The ensemble's three existing models share a structural constraint: they compute expected goal rates as an additive function in log-space. Elo converts rating differences through a logistic. Dixon-Coles estimates log(λ) = attack_home - defense_away + home_adv. The hierarchical Poisson does the same with partial pooling. All three assume that a team's attacking strength contributes independently of which specific opponent they face.
This misses pairwise interactions. A high-pressing team may c…