The Estimation of Vector Multiplicative Error Model on Contaminated Data and Its Applications in Forecasting Volatilities.

摘要

This thesis studies the estimations of vector Multiplicative Error Model (MEM) under different kinds of model mismatches and its application in forecasting. In the first part of the thesis, two estimation methods, Maximum Likelihood (ML) method and Generalized Method of Moments (GMM), which have previously been used on vector MEM, are compared through different situations of data contaminations. From the comparison results it is found that both ML and GMM estimators are suspected to outliers in data. Therefore in the second part of the thesis a novel estimator is proposed: Weighted Empirical Likelihood (WEL) estimator. It is shown to be more robust than ML and GMM estimators in simulations, and also in forecasting realized volatility and bipower volatility of S&P 500 stock index including the current financial crisis period. The forecast ability of vector MEM is further addressed in the third part of the thesis, where an alternative decomposition of realized volatility is proposed, and vector MEM is used to model and forecast the two components of realized volatility. From the realized volatility forecasts of S&P 500, NASDAQ and Dow Jones, this decomposition together with vector MEM are illustrated to have superior performances over three competing models which have been applied on forecasting realized volatility before.