Stock Market Prediction with Stacked Autoencoder Based Feature Reduction

In this study, the hourly movement direction of 9 banking stocks traded on Borsa Istanbul was predicted by Long-Short Term Memory (LSTM) network. In the prediction process raw stock prices, logarithmic scale stock prices and 11 different technical indicators were used. 1-hour samples of stocks were represented with 63 features with technical indicators computed for 5 different time periods. Class labels indicating the hourly movement direction were assigned based on the hourly closing prices of the stocks. Two different Long-Short Term Memory (LSTM) models were proposed in the prediction process. In the training of the first LSTM model, individual stock features were used, whereas in the second LSTM model, the features of all stocks were given as network inputs. The use of all stock features increased the size of the feature space to 567, and stacked autoencoders were used for dimensionality reduction. According to the experiments, the movement directions of 9 stocks were predicted with an average Macro-Averaged F-Measure rate of 0.573. The use of all stock features increased the prediction performance of the stocks by %0.9-1.9. The performance of both LSTM networks outperformed the Random and Naive benchmarking methods.