An Adaptive High Frequency Time Series Prediction Model Based on LSTM and Bayes Model Averaging

Compared with traditional time prediction methods such as ARIMA and GARCH, neural network models such as LSTM can predict high-frequency time series more effectively. In this paper, the original time series information is considered to fit the prediction residual of LSTM model to further improve the prediction accuracy. Specifically, PCA and kernel method are first used to reduce the dimension of the original sequence with fixed lag order to capture the linear and nonlinear information of the original data. Considering the true data distribution and unknown structure, we use the idea of model averaging to adaptively determine the relative importance of linear and non-linear information and select the dimensionality reduction space dimensions of PCA and KPCA. In this paper, OLS, Lasso, ridge regression and principal component regression based on cumulative contribution rate are used as comparative methods to analyze the actual data in three data sets: hourly power change in a state in the United States, hourly PM2.5 content change in Beijing and changes in traffic flow at intersections. The empirical results show that the proposed method has certain competitiveness compared with other methods.