Applications of High-Dimensional Factors and Machine Learning in Behavioral Macroeconomic Forecasting

Abstract

This paper develops a unified macroeconomic forecasting framework that integrates high-dimensional factors, machine learning methods, and behavioral variables, and systematically examines their joint performance in forecasting growth and inflation. First, common high-dimensional factors are extracted from a large panel of macroeconomic indicators in order to compress the shared information contained in data-rich environments. Second, behavioral information, including consumer confidence, expectation measures, sentiment indicators, and uncertainty proxies, is incorporated into the forecasting system to capture the role of beliefs, perceptions, and sentiment in macroeconomic fluctuations. Third, the out-of-sample forecasting performance of benchmark time-series models, factor-based models, and machine-learning-based models is compared within a unified framework. The results show that high-dimensional factors significantly improve upon traditional benchmark models, behavioral variables provide additional forward-looking information beyond common latent factors, and the hybrid framework combining high-dimensional factors, behavioral variables, and machine learning delivers the strongest forecasting performance. Further analysis indicates that the predictive contribution of behavioral variables is strongly state-dependent and becomes especially pronounced at short horizons and during periods of elevated uncertainty. The findings suggest that improvements in macroeconomic forecasting depend not only on high-dimensional information compression and nonlinear modeling capacity, but also on the effective incorporation of behavioral signals related to expectations, sentiment, and uncertainty. This study provides new evidence on the empirical relevance of behavioral macroeconomics for forecasting and offers useful implications for future macroeconomic monitoring and policy analysis.