Optimal GMM Model Averaging 

This paper considers moment conditions model averaging (MA) estimators in the
Generalized Method of Moments (GMM) framework. Optimal weights are chosen so as
to minimise higher-order asymptotic mean squared error (AMSE) of the MA
estimator. We show that the objective function has a closed form expression
and, contrary to other methods, our estimator is applicable in practice.

We derive some asymptotic properties assuming correctly specified models.
We contrast the performance of our averaging approach with existing instrument
selection alternatives, both analytically for a simple IV example and by means
of Monte Carlo experiments with a nonlinear setup, showing that model averaging
compares favourably in many relevant cases.

Co-authored with Luis Martins.