OpenBet
A playful, ambitious experiment in football forecasting.
This is a playful engineering experiment in prediction, not a betting endorsement or financial advice.
Does the model have a clue?
OpenBet starts with a deliberately ambitious question: can newer models say anything useful about the future when their intuition is constrained by real match data?
The project combines a Dixon–Coles Poisson model, an XGBoost classifier, a logistic-regression meta-learner, live bookmaker odds, and an LLM reasoning layer. The point is not to ask an AI for a hunch. It is to see whether context can improve a statistical baseline without being allowed to overrule it.
Four opinions, one bounded answer
The final filter looks for an edge over the market's implied probability rather than ranking matches by confidence alone. Matches can also be rejected as too unpredictable.
Prediction needs honest evaluation
A walk-forward backtest preserves time order so the model is never trained on the future it is being asked to predict. Paper trading then records outcomes with flat one-unit stakes, an equity curve, streaks, and breakdowns by league and pick type.
Tradeoff: the language model can notice context that fixed features miss, but that flexibility is exactly why its influence stays bounded and every result is measured against data that arrived later.
A pipeline that looks after itself
Every six hours the service checks schema drift, stale features, stuck jobs, model compatibility, and orphaned picks before syncing matches, rebuilding features, refreshing odds, generating predictions, and resolving outcomes. Weekly evaluation can trigger retraining when accuracy slips.