Introduction

We argue that macro-models in inflation-targeting central banks are too narrowly focused to provide accurate probabilistic forecasts. Despite the explicit consideration of model uncertainty afforded by Bayesian estimation techniques, the models prominent in central banks devote insufficient attention to ‘uncertain instabilities’. Too much consideration is paid to refining a single preferred but inevitably misspecified model. A product of this oversight is that the 2007-vintage workhorse monetary policy models had little (or nothing) to say about the probability of ‘tail’ events, which now dominate the debate over the causes of, and remedies for, the recent global financial crisis.

In our view, the next generation of macro-modellers should address this deficiency while preserving the architecture of dynamic non-linear modelling. We propose a methodology adapted from the weather-forecasting literature known as ‘ensemble modelling’. In this approach, uncertainty about model specifications – e.g. initial conditions, parameters and boundary conditions – are explicitly accounted for by constructing ensemble predictive densities from a large number of component models. The components allow the modeller to explore a wide range of uncertainties; and the resulting ensemble ‘integrates out’ these uncertainties using time-varying post-data weights on the components.

We provide two economic examples of the ensemble methodology. In the first, we consider a policymaker (recursively) selecting a linear combination of disaggregate predictives to produce an ensemble forecast density for inflation. Each component of the ensemble comprises a univariate autoregressive model using a single disaggregate series.