The naive Poisson model underprices the draw in low-scoring leagues — a known flaw with a known fix. We walk through the full build: data sources, dispersion correction via a Dixon-Coles adjustment, and honest backtesting.
Over three seasons of out-of-sample data, the corrected model achieved positive closing-line value on the draw in two of the five leagues tested. That’s an edge — a small one, which is what real ones look like.
