Hybrid Machine Learning Methods for Demand Forecasting Consecutive application of classification and regression methods for the forecasting of periodic and non-continuous demand

This paper is focused on demand forecasting, where the orders from each customer are generated periodically but in a non-continuous way, so most of the values for each client and temporal instant are zeros. The application of regression models is compared to hybrid methods, where an initial classification is considered in order to identify the temporal instants in which an order has been predicted, and afterwards, a regression is generated to obtain the predicted amount of such order. This procedure is complemented by an application to real demand data, for selecting the proper methods for both phases, as well as comparing to simple regression models.