Soft computing is widely used as it enables forecasting with fast learning capacity and adaptability, and can process data despite uncertainties and complex nonlinear relationships. Soft computing can model nonlinear relationships with better accuracy than traditional statistical and econometric models, and does not make much assumptions regarding the data set. In addition, soft computing can be used on nonlinear and nonstationary time series data when the use of conventional methods is not possible. In this paper, we compare estimates of the nonstationary USD/IDR exchange rates obtained by three soft computing methods: fuzzy time series (FTS), the artificial neural network (ANN), and the adaptive-network-based fuzzy inference system (ANFIS). The performances of these methods are compared by examining the forecast errors of the estimates against the real values. Compared to ANN and FTS, ANFIS produced better results by making predictions with the smallest root mean square error.
