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

This paper introduces a two-layer framework for evaluating possession quality in football, addressing the limitation that traditional possession statistics conflate three distinct aspects: where a team holds the ball, whether that holding threatens the goal, and in what game state it occurs. The authors argue that existing event-based possession-value frameworks (expected threat, VAEP, on-ball value) are fundamentally on-ball: they evaluate the decision of the player in possession, but the question a sterile possession poses is off-ball. The distinguishing evidence is the position of the twenty-one players without the ball: whether the defensive block was compressed and displaced.

Layer 1: Event-side junk-possession index. The paper segments matches into possession sequences and prices each sequence using a value function: value(s) = max(0, xTmax(s) − xTstart(s)) + 0.7·xG(s) + 0.10·⊮[box touch]. Term (i) credits peak threat gain even without a shot; term (ii) uses a gradient-boosted xG model (since the event feed's shot-xG field is a constant placeholder); term (iii) rewards penalty-box touches. Sequences with q < 0.15 are flagged as junk sequences. The headline metric junk-open is the on-ball-event share of junk sequences among a team's possessions in tied-or-losing states — circulation with no excuse, excluding lead-protecting circulation via live scoreline reconstruction.

Validation results. On 206 team-match rows from 103 matches of the 2026 FIFA World Cup: junk-open correlates negatively with expected goals (−0.38), goals (−0.38), points (−0.37), and xG difference (−0.51). Among 71 team-matches above 55% possession, the 8 below median efficiency (junk dominators) averaged 0.62 goals and 1.00 points versus 2.24 and 2.24 for the 63 efficient dominators. A logistic model predicting whether a sequence ends in a shot reaches 5-fold AUC 0.907 ± 0.004.

Central claim: not reducible to on-ball value. When team offensive VAEP and field tilt are held fixed in the same regression, "junk-open is associated with points at p < 10−4 with VAEP not significant (p = 0.34), and efficiency with xG difference at p < 10−4 with VAEP not significant (p = 0.46). Despite surface correlations of ≈0.5 between VAEP and the possession-quality metrics, conditional on the possession-quality metric VAEP is not a significant predictor of the outcome." The paper also shows junk-open is not a repackaging of territory: in a joint regression, junk-open remains significant for points (p = 0.017) while field tilt is not (p = 0.14); the pattern reverses for xG difference.

Out-of-sample evidence. First-half efficiency predicts second-half xG (r = +0.32); first-half junk-open predicts lower second-half xG (r = −0.24). Leave-one-match-out team trait analysis: a team's mean efficiency in other matches predicts this match's points (r = +0.31), and mean junk-open predicts lower points (r = −0.30). Only 9% of tied-or-losing junk is redeemed (regained within 25s and reaching threat), so a dead-only variant improves in-sample prediction slightly but is essentially identical out-of-sample.

Layer 2: Spatial Space-Creation Index (SCI). For event-flagged windows, broadcast video is projected to pitch coordinates via a CPU-only GSR pipeline (PnLCalib calibration, BoT-SORT tracking, colour-based team assignment, shot-change gating, training-free off-screen player imputation). The SCI measures a net pitch-control change capturing whether the possession seized dangerous space or pushed the opponent's block back: SCI = ∆own + ∆opp, where ∆own is the change in the possessing team's attacking-third control share and ∆opp is the change in the opponent's presence in the mirror zone (positive ∆opp means the opponent's shape was pushed back). Operational bins: space creation (SCI ≥ +12), weak progression (+4 ≤ SCI < +12), non-space-creating (SCI < +4).

Spatial results. Across 31 analysable flagged windows from nine purposively selected World Cup matches: 74% of flagged junk is spatially classified non-space-creating, 19% weak progression, and 6% space-creating windows that the event flag alone would score as failure. Case studies include Germany (73% possession, eliminated on penalties to Paraguay): both flagged Germany windows were strongly negative (SCI −24.4 and −14.2), showing the block advanced rather than being pushed back. Conversely, a Paraguay window at 76:01 scored SCI +18.0 — space-creating despite no shot — demonstrating the 6% the two-layer design exists for. Norway beat Brazil despite all four flagged Norway windows being spatially non-creating, showing high junk share and winning are not contradictory. The event layer also separates cases sharing a scoreline: South Africa and South Korea both held 57% and lost 1–0, but with different efficiencies (0.27 vs 0.41), yielding same possession, same result, opposite diagnosis.

Limitations. The event index is a ball-only approximation of an off-ball question; sequence segmentation is coarse; xG uses shot location only; thresholds (q < 0.15; SCI +12/+4) are round operational values; xG-based associations are partly mechanically coupled since the index embeds a shot-xG term; corpus-wide normalisation constants mean out-of-sample splits are cross-fitted only up to these global constants; in-sample outcome associations are descriptive due to score-state feedback; spatial team assignment degrades for low-contrast kits; extra-time windows are excluded; imputation sensitivity was not used as a filtering criterion.

Improvement 1: Two-Layer Possession-Quality Evaluation System for Tactical Analysis

• What it does: An AI system that evaluates football possession quality by combining event-based metrics (junk-possession index) with spatial analysis (Space-Creation Index) to distinguish sterile circulation from threatening possession, going beyond traditional possession stats and on-ball value models.

