Abstract: In this paper, noting temporal order of time series forecasting, we introduce Forward-Validation Model Averaging (FVMA) for the prediction of binary response MIxed DAta Sampling (MIDAS) model. In contrast to the general model averaging prediction method, we allow the frequency of explanatory variables and the sample size to vary among the candidate models. We provide two theoretical justifications for the proposed method. First, we establish the asymptotic optimality of FVMA method in the sense of achieving the lowest possible prediction risk. Second, when the candidate model set includes correctly specified models, we demonstrate that the proposed FVMA asymptotically assigns all weights to the correctly specified models. Our method generally outperforms typical discrete time series prediction methods in both numerical simulation and empirical analysis.