Fractional frequency hybrid model based on EEMD for financial time series forecasting

Abstract

In this paper, we propose a new two-stage methodology, which is a hybrid model based on ensemble empirical mode decomposition (EEMD), to predict the complex financial time series. The hybrid model comprises Multidimensional k-nearest neighbor method (MKNN), Autoregressive moving average model (ARMA) and Quadratic regression model. The main work includes two parts, using weighted Euclidean distance instead of Euclidean distance to measure the similarity in MKNN model and constructing EEMD-FFH model. In this model, an original time series is decomposed to IMFs and a residual wave by EEMD. Then MKNN, ARMA and Quadratic regression models are separately used to predict high-frequency IMFs, low-frequency IMFs and the residual wave. In order to verify the robustness of the EEMD-FFH model, we use daily high price and closing price of four stock indexes for experimental analysis. The experimental results indicate that the prediction performance of the EEMD-FFH model is much better than EEMD-MKNN model, and the best hit rate achieves 87%.