A2032
Title: Changepoint detection under the heterogeneous factor strength
Authors: Catherine Liu - The Hong Kong Polytechnic University (Hong Kong) [presenting]
Abstract: Detecting group structural instability is urgently needed for high-dimensional grouped vector-valued sequences in finance and many other application fields; overlooking it may incur severe risks. The group structure gives rise to mixed factor strengths under the latent factor structure containing both group and global signals. Existing methods fail to detect group structural breaks because of weak group signal strength. The global and group structural abruptness are modeled by a break-pair, which allows for a versatile change profile, including segmentation and ordering of breaks. In the coexistence of strong and weak factors, a new estimation paradigm is developed to guarantee the consistency of quasi-maximum likelihood-type estimators of the break-pair. Specifically, the strong and weak factors are separately estimated and the convergence rate of each estimator is enhanced by removing either the influence of the other or their interaction. This strategy builds upon a three-phase partition of the eigenvalues of the sample covariance matrix of observations, where weak group factors bring about one phase. Simulations demonstrate the unique ability to detect group structural change. The application is illustrated with the HKEX stock and S\&P500 datasets. The proposed method remains applicable, after certain simplification, when the group structure is unknown or absent.
