Factor-Based Forecasting for High-Dimensional Models with Structural Breaks

In factor models, factors are the fundamental drivers of high-dimensional data.Accordingly, the estimated latent factors provide a powerful set of predictors formacroeconomic forecasting. However, structural breaks in macroeconomic data candisrupt factor structures, undermining the reliability and predictive power of estimatedfactors. This research develops a novel "rotated" factor estimator designed to mitigateimpacts of structural breaks on factor estimation, thereby enhancing forecast accuracycompared to traditional estimators.Our study considers both shift and rotational changes, establishing the asymptoticproperties of the rotated factor estimator alongside two alternative. We show that ournew estimator is more robust to the large shift changes and outperforms existingestimators under small rotational changes. In addition, we develop a criterion to combineforecasts generated from different factor estimators using data-driven weights. Thecombined forecast tends to outperform each individual forecast across various breaktypes and magnitudes.Based on our preliminary results, our project plans to explore properties of our newestimator under more general settings. We will analyze the asymptotic properties of therotated factor estimator in the presence of emerging or disappearing factors. We willalso examine forecast combination properties when the number of factors changes.Moreover, we intend to address the problem of unknown breakpoints by adapting ourcriterion to jointly select optimal weights and underlying breakpoints. Furthermore, wewill generalize the rotated factor estimator and analyze its properties under multiplebreaks. Beyond forecasting, we will assess if rotated factors yield improved impulseresponse function estimates in empirical applications.