Beyond xG: The Quiet Rise of Possession Value Models in Football Analytics

Stade Vélodrome, Marseille, March 2024. Real Madrid versus RB Leipzig, knockout round of the Champions League. In the 64th minute, Toni Kroos receives the ball in his own half and plays a 38-yard slipped pass between two pressing midfielders, on the half-turn, into the path of Vinícius Júnior. Vinícius takes one touch, beats his man, and squares for Rodrygo, who finishes from eight yards. The goal is credited to Rodrygo. The assist is credited to Vinícius. Kroos walks back into position with the slight, ironic shrug of a man who has, by every public statistic in the world, done nothing on the play. The model that grades his contribution at +0.31 expected goals — the highest single-action value of the match — exists, but very few people have heard of it.

That gap, between what we measure and what we see, is the project of modern football analytics. Expected goals broke through to mainstream coverage about a decade ago, and the metric is now a fixture of Premier League broadcasts and Champions League graphics. But xG was always a measurement of the final action: the shot. The pass that created the chance, the carry that broke a line, the off-ball run that pulled a defender out of position — those were invisible to xG, and to assists, and to every other public statistic that football coverage has historically relied on. The decade since has been about closing that gap. A generation of possession value models — VAEP, EPV, the StatsBomb On-Ball Value framework — has changed what we can credit a player for, and the writers who use them well have started telling stories that the old data simply could not see.

I have been writing about football analytics from London since 2014, fell into xG via Football Manager around the same time most of the public conversation did, and the shift I have spent the most time thinking about is the one this article is going to walk through. Possession value models — what they measure, how they extend xG, where they break, and how to read them in a match report without sounding like a consultant — is the subject of this article.

The origin: where possession value came from

The first serious work on assigning value to non-shot actions in football came out of the academic and corporate research community in the mid-2010s. The most-cited landmark is the 2018 paper by Tom Decroos and colleagues at KU Leuven, “Actions Speak Louder than Goals: Valuing Player Actions in Soccer.” That paper introduced VAEP — Valuing Actions by Estimating Probabilities — and laid out the basic logic: every on-ball action in football changes the probability of scoring (or conceding) in the next few seconds of play, and that probability change can be measured.

The intuition was simple. A pass that moves the ball from a team’s own half to the final third increases the probability of scoring on the current possession by a measurable amount. A carry that takes the ball from the touchline into the half-space, past a pressing midfielder, increases it further. A miscontrol that turns the ball over decreases it by a meaningful amount, sometimes catastrophically. Sum the per-action value changes over a match, a season, or a career, and you have a single number that captures a player’s total contribution.

By the late 2010s, parallel models emerged in industry. EPV — Expected Possession Value — was the framework that gained traction inside clubs, particularly via the work of Javier Fernández, Luke Bornn, and Daniel Cervone, published while Fernández was at FC Barcelona and Bornn was at the Sacramento Kings (with cross-pollination back to football). EPV took the VAEP idea further by modeling not just the probability change but the full distribution of possible next actions and their consequences. The math was harder; the interpretive power was higher.

StatsBomb, the data provider that has become the dominant force in club-level football analytics, deployed its own version: On-Ball Value (OBV), plus a separate Goalkeeping framework that valued shot-stopping above and below expectation. Opta, Wyscout, and Hudl all developed comparable internal models. By 2022, every serious analytics department in European football had some version of a possession value framework running in the background.

How possession value works: in plain language

The core mechanic is straightforward. Take a match’s worth of event data — every pass, dribble, tackle, interception, shot, save, foul. For each on-ball action, the model estimates two probabilities: the probability that the team in possession scores in the next 10-15 seconds of play, and the probability that they concede. The model is trained on millions of historical event sequences, so it knows that a pass to a player 25 yards from goal in a central zone, with three defenders behind the ball, leads to a goal on roughly 4% of subsequent sequences. After the action, it recalculates. The same player now has the ball 18 yards from goal with one defender to beat. Probability of scoring: 11%. The action’s value: +7% — or, in xG-language, roughly +0.07 possession value.

