Forecasting With Dynamic Factor Models Estimated by Partial Least Squares

2026/04/17 by Samuel Rauhala

paper · doi:10.1002/for.70158

crossref issued 2026/04/17 · crossref published 2026/04/17 · crossref published-online 2026/04/17 · crossref created 2026/04/18 · crossref deposited 2026/08/05 · crossref indexed 2026/08/05 · crossref published-print 2026/09/01

Abstract

ABSTRACT Dynamic factor models (DFMs) have found great success in nowcasting and short‐term macroeconomic forecasting when incorporating large sets of predictive information. The factor loadings are typically estimated cross‐sectionally with principal component analysis (PCA) or maximum likelihood (ML), which ignore whether the factors have predictive power. We suggest two novel alternative approaches using partial least squares to estimate large vector autoregressions (VARs) and DFMs, which take the dynamic dependencies better into account. Our Monte Carlo simulations and forecasting results for the Finnish GDP growth show that these methods generally perform on par with and under certain conditions better than the existing approaches.

Citations

• The three-pass regression filter: A new approach to forecasting using many predictors
• Comparing Predictive Accuracy, Twenty Years Later: A Personal Perspective on the Use and Abuse of Diebold–Mariano Tests
• Nowcasting: The real-time informational content of macroeconomic data
• PLS regression methods
• MAXIMUM LIKELIHOOD ESTIMATION OF FACTOR MODELS ON DATASETS WITH ARBITRARY PATTERN OF MISSING DATA