Bayesian Model Averaging with non-conjugate Priors

This paper considers the case of non-conjugate prior distributions for Bayesian model averaging (BMA). Although the natural conjugate setting is the default choice for BMA, mainly for reasons of analytical tractability, it has come under considerable criticism due to its unrealistic assumptions about prior information, among others. In this study, we extend the literature by considering two special cases of the multivariate Student-𝑡 distribution. We obtain closed-form solutions using Laplace approximations and apply our techniques to a controlled numerical experiment and cross-country growth regressions. Our results show that, under fine tuning of the hyperparameters, the proposed approach has similar performance to the conjugate alternatives on synthetic datasets, whereas in real data it favors, on average, more parsimonious models than the conjugate alternatives and also exhibits superior predictive performance.