Forecasting Exchange Rates using Artificial Neural Networks

In the dynamic global economy, the accuracy in forecasting the foreign currency exchange rates is of crucial importance for any future investment. The use of computational intelligence based techniques for forecasting has been proved extremely successful in recent times. The aim of this study is to identify a neural network model which has ability to predict the US Dollar against Sri Lankan Rupee (USD/LKR) with higher accuracy level. In this study both static and dynamic neural network architectures were considered. Three types of neural network models were employed: (I) Feedforward neural network (FFNN; static neural networks) with the Backpropagation (BPR) algorithm; (II) FFNN with the Scaled Conjugate Gradient (SCG) algorithm; (III) Time Delay neural network (TDNN; dynamic neural network). Best performed models were found from each approach and forecastability of these models were compared to come up with the best model to predict the USD/LKR. It was found that the FFNN trained with the SCG algorithm performs better than FFNN trained with the BPR algorithm. Therefore the best static neural network for predictions is the FFNN trained with the SCG algorithm. The best TDNN outperforms the best static neural network model and finally this model can be proposed as the best model to predict the USD/LKR. The best performed TDNN model contains two hidden layers, three neurons in each layer and six time delays. This model has ability to forecast unseen data with 76% prediction accuracy.