AI model predictions. For entertainment and educational use. Past performance does not guarantee future results. Not financial or gambling advice.
Model Backtest
Only counts predictions where predictions.generated_at < events.starts_at — i.e., the model picked the game BEFORE it was played. Diagnostic confirmed
~99.998% of historical "predictions" in this database were training-time
backtests written after games settled (metadata.actual_value populated, generated_at years after the game).
Pre-filter the page would have shown 33 "+EV" buckets at 75–85% cover —
all in-sample training accuracy, none real edge. Post-filter, only forward
picks survive. Most models won't have any until the prediction pipeline runs
ahead of upcoming slates with the leak fixed upstream.
Per-(model × sport × market) cover rate vs −110 break-even (52.38%). Built on 0 legit forward picks. Once the prediction pipeline starts emitting genuine pre-game forecasts, this page will populate automatically — the nightly grader cron refreshes the MV.
No models meet the threshold
Lower the min picks filter or pick a different sport/market combo.