbers attached.

The body part and the finish

Headers are discounted heavily, because heading is a lower-percentage skill than striking with the foot and because a headed chance is usually contested. Volleys and first-time strikes from awkward heights carry a smaller discount, while a shot taken after a controlling touch is valued slightly lower than one struck first time inside the same geometry for a different reason: the extra touch removes the goalkeeper's hesitation.

The model does not know the identity of the finisher, so these adjustments describe the population, not the individual. A specialist header of the ball will beat the population value consistently, which is one legitimate way a player can post a career-long positive residual against expectation.

The service that created it

How the ball arrived is the richest part of the feature set. A cut-back played across the face of goal produces a shot at a keeper who is moving the wrong way; a low cross into the six-yard area produces a near-tap-in; a lofted pass into a crowded box produces a contested header. Each carries a different value for the same coordinates, and providers that log pass height and speed capture that difference.

This is the ingredient that most rewards tracking data. Without an idea of the ball's path and the defender's position, a provider can only infer service type from the event itself, and the resulting values are coarser rather than wrong. Comparing a tracking-data model with a video-only model is therefore a comparison of resolution, not of accuracy.

Pressure, and what it cannot see

Immediate pressure from a defender changes the value of a shot more than any other non-geometric ingredient, because it affects the time and space available to strike cleanly. Only positional tracking can quantify it, and even then the model must decide how to weight a defender who is close but behind the shooter against one who is further away but on the shooting line.

The honest position is that pressure is the least settled ingredient in public models. Two providers can agree on geometry, body part and service and still disagree on a shot because they treat a recovering defender differently, and that disagreement propagates into season totals in a way that is invisible to anyone reading a single headline figure.

Sequence and game state

A shot is rarely the first act of a possession, and the model can be told how the sequence began. A move that started with a turnover high up the pitch and reached the shot in two passes is valued differently from one that began with a goalkeeper and took twelve passes, because the defence had less time to organise in the first case even though the shot itself looks identical.

Game state enters as a bias rather than a feature. A team behind by two goals takes more low-value shots, which raises its expected-goals total while lowering its quality per attempt. Reading the total without the count of shots is how a chasing performance is mistaken for a dominant one, and the two lines of the shot table are always more informative than the total alone.

Reading a value you disagree with

When a value looks wrong, the productive response is to locate the ingredient that produced it. A header valued near a quarter of a goal is probably being treated as uncontested; a shot from distance valued above the typical range is probably being credited with a deflection or a goalkeeper out of position. Naming the ingredient turns an argument about the model into a specific claim that can be checked.

That habit also explains why the same shot can carry two defensible values from two providers. Neither figure is a measurement of a natural constant, and treating either as ground truth is a category error. The number is an estimate built from named inputs, and the inputs can always be inspected.

Key reference points

• Distance and angle explain most of a shot value before any other ingredient.
• Body-part discounts describe the population, not the individual finisher.
• Service type is the ingredient that most rewards tracking data.
• Pressure is the least settled ingredient across public models.
• Game state biases totals upward, so always read the shot count beside the total.
• When a value looks wrong, name the ingredient rather than blaming the model.

A shot value is a short argument made from named premises, and it can be disputed premise by premise rather than accepted or rejected as a whole.