Averaging estimators for discrete choice by M-fold cross-validation

Considering model averaging estimation under multinomial and ordered logit models, we propose a weight choice method by minimizing an M-fold cross-validation criterion. Unlike Jackknife model averaging in Hansen and Racine (2012), our method deletes one group observations instead of deleting only one observation. Therefore, under the maximum likelihood estimation framework, the computational cost of our method is greatly reduced. We prove the asymptotic optimality of the proposed model averaging estimator under certain assumptions. The simulation results show that our proposed method produces more accurate forecasts than other model selection and averaging methods.