Economic Time Series Predictions and the Illusion of Support Recovery

Abstract

We investigate whether forecast algorithms advanced in statistics and econometrics are capable of detecting the number and identities of relevant predictors for economic time series. Further, we study the relation between support recovery properties and point predictive accuracy. A novel feature of our approach is that we link results from empirical studies with simulation analyses by using the degree of predictability inferred via the former to pin down realistic signal-to-noise ratios for the simulations. While methods that combine feature selection and shrinkage exhibit good support recovery properties in low noise environments, none of the methods unveils the true number and identities of relevant predictors in realistic settings. Nevertheless, producing useful point forecasts is possible, especially when using forecast combinations.

Keywords: Feature selection, Signal-to-noise ratio, Shrinkage, Forecast combinations

JEL Classification: C53, C55