A new decomposition ensemble model for stock price forecasting based on system clustering and particle swarm optimization

Abstract

Accurate forecasting of stock prices has been a challenge in the securities market, while the stock price time series tend to be non-stationary, non-linear, and highly noisy. At present, the traditional method of decomposition and ensemble can improve the forecasting accuracy, but it increases the computational complexity of forecasting, which may introduce additional forecasting errors. In order to address these problems, this paper proposes a novel decomposition and reconstruction model based on system clustering method (SCM) and particle swarm optimization (PSO). Firstly, the system uses the ensemble empirical mode decomposition (EEMD) technique to pre-process stock prices. Furthermore, the complexity of each modality is then measured using the sample entropy method. In addition, the modes are systematically clustered and reconstructed according to three patterns of short, medium, and long-term market changes. Apart from that, subsequences of four different characteristics at coarse granularity (containing the residual term acquired from the mode decomposition) are obtained, thus reducing the complexity of the prediction calculation. Meanwhile, a PSO optimized long short-term memory (LSTM) neural network or support vector regression (SVR) model is used to predict the sub-series with different characteristics separately. The updated model combines decomposition, forecasting, and ensemble steps with a system clustering approach and the idea of “divide and conquer” from a multi-granularity perspective, which not only changes the working mechanism of traditional decomposition-ensemble forecasting models but also realizes a new theoretical model for forecasting stock prices. According to experimental results, the proposed model has higher forecasting accuracy than other current forecasting methods, while it is more consistent with the dynamic stock market’s physical significance.