A stock forecasting method based on combination of SDAE and BP

Abstract

The neural network, as a non-linear system with large-scale parallel distributed processing, has been widely used in the field of prediction. This paper proposes a stock forecasting method based on the combination of Stacked Denoising AutoEncoders (SDAE) and Back Propagation Neural Network (BPNN). Using more than 4100 historical data of Sinopec Group from 2001/8/8 to 2018/8/16 for Seventeen years, then process the data and select six features of them for training and prediction. The experiment first uses SDAE for training in the pre-training stage to obtain the weights. Then the trained weights of SDAE are assigned to the Back Propagation Neural Network (BPNN) for fine-tuning to optimize the entire network structure. Then optimize the network parameters for further step and chose the best optimal hyperparameter combination. Finally, the prediction is performed on the test set, the Mean Absolute Error (MAE), Error Variance (EV), and Absolute Maximum Error (AME) between the predicted value and the actual value are used as evaluation performance evaluation criteria. The results show that this neural network model can achieve high precision, and provides an effective method for predicting the stock market artificial neural network with many influencing factors and unclear mechanism.