A Comparison of the Forecasting Models of the NT$/US$ Exchange Rate

• Main Authors: Wu, Iee-Fung, 吳怡芳
• Other Authors: Goo, Yeong-Jia
• Format: Others
• Language: zh-TW
• Published: 2005
• Online Access: http://ndltd.ncl.edu.tw/handle/62146321424957590164

博士 === 國立臺北大學 === 企業管理學系 === 93 === ABSTRACT A COMPARISON OF THE FORECASTING MODELS OF THE NT$/US$ EXCHANGE RATE by Wu, Iee-Fung July 2005 ADVISOR(S): Dr. Goo, Yeong-Jia DEPARTMENT: DEPARTMENT OF BUSINESS ADMINISTRATION MAJOR：BUSINESS ADMINISTRATION DEGREE：PHILOSOPHY OF DOCTOR Taiwan is a small island, and its economy depends heavily on the international trade. The business or individual investment decisions are made according to the forecasts of future exchange-rate movements. They may make profits if those decisions are made correctly, but lose money if the decisions are made incorrectly. Therefore, it is an important issue to predict the movements of foreign exchange rates. In this paper, we use both the univariate and multivariate models to forecast the dynamic movements of the Taiwanese dollar versus American dollar exchange rate. Then we compare their performance using the criteria of accuracy （MAPE and MAE）and correctness. First, we compare the univariate models: the Genetically Optimized Neural Networks, GARCH, and ARIMA. Here we use up to 5 lagged values of the independent variable as the explanatory variables. Then, the multivariate models are compared. They are the Genetically Optimized Neural Network, TFARMA, and GARCH. Since the genetic algorithm has the ability of riddling significant variables, we put all the variables in the exchange-rate fundamental theories into the Neural Network（NN）. There are 8 macroeconomic indicators, each with 3 lagged values, so we have total of 24 parameters for the Ⅲ multivariate NN. Then the variables riddled from the genetic NN are put into the TFARMA and GARCH models. Empirical results: （1）Among the univariate models, GARCH has the smallest MAPE and MAE, so the GARCH model is the most accurate model, ARIMA is the second best.（2）Among the multivariate models, the GARCH and TRARMA models have the smallest MAPE and MAE respectively. They are equally good in terms of forecast accuracy.（3）The multivariate models beat univariate models in forecasting NT$/US$ exchange rate in terms of accuracy.（4）All the six models are no better than each other in terms of forecast correctness.（5）The hypotheses of H1b、H1c、H1d、H1e、H 2 and H3 are supported. While the hypotheses of H1a、H1f and H 1g are not supported because the coefficients of these parameters are not statistically significant. （6）The most significant macroeconomic indicators that affect the changes of NT$/US$ exchange rate are money supply, GNP, trade balance, foreign investment, and the lagged exchange rate itself. Keywords: Genetically Optimized Neural Network, TFARMA, GARCH, MAPE, MAE