Time series models for realized covariance matrices based on the matrix-F distribution

Jiayuan ZHOU, Feiyu JIANG, Ke ZHU, Wai Keung LI

Research output: Contribution to journalArticlespeer-review

3 Citations (Scopus)

Abstract

We propose a new Conditional BEKK matrix-F (CBF) model for time-varying realized covariance (RCOV) matrices. This CBF model is capable of capturing a heavy-tailed RCOV, which is an important stylized fact, but is not handled adequately by Wishart-based models. To further mimic the long-memory feature of an RCOV, we introduce a special CBF model with a conditional heterogeneous autoregressive structure. Moreover, we provide a systematic study of the probabilistic properties and statistical inferences of the CBF model, including exploring its stationarity, establishing the asymptotics of its maximum likelihood estimator, and giving new inner-product-based tests for model checking. In order to handle a large-dimensional RCOV matrix, we construct two reduced CBF models: the variance-target CBF model (for a moderate but fixed-dimensional RCOV matrix), and the factor CBF model (for a high-dimensional RCOV matrix). For both reduced models, the asymptotic theory of the estimated parameters is derived. The importance of our methodology is illustrated by means of simulations and two real examples. Copyright © 2022 Institute of Statistical Science, Academia Sinica.
Original languageEnglish
Pages (from-to)755-786
JournalStatistica Sinica
Volume32
Issue number2
DOIs
Publication statusPublished - Apr 2022

Citation

Zhou, J., Jiang, F., Zhu, K., & Li, W. K. (2022). Time series models for realized covariance matrices based on the matrix-F distribution. Statistica Sinica, 32(2), 755-786. doi: 10.5705/ss.202019.0424

Keywords

  • Factor model
  • Heavy-tailed innovation
  • Long memory
  • Matrix-F distribution
  • Matrix time series model
  • Model checking
  • Realized covariance matrix
  • Variance target

Fingerprint

Dive into the research topics of 'Time series models for realized covariance matrices based on the matrix-F distribution'. Together they form a unique fingerprint.