Staged tree models are a discrete generalization of Bayesian networks. We show that these form curved exponential families and derive their natural parameters, sufficient statistic, cumulant-generating function as functions graphical representation. give necessary criteria for classifying regular subfamilies discuss implications model selection.
