Does Noise Hurt Economic Forecasts?

Abstract

This paper explores whether variable selection enhances economic forecasting. While economists often remove noise from predictors, we show that economic forecast models are not sparse if the outcome is driven by latent factors. We also prove a compelling result that including noise in predictions yields greater benefits than excluding it.  Empirically, we apply this method to four common forecasting applications including forecasting the U.S. inflation rate and obtain results that surpass many commonly used models that rely on dimension reduction or variable selection. 

Keywords: machine learning, factor model, double descent, dense signals

JEL Classification: C1,G17,E47