Markov Switching Dynamics in REIT Returns : Univariate and Multivariate Evidence on Forecasting Performance

The fact that REIT returns display rich dynamic time series properties, such as conditional heteroskedasticity and time-varying risk premia, has recently come to the forefront of the real estate finance literature. In this paper we document the presence of Markov switching regimes in expected returns, variances, and in the implied reward-to-risk ratio of REIT returns and compare them to properties of non-REIT stock and bond returns. In a Markov switching model conditional means, variances, and covariances depend on an unobservable Markov state variable that can assume a finite number of values characterizing the state of the financial market at each point in time. We analyze the relative in-sample descriptive performance of regime switching and alternative time series models such as multivariate GARCH. Interestingly, our evidence suggests that regime switching techniques may more successfully fit the monthly return series for REITs, stocks, and bonds over the period 1972-2008.When this evidence is extended to a multivariate setting in which REIT, stock, and bond returns are modeled jointly, we find that the data call for the specification of rich structures with as many as four separate states. This results from the absence of complete synchronicity among the regimes that characterize univariate REIT, stock, and bond returns. This finding may have considerable importance as it implies that for REITs there may exist an important de-coupling between the standard notion of (linear) correlation and of “association,” in the sense that REIT returns may co-move with the returns of other asset classes much less than commonly believed on the basis of their correlation coefficients