Expected goals values shots, and only shots. Expected assists values the one pass before a shot, and only that pass. But a football match is thousands of actions, and the overwhelming majority of them are neither shots nor the pass before a shot: the recovery in your own half, the carry that breaks a line, the tackle that ends the other team’s best move of the game. None of those touch the scoresheet, and none of them touch xG either. Possession-value models exist to put a number on all of them.
What xG and xA leave on the table
The expected-goals family is the best thing to happen to football measurement in a generation, but it has a hard boundary: it only ever speaks about chances. A shot gets an xG value; the action that created it gets an xA value; everything else gets nothing. That is fine if all you want to know is who finishes and who teed up the finish, but it leaves enormous swathes of the game unmeasured.
Consider a holding midfielder who, twice a game, intercepts a pass at the edge of his own box and immediately drives forty yards before laying the ball off. Those moves might end in a shot three passes later, or they might not. Under xG and xA, that midfielder’s most valuable contributions (the interception, the carry, the release) are simply invisible unless he happens to be the last man before the shot. The same blind spot hides the ball-playing centre-back, the press-resistant deep-lying playmaker, and the defender whose timing snuffs out danger before it becomes a chance. The box score has no column for "stopped a goal that never happened."
The fix is to stop measuring only the end of moves and start measuring every action by what it does to the probabilities.
The core idea: value = change in scoring probability
Possession-value models (you will also see them called on-ball value [OBV] or expected possession value [EPV]) share one simple premise. At any moment, the team in possession has some probability of scoring in the near future and some probability of conceding. Every action the players take nudges those probabilities up or down. The value of an action is exactly that nudge.
A progressive pass that moves the ball from your own third into the opponent’s half raises your scoring probability a little, so it earns a small positive value. A through-ball that splits the last line and sets a striker free raises it a lot, so it earns more. A carry into the box, a pass that retains possession under pressure, a clearance that ends a dangerous attack: each is scored by how much it shifted the probabilities. Give the ball away cheaply in your own half and you have handed the opposition a better position, so the action is scored negatively.
This is the same logic that drives expected threat (xT), which assigns every pitch location a scoring probability and credits passes and carries with the difference between where the ball ended and where it began. Possession value generalises that idea: instead of valuing only ball location, the richest models look at the full context (where the ball is, how the action moved it, what kind of action it was, even how much time has passed) and ask how the probabilities changed.
VAEP: valuing actions by estimating probabilities
The best-known public framework for this is VAEP, which stands for Valuing Actions by Estimating Probabilities. Its central refinement over a location-only model is that it scores every action on two probabilities, not one: the chance the team scores within the next few actions, and the chance it concedes within the next few actions.
The value of an action is the change in the scoring probability minus the change in the conceding probability. That second term is what makes VAEP fair to defenders and to risk. A flashy pass that splits two players but, if intercepted, would leave your defence horribly exposed is not purely good: it raises your scoring probability but also your conceding probability, and VAEP nets the two. A clean interception that ends an opposition attack lowers your conceding probability, so it scores positively even though no shot was ever taken and no progressive pass was ever made.
For each action, a model estimates two things from a large history of matches: Pscore (the probability the team in possession scores within the next k actions) and Pconcede (the probability it concedes within the next k actions). The value of the action is (ΔPscore) − (ΔPconcede): how much it raised the scoring probability, minus how much it raised the conceding probability. Sum every action a player performs and you have their total VAEP contribution: offensive value, defensive value, and the net.
Because every action gets a number, you can total a player’s VAEP over a match or a season and split it into the part that came from attacking actions and the part that came from defensive ones. That single property, a common currency for a tackle and a through-ball, is what makes possession value so useful for comparing players whose jobs look nothing alike.
A real move, valued action by action
To show how the credit lands, we fitted a deliberately simple possession-value model on the 2022 World Cup’s own events and used it to value one real move. It is our simplified version of the idea, not VAEP itself: VAEP feeds a machine-learned model with the type of each action, its result and the actions before it, while ours knows only where the ball is and which team has it, which makes it closer to the expected-threat end of the family.
