Opening the Black Box: Structural Factor Models versus Structural VARs

In this Paper we study identification in dynamic factor models and argue that factor models are better suited than VARs to provide a structural representation of the macroeconomy. Factor models distinguish measurement errors and other idiosyncratic disturbances from structural macroeconomic shocks. As a consequence, the number of structural shocks is no longer equal to the number of variables included in the information set. In practice, the number of structural shocks turns out to be small, so that only a few restrictions are needed to reach identification. Economic interpretation is then easier. On the other hand, with factor models we can handle much larger information sets - including virtually all existing macroeconomic information. This solves the problems of superior information and fundamentality and enables us to analyse the effects of the shocks on all macroeconomic variables. In the empirical illustration we study a set of 89 US macroeconomic time series, including the series analysed in the seminal paper of King et al. (1991). We find that the system of impulse response functions of these series is non-fundamental and therefore cannot be estimated with a VAR. Moreover, unlike in King et al. (1991), the impulse response functions of the permanent shock are monotonic and therefore more credible if the permanent shock is interpreted as technical change.