ABSTRACT
This research focuses on the performance comparison of classical time-series and the artificial neural network method to forecast rupiah and US dollar exchange rates based on 2008–2017 history data. General characteristic of the data is fluctuating and has heteroskedasticity pattern as typical of financial time-series data. Several ARIMA (p,q) model is verified and ARIMA-GARCH(0,0) gave the best result for a classical time-series analysis. Meanwhile, after testing several network combinations, then backpropagation ANN with two hidden layers, cost of training (0.0001) and epoch (100) were selected for the neural network model. The result from mean square error performance measurement shows if BP-ANN has a better performance compared to classical ARIMA-GARCH in short, medium, and long projections. Although both models demonstrate performance decreasing along with the projection time duration, the ANN has several advanced methods (i.e., LSTM model) to improve long-duration projection performance.
