A comparison of multistep commodity price forecasts using direct and iterated smooth transition autoregressive methods

Abstract

The smooth transition autoregressive (STAR) modeling framework has gained popularity in commodity price analysis due to its ability to capture essential features of complex dynamics. This study addresses the questions of whether the improved in-sample fit of STAR models results in more accurate forecasts compared to linear autoregressive models, and whether direct or iterated multistep STAR methods yield more accurate multistep forecasts. In the STAR framework, either a bootstrap simulation is necessary to numerically approximate iterated multistep forecasts, or a range of horizon-specific STAR models needs to be estimated to generate direct multistep forecasts. The associated computational trade-off underscores the need for a better understanding of advantages one method may have over another. Based on the analysis of 25 agricultural and nonagricultural commodity prices, this study finds that even when the STAR models appear to well approximate complex commodity price dynamics, they offer little advantage, and indeed, in most instances present as inferior alternatives to the basic autoregressive framework for multistep commodity price forecasting.