NBER-NSF Time Series Conference; University of California at Davis, USA

Abstract for "Constrained Factor Models: Parsimonious Estimation for Conditional Covariance Matrices and the Yield Curve" by David S. Matteson

In many applications factor models may prove empirically adequate and provide sufficiently flexible dynamics when estimating time-varying conditional covariance matrices for financial asset returns. The use of prior information, such as an asset’s industry or sector, to impose constraints is shown to provide a more parsimonious model for estimation when the dimension is large. Statistical decompositions of the historical variation in the term structure of interest rates suggest that the level, slope, and curvature of the yield curve are the principal sources of variation. A constrained factor model for the yield curve is considered using these sources to identify appropriate constraints.

Estimation is performed first using constrained principal component analysis; we then consider decorrelation of higher order moments in a constrained independent component analysis and contrast the merits of each approach for these applications. Introduction of constraints allows prior information or economic theory to be incorporated into estimation resulting in simpler forecasting, straightforward interpretation of factors and dimension reduction.

This is based on joint research with Ruey S. Tsay.