Chronological feature store
Every feature is cut off before kickoff and can be regenerated with source timestamps and missingness flags.
MODEL METHODOLOGY
Probability before prediction.
Independent model families measure disagreement and record why a market was approved or rejected.
Dixon–Coles + Elo
A constrained score model produces goal matrices while Elo tracks strength through time.
Dedicated draw intelligence
Low-scoring mass, strength similarity and competition draw rates correct draw underprediction.
Uncertainty and abstention
Lower bounds, data quality and model disagreement determine whether the platform says “no signal.”