We took the 143,492 actions with a location in the 64 matches’ event files from StatsBomb Open Data (StatsBomb Public Data User Agreement): passes, carries, dribbles, shots, interceptions, ball recoveries, clearances, blocks, duels and goalkeeper actions, with the shoot-outs left out. We cut StatsBomb’s 120-by-80 pitch into a 12-by-8 grid of 10-yard cells, oriented so the acting team always attacks to the right. For each cell, P(score) is the share of actions starting there after which the acting team scored within ten actions (this one and the next nine), and P(concede) is the share after which it conceded. The state after an action is the cell where the next action starts, read from the acting team’s side, so a pass that gives the ball away hands the opponent its numbers. An action’s value is the change in P(score) minus the change in P(concede). Shots are left to xG.
The fitted grid behaves as it should. Across all actions, the acting team scored within ten actions 1.01% of the time and conceded within ten 0.19% of the time. The most valuable cell is the 10-yard square on the goal line from y = 30 to 40, where P(score) is 0.193 (from 332 actions); 47 of the 48 cells in a team’s own half sit below 0.01. The highest P(concede), 0.011, is in the central cell in front of a team’s own goal.
The move is South Korea’s stoppage-time winner against Portugal in their 2-1 group win (match 3857262 in StatsBomb Open Data). Moon-Hwan Kim cleared the ball from South Korea’s penalty area, Heung-Min Son picked it up in his own half, carried it 50.8 yards upfield and passed to Hee-Chan Hwang, who scored with a shot worth 0.42 xG, 12.3 seconds after the clearance.
| Action | Player | P(score) before | P(score) after | P(concede) before | P(concede) after | Value |
|---|---|---|---|---|---|---|
| Clearance from his own penalty area | Moon-Hwan Kim | 0.004 | 0.005 | 0.011 | 0.001 | +0.011 |
| Ball recovery in his own half | Heung-Min Son | 0.005 | 0.005 | 0.001 | 0.001 | 0.000 |
| Carry, 50.8 yards upfield | Heung-Min Son | 0.005 | 0.036 | 0.001 | 0.000 | +0.032 |
| Pass to Hwang | Heung-Min Son | 0.036 | 0.147 | 0.000 | 0.000 | +0.111 |
| Shot, goal | Hee-Chan Hwang | 0.147 | Left to xG: a 0.42 xG chance, scored | |||
Read down the value column. Kim’s clearance is worth +0.011, almost all of it defensive: it took the ball out of the cell where a team is most likely to concede and gave South Korea the ball in midfield. Son’s recovery is worth 0.000, his carry +0.032 and his pass +0.111. The pass is the biggest single step, because it moved the ball from just outside the penalty area, at P(score) 0.036, into the cell where Hwang shot, at 0.147. Son’s three actions add up to +0.143, and the four valued actions to +0.154.
The zero for Son’s recovery is the clearest gap between our version and VAEP. Our model knows only where the ball is and who has it, so the action that wins the ball starts from its own team’s numbers and scores about nothing, while the swing in a change of possession lands on the action before it (here, Kim’s clearance; in a turnover, the pass that lost the ball). VAEP’s extra inputs let it pay the player who won the ball, and tell a pass under pressure from a free one. Treat the numbers above as the mechanics of possession value worked through on real events, not as player ratings. Every figure in this section, including the whole fitted grid, is in /data/stats/possession-value-vaep-explained.json, written by our analysis script.
Under xA, only Son’s pass would be credited, because it was the pass immediately before the shot. Under xG, only the shot counts. Possession value pays the clearance, the carry and the pass, each in proportion to how much it changed the probabilities. The progressive actions described in progressive passes and carries are exactly the kind of contributions this rewards, but possession value goes further by attaching a precise scoring-probability number to each one rather than just counting it.
