This paper introduces a multi-stage optimization approach （MSOA） used in genetic algorithm （GA） for training neural networks to forecast the Chinese food grain price. We divide the training sample of neural networks into two parts considering the truth that the recent observations should be more important than the older ones. Firstly, we use the first training sample to train the neural network and achieve the network structure; Secondly, we continue to use the second training sample to further optimize the structure of neural network based on the previous step. Aiming at the characteristics of neural network structure, a model using a hybrid GMDH and artificial neural network is established. It can make the selection of input-lay units easy and improve the ability of rate of studying and the adaptability of neural network. Empirical results show that the neural networks based on MSOA can improve greatly the global convergence ability and convergence speed of most networks. Furthermore the result indicates that the combined model can be an effective way to improve forecasting accuracy