- 學位論文
亞洲單一貨幣修正成份及傳染效果成因研究 ─運用類神經網路
The Study of Modified Component of Asian Currency Unit (ACU) and the Causes of Contagion Effect ─ An Application of Artificial Neural Network
摘要
本研究以亞洲十國(日本、印度尼西亞、馬來西亞、香港、南韓、新加坡、泰國、菲律賓、台灣和中國大陸) 為樣本國，探討亞洲單一貨幣(ACU)的建制過程以及預測分析，第一部分提供亞洲單一貨幣兌美元匯率預測績效的實證分析，運用倒傳遞類神網路(BPN)、回饋式類神經網路(RNN)、時間稽延回饋式類神經網路(TDRNN)以及GARCH作為分析模型，期間由1992年3月至2005年6月的匯率週資料為樣本，研究結果顯示類神經網路的預測績效優於GARCH並且倒傳遞神經網路(BPN)在多種不同樣本切割情形下呈現最佳的預測績效。 第二部分驗證亞元(ACU)中心價值的修正成份，使用倒傳遞類神經網路(BPN)進行修正成份分析，並與輻狀基底函數類神經網路(RBFNN)比較之，我們建構八個模型用以衡量總體經濟變數對亞元中心匯率的衝擊程度，實證證據支持本研究中的十個亞洲國家在建構亞元中心匯率時，將國外直接投資、外部負債以及銀行私部門所有權納入考慮，可以更有效地陳述亞元中心價值。 在驗證亞洲金融風暴的傳染成因部分，資料期間為1996年到1998年，利用調適性網路模糊推論系統作為實驗模型(ANFIS)，實證結果指出傳染效果受到緊密金融連結以及相似地總經條件的影響程度較大，此發現可提供管理當局藉由控制國家銀行間資金流動的金融聯繫和基礎地總經相似度來減少傳染發生的機率。
並列摘要
This research discusses constitution process and forecasting issue for Asian currency units (ACU) in terms of ten Asian countries including Japan, Indonesia, Malaysia, Hong Kong, South Korea, Singapore, Thailand, Philippine, Taiwan and China. The first, this article provides the analysis of forecasting performance for ACU against U. S. Dollars by employing variant of methods, e.g. Back-propagation neural network (BPN), recurrent neural network (RNN), time-delay recurrent neural network (TDRNN) and GARCH. The weekly exchange rate data for ACU was covered from March 1992 to June 2005. The results showed that ANNs has better performance than GARCH and BPN model presents prominent forecasting performance in most of division conditions. The second part is to identify the modified components for central value of Asian currency units (ACU). We utilize back-propagation neural network (BPN) to implement the optimal components analysis and compare the results with radial basis function neural network (RBFNN). We constitute eight models to evaluate the impacts of macroeconomic variables on central value of ACU. The empirical evidence supports that our sample countries need to take foreign direct investment, external debt and bank’s claim on private sector into account for expressing the central value of ACU more effectively. To verify contagion causes for Asia flu, the data was covered from 1996 to 1998. The experimental model is adaptive network-based fuzzy inference system (ANFIS). The empirical result was indicated that the contagion effect would most likely be influenced by tight financial linkage and similarly macroeconomic condition. This finding could provide solution for authority to diminish the probability of contagion via controlling financial correspondence in inter-bank flows and fundamentally macroeconomic similarity.