Player information in football prediction
Does the recent form of the announced starters improve 1X2 predictions beyond the bookmaker's closing odds? On a season held out until the end, no.
- Held-out matches
- 1,163
- Market only
- 0.9567
- Plus player form
- 0.9580
- Change
- 0.143% worse
- 95% interval
- −0.51 to +0.23%
Change in log loss from adding player features, with paired 95% interval
- Plus player form−0.143%
Question
Closing odds summarise what the market knows shortly before kickoff, line-ups included. Recent individual form could still carry something prices miss. The paper asks whether the recorded performance of each starter in earlier matches improves three-way outcome probabilities beyond a recalibrated closing market.
The baseline
A model fitted on bookmaker probabilities can beat raw odds just by learning to recalibrate them, even if every football feature it receives is useless. So the comparison is not against the raw market. It is a pooled multinomial logistic model on recalibrated closing probabilities, against the same model with four recent player-performance features added. Nothing else differs.
Data
- 7,801 eligible top-division matches from Belgium, the Netherlands, Portugal, Scotland and Turkey.
- Greece was built and audited, then excluded: its first two seasons fail the declared starting-lineup coverage rule.
- A match counts only with valid closing odds and exactly eleven identified starters per team. Every exclusion is recorded.
- Player form uses only earlier appearances, and team histories update only from earlier matches in the same league.
Fixed before the test
Development used 6,638 matches from 2020/21 to 2024/25, with three expanding-window folds. Feature choices, model settings, the comparisons and the decision not to apply post-hoc calibration were all fixed before 2025/26 was opened. The final evaluator checks the settings and a checksum of the selected features before fitting, and refuses to overwrite an existing final result.
Result
Log loss was 0.9580 with player features against 0.9567 for the recalibrated market, 0.143% worse, with a paired 95% interval from 0.511% worse to 0.234% better. Brier score and ranked probability score show the same pattern. Country-level effects were small, mixed and uncertain.
Historical frequencies, Dixon-Coles with and without player information, a logistic model without market inputs and LightGBM were run as further comparisons.
What it means
These four summaries of recent form add no reliable value over strong closing odds. That is a claim about these features against this market, not about player information in general. It is also the reason the market-only model has to be the baseline: against a weaker one, the same features would have looked useful.