Due to the non-stationary and non-linearity behaviors of exchange rate data, an appropriate forecasting model that can capture these behaviors is crucial. This paper comparing the performance of modified empirical mode decomposition (EMD) and autoregressive integrated moving average (ARIMA) named as MEMD-ARIMA and modified empirical mode decomposition (EMD) and least squares support vector machine (LSSVM) named as MEMD-LSSVM in forecasting daily USD/TWD exchange rate. EMD technique is firstly used to decompose the exchange rate data that resulting in few intrinsic mode function (IMF) and one residual. In order to improve the result of the EMD so that more effective input can be provided to the forecasting models which are LSSVM and ARIMA, they are clustered into several groups via permutation distribution clustering (PDC). The successfulness of LSSVM in forecasting is depending on the input number selection. The problem is the input number selection is not based on any theories or techniques. Therefore, partial autocorrelation function (PACF) is used in this paper in determining the best number of input for LSSVM. This paper finds that the implementations of PDC has improved the performance of EMD-LSSVM and EMD-ARIMA and also suggest the PDC is suitable either for linear or non-linear model.
展开▼
