🤖 machine learning

Space-Creating versus Dead Possession: An Off-Ball Possession-Quality Index for Broadcast Football

This paper introduces a two-layer Off-Ball Possession-Quality Index that distinguishes between sterile "dead" possession and space-creating circulation by combining a threat-based junk-possession flag with a video-derived Space-Creation Index, demonstrating that this metric captures critical offensive value beyond traditional on-ball models.

Original authors: Seongjin Choi

Published 2026-08-11

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Seongjin Choi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

=== SUMMARY ===
Imagine you are watching a soccer match on TV. The commentators are buzzing about one number: "Possession." They say Team A has the ball 65% of the time, so they must be dominating. But have you ever watched a game where one team kept the ball, passed it around their own backyard, and still lost? That's the problem with the old way of counting. It treats all ball-holding as equal, whether it's a dangerous attack or a boring circle in the middle of the field.

To fix this, sports scientists have developed "event-based" tools. Think of these like a video game scorecard that only counts what the player with the ball does. If they pass, they get points. If they shoot, they get more points. These tools are great at judging the person holding the ball, but they are blind to the other 21 players on the field. They can't see if the team holding the ball is actually pushing the other team back, or if the other team is just standing still, letting them pass the ball around harmlessly. This paper asks a simple but tricky question: How do we tell the difference between a team that is creating space for a goal and a team that is just wasting time with the ball, especially when we only have the TV broadcast and not high-tech sensors on every player?

The "Junk" Possession Detector

This paper introduces a two-layer system to solve the mystery of "sterile" possession. The author, working with data from the 2026 FIFA World Cup, built a tool that acts like a detective, first looking at the game's "paperwork" (the event data) and then zooming in on the video to see what's really happening.

Layer 1: The "Junk" Flag
First, the author created a "Junk-Possession Index." Imagine you are grading a student's homework. Most systems just count how many pages they filled out. This new system looks at what they wrote. It breaks the game down into small chunks of time where one team has the ball. It asks: "Did this chunk of time actually threaten the goal?"

They give points for reaching dangerous spots on the field, even if no shot was taken. But here's the catch: they only care about this when the game is tied or the team is losing. If a team is winning and just passing the ball to kill time, that's smart tactics, not "junk." But if a team is losing or tied and they are just passing the ball around without making the other team sweat, the system flags it as "Junk."

When they tested this on 103 World Cup matches, the results were clear. Teams with a high amount of "Junk" possession (passing without threat) actually scored fewer goals and won fewer games. In fact, among teams that held the ball more than 55% of the time, the ones with "Junk" possession averaged only 0.62 goals per game, while the efficient ones averaged 2.24 goals. The paper proves that this "Junk" flag isn't just a fancy way of saying "they didn't score"; it captures something the old "on-ball" value systems miss. Even when you account for how good a team's passes are (using a metric called VAEP), the "Junk" flag still predicts whether a team will lose. It's a unique signal that says, "This team is holding the ball, but they aren't doing anything with it."

Layer 2: The "Space-Creation" Zoom
The first layer tells you that a possession was bad, but not why. Was the ball just going in circles? Or was the team actually pulling the defense apart, but just got unlucky? To answer this, the author built a second layer using the actual TV broadcast video.

They used a special computer pipeline to turn the video into a map of the field, tracking where the players are even when they aren't on the screen (a bit like guessing where the missing pieces of a puzzle are). They calculated a "Space-Creation Index" (SCI). This measures if the team with the ball managed to push the opponent's defense back or seize control of the dangerous area near the goal.

They looked at 31 specific moments where the first layer had flagged a possession as "Junk." Here is the surprising part:

74% of these "Junk" moments were truly dead. The ball moved, but the defense didn't budge. The team was just spinning their wheels.
19% were "weak progression," where they moved a little but didn't really break the defense.
6% were actually space-creating, even though the first layer said they failed!

The Real-World Proof
The paper gives a vivid example from a match between Germany and Paraguay. Germany had 73% of the ball but lost in a penalty shootout. The system flagged two of their possessions as "Junk." When they looked at the video, the Space-Creation Index showed a massive negative number: Germany was moving the ball, but their defense was actually getting pushed forward by Paraguay. They were dominating the ball but losing the space.

In the same match, Paraguay had a moment flagged as "Junk" because they didn't shoot. But the video analysis showed they had actually seized control of the dangerous area (a positive Space-Creation Index). They created space but just didn't convert it into a goal. The first layer would have called this a failure, but the second layer revealed it was actually a success in terms of strategy, just a failure in execution.

The Bottom Line
This paper doesn't just say "possession percentage is bad." It gives us a way to separate "sterile" possession (dead, boring, and useless) from "space-creating" possession (dangerous, strategic, but maybe unlucky). It shows that you can have a team that looks like they are dominating the ball but is actually losing the game, and another team that looks like they are struggling but is actually tearing the opponent's defense apart. By combining the quick data scan with a deep video look, the author found a way to see the invisible battle for space that happens when the ball isn't being kicked.

