Long-Run Performance of Bayesian Model Averaging