Bayesian Analysis of Linear Factor Models with Latent Factors, Multivariate Stochastic Volatility, and APT Pricing Restrictions

We analyze a new class of linear factor models in which the factors are latent and the covariance matrix of excess returns follows a multivariate stochastic volatility process. We evaluate cross-sectional restrictions suggested by the arbitrage pricing theory (APT), compare competing stochastic volatility specifications for the covariance matrix, and test for the number of factors. We also examine whether return predictability can be attributed to time-varying factor risk premia. Analysis of these models is feasible due to recent advances in Bayesian Markov chain Monte Carlo (MCMC) methods. We find that three latent factors with multivariate stochastic volatility best explain excess returns for a sample of 10 size decile portfolios. The data strongly favor models constrained by APT pricing restrictions over otherwise identical unconstrained models.