A novel decomposition approach is presented for a class of on-line problems of Wagner and Whitin’s economic lot-sizing type. The decomposition is based on the fact that optimal plans contain regeneration points and that the plan between two regeneration points is independent of the rest of the plan. This property is exploited in the following way: first, estimate the next regeneration point and second, determine an optimal plan up to that point. The estimation of a next regeneration point can be done with an artificial neural network or with a statistical classification method. We present extensive experimental comparisons of the novel approach with more classical ones. The main conclusion from the results is that our approach dominates all other approaches with respect to robustness, performance, and data efficiency. Only in those cases where the demand is known for a large number of periods in advance some classical approaches perform better.