The stock comprehensive decision indexes are extracted by kernel principal component (KPC) analysis method. An effective stock decision model based on optimized neural network algorithm is presented. The weight and threshold value of the proposed BP neural network with KPC input and stock return output are adjusted and optimized by genetic algorithm. The empirical test results of CSI 300 stock sample show that the proposed stock decision model improves the operational efficiency and possesses strong learning ability with high prediction accuracy. This paper develops a bottom-up stock decision model to conduct data mining in stock returns and risks in order to select stock portfolio from individual companies with high quality.