Tests of conditional predictive ability
|Speaker:||Raffaela Giacomini, University of California, San Diego|
|Date:||Wednesday 20 November 2002|
|Location:||Lecture Room D, Streatham Court|
The current framework for predictive ability testing (e.g., Diebold and Mariano, 1995; West, 1996) neglects important sources of forecast error and, as we argue, is not necessarily useful for deciding which model will give better forecasts in the future. We propose a general framework for predictive ability testing which represents a more realistic setting for economic forecasting. We argue that the way to think about the sources of forecast error econometrically is to use conditional inference rather than the unconditional inference embedded in the current framework. Some of the advantages of this approach are that we allow for heterogeneous, rather than stationary time series; that the tests can be used to compare both nested and non-nested models and that the test statistics have canonical limiting distributions. To illustrate the usefulness of the proposed tests, we generate forecasts for eight macroeconomic variables using a large number of predictors and compare the forecast performance of different parameter-reduction methods: a general-to-specific model selection approach, the "diffusion indexes" approach of Stock and Watson (2002) and the use of Bayesian VAR.