New hybridization of empirical mode decomposition and least squares support vector machine model in forecasting Malaysia exchange rates

Abstract

This paper proposed a modification of the hybridization of empirical mode decomposition (EMD) and least squares support vector machine (LSSVM) model named as MEMD-LSSVM in forecasting daily Malaysia exchange rate. The non-linear and non-stationary behavior of the exchange rate data are decomposed first via EMD where several intrinsic mode function (IMF) and a residue are produced. Then, the components are reconstruct by using permutation distribution clustering (PDC) in order to improve them before being fed to the LSSVM. The components are clustered into groups based on their similarities. After that, LSSVM is used to forecast each of the groups and the forecasting values are added in order to obtained the exchange rate actual forecasting value. The best number of input for LSSVM is determined by using partial autocorrelation function (PACF). The performance of the proposed model is compared with LSSVM, ARIMA, EMD-LSSVM and EMD-ARIMA The result shows that MEMD-LSSVM outperforms the other models and proves that PDC is effective in improving the LSSVM input and resulting in more accurate forecasting result.