Cause of BLR–GP performance differences across nowcast and forecast horizons

Establish whether the observed pattern in the mixed-frequency MIDAS nowcasting/forecasting exercise—Bayesian linear regression (BLR) ranking first for nowcasts while Gaussian process (GP) models often perform better for forecasts—is explained by a decline in the importance of nonlinearities as more within-quarter high-frequency information becomes available, thereby reducing the need for nonparametric flexibility in the conditional mean function.

Background

The paper compares linear and nonparametric mixed-frequency regression frameworks for nowcasting and forecasting U.S. GDP growth and GDP deflator inflation using quarterly targets and monthly predictors. The conditional mean is modeled using (i) Bayesian linear regression (BLR) with global–local shrinkage, (ii) Gaussian processes (GP), and (iii) Bayesian additive regression trees (BART), with various MIDAS lag-weighting schemes and both homoskedastic and stochastic volatility error structures.

Empirically, the authors find that BLR frequently attains the best performance for nowcasts, particularly for output, whereas GP models often perform better for forecasts at longer horizons. They explicitly conjecture that the difference may arise because, as additional within-quarter information becomes available approaching the nowcast, the role of nonlinearities in the conditional mean diminishes, making flexible nonparametric specifications less necessary. Determining whether this mechanism indeed explains the observed performance pattern remains unproven in the paper.