A football prediction benchmark is only as trustworthy as its resistance to data leakage. If a model can peek at information from after kickoff, or if different models receive different inputs, the resulting scores are meaningless. AlphaBench prevents data leakage by freezing every prediction at a pre-kickoff snapshot, enforcing input parity across models, controlling odds timing, and auditing the entire pipeline. This article explains how those safeguards work and why they matter for metrics like Brier score, log loss, calibration, ROI, and maximum drawdown.
Pre-kickoff snapshots only
Data leakage in prediction benchmarks usually means one thing: the model had access to information it would not have had at prediction time. In football, the most obvious leak is the final score. But leakage can also come from in-game events, post-match statistics, or even late team news. AlphaBench avoids this by taking a single pre-kickoff snapshot for every fixture. All features, odds, and context are frozen at a fixed cutoff before the match starts. For a fixture like Khovd Western vs Deren in the Premier League, the snapshot is taken before kickoff; the result is used only after the match is settled. Models such as deepseek-flash, gpt-5-mini, and grok-build-0.1 each produce a prediction from that same frozen snapshot. No model gets to see the final score, the halftime score, or any event that occurs after the cutoff. This is the foundation of leakage prevention: if the information was not available at the cutoff, it cannot be used.
Input parity across models
Even with a pre-kickoff cutoff, leakage can occur if models receive different inputs. One model might get an updated injury report, while another gets an older one. One model might receive odds from a later timestamp, while another receives earlier odds. AlphaBench enforces input parity: every model sees exactly the same data, in the same format, with the same timestamps. That includes the same feature set, the same odds snapshot, and the same metadata. This parity is essential for fair comparison. If Model A and Model B are scored on different inputs, their Brier scores or log losses are not comparable. Input parity also makes it harder for a model to exploit a privileged data source that others cannot access. The methodology page describes the full input contract. In practice, parity means that when you see a leaderboard ranking, you can trust that every model faced the same information set.
Odds timing
Odds are a powerful predictor, but they are also a common source of leakage. Closing odds, in particular, incorporate information from many market participants right up to kickoff. They may reflect late team news, weather, or sharp money. If you feed closing odds into a model as an input, you are leaking post-cutoff information. AlphaBench uses a fixed pre-kickoff odds snapshot for all models. The timing is consistent: every model gets odds from the same moment before the match. This matters for evaluation. CLV (closing line value) is a useful benchmark metric: it measures whether a model's probabilities imply value relative to the closing line. But CLV should be computed after the fact as a score, not fed to the model as an input. Mixing the two would create leakage. For more on how odds timing interacts with scoring, see Brier score for football predictions. The key rule is simple: odds used as inputs must be from the pre-kickoff cutoff; odds used for evaluation can be closing odds, but only after predictions are locked.
Auditing
No leakage-prevention system is complete without auditing. AlphaBench logs the exact input snapshot, the timestamp of every feature, the model version, and the prediction for each fixture. That log makes it possible to reproduce any score and to check for anomalies. Automated checks flag any feature with a timestamp after the cutoff. If a model's prediction correlates suspiciously with post-kickoff information, the audit can investigate. Auditing also covers the scoring pipeline. Metrics like Brier score, log loss, and calibration are computed only on settled fixtures, after the fact. ROI and maximum drawdown are tracked for a notional staking strategy as benchmark diagnostics, not as betting advice. The how AlphaBench scores models article explains the scoring formulas. Auditing is what turns a claim of "no leakage" into evidence. It also supports continuous improvement: when a new leakage risk is identified, the audit trail shows which historical predictions might be affected. You can see the current results on the leaderboard.