Conditional Model Confidence Sets with an Application to Forecasting Models

An new approach for model selection which nests as special cases Hansen, Lunde and Nason's model confidence set procedure (MCS) and Giacomini and White’s conditional testing framework (CPA) and contains the models which are expected to provide the best future forecasts given certain conditional information.

Abstract

In this paper we develop an new approach for model selection which nests as special cases Hansen, Lunde and Nason’s (2004) model confidence set procedure (MCS) and Giacomini and White’s (2004) conditional testing framework (CPA). The resulting conditional model confidence sets contain the models which are expected to provide the best future forecasts given certain conditional information. Our approach allows the comparison of a large number of possibly nested models and is also robust to heterogeneity in the data. In an empirical application we evaluate the relative performance of several forecasting models applied to UK inflation and growth data. Our results illustrate the fact that the use of conditional information may be useful in situations where the relative forecast performance is dependent on the state of the economy.