Title: A new Bayesian MIDAS approach for flexible and interpretable nowcasting
Authors: Galina Potjagailo - Bank of England (United Kingdom) [presenting]
David Kohns - Heriot-Watt University (United Kingdom)
Abstract: The T-SV-t-BMIDAS model is proposed for nowcasting quarterly GDP growth. The model incorporates a long-run time-varying trend (T) and t-distributed stochastic volatility accounting for outliers (SV-t) into a Bayesian multivariate MIDAS. To address the high dimensionality of the model, to account for group correlation in mixed frequency data, and to make the model interpretable to the policymaker, we propose a new group-shrinkage prior combined with a sparsification algorithm for variable selection. The prior flexibly accommodates between-group and within-group sparsity and allows communicating the importance of predictors over the data release cycle. We evaluate the model for UK GDP growth nowcasts over the period 1999 to 2021. Our model is competitive before the pandemic relative to various benchmark models, while yielding substantial nowcast improvements during the pandemic. First, accounting for a long-run trend and t-distributed stochastic volatility substantially improves forecast performance relative to a simple BMIDAS. Second, the shrinkage prior enhances nowcast performance by inducing group-wise sparsity while enabling the model to shift flexibly between signals. During the Covid-19 pandemic, the model reads stronger signals from indicators for services, which reflected spending shifts related to lockdowns, and less from production surveys. This helps to nowcast the recovery from the shock precisely, and to update the nowcast for the pandemic-related trough sooner.