• Specific capabilities:

• Automatically flags junk sequences (q < 0.15) in tied-or-losing states using a value function that combines peak threat gain, shot xG, and box touches, while excluding lead-protecting circulation via live scoreline reconstruction.

• Integrates a CPU-only GSR pipeline (PnLCalib calibration, BoT-SORT tracking, colour-based team assignment) to compute net pitch-control changes (SCI) for flagged windows, classifying them as space-creating (SCI ≥ +12), weak progression (+4 to +12), or non-space-creating (< +4).

• Provides a joint diagnosis that separates cases like South Africa vs. South Korea (same possession %, same result, opposite efficiency: 0.27 vs 0.41), enabling coaches to identify whether a team's possession is genuinely threatening or merely sterile.

Improvement 2: Score-State-Aware Possession Quality Metric for Match Outcome Prediction

• What it does: An AI system that predicts match outcomes (points, xG difference) using a possession-quality metric that is conditional on game state, rather than raw possession share or on-ball value.

• Specific capabilities:

• Reconstructs live scorelines to compute junk-open — the on-ball-event share of junk sequences in tied-or-losing states — and uses it as a predictor that remains significant (p < 10−4 for points, p < 10−4 for xG difference) even when VAEP and field tilt are held fixed in regression.

• Achieves 5-fold AUC 0.907 ± 0.004 for predicting whether a possession sequence ends in a shot, using a gradient-boosted xG model that corrects for the event feed's constant placeholder shot-xG field.

• Demonstrates out-of-sample validity: first-half junk-open predicts lower second-half xG (r = −0.24), and leave-one-match-out team traits (mean junk-open in other matches) predict this match's points (r = −0.30), enabling pre-match scouting of opponent possession tendencies.

Improvement 3: Off-Ball Spatial Threat Quantification via Broadcast Video Analysis

• What it does: An AI system that quantifies off-ball spatial threat creation from broadcast video, addressing the limitation that event-based possession-value frameworks (expected threat, VAEP) only evaluate on-ball decisions.

• Specific capabilities:

• Projects broadcast video to pitch coordinates using a training-free pipeline with shot-change gating and off-screen player imputation, enabling analysis of the 21 players without the ball.

• Computes SCI as the sum of changes in the possessing team's attacking-third control share and the opponent's presence in the mirror zone, capturing whether possession seized dangerous space or pushed the opponent's block back.

• Identifies the 6% of flagged junk windows that are actually space-creating (e.g., a Paraguay window at 76:01 scoring SCI +18.0 despite no shot), preventing false negatives that pure event flags would produce — critical for tactical analysis where a team may create space without a direct shot attempt.

Improvement 4: Cross-Validated Team Trait Profiling for Scouting and Match Preparation

• What it does: An AI system that generates team-level possession-quality profiles using leave-one-match-out cross-validation, separating stable team traits from match-specific noise.

• Specific capabilities:

• Computes a team's mean efficiency and mean junk-open across other matches to predict this match's points (r = +0.31 and r = −0.30, respectively), enabling pre-match predictions of opponent effectiveness in sterile vs. threatening possession.

• Distinguishes junk dominators (teams with >55% possession but below-median efficiency) — these averaged 0.62 goals and 1.00 points versus 2.24 and 2.24 for efficient dominators — allowing scouts to flag teams that over-possess without creating threat.

• Handles score-state feedback by excluding lead-protecting circulation, so the metric reflects genuine attacking intent rather than game-management, making it robust for in-game adjustments and post-match review.

Ball possession is the most-cited and most-misleading number in football: 60% recycled in one's own half is not 60% spent pinning the opponent back. Existing event-based possession-value frameworks (expected threat, VAEP, on-ball value) price on-ball actions but ignore the off-ball question a sterile possession poses: did holding the ball create space, or was the circulation dead? We answer this in two layers. First, an event-side junk-possession index prices each possession sequence by its peak threat gain under an expected-threat grid and -- after reconstructing the live scoreline to exclude lead-protecting circulation -- flags low-threat sequences in tied-or-losing states. On the 2026 FIFA World Cup (103 matches, 206 team-matches) the flag correlates negatively with points (r=-0.37) and xG difference (r=-0.51, partly index-coupled). It is not a repackaging of on-ball value: with team offensive VAEP and field tilt held fixed, the junk flag stays strongly negatively associated with points (p<0.0001, also match-clustered) while VAEP is not significant -- in this same-match (descriptive) regression it adds information beyond this on-ball action-value model. Second, for a flagged window we resolve whether it was spatially dead or space-creating by projecting broadcast video to pitch coordinates and measuring a Space-Creation Index (SCI): a net pitch-control change capturing whether the possession seized space or pushed the opponent's block back. Across 31 of 35 flagged windows from nine World Cup matches (a purposive sample), 74% are spatially non-space-creating, 19% weak progression, and 6% space-creating windows the event flag alone would score as failure -- including a side with 73% of the ball that exited on penalties (two non-creating windows). The two layers separate space-creating-but-unconverted from sterile possession, a distinction event-only on-ball value cannot make.