Alternative bvar models for forecasting inflation

Abstract

This paper investigates the use of different priors to improve the inflation forecasting performance of BVAR models with Litterman’s prior. A Quasi-Bayesian method, with several different priors, is applied to a VAR model of simulated data as well as to the Australian economy from 1978:Q2 to 2006:Q4. A novel feature with this paper is the use of g-prior in the BVAR models to alleviate poor estimation of drift parameters of Traditional BVAR models. Some results are as follows: (1) In the Quasi-Bayesian framework, BVAR models with Normal-Wishart prior provide the most accurate forecasts of Australian inflation; (2) Generally in the parsimonious models, the BVAR with g-prior performs better than BVAR with Litterman’s prior; (3) In simulated data, the BVAR model with g-prior produces more accurate forecasts of driftless variable in the long-run horizons (first and second year forecast horizons).