Research on Exchange Rate Forecasting Based on Information System Algorithm

• Main Authors: SUN AI, 孫艾
• Other Authors: Jui-Fang, Chang
• Format: Others
• Language: en_US
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/65v4u5

博士 === 國立高雄應用科技大學 === 國際企業研究所 === 105 === Abstract Along with the rapid development of financial globalization, our country faces complicated financial risks and foreign exchange risks. The subprime mortgage crisis, the sovereign debt crisis in areas with the euro, etc., spurred a global financial crisis and an economic recession and hence caused exchange rate prediction to evolve into an important economic issue, drawing wide attention. However, the foreign exchange market is a non-linear system with multiple variables, in which correlations between all factors are perplexing, exacerbating the difficulty of exchange rate prediction. As a complex non-linear system, exchange rate prediction methods have developed into a time series prediction from a parametric regression. However, in real applications, exchange rate fluctuations and varying trends are very complex, and the execution speed of the algorithm must surpass the variation speed of exchange rate at the same time as the exchange rate is precisely predicted. Although numerous studies pertaining to exchange rate prediction methods are currently available, the majority of the algorithms have been constrained by their complexity, and relevant research analysis has not been conducted on the applicability to data sets of the algorithms commonly used in exchange rate prediction. On account of this, three major method types are selected in this dissertation as the methodological basis of the research: the algorithm based on the empirical risk minimization principle, the algorithm based on the structural risk minimization principle, and the statistical filtering algorithm. Methods representative of algorithms theoretically applicable to exchange rate prediction are separated from the three major methods, namely, the Radial Basis Function Neural Network (RBFNN), the Least Squares-Support Vector Machine (LS-SVM), and the Kalman Filter (KF). The three methods mentioned above are selected in this dissertation to represent the three major methods, and explore their precision, efficiency, and applicability concerning exchange rate prediction. In addition, we contrast the three major types of algorithms according to test results, analyze the applicability of the different algorithms to data sets, and offer a novel train of thought and technological research on solving the problem of exchange rate prediction. The main sections of the dissertation are as follows: 1. The widely-used type of neural network, RBFNN, is introduced into the field of exchange rate prediction based upon the empirical risk minimization principle. This method both inherits the empirical risk minimization principle and introduces the kernel functions of RBF, has a higher prediction accuracy, simple structure, fast training speed, and different from the ordinary feedforward neural networks, with the best approximation performance and overall optimization. 2. This dissertation takes LS-SVM to represent the methods based on the structural risk minimization principle used for exchange rate prediction, since the methods based on the empirical risk minimization principle have lower prediction accuracy in circumstances of insufficient data. Addressing the issue of slower computation and convergence speeds of the traditional SVM algorithm, this method solved the problem of quadratic programming with LS on the premise of ensured minimal structural risks. Therefore, adopting this method may ensure the accuracy of the algorithm in cases of small sample size, as well as completing the prediction faster. 3. Addressing the deviation existing in both the prediction results from each type of method and in the exchange rate data, this dissertation proposes an exchange rate method based on the Kalman Filter. This method is representative of statistical filtering algorithms and may internally reduce noise in the two models to acquire more accurate prediction values. Therefore, adopting this method may effectively utilize the accuracy of the two models and allow the acquisition of more precise prediction values by statistical means.