Skip to main content Skip to navigation

David F. Hendry

Resolving Five 'Intractable' Problems by Automatic Model Selection

 

The five problems are: unbiased estimation despite selection; selecting from a set of perfectly collinear variables; selecting if more variables than observations; selecting a non-linear model; and selecting a simultaneous equations model, when there are no a priori restrictions but multiple endogenous variables. All five seem intractable in conventional approaches, but all five yield to automatic model selection. The important special case of selecting a regression saturated by impulse dummies (one indicator for every observation) will be discussed, where we can derive the sampling distributions of estimators of the mean and variance in a simple location-scale model under the null that no impulses matter. The results are confirmed by Monte Carlo, which also shows power against alternatives of interest. An example will illustrate the latest version of PcGets to highlight our future research directions as more empirical modelling decisions can be explored automatically.

Incorporating research with Jennifer Castle, Soren Johansen, Hans-Martin Krolzig and Carlos Santos.