1. Define eligible matches
Record competition status, match length, venue, substitution rules and whether the fixture was closed-door. Exclude or flag matches whose format cannot be reconciled.
2. Measure player exposure
Team totals conceal academy-heavy halves and staggered substitutions. Store lineups and minutes, then calculate features only for comparable exposure. Missing lineup data should be visible, not silently treated as full strength.
3. Adjust opposition and objective
Use a pre-match opponent rating and distinguish conditioning fixtures from near-competitive lineups where documented. Do not infer intent from the final score alone.
4. Build an incremental test
Fit a baseline from prior competitive matches, squad continuity and venue. Add pre-season result, xG or lineup features, then score both on the first declared block of competitive fixtures. TimeSeriesSplit provides a chronological evaluation pattern.
5. Report all seasons
Do not publish only a memorable successful summer. Football benchmark research shows why dataset and evaluation design control model conclusions.
| Field | Release requirement |
|---|---|
| Friendly scope | Declared competitions and formats |
| Lineup completeness | Missingness count |
| Baseline | Competitive-data model without friendlies |
| Test period | Fixed before model selection |
| Decision | Keep, revise or remove feature |
StatsBomb's open repository illustrates the event and lineup structures a reproducible test needs, although its coverage is not a universal friendly dataset.
Build an auditable pre-season row
For every friendly, store opponent, venue, date, format, player minutes, substitution limits and the stated competition phase. Mark closed-door or shortened matches. Keep penalties after a draw separate from regulation scoring.
| Signal | More useful form | Main caveat |
|---|---|---|
| Result | Goal and chance process by lineup | Objectives can differ sharply |
| Team xG | Split by likely starters and reserves | Provider coverage may be incomplete |
| Player output | Per minute with role and opposition | Exposure is often tiny |
| Tactical shape | Repeated use with personnel | Coaches may be experimenting |
| Fitness | Verified availability and minutes | Workload plans are partly private |
Incremental validation
At each historical league opener, build the normal pre-season baseline first. Add the friendly features using only matches available before that date. Compare the models through the opening weeks across multiple seasons and teams. Report seasons where the feature hurts as well as helps.
Avoid choosing friendly weights from the same opening fixtures used for evaluation. If the data is sparse or inconsistently collected, an explicit missing flag and conservative fallback are more honest than manufacturing precision. Pre-season information can be contextually useful without supporting a standalone predictive claim.
For readers, show the likely-starter minutes beside any team total. That one addition prevents a reserve-heavy friendly from being read as if it represented the expected league lineup. Also state whether the match used standard duration and substitution rules.
Related resources
Use form tables for competitive windows or model training and testing for the release design.
Continue learning
- Next guide: Defensive Football Metrics
- Related guide: Expected Assists (xA)
Assumptions and limitations
No claim is made that pre-season data always helps or never helps. Friendly coverage is often selective, and transfer activity or late squad changes can invalidate early estimates.
GambleAware
GamCare
Gambling Therapy
GAMSTOP