Iliyan Georgiev (Università di Bologna)
Abstract:
A number of time-series features that are difficult to detect and model (e.g., near unit-root nonstationarity and regime changes) make it desirable to perform inference in models constructed conditionally on the problematic series. The adverse features, however, often reappear in the limit distributions of test statistics, potentially compromising the usefulness of the conditioning approach. A possible solution is to approximate the relevant limit distributions by means of a fixed-regressor bootstrap. First, we introduce convergence concepts that formalise the asymptotic validity of the bootstrap approximation and correct errors in the existing literature where the use of an inadequate probabilistic framework has lead to improper understanding of the bootstrap. Second, we discuss econometric examples of inference related to predictive regression and long-run relationships.
Contact person
Giuseppe Cavaliere e Alessandra Luati