What an in-play win probability is actually modelling

A statement about situations, not about teams

An in-play probability answers a narrow question, which is how often matches reaching a comparable state have finished in each possible way. It does that by locating the current state within a reference set of previous situations and reporting the distribution of their eventual outcomes. Because the reference set is built from many matches, the number describes a class of situations rather than the specific one being watched.

Anything unusual about this particular contest is therefore absent from the estimate unless the model has an input that captures it explicitly. This is the source of the common complaint that the figure fails to reflect what a viewer can plainly see happening.

The inputs are coarser than they look

Most published models take a small number of inputs, typically the state of the score, the time remaining, and some measure of relative strength. Additional inputs can be added, but each one fragments the reference set and reduces the number of comparable historical situations available to estimate from. That trade-off between richness and sample size is the central design constraint, and it usually resolves in favour of fewer inputs.

As a consequence the model cannot see momentum, personnel changes or conditions unless those have been deliberately encoded as features. A coarse model is not a bad model, but it is answering a coarser question than the confident presentation of a single percentage suggests.

Why it moves the way it does

Sharp movements in the number usually accompany changes to whichever input the model weights most heavily, which is almost always the score state. Between such changes the figure drifts steadily as time remaining shrinks, since the same deficit becomes harder to overturn as opportunity contracts. That drift is a structural feature of the model rather than a reading of the play, and it continues regardless of what is happening on the field.

Viewers interpreting drift as an assessment of performance are reading the passage of time and mistaking it for analysis. Understanding which movements are structural and which are informational is most of what it takes to read the number sensibly.

Calibration is the honest test

The useful question about any probability model is whether situations it assigns a given likelihood actually resolve that way at roughly that rate. A well-calibrated model can look wrong on any individual match while being right across the full set of matches it has ever assessed. That property makes single-match criticism close to meaningless, since one outcome cannot distinguish a good model from a poor one.

It also means published models should be judged over long periods, which is rarely how audiences or broadcasters actually engage with them. Providers who publish calibration behaviour are making a claim that can be checked, and that transparency is itself informative.

What it cannot be used for

A win probability is not a rating of the teams, since two evenly matched sides in an unbalanced state will produce a lopsided number. It is also not a measure of how the match is being played, because it responds only to inputs and is blind to everything else. Using it to argue that a side deserves more than the situation reflects confuses an outcome estimate with a performance assessment.

The appropriate use is narrow, which is to quantify how much a particular state constrains the range of endings still available. Held to that narrow use, it is genuinely informative, and stretched beyond it, it produces exactly the frustration viewers routinely express.