This dissertation has three independent chapters, Chapters 2, 3, and 4. Chapter 2 introduces similar test statistics for the structural parameter in a linear model. The tests are based on Generalized Empirical Likelihood (GEL) techniques. Chapter 3 is concerned with the bias-reduced estimation of the long-memory parameter in a long-memory time series. The final chapter introduces a flexible approach to combine econometric models. The approach can be viewed as a generalization to the Markov-switching literature. As an empirical application, the model is applied to the prediction of conditional variance in stock market time series.; Chapter 2 introduces two new statistics that can be used to test simple hypotheses involving the structural parameter vector in a linear single equation instrumental variables model. The finite sample size properties of tests based on classical statistics such as the Wald or likelihood ratio statistic depend crucially on the strength or weakness of identification. The main feature of the new statistics is that they have correct finite sample sizes independent of the strength or weakness of identification. The statistics are based on GEL techniques. The first statistic equals the criterion function of the GEL estimator and has a likelihood ratio interpretation. The second statistic is given as a quadratic form in the first order condition of the GEL estimator and has an interpretation as a Lagrange-multiplier statistic.; Chapter 3 proposes a new bias-reduced semiparametric estimator of the long-memory parameter. Existing semiparametric estimators such as the Geweke and Porter-Hudak estimator, approximate the short-run component of the spectrum by a constant around the origin. This can result in a considerable finite sample bias. The new estimator reduces the bias by replacing the constant in the approximation by an even polynomial. Under weak assumptions, it is shown that the rate of convergence to zero of the asymptotic RMSE is faster than the rate for the GPH estimator.; Chapter 4 introduces a general method to link variable length Markov chain models for discrete time series with real-valued time series models. The resulting model class, called Dynamic Combination of Models (DCM), incorporates the idea of model mixing and model switching. The transition probabilities for the regime in the next period depend on a possibly long history. Unlike the Markov-switching literature the length of the history is not pinned down, but estimated together with the parameters of the model. As an empirical application, a particular GARCH(1,1) DCM is used for the prediction of the conditional variance in stock market time series.
展开▼