STOCK PRICE MOVEMENT PREDICTION BASED ON RE-EXTRACT FEATURE LSTM

As one of the most important stock prediction methods, the Long Short-Term Memory (LSTM) does produce good results. However, the pursuit of higher accuracy has always been the goal of scholars. In order to obtain a more precise result than LSTM, this paper presents a Re-Extract Feature LSTM (RE-LSTM) for the stock price movement prediction based on LSTM, convolution and max-pooling operation. Firstly, the LSTM layer takes a group of stock data as training input and produces the cell state and hidden state. Then, the convolutional layer and the max-pooling layer are employed to re-extract the features of the hidden state. Ultimately, an LSTM layer is executed again to achieve the stock price movement prediction. The results of experiments on stock index price of China Securities 300 Index (CSI300) and Korean Composite Stock Price Index (KOSPI) show that RE-LSTM outperforms naive LSTM and achieves excellent result.