The basic state-change idea
A simple grid model divides the pitch into zones and estimates the probability that possession in each zone eventually leads to a goal. Moving the ball from a zone worth 0.02 to one worth 0.07 creates an illustrative value change of 0.07 - 0.02 = 0.05. A failed action may receive a cost, depending on the implementation.
StatsBomb's contextual xT paper documents one richer approach and shows that xT is a model family rather than a universal provider field.
Inputs and transformations
| Decision | Example choices | Why it changes output |
|---|---|---|
| State | Grid cell, possession context, pressure | Defines what situations are treated as similar |
| Actions | Passes and carries, or broader events | Changes credited action population |
| Outcome | Next goal, eventual shot, retained possession | Changes the target being estimated |
| Failure cost | Zero, turnover value, opponent threat | Changes risky-action treatment |
StatsBomb's open data provides event coordinates and action fields for a reproducible prototype. A model should record coordinate orientation, possession segmentation, train-test split and any smoothing for sparse zones.
Worked action comparison
Suppose Pass A moves possession from 0.03 to 0.08 and completes: its simple change is 0.05. Pass B moves from 0.06 to 0.12 but succeeds only in the observed event: the realised state change is 0.06, yet a decision model may also estimate completion risk. Ranking actions by realised xT alone can reward risky passes selectively when failures are excluded.
Validation
StatsBomb's contextual xT research documents the state-value choices and validation questions used below.
A useful xT model should be tested on later matches and compared with simpler baselines. Check whether higher predicted state values correspond to higher future scoring rates, whether calibration changes by league, and whether player rankings are robust to role and minutes.
xT and xG answer different questions
StatsBomb's contextual xT research documents the state-value choices and validation questions used below.
xG evaluates recorded shots. xT can assign value before a shot exists, but it depends more heavily on possession and spatial assumptions. Neither metric alone establishes player quality or a future market probability.
Grid design changes the answer
StatsBomb's contextual xT research documents the state-value choices and validation questions used below.
An xT model first assigns values to pitch states, then credits an action for moving the ball between them. A coarse grid can hide meaningful differences inside a cell; a very fine grid can produce unstable estimates where few actions occur. Possession-ending actions also need explicit treatment.
Use a sensitivity table when implementing the metric:
| Design choice | Question to record |
|---|---|
| Pitch partition | Are cells fixed, learned or continuous? |
| Action set | Do carries, passes and crosses use the same transition logic? |
| Turnovers | Is lost possession a negative value and how large is it? |
| Possession horizon | Does value mean the next action, possession or later outcome? |
| Smoothing | How are sparse origin-destination pairs handled? |
Player-level interpretation
Total xT can reward players with more possession opportunities. Report minutes, touches or attempted actions alongside the total, and split progression from loss where possible. A full-back repeatedly moving the ball from a low-value build-up zone into a slightly better zone may accumulate value differently from a winger creating one large jump near goal.
scikit-learn's TimeSeriesSplit guidance supports the chronological validation rule used in the next step.
When comparing model versions, freeze the action data and score it under both implementations. Investigate which actions change rank and why. For forecasting, train the downstream target only on xT information available before the prediction cutoff and test whether it adds value beyond simpler territory and shot features.
Related resources
Read xG explained for shot quality or player metrics for a multi-metric view.
Continue learning
- Next guide: Fixture Congestion in Football
- Related guide: Football Referee Statistics
Assumptions and limitations
All numbers are illustrative state values. Model output depends on event accuracy, competition coverage, possession definitions and the chosen target. Comparing xT from different implementations without harmonising those choices is not valid.
GambleAware
GamCare
Gambling Therapy
GAMSTOP