Predictive model averaging with parameter instability and heteroskedasticity

This paper proposes a frequentist model averaging approach in the presence of parameter instability and heteroskedasticity. We derive optimal weights combining the stable and break specifications of a predictive model, with the weights from minimizing the leave-one-out cross-validation information criterion (CV). We characterize the asymptotic distribution of the CV and provide the analytical expressions of the feasible optimal CV weights. Our simulations and applications forecasting the US and Taiwanese GDP growth demonstrate the superior performance of the CV model averaging relative to other methods such as the Mallows averaging, the approximate Bayesian averaging, and equal weighting.