Describing the Dividend Process

Dividend payments are the primary reason that most stocks have value. We construct a model of the time evolution of dividends for individual securities that ties that evolution to one or more unobserved aggregate factors. Specifically, we use a state space model in which the aggregate factors are unobserved states and the dividends on individual securities are signals that depend on the estimated states. Because dividends cannot be negative, and zero dividends are quite common, the usual linear, Gaussian Kalman filter cannot be applied directly. Instead, we use a Bayesian approach with diffuse priors in a Gibbs sampler that combines a Tobit block for a latent desired dividend level with a standard Kalman filter. While in principle straightforward, the large number of observed stocks makes estimation computationally challenging. Joint work with Kwok Ping Tsang.