Abstract:
This paper evaluates a novel sampling algorithm, called shotgun stochastic search (S³), for Bayesian model averaging in the context of finding predictors for inflation when the set of potential predictors is large. This is a relevant case in the forecasting literature, where often hundreds of predictors are compared with autoregressive distributed lag models for inflation. With such a large model space, standard Bayesian approaches like MCMC model composition (MC³) tend to converge slowly. On the other hand, S³ systematically searches in the neighborhood of good models and concentrates on regions of high posterior probability in the model space. We carry out a Monte Carlo simulations to compare the computational efficiency of S³ to MC³, based on standard data generating processes from the literature. When many potential predictors are available, S³ outperforms MC³. In an empirical exercise, we apply the two algorithms to find predictors for US inflation from a set of about one hundred indicators and their lags. S³ absorbs posterior mass much quicker than MC³ and makes Bayesian estimation of the standard inflation equations with many predictors computationally feasible.