Run that calculation over every action of every match, and a player accumulates a total value contribution. A central midfielder making 60 passes a game, most of them small probability swings of +0.005 or +0.01, can quietly post 0.4 to 0.6 possession value per 90 minutes. A forward who scores one goal and creates two big chances might post 0.5 in a match, with most of the value concentrated in three or four actions. A defender who breaks up a counter-attack at the edge of his own box can post a single action worth +0.20, because the action prevented a sequence with high scoring probability for the opponent.

The transformation, conceptually, is that every touch is now a number. A back-pass under no pressure is essentially zero value. A line-breaking through-ball is sometimes the most valuable action of a match. The model does not care whether the action led to a goal in this specific possession. It cares about the probability that an action like this, in this context, leads to a goal across the historical data. The result is a measurement of process rather than outcome, which is, ultimately, what analytics was always supposed to provide.

The critical component: separating creator from finisher

The single most important application of possession value models is in distinguishing the player who creates a chance from the player who finishes it. Assists, the historical proxy, are a profoundly limited measurement — they only credit the immediately preceding pass, and only when the chance is converted. A central midfielder who plays the killer pass two passes before the assist is invisible. A winger who beats his man and squares for an empty-net tap-in receives less statistical credit than the striker who taps it in.

Possession value models fix this directly. A through-ball that splits a defensive line, played from 35 yards out, is credited with the value differential between the pre-pass state and the post-pass state, regardless of what happens after. If the receiving forward squanders the chance, the through-ball’s value is unchanged. If the chance is converted, the goal-scorer’s contribution — the actual shot — is credited separately, usually with a small fraction of the total possession value, because the chance was already very high quality.

Possession value vs the alternatives: a comparison

The football analytics public toolkit has grown crowded in the last decade. A short comparison:

The honest reading of a modern player profile uses three or four of these in concert. xG and xA describe the final-action quality. Possession value describes the chain that produced it. Progressive passes describe pure ball-movement output. Goals plus assists describe the box-score outcome. A scouting report that quotes only one is, almost by definition, missing something.

What the data needs: inputs

Possession value models are data-hungry. The minimum inputs are event data — every on-ball action of every match, with location, time, outcome, and identity of the player involved — and a training corpus large enough to calibrate the probability distributions. The training data is usually two to three seasons of league-wide events, ideally tens of thousands of matches.

The most sophisticated public-facing models also include tracking data: the position of every player on the pitch, at every moment, captured by camera arrays in stadiums. Tracking data lets the model know not just that a pass was attempted from a particular zone but how many defenders were nearby, where the receiving forward was relative to the offside line, and what the pressing structure looked like. The added precision is significant. The data is also expensive, proprietary, and only available to clubs and major data providers.

For the public conversation, FBref (powered by StatsBomb data for top leagues) publishes per-90 progressive passing and progressive carrying numbers, plus their own version of expected possession contribution. Understat, the OG public xG site, publishes match-level xG and xA. Twenty Minutes’ work and the writing at The Athletic occasionally surfaces club-level VAEP results when analysts share them. The full possession value picture is, for now, mostly behind paywalls or inside clubs.

Building the analysis: a working framework

The practical workflow for using possession value models in football writing:

1. Start with the match xG and xA totals. These are the simplest entry point. Did the team out-create their opponent? Did the result match the chance quality?
2. Layer in progressive passes and carries for the midfielders. A team that out-progresses its opponent through midfield is usually doing something structurally right that the shot total may not capture.
3. If you have access to VAEP or OBV data, check the per-90 totals for both teams’ starting elevens. The top three or four contributors are often the players the eye test has already picked out. The interesting moments are when the data and the eye disagree.
4. Cross-check defensive contributions. A center-back’s possession value comes largely from interceptions and recoveries in high-value zones. A midfielder who breaks up counters at the edge of the box can be quietly worth more than a flashy attacker on the same night.
5. Watch the match again with the data on hand. The numbers tell you where to look. The video tells you what actually happened.

