Any data-consistent macroeconomic model must depend, to a greater or lesser extent, on the goodness-of-fit of predictive regressions. But a long-standing feature of macroeconomic forecasting has been that a wide variety of multivariate models have struggled to out-predict univariate representations. We seek an explanation of this feature in terms of population properties. We ask: if we just observed the time-series history of the variable of interest, what would this tell us about the properties of any multivariate structural model, and its associated predictive regression, that is consistent with this history? We illustrate our analysis using data on U.S. inflation. We find that, especially in recent years, univariate properties of inflation dictate that a) any predictor of inflation must be near-IID; and b) even the best possible multivariate predictive model would struggle to beat a univariate model.

Joint work with Donald Robertson (Cambridge) and Stephen Wright (Birkbeck, London)