Quantitative Easing on Asymmetric GARCH Value at Risk of NTD/USD

• Main Authors: Ping-Yen Chung, 鍾秉諺
• Other Authors: 蘇永成
• Format: Others
• Language: en_US
• Published: 2013
• Online Access: http://ndltd.ncl.edu.tw/handle/66855688281567616115

碩士 === 國立臺灣大學 === 財務金融學研究所 === 102 === The ultimate goal of this research is to find out which model fits the NTD/USD continuous rate of return in the U.S. QE period data best. Symmetric GARCHM and two asymmetric GARCHM models are utilized: one is GJR-GARCHM (for rotation) and the other is NA-GARCHM (for shift). AR(1), MA(1), or ARMA(1,1) is taken into the conditional mean equation. VaR forecasts under 99% and 95% confidence level are obtained and the forward test with violation analysis is done to examine the capability in market risk management and efficiency in capital reserves of each GARCHM model. The main discoveries are as follows: (i) If judged from the violation numbers under 99% confidence level, symmetric GARCHM models have the fewest, but GJR-GARCHM models are almost as good as symmetric GARCHM, and NA-GARCHM models are the worst. (ii) Considering the absolute mean VaR, symmetric GARCHM models have the largest value, GJR-GARCHM models are in the middle, and NA-GARCHM models have the smallest. Larger absolute mean VaR value means fewer violations. Fewer violations are safer, but the price is the stricter VaR and more capital reserves. Compared with symmetric GARCHM models, GJR-GARCHM models have somewhat more violations but somewhat less absolute mean VaR, which means fewer capital reserves are required in GJR-GARCHM models but their violations just increase a little bit and still in the allowance range of Basel Accord. Therefore, GJR-GARCHM models fit the NTD/USD continuous rate of return data in the U.S. QE period best and most efficiently. (iii) Because of the smallest absolute aggregate violation and absolute mean violation, MA(1)-GJR-GARCHM(1,1) is the best under 99% confidence level. GJR-GARCHM(1,1) is the best under 95% confidence level. However, the original GJR-GARCHM(1,1) model is still useful under 99% confidence level and symmetric GARCHM models are still full of reference values. (iv) In this research, there are two exceptions: AR(1)-NA-GARCHM(1,1) and ARMA(1,1)-NA-GARCHM(1,1). Their results are often different from others. NA-GARCHM models are the worst for the data in this research. (v) In this research, the involvement of AR(1) and ARMA(1,1) in the conditional mean equation causes the violation number, absolute aggregate, absolute mean, and absolute maximum violation to increase and absolute mean VaR to decrease. The possible reasons may be that NTD/USD is controlled in some degree by the Central Bank of Taiwan, and the Central Bank of Taiwan may intentionally does some interventions to make the NTD/USD market somewhat out of order and out of forecast ability. The interventions by the Central Bank of Taiwan may be the reason that the AR(1) is not capable of being fully effective in measuring the two-period serial correlation of Rt of the NTD/USD.