# Does Barcelona's possession create danger?

**Data: StatsBomb Open Data. Research analysis; no tactical intent is inferred.**

Pinned source: `4b73468fc5b0f1950f9f66fada70ad3a4f9327cb`. Coverage: 240 fixtures; 192 training matches and 48 held-out matches. Latest 20% per competition held out, rounded up. All actions from a match stay in one split.

In the 6 held-out Barcelona fixtures, Barcelona averaged 68.4% of active possession time, 94.0% of final-third completed passes, and 88.2% of positive move xT.

Barcelona generated 3.17 positive move xT per 100 possessions; its opponents in those same fixtures generated 0.45. Barcelona's rate was higher in this selected held-out schedule. The paired difference was +2.72 per 100 possessions (95% paired fixture-bootstrap interval +2.42 to +3.01; 2,000 resamples, seed 2026). This interval omits model-fitting uncertainty and may be unstable with few fixtures.

These are descriptive comparisons under one location-only model. They do not establish that possession causes danger, nor that the team dominates every opponent or every score state.

The longest held-out possession with no shot and less than 0.01 positive move xT belonged to Real Madrid CF W: match 3911641, period 2, possession 147. It lasted 70.2 seconds, contained 17 passes, and began while winning. This describes its immediate output, not its tactical usefulness.

## Validation

All 240 final scorelines reconcile with event goals, including own-goals. xT converged in 83 iterations (Bellman residual 6.6e-11). On 83,472 held-out actions, shot-choice Brier score was 0.00997, against 0.01447 for a constant training-rate baseline; lower is better. 0.0% of held-out actions began in unseen training cells. This checks one-step shot choice, not full xT calibration.

Duration cleaning flagged 1 tiny negative durations. Values in [-0.1, 0) seconds contribute zero time, with original values and affected records retained in the audit output. Larger negative durations fail validation.

## Definitions and limits

Possessions follow StatsBomb IDs, split at period boundaries and whenever the other team performs an attacking on-ball action (pass, carry, shot, dribble, miscontrol, dispossession or shield). The provider's possession_team names the starter and can persist through opposition attacks; it is retained separately. Defensive pressures and tackles do not switch the analytical owner. A sequence requires at least one on-ball anchor. All source passes, shots and shot xG reconcile to sequence totals. Elapsed span runs from first owner on-ball action to last action end, capped at the next sequence. Active time unions owner action-duration intervals, excluding idle gaps and overlaps. Field tilt is each team's share of completed passes starting at x ≥ 80 on the 120 × 80 pitch.

The 12 × 8 xT grid uses empirical non-penalty goals and transitions from passes/carries. Failed moves, miscontrols, dispossessions and failed dribbles absorb at zero. Values solve V = r + TV. Signed successful-move value is V(end) - V(start). Positive move xT sums only increases; it can count repeated advances and is not a possession-goal probability. Penalties are excluded from fitting and xT but included in observed shot/xG totals.

Score state is reconstructed before each event; a goal shot belongs to the pre-goal state. A possession's start state describes the whole chain, while score-state tables use action-level state. StatsBomb xG is supplied, not trained here; its original training data is unknown, so it is a descriptive supporting metric rather than a leakage-free model benchmark.

A low-threat possession may protect a lead, rest a team or draw opponents out. No off-ball tracking or causal evidence here can verify those explanations. No opponent-strength adjustment or cross-league equivalence is claimed. xT can reflect finishing luck and small cell counts; grid sensitivity and external-season validation remain future work.

## Reproduce

See README.md for the exact sample and full-league commands. `manifest.json` records source URLs, revision, cached-file hashes, split IDs, package versions, code hashes and output hashes. The PNG/SVG chart is regenerated by the same pipeline.

## Sources

- [StatsBomb Open Data and terms](https://github.com/hudl/open-data)
- [Official five-league release](https://www.hudl.com/blog/statsbomb-free-womens-data-wsl-ligaf-bundesliga-seriea-nwsl)
- [Karun Singh: Introducing Expected Threat](https://karun.in/blog/expected-threat.html)
- [socceraction: expected threat](https://socceraction.readthedocs.io/en/stable/documentation/valuing_actions/xT.html)