A VAEP-style model on all 64 matches: what knowing the action adds
The model above knows only where the ball is. VAEP, as Tom Decroos, Lotte Bransen, Jan Van Haaren and Jesse Davis set it out in “Actions Speak Louder than Goals: Valuing Player Actions in Soccer” (KDD 2019), also knows what each action was and whether it worked. To see what that buys, we fitted a second model on the same 143,492 actions and the same ten-action labels, and changed one thing: the state after an action is now described by the action itself (its type, whether it worked, and the 10-yard cell where it left the ball), which gives 1,149 distinct states. It is still our simplified version. It counts frequencies rather than training a classifier, and it looks at one action, where VAEP looks at three. Rare states borrow strength in two steps: each state’s frequency is blended with the same action type and result in the same 20-yard band of the pitch, and that band with the same type and result anywhere, each with a weight of 20 actions. An action’s value then follows VAEP’s formula: the change in the acting team’s P(score) from the state before the action to the state after it, minus the change in its P(concede). Shots are again left to xG.
It forecasts better on matches it has not seen. We fitted both models on the 48 group matches and scored them on the 16 knockout matches (38,657 actions) with the Brier score, the mean squared error of a probability forecast. Against a forecast that gives every action the group-stage base rate of 0.96%, the location-only model improves the Brier score for “scores within ten actions” by 3.6%, and the action-aware model by 14.4%. For “concedes within ten actions”, a much rarer event (0.18% of actions), neither model does much: 0.3% and 0.5%.
| Action | Actions | Value per 100, location-only | Value per 100, action-aware | Valued positive, action-aware |
|---|---|---|---|---|
| Pass | 68,515 | 0.00 | 0.19 | 57.6% |
| Carry | 53,764 | 0.02 | 0.04 | 49.6% |
| Ball recovery | 5,388 | 0.00 | 1.50 | 81.3% |
| Duel (tackles and lost aerials) | 4,389 | −2.29 | 0.82 | 57.0% |
| Clearance | 2,684 | −1.37 | −0.18 | 26.7% |
| Block | 2,386 | −1.49 | −0.27 | 47.9% |
| Dribble | 1,793 | −0.93 | −0.26 | 38.9% |
| Goalkeeper action | 1,749 | −0.99 | 0.16 | 48.4% |
| Interception | 1,371 | −0.62 | −0.32 | 48.1% |
The credit moves to whoever wins the ball. Ball recoveries go from worth nothing on average under the location-only model to +1.50 per 100 under the action-aware one, and duels from −2.29 to +0.82: knowing that a team has just won the ball changes its prospects, and a model that sees only where the ball is cannot know that. Clearances and blocks are no longer heavily penalised. South Korea’s winner against Portugal, worked through above, shows it in one move. Heung-Min Son’s recovery, worth 0.000 under the location-only model, is worth +0.031; Moon-Hwan Kim’s clearance falls from +0.011 to −0.016, because the model now judges the clearance by what usually follows a clearance like it and gives the recovery that came next to Son. Son’s carry falls from +0.032 to +0.009, and his pass to Hwang is still the biggest step, +0.098 against +0.111.
| Player | Team | Minutes | Value per 90 | Attacking part | Defensive part | xA per 90 | xA rank |
|---|---|---|---|---|---|---|---|
| Piotr Zieliński | Poland | 362 | 0.437 | 0.394 | 0.043 | 0.11 | 16 |
| Antoine Griezmann | France | 613 | 0.432 | 0.367 | 0.065 | 0.37 | 1 |
| Lionel Messi | Argentina | 791 | 0.427 | 0.416 | 0.011 | 0.28 | 2 |
| Youssouf Sabaly | Senegal | 406 | 0.373 | 0.271 | 0.102 | 0.13 | 11 |
| Bruno Fernandes | Portugal | 396 | 0.366 | 0.324 | 0.042 | 0.27 | 4 |
| Ismaïla Sarr | Senegal | 384 | 0.362 | 0.321 | 0.041 | 0.06 | 35 |
| Ousmane Dembélé | France | 468 | 0.346 | 0.329 | 0.018 | 0.25 | 5 |
| Luka Modrić | Croatia | 705 | 0.296 | 0.251 | 0.045 | 0.08 | 27 |
| Theo Hernández | France | 576 | 0.284 | 0.241 | 0.043 | 0.15 | 7 |
| Young-Gwon Kim | South Korea | 387 | 0.272 | 0.169 | 0.103 | 0.05 | 40 |
It finds players xA misses, and misses one xA loves. Across the 94 players, value per 90 and xA per 90 correlate at r = 0.57, and only 5 of the top 10 by value are in the top 10 by xA. The names that xA ranks low are telling: Piotr Zieliński is 16th by xA, Ismaïla Sarr is 35th by xA, Luka Modrić 27th and the centre-back Young-Gwon Kim 40th, yet all four are in the top ten here for raising their team’s chance of scoring, or cutting its chance of conceding, without setting up many shots themselves. The reverse case is Harry Kane: third of the 94 by xA at 0.27 per 90, and last of the 94 by value at −0.029 per 90. The model does not see his shots, and the rest of his actions came out slightly negative; that is a reminder of what leaving shots to xG does to a centre-forward, not a judgement on Kane.
