Artificial neural networks (ANN) like learning vector quantization (LVQ) are suitable for solving classification tasks. Several enhancements of standard algorithms for improving convergence or accuracy exist. They show promising results, at least when applied to simple problems. In this paper, a new approach of evolutionary optimized LVQ is proposed. Its classification accuracy in a complex economic task is examined. The analyzed real-world problem is the classification and evaluation of retail stores in terms of sales volume. Given data reflect macroscopic external infrastructure and microscopic internal aspects of existing stores. Results of numerous computational experiments with a parallelized Implementation in a PC network are compared with results of some standard neural networks which are dominated. Finally, the results are interpreted as support for investment decisions. New stores can be established, or existing stores without prospective profits can be shut down. Alternatively, their in-store design or assortment of goods can be modified.