Fundamental Drivers of Dependence in REIT Returns

Evidence for this stylised fact is found in various episodes of market turmoil, see, e.g., Ang and Bekaert (2002), Ang and Chen (2002), Ang et al. (2006), Hong et al. (2007), Longin and Solnik (2001), Patton (2004). Similar return behaviour may be observed in real estate securities, as discussed, e.g., in Gordon (2009).

Clustering of poor returns occurs frequently in downturns and can significantly affect portfolio performance and asset prices (Ang and Bekaert 2002; Ang et al. 2006; Longin and Solnik 2001; Patton 2009).

This decomposition is commonly referred to as the Edgeworth series expansion. For details, see Hall (1992).

Zhou and Gao (2012) provide a lucid discussion of the shortcomings of correlation as a measure of dependence.

We could also consider test statistics for structural breaks in dependence patterns. However, tracing test statistics for structural breaks in existing dependence measures does not allow us provide a continuous measure of dependence patterns that can easily be estimated as a function of firm characteristics.

The linear component of dependence in any bivariate distribution is captured by the OLS slope coefficient estimator, (X ^T X)⁻¹ X ^T Y. The adjusted J statistic thus controls for linear dependence by filtering with this estimator. By doing so, the adjusted J statistic controls for the first moment and first co-moment within the Edgeworth expansion of joint returns, ensuring that the adjusted J captures economically relevant and theoretically motivated information about asymmetric dependence in a statistically rigorous manner. The estimator for the CAPM beta “just happens” to be the same estimator as the OLS beta estimator, and so by filtering through the linear control we are also filtering through the CAPM beta. Filtering through a set of firm-characteristic of macro factors in addition to the CAPM beta might provide additional insights into the role played by these factors and their relationships with asymmetric dependence and thus shed additional light on any return premium generated by these factors. An analysis of this question however requires a structural alteration of the Adjusted J statistic and is therefore beyond the scope of this study. We thank an anonymous referee for pointing this out.

The downside of this approach is that we introduce autocorrelation, potentially over four quarters, in the Adjusted J statistic. However, this is easily remedied by including up to four lags of the Adjusted J statistic in the regression model. Our results are robust to controlling for these additional lags and are available on request.

All of our findings are qualitatively equivalent when replacing the policy-determined federal funds rate with a market-determined benchmark bond yield. Results are available on request.

For robustness, we employ the Russell 2000 small-cap index as an alternative benchmark. Our results remain unchanged and are available upon request.

Our findings are consistent across the policy rate (federal funds rate) and the market-determined interest rate (10-year Treasury rate).

See, for example, Besanko and Thakor (1987a, 1987b), Bester (1985), Boot and Thakor (1994), Chan and Kanatas (1985), Chan and Thakor (1987)

Our findings are consistent across the policy rate (federal funds rate) and the market-determined interest rate (10-year Treasury rate).