| Summary: | 碩士 === 國立臺灣科技大學 === 資訊管理系 === 93 === After 90’s several famous failed institutions which boomed up due to manipulate derivatives improperly, many countries in the world are paying attention to the risk management progressively. In the early 1990s, three events popularized VaR as a practical tool of risk management for financial institutions: In 1993, Group of Thirty published a groundbreaking report on derivatives practices. In 1994, JP Morgan launched its free RiskMetrics service. In 1995, the Basle Committee on Banking Supervision implemented market risk capital requirements for banks. As a consequence, VaR plays an important role in risk management.
The study is aimed at TSEC Weighted Stock Price Index in different trading periods. It uses different volatility forecasting models to improve the performance of VaR estimates, and discusses the predictability of different sample frequency and different volatility forecasting models. The study adopts ARFIMA-FIGARCH models to capture returns with long memory based on three kinds of density assumption including normal, Student-t, and skewed Student-t distributions. In addition, artificial intelligence methods regarding Adaptive Neuro-Fuzzy Inference System (ANFIS) is adopted to estimate Value at Risk. The empirical results of the study are summarized as the followings:
1. In the sampling periods, no models’ performance is superior to others due to different asset return’s distribution.
2. In terms of the overall performance of VaR model for the long trading positions, ANFIS models are the best. The second and successive models are ARFIMA(1,d,1)-FIGARCH(1,ξ,1), and GARCH(1,1).
3. In respect of the overall performance of VaR model for the short trading positions, ARFIMA(1,d,1)-FIGARCH(1,ξ,1) models are the best. The second and successive models are ANFIS, and GARCH(1,1).
4. ANFIS models are more stable than other models in each period. The asset returns’ fat-tailed and kurtosis or long memory property aren’t liable to influence ANFIS models.
|
|---|
