Identifying Hedge Fund Skill by Using Peer Cohorts

Abstract

We propose a cohort model that evaluates hedge funds against peer groups executing similar investment strategies formed by using return correlations. Our method improves the identification of skilled managers, as evidenced by a strong ability to explain hedge fund returns out-of-sample, with cohort alpha being more persistent than alpha based on the widely accepted seven-factor model. A hedge fund-of-funds analysis found significant performance enhancement from exposure to the best funds within each cohort. The cohort approach can be used to enhance the construction of hedge fund-of-funds portfolios by isolating strategy groupings as well as the best managers within each group.

Disclosures: The authors have no conflicts of interest to declare. The content of this article reflects the views of the authors and not necessarily the views of BlueCove Limited.

Editor’s Note:

Submitted 2 September 2020.

Accepted 6 January 2021 by Moshe Arye Milevsky

This article was externally reviewed using our double-blind peer-review process. When the article was accepted for publication, the authors thanked the reviewers in their acknowledgments. Nicole M. Boyson and one anonymous reviewer were the reviewers for this article.

Notes

¹ A few mutual fund studies have also investigated benchmarking against peers in other ways. Cohen, Coval, and Pástor (2005) identified skilled managers by comparing holdings and trades with the better-performing funds. Hunter, Kandel, Kandel, and Wermers (2014) defined peer groups through matching returns to index benchmarks and included variations from benchmark performance as an additional factor in evaluating mutual fund returns. Hoberg, Kumar, and Prabhala (2018) formed customized peer groups against which performance was evaluated by using exposures to holdings-based style characteristics.

² Agarwal and Naik (2000a) extended Sharpe’s (1992) style analysis to hedge funds by easing the constraint that style exposures be long only and must sum to 1.0. They found that their eight style indexes explained 44%–84% of return variance for eight hedge fund strategy indexes.

³ Since the seven-factor model was introduced in Fung and Hsieh (2004), it has been utilized in several studies, including Aiken, Clifford, and Ellis (2013); Bali, Brown and Demirtas (2013); Boyson (2008); Brandon and Wang (2013); Brown, Grundy, Lewis, and Verwijmeren (2012); Buraschi, Kosowski, and Trojani (2014); Fung, Hsieh, Naik, and Ramadorai (2008); Jagannathan et al. (2010); Kosowski, Naik, and Teo (2007); Nohel, Wang, and Zheng (2010); Ramadorai (2013); Sadka (2010); Sun, Wang, and Zheng (2012); Teo (2011); Titman and Tiu (2011); and Yin (2016).

⁴ This step was necessary because funds often exist in different classes.

⁵ Because many firm names are reported with small differences in different databases, we allowed for some contrast in the firm names when performing the matching. All the matches were then manually confirmed.

⁶ Agarwal et al. (2009) based the hurdle for performance fees on LIBOR and made assumptions about the tracking of returns in estimating high-water marks. These assumptions create the potential for error because of the wide variety of performance fee hurdles (coupled with incomplete disclosure of high-water mark terms) in our databases.

⁷ The eVestment database had a few cases in which the date that the fund was added to the database was reported incorrectly. In those cases, we removed the first 24 months of reported returns.

⁸ When a fund did not report one of these variables, we either replaced the missing value with the median presented in Table 1 or dropped the fund, depending on the test.

⁹ The seven-factor model is available for download at http://faculty.fuqua.duke.edu/~dah7/DataLibrary/TF-FAC.xls.

¹⁰ We use the acronym PAR (peer-adjusted return) rather than cohort-adjusted return, or CAR, to avoid confusion with the use of CAR to represent cumulative adjusted return in the finance literature.

¹¹ Under the 0.25 distance measure, 22% of funds were not assigned to cohorts at some stage.

¹² For comparison, factor loadings for the 15 cohorts with the highest adjusted R²sare provided in the Supplemental Online Material. They reveal that only five cohorts had an R² in excess of 0.70. The R² dropped to 0.60 for the 13th cohort.

¹³ This approach follows Tashman (2000). Swanson and White (1997) argued that a rolling window is preferable to an expanding window for estimating multiperiod model accuracy.

¹⁴ The control variables are provided in the notes to Figure 4.

¹⁵ Eling (2009) provided a literature review of hedge fund persistence and documented that more than half of previous studies were unable to find persistence at a one-year horizon.

¹⁶ These R² values are in line with the analysis by Fung and Hsieh (2004) of hedge fund indices, for instance.

¹⁷ Initially unassigned funds were allocated to cohorts with an average correlation of 0.57 during the second step.

¹⁸ The Supplemental Online Material presents quartile analysis results for the initially unassigned funds and for the complete fund sample when these funds were allocated to the cohort with which they had the highest correlation.

¹⁹ We also allocated all initially unassigned funds to the cohort with which they had the highest correlation. This step resulted in moderately stronger persistence, which could be the result of either skill or persistence in exposure to omitted factors.

²⁰ Refer to the Supplemental Online Material for further details on this model.