Frequentist Model Averaging with Nash Bargaining: A Stochastic Dominance Approach

Within the Frequentist Model Averaging framework for linear models, we introduce a multi-objective model averaging methodology that extends both the generalized Jackknife Model Averaging (JMA) and the Mallows Model Averaging (MMA) criteria. Our approach constructs estimators based on stochastic dominance principles and explores averaging methods that minimize multiple scalarizations of the joint criterion integrating MMA and JMA. Additionally, we propose an estimator that can be interpreted as a Nash bargaining solution between the competing scalar criteria. We establish the asymptotic properties of these estimators under both correct specification and global misspecification. Monte Carlo simulations demonstrate that some of the proposed averaging estimators outperform JMA and MMA in terms of MSE/MAE. In an empirical application to economic growth data, our model averaging methods assign greater weight to fundamental Solow-type growth variables while also incorporating regressors that capture the role of geography and institutional quality.