Date of Graduation
Statler College of Engineering and Mineral Resources
Industrial and Managements Systems Engineering
Robert C Creese
This dissertation examines the forecasting performance of multi-layer feed forward neural networks in modeling five weekly foreign exchange rates: Australian Dollars/U.S. Dollars (AUS/USD), Euro/U.S. Dollar (EUR/USD), Swiss Franc/U.S. Dollar (CHF/USD), British Pound sterling/U.S. Dollars (GBP/USD), and Japanese Yen/U.S. Dollars (JPY/USD). There are five objectives to accomplish. The first is to determine the key modeling factors that should be considered in topology determination. The second is to compare the performances of Genetic Algorithm (GA) and Modified Tabu Search (TS) in choosing the topology for Neural Networks (NN) implementation. The third is to investigate the suitable learning algorithm for NNs for time series forecasting by comparing Back Propagation (BP) with GAs and TS. The fourth is to conduct computational studies for multi-step ahead forecasting for GBP/USD and EUR/USD, as well as to study other accuracy forecasting issues. The last is to study the implementation of multivariate time series forecasting using NNs.;The results of the experiments performed indicate that one should examine the correct topology, especially the three most important factors (number of input nodes, hidden nodes, learning rate) prior to using NNs for time series forecasting.;The comparison performance of topology suggested using GA, TS, and benchmark led to the conclusion that neither GA nor TS is guaranteed to provide better results, especially in terms of percentage of true directional changes (DIR). However, if there is no prior knowledge of the problem, GA searches for topology determination are favored and provide reasonably good performances. GA is also preferred for NN training. Compared to BP, GA guarantees better performance in terms of Mean of Absolute Percentage Error (MAPE) and, most of the time, performs better in terms of Mean of Square Error (MSE).;Caution should be taken in adopting the results, since the study of time periods indicated that the best topology for forecasting a specific foreign exchange is "data specific"; hence the best for a certain period is not always the best to forecast other periods. However, the chosen topology is reasonably useful for up to three steps ahead forecasting.;The trivariate time series, which incorporate interest rates of the two countries involved, did improve the results. Multivariate time series forecasts for monthly JPY/USD, as well as for monthly EUR/USD, produced a higher level of success than the one achieved in the previous experiment. The NNs were programmed using MATLABRTM.
Setyawati, Bina R., "Multi-layer feed forward neural networks for foreign exchange time series forecasting" (2005). Graduate Theses, Dissertations, and Problem Reports. 4189.