Accurate stock price prediction is very difficult due to the high noise, non-linearity, and non-smoothness of stock price data. A wavelet-long and short-term memory network (LSTM-Wavelet) model was proposed for stock price prediction. First, wavelet decomposition was used to reduce the instability of financial time series, and the detailed features of wavelet coefficients were analyzed. Then, taking advantage of the long and short-term memory network (LSTM) model, the long-term dependencies in the wavelet coefficients were mined deeply and the predictions were modeled separately for each layer of the decomposed wavelet coefficients. Finally, the data reconstruction of the predicted wavelet coefficients was performed. Based on the empirical analysis of PetroChina’ s stock price data for the past two years, the opening price, high price, low price, and trading volume of each trading day were used as the characteristic inputs to predict the closing price of PetroChina on that day. The results showed that the proposed LSTM-Wavelet model had better prediction results compared with the standard LSTM model and the wavelet-ARIMA (ARIMA-Wavelet) model; the complex stock data were decomposed into wavelet coefficients easily recognized by Long Short Term Memory (LSTM) network through wavelet analysis, the stratified prediction was performed based on the data characteristics with different wavelet coefficients in each layer, and the prediction accuracy has been improved.