Is average correlation related to expected returns: evidence from global markets

ABSTRACT

We examine whether average country-level stock market correlation is related to global equity returns. Previous research focusing on the U.S. suggests that average firm-level correlation captures some of the risk not accounted for by other variables and is positively related to returns on the broad stock market. In contrast, we find that average country-level correlation does not appear to be related to global returns, and that the Roll (1977) critique may not be responsible for this lack of relation. Additionally, average correlation does not help forecast returns out-of-sample.

KEYWORDS:

• Average correlation
• forecasting
• predictability
• Roll critique

JEL CLASSIFICATION:

• G11
• G15

Acknowledgments

We thank the referee as well as the seminar participants at Brooklyn College for very helpful comments. All errors are ours.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

¹ The time zones relative to GMT are for northern hemisphere’s summer months.

² This relationship holds in the limit, as the number of stocks increases and idiosyncratic risk is diversified away.

³ Due to the volatility feedback effect (see French, Schwert, and Stambaugh (1987) and Campbell and Hentschel (1992)), market returns are negatively correlated with observed market variance and average correlation over the same period. This is not indicative of the true relation between expected returns and risk.

⁴ A closely related multivariate GARCH model is proposed in Engle (2002). For a survey of the literature on multivariate GARCH, see Bauwens, Laurent, and Rombouts (2006).

⁵ Regression diagnostics, such as tests for normality and autocorrelation of the error term as well as the Ramsey RESET test, indicate that our models are not misspecified.

⁶ Based on the suggestion of a referee, we examined the causal relationship between AC and market variance. A vector autoregression model indicates that market variance Granger-causes AC (p-values are below 0.05 for all definitions of market variance and AC), but AC does not Granger-cause market variance.

⁷ The European factor is based on the 16 European countries from the original list of 23.

⁸ We also conducted a split-sample analysis by breaking up the 18-year sample into two 9-year subsamples and running all the monthly regressions for each subsample on both the global and European data. None of the coefficients on AC are statistically significant. As an additional robustness check, we fit a Markov-switching model for AC. Our results indicate that, conditional on other factors, equity returns are not statistically different in the two states; p-values for all specifications are above 0.146.