03Events

Wyscout academic event dataset

About 1,941 matches and 3.25 million on-ball events from a season of Europe's five big leagues plus two international tournaments, under CC BY 4.0.

  • Events
  • 1941 matches
  • CC BY 4.0
3.25 million tagged events. Each one is an instant; the possession lives in the gaps.
Modality
Tagged on-ball events with coordinates and timestamps; match, team and player metadata
Scale
About 1,941 matches, 3.25 million events
Competitions
Five top European leagues 2017/18, Euro 2016, World Cup 2018
Possession labels
Acting team per event. No spells
Video
None
Calibration
None
Licence
CC BY 4.0
Access
Open download from Figshare

What it is

Released by Pappalardo and colleagues with a Scientific Data paper. Wyscout’s analysts tag every on-ball event from video; this is the public slice: a complete season for five leagues and two international tournaments, as JSON, with coordinates, timestamps and descriptive tags.

What it teaches

Signal What you get
Possession The team behind each event
Tracking Nothing
Events Passes, duels, shots, fouls, free kicks and interruptions at scale
Calibration Nothing

Where it wins

  • Size. Thousands of matches means transition statistics that are stable, per league and per game state.
  • Open licence. CC BY 4.0. Derived statistics can be published and shipped.
  • Restarts are explicit. Free kicks, corners, throw-ins and interruptions are event types. That is a direct view of how play stops and starts.

What it costs

  • No positions between events. No tracking, no video.
  • Human timestamps. Event times are tagged by people watching video. They are approximate, not frame-accurate.
  • Acting team is not ownership. A duel has two actors. Who came away with the ball is in the next event.

How it plugs into possession

Base rates. How many possessions does a match contain? How long does a spell last? How often does a throw-in go back to the team that took it? A decoder with a cost for changing state is making a statement about those numbers, and this corpus is where they can be measured across thousands of matches instead of assumed.

It is also a check on any spell-level metric. If a model’s output for a match contains twice as many possessions as elite football produces, the model is fragmenting. No frame-level label is needed to see that.

Our read

Scale and a clean licence. When a model needs to know what football usually does next, this is where the base rates come from.

Sources

  1. Pappalardo et al., Scientific Data (2019)nature.com
  2. Dataset on Figsharefigshare.com
  3. Community parser and mirrorgithub.com