Title: Forecasting with Bayesian adaptive penalized mixed-frequency regressions
Authors: Clement Marsilli - Banque de France (France) [presenting]
Matteo Mogliani - Banque de France (France)
Abstract: A new approach is proposed to forecast with mixed-frequency regressions (MIDAS) that address the issues of estimation and variable selection in presence of a large number of predictors. Our approach is based on adaptive penalized regression models (Lasso, Group Lasso, and Elastic-Net) and relies on Bayesian techniques for estimation. In particular, the penalty parameters driving the model shrinkage are automatically fine-tuned via an adaptive MCMC algorithm, which is computationally efficient compared to the standard EM algorithm. Simulations show that the proposed penalized MIDAS models significantly outperform a benchmark represented by an optimal combination of single-predictor MIDAS regressions. When applied to US GDP, the results suggest that our models produce significant out-of-sample predictive gains compared to several alternative models.