Stability-Adjusted Portfolios

Abstract

The conventional approach for addressing estimation error in portfolio construction is to devise techniques for reducing the errors, such as compressing all the estimates toward a cross-sectional average or some other prior belief. In this article, the authors propose an alternative approach for dealing with estimation error, arguing that some estimates may be more or less stable than others. The authors propose that rather than attempting to mitigate estimation error by making the estimates more similar to each other, portfolio managers should measure their relative stability and form portfolios that explicitly account for this feature, thus potentially making them less similar to each other. The authors focus on measures of risk rather than means, because portfolio managers typically extrapolate historical covariances but estimate expected returns based on subjective views rather than historical averages. Moreover, many important investment applications, such as index replication, focus exclusively on risk mitigation. The authors show that portfolios that explicitly account for stability in their construction have substantially different allocations and more stable risk profiles than portfolios that are blind to estimation error, as well as those that rely on Bayesian shrinkage.