Dataset of 752 consecutive monthly international prices of maize released by the World Bank, were regarded as discrete price time series, and the wavelet decomposition method based on Mallat's algorithm was used to split this sequence signal into several high-frequency components and one low-frequency component, each one was then imported into the recurrent neural network, after that the predicted values of each component from the networks were accumulated as the final predicted price. Research experiments show that the neural network model using wavelet decomposition can flexibly capture high-frequency and low-frequency signals in the price time series of maize, and accurately fit and predict the values of those parts, which tells that the method has practical application significance for scenarios with frequent and violent price fluctuations.