Factor models of security returns aim to decompose an asset return covariance matrix into a systematic component and a specific risk component. Standard approaches like PCA and maximum likelihood suffer from several drawbacks including a lack of robustness as well as their strict assumptions on the underlying model of returns. We survey some modern, robust methods to uniquely decompose a return covariance matrix into a low-rank component and a sparse component. Surprisingly, the identification of the unique low rank and sparse components is feasible under mild assumptions. We apply the method of Chanrasekaran, Parillo and Willsky (2012) for latent graphical models to decompose a security return covariance matrix. The low rank component includes the market and other broad factors that affect most securities. The sparse component includes thin factors such as industry and country, which affect only a small number of securities, but in an important way. We illustrate the decomposition on simulated data, and also an empirical data set drawn from 25,000 global equities.
- Start date: 2016-04-12 11:00:00
- End date: 2016-04-12 12:30:00
- Venue: 639 Evans Hall at UC Berkeley
- Address: 639 Evans Hall, Berkeley, CA, 94720