Review of guidelines for the use of combined forecasts

A large literature has evolved in the thirty years since the seminal work on combining forecasts. Despite this, when evaluating performance we only look at measures of accuracy and thus ignore most of the rigour of time series analysis. Furthermore, the output from a combination of forecasts is just a single point estimate which is insufficient for business planning models that take explicit account of risk and uncertainty. In this paper, we review evidence on the performance of different combining methods with the aim of providing practical guidelines based on three properties of the forecast errors: variance, asymmetry and serial correlation. The evidence indicates that using different criteria leads to distinct preferences, and that the properties of the individual forecast errors can strongly influence the characteristics of the combination's errors. We show that a practical approach to combining also requires a degree of judgement on the attributes of error specification.