Availability is a changing state
A player can move from reported doubt to predicted starter, confirmed substitute and actual participant. Historical data often preserves only the final state, which is unsafe for an earlier forecast. Sportmonks' lineup documentation distinguishes predicted and confirmed lineups and describes sidelined-player data; a model should retain those states and retrieval times.
Minimum availability record
| Field | Example | Purpose |
|---|---|---|
| player and team ID | Stable provider IDs | Avoid namesake and transfer errors |
| status | reported doubt | Preserve uncertainty |
| evidence type | club report, feed, lineup | Separate sources |
| effective time | report publication | Know when information became usable |
| retrieved time | system receipt | Measure feed latency |
| expected role | starter, substitute, unknown | Connect status to model input |
| confidence or state | predicted, confirmed | Avoid false certainty |
Modelling choices
The simplest model can use squad-level availability counts or expected minutes lost. A deeper model can estimate a lineup distribution and average forecasts across plausible lineups. In either case, player impact must be learned from prior data, regularized for small samples and separated from team strength. A missing player should not automatically receive a large negative adjustment because of reputation.
Cutoff-safe workflow
- Choose the forecast time, such as 24 hours before kickoff.
- Retrieve the availability snapshot that existed then.
- Map players to the correct club and competition date.
- Preserve unknown status instead of filling from the final lineup.
- Generate one forecast or a weighted set of lineup scenarios.
- Store the snapshot beside the prediction.
- Re-run after confirmed lineups only as a separately timed forecast.
scikit-learn's leakage guidance applies directly: a final lineup cannot be part of training examples that represent an earlier cutoff.
Congestion is context, not a diagnosis
A systematic review of fixture congestion and injury summarizes time-loss injury evidence in professional male football. It does not identify an individual player’s status or justify treating every congested schedule as an injury. Congestion, recent minutes and travel can be contextual inputs while medical status remains a separate evidence stream.
Evaluate uncertainty
Compare performance for forecasts made before and after confirmed lineups. Report coverage and calibration separately when key statuses are unknown. Calibration guidance supports testing whether probabilities behave reliably within those evidence states.
Illustrative scenario average
If a player has a 0.60 estimated chance of starting, a model can combine a start scenario and a non-start scenario using declared weights. That is an uncertainty calculation, not evidence that 0.60 is the correct chance. The probability source and later calibration still need validation.
Compare forecasts at declared information states
Create separate evaluations for an early pre-match snapshot, a later team-news snapshot and confirmed lineups. Use the same target and later fixtures, but never replace the earlier state with the confirmed record.
| Snapshot | Reader question | Coverage report |
|---|---|---|
| 24 hours before kickoff | How useful are reports and squad status? | Unknown and conflicting reports |
| Two hours before kickoff | What changed with later evidence? | Predicted-lineup availability |
| Confirmed lineup | What is the final pre-match update? | Missing or delayed confirmations |
Sportmonks' lineup documentation supports preserving predicted and confirmed states. Compare score and calibration at each snapshot rather than assuming later is always better; a late pipeline with poor latency can lose its informational advantage.
Audit player-impact estimates
Estimate impact from historical participation and team context, then shrink unstable player effects. Test whether the adjustment helps on later fixtures and whether it fails for transfers, changing roles or small minute samples. Do not infer medical severity from an estimated football effect.
When statuses conflict, retain source and retrieval time and produce scenarios or withhold the adjustment. A deterministic “injured” flag hides uncertainty that the model and reader need. Sportmonks' lineup documentation supports retaining predicted, confirmed and sidelined-player evidence states.
Continue the workflow
Use the fixture-congestion analysis to keep schedule load, recent minutes and medical availability as separate evidence streams.
Continue learning
- Next guide: Machine Learning for Football Prediction
- Related guide: Neural Networks for Football Prediction
Assumptions and limitations
Medical details may be private, incomplete or wrong. Predicted lineups are not confirmations, and actual participation is not proof that a player was fully fit. This page describes data controls, not a clinical assessment or a universal player-value formula.
GambleAware
GamCare
Gambling Therapy
GAMSTOP