Limits. One action of context is not enough to judge every duel: a lost duel still averages +0.44 per 100 actions in this model (+0.86 when it comes straight after an opponent’s action), credit that a model reading the actions before it would not give. The player values are in-sample, because the model was fitted on the same 64 matches; 360 minutes is four full matches; and the conceding label is thin, since a team concedes within ten actions after only 0.19% of them. Treat the ranking as a description of one tournament under one simple model, not a verdict on players.
Reproduce it. Our analysis script fits both models, runs the group-stage test and computes every figure in this section, and draws the chart, from the June 2026 event files. Its output, including all 94 players, is in /data/stats/possession-value-vaep-explained-players.json.
How it relates to xT and xGChain
It helps to see possession value as the most general member of a family you already know. Expected threat is a possession-value model that uses ball location as its only input: elegant, transparent, and easy to compute, but blind to anything location does not capture. VAEP and on-ball value are richer models that fold in the type of action and its context, and crucially add the conceding side so defensive actions get scored too.
xGChain and xGBuildup attack the same problem, crediting players who build moves rather than finish them, but with a blunter instrument: they hand every player in a shot-ending possession the full xG of the chance, regardless of whether their individual touch helped or hurt. Possession value is finer-grained: instead of splitting one chance’s xG evenly across everyone who touched the ball, it asks what each specific action did to the probabilities, so a meaningless square pass and a line-breaking carry in the same move are not paid the same.
The caveats: read these before you rank anyone
Possession value is powerful, but it is more model-dependent than xG, and you should treat its outputs with more caution. Three things in particular.
First, it is only as good as the model and the data behind it. Two providers’ possession-value numbers are not directly comparable, because they trained different models on different event data with different definitions of "the next few actions." A player’s VAEP per 90 means something only relative to others computed the same way.
Second, it needs rich event data. A credible model wants carries, pressures, and accurate locations, not just passes and shots. On thin data it degrades quietly: it will still produce numbers, but those numbers lean harder on the model’s assumptions than on what actually happened.
Third, it can over-credit volume. A player who simply touches the ball constantly accumulates value almost mechanically, just as raw xGChain rewards being on a high-possession team. The standard guardrails apply: read the numbers per 90, think in possession-adjusted terms, and never let a single composite number stand in for watching the player. Used that way (as a diagnostic that points you toward contributions the box score hides, not as a final verdict), possession value is the closest the public game has come to a single currency for everything that happens on the ball.
Where these numbers come from
- StatsBomb: research and methodology on on-ball value and possession-value modelling.
- StatsBomb Open Data (StatsBomb Public Data User Agreement): event-level data with the carries, pressures, and locations a possession-value model needs; the 2022 World Cup event files (64 matches, retrieved June 4, 2026) behind our simplified model and the worked move. Our results: /data/stats/possession-value-vaep-explained.json.
- Tom Decroos, Lotte Bransen, Jan Van Haaren and Jesse Davis, “Actions Speak Louder than Goals: Valuing Player Actions in Soccer”, Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2019), pages 1851-1861: the paper that introduced VAEP.
- The action-aware model and the player values (added November 8, 2026): the same StatsBomb Open Data event files (StatsBomb Public Data User Agreement). Our results: /data/stats/possession-value-vaep-explained-players.json.
- FBref: progression and shot-creation data that complement a possession-value view.
- Karun Singh’s blog: the original public write-up of expected threat, the simplest possession-value model to learn from.
Published by SoccerAnalytics from public data and the sources named in this piece. How this site is made.