Recently many statistical methods for the design of neural networks were proposed. In this paper we present an integrated approach for building high performant classifiers for time series prediction. The classifier is based on a committee of neural networks, which are constructed to be as independent as possible. Each network is trained using a Bayesian learning rule. The training process is intertwined with the optimization of the network topology, especially the input structure. The mutual information of the input-output relation is exploited, to reduce the input dimension to the most possible limit. An evolutionary algorithm serves as search heuristic for the optimization process. The benefits of the approach are demonstrated on the development of a time series prediction system, that recently replaced its predecessor, which was trading online successfully since April 1996 at the stock market in Frankfurt, Germany. We conclude that by combining these efficient methods, one can increase considerably the performance of classifiers for time series prediction compared to conventional approaches.