The Invisible Run: Quantifying Off-Ball Movement in Football

There's a moment every football analyst has seen and never quite known what to do with. A forward makes a darting run toward the near post, drags a centre-back three meters out of position, and a midfielder receives the ball in a pocket of space that didn't exist two seconds ago. The commentator says "great movement from the striker there" and moves on.

You know something real just happened. You have no idea how to measure it.

That's what I've been trying to capture.

The question

What I wanted to know, specifically, was this: when a player runs without the ball in the final third, does the run actually do anything — does it directly create space for the receiver — or is it just movement that happened to coincide with a pass? And separately: how often do players find space in the box and nobody plays it to them?

To answer it, I used Gradient tracking data – released for free – from all 64 matches of the FIFA World Cup 2022 at 25 frames per second. Every off-ball run in the three seconds before a pass, cross, or carry in the final third. I built, with help from Claude, which was also used to help me write what you’re reading, two components — V8 for space-creating runs on passes into the box, V9 for ghost runs, players who found space and were ignored — and they share a common detection engine, which is worth explaining before the findings.

The shared engine

Both components use the same methodology to identify qualifying runs. For every triggering event, I look back three seconds and find every attacking player (excluding the passer and goalkeeper) who moved at least 3 meters, didn't retreat, ended up in the final third (within 35m of the opponent's goal — this filter matters and I'll come back to it), had a defender within 10 meters at the start of the window, and demonstrably dragged that defender with them — measured by cosine similarity between the runner's movement vector and the defender's, threshold of 0.7, with the defender covering at least 50% of the runner's distance.

That last criterion is doing a lot of work. It's the closest proxy I have for a defender being genuinely engaged by a run rather than just moving in the same direction coincidentally — a defensive line shift would produce similar numbers, which is a real limitation. I don't think it voids the finding, but it's worth being honest about.

The final third filter is one I added after the first pass at the data. Without it, defenders who push forward from deep — Stones, Lovren, Gvardiol — were appearing prominently because their runs engaged defenders, just in midfield rather than where it matters. Restricting to runs that end within 35m of goal keeps the analysis focused on movement that could actually affect the danger zone.

Part 1: Space-creating runs (V8)

V8 triggers on passes into the penalty box — 963 of them across 64 matches. For each pass, every qualifying runner detected by the shared engine gets credited. The classification question is then: did the defender that was dragged happen to be one of the three closest to the pass target at the start of the window? And did they move at least 1 meter further away from the target by the time the pass was played? If yes — the run directly freed the receiver. That's a causal run.

First finding, which reframed how I thought about the rest of the analysis: space creation is collective. 3,142 qualifying runs across 963 passes, an average of 3.3 runners per pass. Not one player making one brilliant diagonal — three or four people moving simultaneously, each engaging a different defender. The histogram is skewed but the 3-4 runner range is the modal situation, not an outlier.

Of those 3,142 runs, 405 (12.9%) were causal.

The space metrics on causal versus non-causal are stark. Causal runs drag the key covering defender 3 meters further from the target on average — non-causal runs actually see that defender move slightly closer. The three nearest defenders end up about a meter further from the receiver after a causal run. Density within 5 meters barely changes for causal runs; it drops significantly for non-causal, because non-causal runs are pulling defenders sideways or deepening the defensive shape without clearing the specific space the receiver needs.

Here's what I found most useful from a practical standpoint: causal runs don't make passes easier to complete. Pass completion into the box is similar whether or not a causal run preceded it. The run doesn't help the ball get through — it helps what happens after. Chance creation rate on completed passes is 19.2% with a causal run preceding it versus 14.8% without. The covering defender was displaced. The receiver has a fraction more time and space. That's where the +4.4 percentage points comes from.

Individual profiles

Sample sizes at the player level are 2-7 matches per player depending on how far their team went. Treat everything here as directional.

Messi has the highest causal rate in the entire top 30 at 30% – always a good signal when Messi pops up. His runs aren't the longest — 9.56m average displacement — but they're surgically precise. Nearly one in three directly frees a teammate. That's not explosive athleticism creating space, it's reading the defensive structure and moving into exactly the gap that matters. Kane has the highest causal rate among high-volume runners at 25%, which tells a similar story — deliberate, economical movement that consistently produces the right outcome. Mbappé's 10.82m average displacement is the longest in the dataset — he's covering enormous ground, reorganising entire defensive structures with his runs rather than exploiting specific gaps.

Giroud leads total volume at 37 runs with a 20% causal rate and the highest positional value among high-volume players at 0.35 — pulling centre-backs out of dangerous central zones.

