Group fused Lasso for large factor models with multiple structural breaks

This paper reformulates the identification of multiple structural breaks in factor loadings as a problem of detecting structural breaks in a factor regression, where the estimated pseudo factor corresponding to the largest eigenvalue is regressed on the remaining estimated factors. As a result, a group fused Lasso based estimation procedure is proposed to identify the break dates. Our procedure is practically easy-to-implement with standard statistical packages, overcoming the drawbacks of the existing methods that they often involve multiple tuning parameters and are computationally demanding in dealing with multiple unknown breaks. Theoretical properties of the proposed estimators are established, with a data driven choice of tuning parameter in the procedure. The Monte Carlo simulations and a real data application demonstrate that our procedure is fast implementable with desirable accuracy performance, and thus enjoys practical merits.