Thick modeling

The standard method of model building is to consider just a single specification and then discuss its estimation, interpretation, and output. In practice many specifications are available so the usual technique is to choose the best, according to some criterion such as by testing, and then use that to produce an output. However, this means that any information in the alternative specifications is not being utilized but theories such as those concerned with portfolio selection and the combining of forecasts suggest that just using the best asset or forecast is sub-optimal. The alternative is to keep all close specifications, find their outputs that relate to the purpose of the models such as essential parameter estimates (elasticities, say), impulse responses, policy simulations, hypothesis tests or forecasts, and pool these values. If there are many models involved a simple method of combining is to give equal weights after removing a few outliers. To form confidence intervals for the combination alternative specifications can be found using bootstrap techniques. Examples are provided. (C) 2003 Elsevier Science B.V. All rights reserved.