Where this gets weird: common mistakes

Possession value models have known failure modes, and the writers who quote them well usually name those failures out loud.

Model variance is real. Two different VAEP implementations, trained on different leagues and different time windows, can produce noticeably different player rankings. A player who grades out as top-five in expected contribution under one model can rank fifteenth under another. The difference is usually attributable to data filtering, training corpus, or weighting of defensive versus offensive actions. Always check the source.

Sample size matters more than fans realize. A player’s possession value over five matches is meaningful but not stable. Ten to fifteen matches is when the per-90 numbers start to converge to a defensible estimate. Career-spanning numbers are, of course, the strongest argument. Single-match VAEP totals are a data point, not a verdict.

The metric over-credits volume. A central midfielder who touches the ball 80 times per match has more opportunities to accumulate small per-action values than a forward who touches it 40 times. Some implementations normalize for this; many do not. A scouting report comparing a Toni Kroos to a Vinícius Júnior on raw VAEP per match is comparing two different jobs.

Defensive actions are still undermeasured. The frameworks were built to value progression toward goal. They handle defensive actions, but the public versions remain less robust on the defensive side. A center-back who organizes a high line, communicates a switch in coverage, and prevents a chance from ever materializing produces invisible value in most models. The good public analysts know this and combine VAEP with traditional defensive watch.

When possession value shines: use cases

The strongest applications:

Midfielder evaluation. The position the public has the hardest time measuring, in any sport, is the central midfielder who shapes the game without scoring or assisting. Possession value is, currently, the cleanest tool for capturing what these players actually do. Bruno Guimarães at Newcastle, Mikel Merino at Real Sociedad in his pre-Arsenal years, Rodri at Manchester City — the public assists-and-goals data understates them dramatically. VAEP and OBV catch the gap.

Transfer scouting. Clubs use possession value as one of the first filters in identifying targets, particularly for players who don’t accumulate goals and assists in volume. The math lets a scouting department compare a Bundesliga 2 central midfielder to a Premier League incumbent on equal terms — the same actions are valued under the same framework. The success rate of analytics-led scouting decisions in the last decade is, in part, a story about possession value.

Match analysis. A 1-0 loss in which one team out-possesses, out-progresses, and out-VAEPs the other is a different match than a 1-0 loss in which they were out-played. The result is the same. The underlying performance — and, by implication, the likely outcome of the next match — is opposite. Possession value gives the writing a way to separate result from performance with more granularity than xG alone.

Coaching evaluation. A manager whose team consistently posts elite possession value numbers despite a roster that the public considers ordinary is, at minimum, doing something tactically distinctive. The math has helped surface coaching reputations — Roberto De Zerbi at Brighton was the canonical example pre-Marseille — that the win-loss record alone would have under-credited.

The limits: what possession value cannot tell you

The honest version of this writing names the limits.

Possession value cannot tell you who is going to win on Saturday. It can tell you who has been the better team in the matches already played. The translation is non-trivial, especially in football, where high variance and small samples make week-to-week predictions noisier than in baseball or basketball.

Possession value cannot, on its own, distinguish between a tactical system and a player. A central midfielder posting elite VAEP numbers in a Pep Guardiola team is producing those numbers in part because the system around him generates the conditions for high-value passes. The same player in a relegation battle, with worse forwards and worse spacing, may not replicate the same per-90 numbers. The metric is real. The role it was earned in is also real. Both have to travel together.

Possession value cannot capture the off-ball game completely. A forward who pulls a center-back out of position by making a run, opening the lane for a teammate’s carry, contributes value that does not show up in his own ledger. Some tracking-data-enabled models attempt to capture this. The public versions, mostly, do not.

Possession value cannot replace watching football. It tells you which players to pay attention to. It tells you which decisions to interrogate. It does not tell you whether the match was beautiful, which is, in the end, why most of us watch football at all.

