Marcellino, Hauzenberger, Pfarrhofer and Stelzer develop a new model for mixed frequency data

Machine learning (ML) methods have emerged as flexible tools for capturing nonlinearities in macroeconomic and financial relationships. Contributions in the macroeconometric ML literature suggest that such approaches can improve predictive accuracy by uncovering complex interactions and nonlinear patterns among variables, while offering robustness to various kinds of misspecification. The abundance of potential predictors in a “Big Data” environment is another key aspect — and relevant predictors are often measured at different frequencies, which calls for a framework capable of handling frequency mismatches.

A recent paper coauthored by Massimiliano Marcellino (in photo), director of Baffi Centre and Asset and Risk Management (ASSET) Research Unit Director, explicitly acknowledge the mixed frequency nature of economic data, thanks to a direct (horizon-specific) Gaussian Process (GP) regression. The paper, entitled “Direct gaussian process predictive regressions with mixed frequency data” was written by Massimiliano Marcellino, Niko Hauzenberger, Michael Pfarrhofer and Anna Stelzer and has just been published in the BAFFI Working Papers Series (available on SSRN and REPEC).

They leverage techniques for handling large crosssections of predictors sampled at higher frequencies (HF) than that of the target variable. This involves splitting each HF variable into multiple low frequency (LF) ones, an approach also known as blocking. Rather than assuming a parametric functional form with priors on the parameters of blocking-based regressions, they place a Gaussian Process (GP) prior directly on the conditional mean function. Their main empirical focus is on out-of-sample predictive accuracy and robustness for nowcasting and short-horizon predictions. Their proposed framework is complementary to existing more structural approaches, offering robustness, good predictive accuracy, and the opportunity to account for nonlinearities of unknown form with relative ease.

The main contributions of the paper are threefold. First, the authors propose a framework to directly model nonlinear relationships in mixed frequency datasets with GPs, without the need for computationally costly filtering and smoothing algorithms. Second, they discuss how their proposed blocking, frequency aggregation, and compression techniques for HF predictors yield distinct implications for kernels that address matters of correlation and provide a form of regularization in (groups of) HF predictors. Third, they demonstrate in a thorough out-of-sample application with US data that our approach is competitive and can improve short-horizon predictions (including backcasts and nowcasts) of GDP growth and inflation relative to several parametric and nonparametric benchmarks. Predictive performance for the proposed models remains remarkable robustness across a range of specification choices and information sets.