| My dissertation mainly addresses two issues in econometric modeling: the in-sample properties of a model when the sample size is finite and a model's out-of-sample predictive ability when the sample size is large. While the two issues at a first glance seem to be isolated, they both in fact hinge upon asymptotics. It is the second-order asymptotic results that we are interested in to conduct finite sample inference. On the other hand, we can use the first-order asymptotic results to investigate financial series. In essence, this difference depends on the sample size at hand. Both approaches can be used for the purpose of model selection. Chapters 1 to 4 of this thesis discuss the finite sample inference part and Chapter 5 and Chapter 6 present some interesting results in applied finance.; In Chapter 1, I present the analytical general results on the first two approximate moments of root-n-consistent econometric estimators in models with non-IID observations identified by some suitable moment condition. As an illustration, the application to some commonly used econometric models is given. The next three chapters can be regarded as special cases of the first chapter. In particular, Chapter 2 deals with time series models, Chapter 3 deals with spatial models, while Chapter 4 discusses the bias result for a particular time series model, the Value-at-Risk (VaR) model. The results indicate that small sample bias may be quite significant in these models and the asymptotic-theory based inference in finite samples may be misleading. Next, in Chapters 5 and 6, we conduct a comparison of the predictive performance of various VaR models for five Asian emerging markets and density forecast models for the S&P 500 and Nasdaq return series. The out-of-sample comparisons are implemented through a statistically rigorous test, which accounts for the data-snooping problem. Empirical findings in Chapter 5 suggest that filtering is crucial in risk forecasting and Chapter 6 demonstrates that the choice of conditional distribution specification is a much more dominant factor in determining the quality of density forecasts than the choice of volatility specification. |
