Financial innovation is continuously testing the asset selection models, which are the key both for building robust portfolios and for managing diversified risk. This paper describes a novel evolutionary based scheme for the asset selection using Robust Genetic Network Programming(r-GNP). The distinctive feature of r-GNP lies in its generalization ability when building the optimal asset selection model, in which several training environments are used throughout the evolutionary approach to avoid the over-fitting problem to the training data. Simulation using stocks, bonds and currencies in developed financial markets show competitive advantages over conventional asset selection schemes. © 2010 IEEE.