This chapter examines the effective roles of probability theory in econometric research. The examination reveals that empirically viable econometric models actually dispense with the widely claimed probability foundation, i.e. joint distributions of all the relevant variables. Theoretically, formal representations of economically causal claims are mostly logic based and distribution free. Correspondingly, econometric models built from, and to be used in, an open world environment are predominantly discriminative rather than stochastically generative. Therefore, the primary role of probability is to assist the process where those models are discriminatively learnt through effective synthesis of prior knowledge and data information.