A working example: the 2023-24 Arsenal title race

The 2023-24 Premier League season provided one of the cleanest possession-value case studies of the decade. Arsenal led the league in expected goals, expected goals against, and per-action possession value for most of the season. They finished second to Manchester City by two points. The result — runners-up — flattens what was actually happening, which is that Arsenal were, by every underlying metric, marginally the better team across the full season. The xG difference was within a goal. The OBV difference favored Arsenal at multiple key fixtures. Manchester City’s slim title margin was, in retrospect, narrowed almost entirely by clinical finishing in three specific matches where City converted high-pressure chances that the model treated as near-equal probabilities.

That gap between performance and result is exactly what possession value frameworks were built to surface. A title race decided by two points is not a story about one team being meaningfully better than another; it is a story about variance at the margins. The post-season retrospectives largely cast City as the dominant team and Arsenal as the over-achievers who fell short. The possession value data tells a different story — two roughly equal teams, with City’s razor-thin advantage compressed into the moments that produced goals. Both readings can be true at once. The math reminds us not to overcount the trophy.

The final caveat is one that football analytics has been slower to confront than basketball: cross-league comparison. Possession value numbers from the Premier League are not directly comparable to numbers from Ligue 1, because the training data, the action distributions, and the defensive intensities differ. A player who posts elite VAEP in a top-five European league will not necessarily replicate those numbers in another. Recruitment departments know this. Public coverage often does not. The cleanest possession value writing names the league context before comparing players across competitions.

Frequently asked questions

What’s the difference between xG and possession value?

xG measures the quality of a shot. Possession value measures the change in scoring probability across every on-ball action — passes, carries, dribbles, defensive recoveries, and yes, shots. xG is one specific application of the broader probability-change framework; possession value generalizes the idea to the rest of the game. A player can post high possession value without ever shooting, by progressing the ball through midfield consistently.

Why don’t we see VAEP scores on Sky Sports?

Two reasons. First, the data is mostly proprietary and held by clubs or data providers like StatsBomb. Second, possession value is harder to explain to a casual viewer than xG. A “0.31 VAEP” doesn’t have the intuitive feel of “0.45 xG” — fans understand “this was a good chance” more readily than “this action raised scoring probability by a measurable amount.” Broadcasters tend to surface the metrics fans can absorb quickly.

Are these models the same across providers?

No. StatsBomb’s OBV, Opta’s possession value, KU Leuven’s VAEP, and various proprietary club models all use the same core idea but implement it differently. The training data, the action types, the weighting of probabilities, and the handling of defensive actions all vary. Two providers can rank the same player differently. When you see a possession value number in a piece of writing, ask which model it came from.

How can I see this kind of data myself?

FBref publishes a public version of expected possession contribution for top European leagues, courtesy of StatsBomb. Understat covers xG and xA for the major leagues. The Athletic and Twenty Minutes occasionally publish club-level VAEP results in their longer-form pieces. The full per-action data is, for now, mostly behind a paywall or inside clubs.

Sources and further reading

• “Actions Speak Louder than Goals: Valuing Player Actions in Soccer” — Decroos, Bransen, Van Haaren, Davis (2018) — the foundational academic paper for VAEP, still the cleanest articulation of the framework.
• StatsBomb’s introduction to On-Ball Value — the industry-leading commercial implementation, explained by the team that built it.
• FBref.com — the most accessible public data hub for top-flight football analytics, including possession-derived metrics.
• Understat — the longest-running public xG and xA site, indispensable for match-level chance quality analysis.
• The Athletic’s football analytics writing — particularly Tom Worville, Mark Carey, and the work that has translated this generation of models into prose.

The Kroos pass that opened this article — half-turn, 38 yards, between two pressing midfielders — was worth +0.31 in OBV terms, the highest single-action value of that match. Vinícius’s carry that followed was worth +0.18. Rodrygo’s finish, the goal itself, was worth roughly +0.10 — the chance had been engineered into a near-certainty by the time he received it. The headline credited Rodrygo. The math credited Kroos. Football, at its best, is large enough to hold both readings at once. For the foundation that makes this kind of writing possible, our piece on expected goals is the natural starting point.