Evaluating Hedge Funds with Machine Learning-Based Benchmarks

Abstract

Evaluating hedge fund performance is notoriously difficult. The flexibility of hedge fund strategies in terms of asset class exposures and leverage, along with the lack of operational transparency, makes this a particularly difficult task. This paper adopts a Bayesian machine learning approach (Bayesian Additive Regression Trees (BART)) to address the challenge of hedge fund performance evaluation. We illustrate the advantages of the methodology over the conventional factor model approach in several contexts including a real-time fund selection and investment strategy, and fund failure prediction. Importantly, the BART framework can successfully characterize the risks of funds that have near-zero R2 values with respect to traditional performance attribution models, as well as funds with short return histories. A key reason for methodology's success is its ability to account for the nonlinearities and higher-order interaction effects among risk factors that determine hedge fund strategy payoffs. To further illustrate the methodology's advantage over conventional performance measures, we re-examine a well-known result regarding the link between strategy distinctiveness and fund performance. Our results suggest that the documented positive relation between strategy distinctiveness and fund performance is largely an artifact of benchmark model error that contaminates inference based on the widely used Fund and Hsieh (2004) model.

Keywords: Hedge Fund, Investment, Bayesian Inference, Machine Learning, Regression Trees

JEL Classification: G11, G23