In view of the non-stationary and non-linear characteristics of stock data, traditional statistical models cannot accurately predict the future trend of stock prices. To address this problem, a hybrid deep learning method is constructed to improve prediction performance. Firstly, the distance algorithm is modified to DTW (dynamic time warping) by expanding the K-means clustering algorithm to K-means-DTW, which is more suitable for time series data, to cluster securities with similar price trends. Then, the LSTM (long short-term memory)model is trained through clustering data to predict the price of a single stock. Experimental results show that the hybrid model K-means-LSTM shows better prediction performance and its prediction accuracy and stability are better than the single LSTM model.