By: Ramin Baghai-Wadji & Stefan Klocker, Vienna University of Economics and Business Administration
Published: May 20, 2006

Assuming hedge fund beta exists, determining the amount of alpha produced by a manager requires one to know what particular hedge fund beta a manager is leveraging.  So the identification of a hedge fund as being say, ”merger arb” or “distressed” is critical in determining value added by the manager.

Problem is, asking managers to self identify may not always be the best strategy.  Managers would face a fundamental conflict as their choice might make them look like a hero or a dog.

This paper was first presented in October 2006 at the Annual Meeting of the German Finance Association.  It proposes a new methodology for identifying the strategies of hedge funds by grouping them together into natural clusters using a “self-organizing map” (a.k.a. “a neural network”).

By placing hedge funds into buckets according to their return histories, not their self-reported strategies, the authors contend that they can better identify returns attributed to manager skill and returns attributed to style benchmarks.  In addition, since the methodology uses actual peer-group benchmarks within each cluster, it naturally captures any non-linearity inherent in each strategy.  In their words:

“…the explanatory power of our SOM-based model exceeds all other analyzed performance attribution models in the case of our randomly chosen Regressions Sample comprising 1,000 funds.”

They propose the model be used by consultants and funds of funds:

“…our SOM-based performance attribution model can contribute valuable information to the portfolio selection process of individual investors. Moreover fund of hedge funds managers could use our model for the selection of single strategy hedge funds.”

Unfortunately, they come to the same conclusion as many of their hedge fund replication compatriots:

“All tested performance measurement models lead to the conclusion that the average fund of hedge funds does not create enough excess return to compensate for the additional layer of fees.”

But perhaps the best part about this paper is that it provides a clear overview of the state of the hedge fund replication business today.  We include an excerpt below, but recommend you peruse the paper for more useful background on this topic:

“Measures used to evaluate the performance of mutual funds have evolved from simple CAPMbased models to more elaborate multi-factor models, such as the 8-factor model by Grinblatt and Titman (1988), the asset class factor model developed by Sharpe (1992), the Fama and French (1993) 3-factor model and the 4-factor model by Carhart (1997). These models perform reasonably well with static buy-and-hold long only investments, which are common in the mutual fund universe, but are of limited value for the performance attribution of hedge funds. This is due to the unique features of hedge funds distinguishing them from ordinary mutual funds: Hedge funds employ dynamic trading strategies including alternating long and short positions. Moreover extensive leverage is much more common in the hedge fund industry than in the mutual fund universe. These stylized facts of hedge funds make the models originally designed for mutual funds inappropriate for hedge funds.

“More recently, considerable effort went into devising approaches specific to the hedge fund universe (see for example Fung and Hsieh (1997, 2004), Schneeweis and Spurgin (1998), Brown et al. (1999), Ackermann et al. (1999), Agarwal and Naik (2000), Liang (1999), Lhabitant (2001)). The large variety of models most commonly used in performance measurement shows that researchers and practitioners are far from accepting a particular model as the standard.

“In this paper, we present a new approach for evaluating the performance of hedge funds which is based on a classical multi-factor model in the tradition of Sharpe’s (1992) asset class factor model but uses neural network derived hedge fund indices.”

Read Full Paper