Bayesian learning for sales rate prediction for thousands of retailers

Every newspaper publisher has to solve the problem of printing a large number of copies and distributing them to the retail traders trying to keep the return quote as low as possible. To solve this task he needs to estimate as accurately as possible the sales rates for each retail trader. In this paper, we want to show how a prediction system for many thousands of retail traders can be built based on the prediction of the individual sales rates. This prediction is based on a neural network approach, We use a Bayesian learning algorithm to regularize the networks automatically. Furthermore, a top down search based on mutual information is used to optimize the input structure of the networks. The neural network approach reduces the return quota significantly. We conclude with the observation that several data sets are hard to predict and give reasons for that behaviour. (C) 2002 Elsevier Science B.V, All rights reserved.