Forecasting with a parsimonious subset VAR model

Abstract

This paper suggests using a unit t-value criterion in imposing restrictions on lags to formulate a subset vector autoregressive (VAR) model for the purpose of point forecasts. Among any other alternative models nested to the initial VAR model, this less restrictive modeling strategy produces the smallest log determinant of the residual covariance matrix adjusted by degrees of freedom. Each equation of the finally derived subset VAR model has a maximized R̄2 adjusted by degrees of freedom in samples and consequently a minimized 1-step-ahead prediction error in out-of-samples. The applicability of this modeling strategy is excised to the case of a bivariate VAR model for output growth and inflation.