03Events
StatsBomb Open Data
The richest free event data in football: every on-ball action with its team, outcome and possession sequence, plus 360 freeze frames of surrounding players on selected matches.
- Events
- 360 freeze frames
- Open JSON
- Modality
- Event data, lineups, match metadata, 360 freeze frames on selected matches
- Scale
- Many competitions and seasons
- Possession labels
- A possession sequence on every event. Not per frame
- Video
- None
- Calibration
- None
- Licence
- StatsBomb open-data terms, attribution required
- Access
- Public GitHub repository, JSON
What it is
StatsBomb, now part of Hudl, publishes a slice of its commercial event data for free. Each match is a list of on-ball events with location, player, team, outcome and a possession counter that groups events into sequences. For selected matches a 360 freeze frame is attached to events: the positions of the players visible on camera at that instant.
What it teaches
| Signal | What you get |
|---|---|
| Possession | An event-level possession sequence with the team that owns it |
| Tracking | Snapshots only: 360 freeze frames at event times |
| Events | A deep, consistent vocabulary of on-ball actions |
| Calibration | Nothing |
Where it wins
- Semantics. Turnovers, restarts and set pieces are typed, not inferred. It shows what a careful definition of possession looks like.
- Breadth. Many competitions and seasons in one schema.
- Context at the event. Freeze frames show pressure and support around the ball without needing tracking.
What it costs
- Sparse in time. Events are instants. Between two of them the data is silent, so a per-frame label cannot be read off.
- No dead-ball clock. A foul and the free kick that follows are both events. The stoppage between them has no explicit duration.
- Partial 360 coverage. Freeze frames exist for selected matches and show only the players in shot.
How it plugs into possession
Use it to learn the grammar, then enforce the grammar on frames. Event data tells you how possession really behaves: how long sequences last, which actions end them, how often a duel changes ownership. Those statistics are priors for a sequence decoder. They set what a plausible spell length is and how expensive a change of state should be.
It also keeps a frame-level model honest about definitions. A deflection that comes straight back to the same team is still one possession under most event definitions. A model that splits it in two has not made a tracking error. It has made a definition error.
Our read
The dictionary of possession. It defines the vocabulary, turnovers, restarts, pressure, carries, with a rigour no tracking dataset matches. It just speaks in events, not frames.