AbstractIn this paper, using composite predictors we examine whether the use of survey information on consumer confidence would have helped to predict fluctuations in economic activity. We also consider the implications of the new literature on time‐series modelling when the underlying processes are not stationary. We then examine what implications this has for the construction of composite predictors. We find that it is essential that any forecast—used as part of a composite predictor—is co‐integrated with the outcome. It is likely that this will hold in practice, but if it does not then the forecast errors will be non‐stationary and the interpretation of the composite predictor