Technical Summary: Space-Creating versus Dead Possession

Problem Statement
Traditional football analytics rely heavily on possession percentage, a metric that conflates three distinct tactical realities: the location of ball retention, the threat level of that retention, and the game state in which it occurs. Consequently, a team holding 60% possession while recycling the ball in their own half is indistinguishable from a team pinning an opponent in their defensive third. Existing event-based frameworks (e.g., Expected Threat, VAEP, On-Ball Value) refine possession by pricing on-ball actions but remain "blind" to the off-ball question: did the possession create space, or was it sterile? These models evaluate the decision of the player with the ball but cannot detect whether the defensive block was displaced or compressed, a distinction that requires spatial data often unavailable in standard broadcast feeds.

Methodology
The paper proposes a two-layer framework to distinguish "space-creating" possession from "sterile" (dead) possession, utilizing a macro/micro split to balance computational efficiency with spatial depth.

Layer 1: The Event-Side Junk-Possession Index
This layer processes full-match event data to flag low-threat possession sequences.

Sequence Valuation: Matches are segmented into possession sequences. Each sequence $s$ $s$ is assigned a value $q(s)$ $q (s)$ based on:
1. Peak Threat Gain: The maximum increase in Expected Threat (xT) achieved during the sequence, credited even if no shot occurs.
2. Realized Shot Quality: The xG of the best shot in the sequence (using a custom gradient-boosted model).
3. Box Touch: A binary credit for any touch inside the 18-yard box.
Normalization: Sequence values are normalized against the 90th percentile of the corpus. Sequences with $q < 0.15$ are classified as "junk."
Game-State Filtering: To avoid penalizing tactical lead protection, the metric junk-open isolates junk sequences occurring only when the possessing team is tied or losing.
Validation: The index is tested against match outcomes (points, goals, xG) and compared against on-ball value models (VAEP) and territory metrics (Field Tilt) to ensure it captures unique variance.

Layer 2: The Spatial Space-Creation Index (SCI)
This layer operates on short broadcast video windows (10–60s) corresponding to event-flagged "junk" sequences to determine the spatial nature of the possession.

Pipeline: A CPU-only Game-State Reconstruction (GSR) pipeline projects broadcast video to pitch coordinates using camera calibration, multi-object tracking, and color-based team assignment. Off-screen players are imputed using a training-free "ghosting" layer.
Metric Calculation: The SCI measures the net change in pitch control share between the start and end of a window across two zones: the attacking third (A) and the opponent's attacking third (Z).
$SCI = \Delta_{own} + \Delta_{opp}$
Where $\Delta_{own}$ is the change in the possessing team's control in zone A, and $\Delta_{opp}$ is the change in the opponent's control in zone Z (a decrease in opponent control implies the block was pushed back).
Classification: Windows are binned into:
- Space-Creating: SCI $\ge$ +12 (seized dangerous space or pushed the block back).
- Weak Progression: +4 $\le$ SCI < +12.
- Non-Space-Creating: SCI < +4 (static or regressive possession).

Key Results

Correlation with Outcomes: On 206 team-match rows from the 2026 FIFA World Cup, junk-open shows a strong negative correlation with points ( $r = -0.37$ ) and xG difference ( $r = -0.51$ ). Teams with high possession (>55% on-ball event share) but low efficiency (high junk-open) averaged 0.62 goals and 1.00 points, compared to 2.24 goals and 2.24 points for efficient dominators.
Independence from On-Ball Value: In regression models controlling for VAEP and Field Tilt, junk-open remains a significant predictor of points ( $p < 10^{-4}$ ), whereas VAEP is not significant. This demonstrates that the index captures outcome-relevant variance not explained by on-ball action value alone.
Spatial Distinction: In a purposive sample of 31 flagged junk windows from nine matches:
- 74% were spatially classified as non-space-creating (confirming the "sterile" diagnosis).
- 19% were weak progression.
- 6% were space-creating despite being flagged as low-threat by the event layer (e.g., a Paraguay window that seized the attacking third without generating a shot).
Case Studies: The framework successfully identified "sterile domination" in Germany's Round of 32 loss to Paraguay (73% on-ball event share, negative SCI, elimination) and distinguished it from a space-creating but unconverted Paraguay window in the same match.

Significance and Claims
The paper presents a two-layer framework that separates "space-creating but unconverted" possession from "sterile" possession using broadcast data. Its primary contribution is demonstrating that:

Possession Quality and Territory are Distinct: High field tilt (territory) does not guarantee points; the quality of possession (threat generation) is the critical differentiator.
Off-Ball Evidence is Necessary: Event-only models cannot distinguish between a team that successfully displaces a defense but fails to convert, and a team that circulates the ball without moving the defense. The two-layer approach bridges this gap by using event data for macro-flagging and broadcast video for micro-spatial adjudication.
Practical Scalability: The macro/micro split offers a practical solution for analyzing off-ball spatial evidence at the scale of a whole tournament without requiring full-match tracking data, which is rarely available for broadcast matches.

The author explicitly notes limitations, including the reliance on imputed off-screen players, the use of corpus-wide constants for thresholds, and the descriptive nature of in-sample score-state associations. However, they assert that the out-of-sample splits and the independence from VAEP provide robust evidence that the "junk" flag identifies a distinct, outcome-relevant phenomenon.

The "Junk" Possession Detector

Technical Summary: Space-Creating versus Dead Possession

More like this