The midfielders tell a different story. Kimmich has 28 qualifying runs at 7.7% causal. Eriksen has 19 at 0%, with a 76% pass completion rate for the passes. That combination is interesting — his runs help the ball arrive even when they don't directly free anyone.

Part 2: Ghost runs (V9)

V9 uses the same detection engine but flips the question. Instead of asking whether a run created space, it asks: was there already a teammate open in the box at the moment of a final-third event, and did anyone play it to them?

The trigger is every final-third event — 14,017 passes, crosses, and carries. The check at the moment of each event is whether any teammate is in the penalty box, ahead of or level with the ball, with at least 3 meters of space from the nearest defender. Thresholds were derived from the actual distribution of passes into the box: 3-5 meters as an "opportunity" and 5+ as "clear."

Across 64 matches: 2,603 instances of a teammate being open. 318 were served. 2,285 weren't.

For every pass that finds an open player in the box, 7.2 equally open teammates were ignored.

The part that makes this finding genuinely uncomfortable rather than just large: unserviced players average 0.271 positional value. Serviced players average 0.238. Passers are choosing the less dangerous option — systematically, not occasionally — and it holds in both the opportunity and clear tiers. It's not noise but it’s also somewhat okay given what we’re trying to capture is quite complex, and you should also consider players could be covered by a defender between him and the passer, or multiple ones.

Giroud is the most striking individual case. Open 24 times in the box, average positional value of 0.41, the highest among top players, zero instances of a credited runner creating the space for him. He's finding dangerous central positions through his own movement intelligence and nobody is serving him. Youssef En-Nesyri sits just below at 0.39 PV. Müller is in the top 20. The Raumdeuter — the "space interpreter" — appearing in the metric designed to detect invisible space-finding is either a pleasing validation or a very good coincidence.

The passer's blindspot

Flipping the analysis to the passer produces a vision metric: for every player, how many open teammates existed during their final-third events, and what percentage did they actually find? Some passers with 20+ open teammate opportunities service them 30% of the time. Others with the same volume find them less than 5%.

The missed value metric — positional value of the open teammate multiplied by times ignored — quantifies the total cost of a passer's blind spots. Ignoring 30 teammates in low-value positions is less costly than ignoring 15 in prime scoring zones. Messi sits in the high-volume, good-vision quadrant. Antony has 18 open teammate detections and a 0% service rate.

Four skills, not one

The thing the framework made clearest is that "off-ball movement" isn't a single ability. It's at least four distinct things.

Whether you stretch or overload — do you move away from the receiver (pulling defenders out, creating space elsewhere) or toward them (crowding the zone, creating confusion through overload)? Griezmann stretches; Müller overloads. Both can be causal, but the tactical signature is entirely different.

Defender magnetism — how tightly do defenders track you? Some players are so threatening that defenders always follow; others move freely because they're not perceived as dangerous. The elite runners are tightly marked and still causal — that's the intersection.

Pass completion influence — does your run help the pass get through? Eriksen at 76% pass completion and 0% causal is the purest example of a player whose movement helps without directly attributing to any single receiver. Foden at 63% and 21.1% causal combines both effects – the image you’ve seen already.

Positional value of created space — two players with identical causal rates might be creating space in front of goal versus at the edge of the box. Musiala's unserviced opportunities average 0.44 positional value. Saka produced one at 0.66 — practically on the goal line. The rate tells you how often, the positional value tells you what it was actually worth.

What I'd want to build next

The positional value model here is geometric — distance and angle to goal, weighted 70/30, validated against actual chance creation (correlation ~0.26, chance rates climbing from 3.8% at low PV to 28.6% at medium-high). It works well enough to be useful but a proper xG integration — accounting for defensive pressure, body orientation, pass difficulty — would sharpen everything considerably.

The passer's blindspot analysis also deserves more than I've given it here. The raw numbers are interesting but the sample problem bites hard, and I'd want to look at this across a full league season before saying anything confident about individual passers.

The fundamental limitation is one I can't engineer around: this is one tournament. 64 World Cup matches from a single competition where teams meet a maximum of seven times, players are in varying levels of familiarity with each other's movement, and the tactical context is different from week-to-week club football with practiced patterns and a manager drilling specific runs. Everything here should be read as proof of concept — the framework works, the methodology is defensible, the findings are consistent. But I'd want considerably more data before attaching strong conclusions to specific players.

That said: off-ball movement can be quantified. It requires caring about what happened in the three seconds before the pass, not just the pass itself, but there’s also a problem of context which I did not consider here, are you facing a deep block or was it a transition? There’s plenty of room for improvements, but it was still fun to try and capture that specific skill which is a hybrid between team and player.

Data: Gradient tracking data from the FIFA World Cup 2022. All code and methodology notes available on request.