Expected goals gets all the attention, but among people who actually work with football data, the more useful tools are usually possession value models. They go by various names — expected threat, expected possession value, VAEP and others — and they all attempt the same thing.
The core idea is simple enough to explain without any maths at all.
The basic idea
Every location on the pitch has a value, defined as the probability that a possession starting from there ends in a goal.
The centre circle is worth very little. The penalty spot is worth a lot. The corner of your own six-yard box is worth close to nothing, and might even be negative if you account for the chance of conceding.
Now, any action that moves the ball from one location to another changes the value of the possession. A pass from the halfway line into the opposition box moves the ball from a low-value area to a high-value one. The difference between those two values is the credit that action earns.
Do this for every action in every match, add it up per player, and you have a measure of how much each player contributed to moving his team towards scoring — regardless of whether a goal actually resulted.
Why it beats counting
The advantage over traditional statistics is that it weights actions by their consequence rather than counting them equally.
A defender making a five-yard square pass and a midfielder threading a ball through a defensive line both get one "pass completed" in a conventional stat sheet. In a possession value framework, one is worth almost nothing and the other is worth a lot.
It also captures contributions that end in nothing. A brilliant pass that puts a striker through, who then shoots wide, produces no goal, no assist, and no entry in any traditional record. In a possession value model, the passer gets full credit for the value he created; the striker gets penalised for destroying it.
Over a season this makes the metric far more stable than goals or assists, because it's accumulating many small signals rather than a handful of rare events.
The variants
The models differ mainly in what they consider and how they assign credit.
Expected threat in its common form is grid-based. Divide the pitch into zones, compute each zone's value from historical data, and credit any action by the change in zone value. Simple, transparent, easy to implement, and it ignores everything about the situation except location.
VAEP and similar are richer. They consider the game state — the sequence of preceding actions, the score, the time — and they explicitly model both the probability of scoring and the probability of conceding from a given state. This means defensive actions get valued properly, which grid-based approaches struggle with.
Tracking-based models use the positions of all players, so the value of a location depends on where the defenders are, not just where the ball is. These are the most sophisticated and the most data-hungry.
Where they break
Several important limitations, and practitioners are generally quite open about them.
Credit assignment is arbitrary. If a possession involves eight actions and ends in a goal, how should the credit be split? Every model makes a choice, and the choices are defensible rather than correct. Different reasonable choices produce meaningfully different player rankings.
Off-ball contribution is invisible in event-based versions. The run that created the space for the pass gets nothing. This is the biggest gap and it systematically undervalues certain kinds of player — particularly forwards whose main job is to occupy defenders.
They're descriptive, not causal. A player with high possession value added moved the ball into dangerous areas. Whether he would do that in a different team, with different teammates and different instructions, is a completely separate question the model doesn't address.
They reward volume in possession-dominant teams. A midfielder who touches the ball a hundred times has a hundred opportunities to add value. One who touches it forty times has forty. Per-touch normalisation helps but introduces its own distortions.
How to read the outputs
A few practical notes for anyone encountering these numbers.
They're best used comparatively within a role. Comparing a centre-back's possession value added to a winger's is close to meaningless, since the two operate in areas with completely different baseline values.
They're most informative in aggregate over a season and least informative for a single match.
Negative values are normal and don't indicate a bad player. Anyone who takes risks will destroy value regularly. A player with zero negative actions has probably never attempted anything difficult.
And the model matters. "Expected threat" from one provider and from another are different numbers with the same name, built on different data with different assumptions. As with everything else in this field, the label is less informative than it looks.
Why I like them anyway
Despite the caveats, this framework changed how I look at matches, and I think that's the strongest recommendation available.
Once you start thinking about football as a continuous process of moving a possession up and down a value gradient, a lot of things become clearer. Why sideways passing is sterile. Why a ball carrier who beats a man is so valuable. Why the pass before the assist often matters more than the assist.
None of those insights require the model. But the model makes them precise, and precision is how arguments get settled.