Define the event and window before analysis
| Design choice | Required definition |
|---|---|
| Index goal | First qualifying goal, every goal, or another predeclared event |
| Follow-up event | Next qualifying goal by either team or a named team |
| Window | Match-clock or elapsed-time interval after the index goal |
| Exposure | Actual minutes in which another goal could occur |
| Censoring | Half-time, full-time, abandonment, second goal or another event |
| State | Score, player count, teams, competition, period and venue |
The Wyscout open-data paper documents a public spatio-temporal football event dataset suitable for reproducible event sequencing. Researchers still need to verify competition coverage, event definitions and licence terms for their use.
Use exposure rather than raw counts
Ten follow-up goals observed in many thousands of post-goal minutes do not imply a higher rate than eight goals in a much smaller comparison exposure. Calculate events per minute at risk or use a time-to-event method. The IFAB duration law also matters because added time varies; fixed nominal halves are not identical exposure windows.
Score state is a major confounder
After a goal, one team leads and the other trails. Teams may change risk, territory, substitutions and pace because of that state. Remaining time is also shorter than at kick-off. A comparison that ignores score state can relabel ordinary conditional match behaviour as momentum.
The peer-reviewed Bayesian in-play model demonstrates why outcome probabilities can update with match information. It does not establish a causal goal-after-goal mechanism, but it provides a model-based reason to condition on the live state.
Illustrative analysis table
| Cohort | Follow-up goals | Minutes at risk | Rate calculation |
|---|---|---|---|
| Post-goal windows | 24 | 1,200 | 24 / 1,200 = 0.020 per minute |
| Matched control windows | 18 | 1,000 | 18 / 1,000 = 0.018 per minute |
The counts are invented. The raw difference is 0.002 events per minute, but no conclusion follows without uncertainty, matching quality, repeated-match dependence and later-sample validation.
A defensible study sequence
- Freeze event definitions, window length and exclusions.
- Validate goal timestamps and disallowed-goal treatment against the documented Wyscout event-data structure or the chosen provider's equivalent.
- Create at-risk intervals with actual elapsed time.
- Match or model score, time, team strength, player count and competition.
- Cluster uncertainty by match or team as appropriate.
- Test alternative windows and definitions without selecting the most favourable result.
- Evaluate the predeclared result on later competitions or seasons.
Do not convert association into a live rule
The PLOS ONE momentum study shows one explicit way to operationalise football momentum from sustained changes in win probability. Its existence underscores the need to define the construct rather than infer it from a memorable sequence. Even a reproducible post-goal association would not establish a favourable live price.
Continue with Momentum Betting for a broader operational test of live narratives.
Continue learning
- Next guide: Half-Time In-Play Analysis
- Related guide: How Football Odds Move During a Match
Assumptions and limitations
The worked data are illustrative. Event datasets can omit competitions, added-time detail, reviews and corrections. Observational designs remain vulnerable to unmeasured tactical and personnel changes, and findings from one competition or period may not generalise.
GambleAware
GamCare
Gambling Therapy
GAMSTOP