Comparison Between Methods Based on LSTM and BLS for Stock Analysis and Prediction

Abstract

Broad Learning System (BLS) is a new model to expand the neural network from a horizontal perspective in recent years. Based on its outstanding performance in various aspects, this paper intends to apply it to the analysis and prediction of the stock market. Stock market data is a kind of time series data with strong correlation in time. After analyzing and summarizing the advantages and disadvantages of various existing time series prediction and analysis methods, this paper selects the LSTM (long - short term memory network) model, which is outstanding in time series prediction, to compare with the BLS model, and to examine the ability of these two methods in stock market data analysis and prediction. The experimental results show that the cyclic BLS system also has good performance in the prediction of time series data, especially in reducing the training time. The main reason is that the weight of each layer of LSTM is updated by gradient layer by layer, while the weight from hidden layer to output layer in BLS model is solved by pseudo-inverse calculation, which avoids the gradient update method and ensures the efficiency of network training. In addition, the BLS-based loop structure is adopted in this paper. The nodes in the feature layer or enhancement unit are connected circularly. The nodes can capture the dynamic characteristics of the time series, and well acquire and calculate the time-related information in the time series.