Football Possession Value and Expected Goals
What is expected possession value in football?
Expected possession value, or EPV, measures how close a team is to scoring the next goal from the exact spot where it has the ball. Given the position on the pitch, the clock and the situation, how much has the team in possession improved its chance of scoring before the half ends? It runs roughly from -1 to +1. At +1 a goal for the team in possession is effectively locked in, at -1 the opponent’s next goal is, and at 0 the situation is neutral.
Every tracked action (pass, shot, tackle, dribble, clearance) starts from an EPV value. The change from one action to the next is the value that action created or destroyed. We share each change among the players involved to build a running ledger of who is creating value and who is losing it.
How is EPV calculated?
EPV comes from an XGBoost model, which combines many small decision trees. It reads 14 features describing the state just before each action and returns a value for the rest of the half. It is trained on 15 leagues from 2020-21 to 2023-24.
The model works at the level of single events, which we convert into a common format (a standard way of describing every action as a start, an end, a player and a result) before scoring.
Credit is not simply the change in EPV. We apply rules so that the numbers behave sensibly:
- Shots take the shot’s own xG as their value, so a striker is not blamed by the model’s general prior that most shots miss. A goal earns 1 minus the xG, and a miss loses the xG.
- Turnovers split blame and credit 50/50 between the player who lost the ball and the player who won it.
- Passes split credit between passer and receiver, and a difficult pass that succeeds credits the passer more than an easy lay-off. We estimate difficulty with a separate pass-completion model, xPass.
- Defensive actions (clearances, interceptions, tackles and ball recoveries) get 1.5 times the credit, because the raw change chronically undervalues stopping an attack.
- Passes from deep are scaled down so routine possession in our own half cannot out-credit a creative pass in a dangerous area.
- Duels are zero-sum: the winner gets half the change and the loser loses half.
These multipliers are judgements we set by hand, not fitted values.
Per-player totals are available per match and per season. They feed the EPR rating described in football ratings.
What are xG and xGOT?
xG (expected goals) is the chance a shot becomes a goal, judged only from where it was taken, how it was struck and what kind of chance it was. It is a probability between 0 and 1. A tap-in from a yard out is close to 1, and a speculative 30-yard strike is close to 0.
Our xG model is an XGBoost classifier with 14 inputs: location, distance and angle to goal, whether the shot was in the penalty area or six-yard box, header or right or left foot, open play, set piece, corner or direct free kick, and whether the data provider flagged it a big chance. It is trained on shots from 15 leagues, 2020 to 2024.
Penalties are a special case. A spot kick is a different kind of event from an open-play shot, so the model never sees one in training. Instead every penalty is given a fixed xG of 0.80. In English top-flight data from 2021 to 2024, 251 of 306 penalties were scored (0.82), and the long-run figure for top leagues is about 0.78, so 0.80 sits between them.
xGOT (post-shot expected goals) answers a narrower question: given that a shot was on target and crossed the goal line at a particular point in the frame, how likely was that placement to beat the keeper? It uses the xG inputs plus where the ball crossed the line. Only on-target shots have an xGOT. We train it on 2021-22 onward only, because goalmouth coordinates were recorded for only about 55% of shots before then.
The two numbers give a useful split. Goals minus xG equals placement (xGOT minus xG, the shooter finding corners) plus goals minus xGOT, which is the keeper’s work plus luck. We treat placement as the finishing skill we can measure and goals minus xGOT as not shooter skill.
What is win probability added?
Win probability added (WPA) measures how much an action moved the chance that the acting team wins the match. A pass that keeps the ball and nudges confidence up one percentage point is worth +0.01. A goal is worth a large swing, often 0.2 to 0.5 or more.
The win-probability model looks at the score, the expected-goal difference, time remaining, red cards and home advantage, and predicts the chance that the team currently in possession goes on to win. A draw counts as half a win. WPA is centred so it sums to roughly zero over a match, so it is a ledger of who moved the needle, not a quality score. It measures the stakes of this particular match, not repeatable skill, so late actions in close games swing hardest. It is calculated after the match is complete and is not the live win-probability bar.
Where can I see it?
- Match chains: every possession chain in a match, with the players involved.
- Match events: per-action EPV credit, shot maps and match totals.
- Match page: xG for each shot, calculated live during a game, and live win probability.
- Football stat definitions: the glossary.
What are the limits?
- EPV and xG only know what the event data records. Defensive pressure, off-ball movement and the quality of the keeper are invisible.
- The xG model does not know about penalties by design. A shot taken from the penalty spot gets the fixed 0.80 value, whatever the situation.
- Goals minus xGOT mixes keeper quality with luck. Do not use it to rate strikers.
- Placement and the other xGOT-based figures can lag by a few days after a match, because their source data is rebuilt weekly.
- Credit rules, including the 1.5 defensive multiplier, are our judgement and can be argued with.
How accurate is it?
We score each new shot model on shots it never saw in training, using log-loss (how surprised the model is by which shots went in; lower is better). On 247,403 shots from the 2025-26 season, the xG model introduced in September 2026 scored 0.2476 in testing, against 0.2511 for the model it replaced. The matching xGOT model scored 0.390 against 0.403 for its predecessor on a season neither was trained on. The live win probability was checked on 1,747 games from the 2024-25 season in the big five leagues: on average its stated chances sat 1.3 percentage points from how often those outcomes actually happened.