The Negative Binomial Distribution (NBD) has been widely used as a stochastic model for purchase frequency under the assumption that the purchase process is stationary. This paper proposes a method for estimating the parameters of the NBD by utilizing the entire past purchase history rather than using only the purchase record from one prior period, in order to model the non-stationary phenomenon of purchasing behavior. As a method for estimating the NBD conditional expectation using purchase records over the past t periods, an optimal weighting procedure that blends predicted and actual values is presented. Once the conditional expectation and the proportion of non-buyers are estimated through this procedure, the parameters of the non-stationary NBD are calculated using the series estimation method proposed by Morrison (1969b), and then this NBD is used to predict the average number of buyers and the non-purchase probability, yielding purchase probabilities corresponding to the number of purchases. Finally, the predictions from the model presented in this paper are compared with those from a model assuming stationarity.