LSTM Framework Design and Volatility Research on Intelligent Forecasting Model for Solving the Parallel Dislocation Problem

Abstract

The yield of treasury bonds is the benchmark interest rate in the financial market which is worth predicting and judging. Based on the Long Short-Term Memory (LSTM) neural network model in deep learning, combined with the vector autoregression method (VAR), this paper creatively constructs the VAR-LSTM framework and uses the predicted values of macroeconomic variables and lagged value of the time sequence as input factors to solve the problem of “parallel dislocation” of the fitting results of the traditional LSTM model which significantly improves the prediction accuracy. In order to meet the requirements of active quantitative investment for high precision prediction of stock market index, adaptive noise complete ensemble empirical mode decomposition (EMD) is introduced into the modeling of stock market index prediction. Combined with the efficient modeling ability of long-term and short-term memory network for medium- and long-term dependence of complex series, using the idea of “decomposition-reorganization-prediction-integration”, an integrated prediction method of stock market index CEEMDAN-LSTM is proposed. CEEMDAN is used to decompose and reconstruct the index to obtain its high and low frequency components and trend items. The LSTM prediction models of each component are constructed respectively and the IMF reorganization mode of high frequency subseries is optimized. Then the overall predicted value of the index is obtained by adding and integrating the predicted values of each component. Taking five representative stock market indexes as test data, the prediction results of CEEMDAN-LSTM and mainstream financial time series machine learning modeling methods are compared systematically. The results show that for treasury bond yield series, the prediction accuracy of ARIMA model is higher than that of general LSTM method, while VAR-LSTM model is better than ARIMA model. The prediction error in the training set and the test set is reduced by about 55% and 50% respectively, and the prediction accuracy of the change direction is improved by about 5% and 8% respectively, which has higher application value. The prediction performance of CEEMDAN-LSTM is consistently better than that of existing modeling methods, and has less prediction error and lower lag.