Does a model based on factors extracted from investable exchange-traded funds (ETF) work as well in evaluating fund alphas as models based on factors from more conceptional portfolios? In their October 2016 paper entitled “Bringing Order to Chaos: Capturing Relevant Information with Hedge Fund Factor Models”, Yongjia Li and Alexey Malakhov examine a hedge fund performance evaluation model that identifies risk factors dynamically based on the universe of index-tracking ETFs, focusing on data since 2005 when more than 100 ETFs become available. They first suppress redundancy among these ETFs via cluster analysis, iteratively grouping ETFs by return series similarity (up to 100 clusters) to find the best set of clusters and selecting the ETF most representative of each cluster. They apply regression techniques to identify each year the optimal set of factors (weighted ETFs) for explaining hedge fund returns over the prior 24 months. They compare the power of the ETF-based factor model to explain (in-sample) hedge fund returns with the predictive powers of seven published hedge fund return models that have fixed sets of 1 to 15 factors. They also test hedge fund performance persistence based on out-of-sample performance of funds ranked by in-sample alphas for ETF-based and conventional factor models. Using net monthly returns and descriptions for 10,506 unique hedge funds (2,404 live and 8,102 dead), excluding the initial 24 months of reported returns for each fund, and monthly returns for all index-tracking ETFs with at least 24 months of history as available during 2003 through 2012, they find that:
- From in-sample tests during 2005-2012:
- The ETF cluster analyses/regression techniques identify an average of 2.5 key explanatory factors per year, much simpler than many conventional factor models, suggesting that investors could easily hold a hedge fund clone portfolio.
- Hedge funds generate average monthly alpha 0.04% for the ETF-based factor model, much lower than alphas for conventional models (0.19% to 0.64%), indicating that the ETF-based model offers superior explanatory power for hedge fund returns.
- From out-of-sample tests during 2005-2012, the tenth of hedge funds with the highest (lowest) in-sample ETF-based alphas, reformed annually:
- Generate average monthly return 0.72% (0.48%), with monthly Sharpe ratio 0.29 (0.13). Top decile performance statistics are higher than those for all seven conventional factor models.
- Generate average net monthly alpha 0.53% (0.18%) relative to the seven conventional factor models.
- Have average annual attrition rate of 7.6% (20.7%).
- Returns for hedge fund clone portfolios constructed with ETF-based factor model beta weights have substantially higher out-of-sample correlations with returns of underlying hedge funds than do comparable clone portfolios constructed from conventional factor models.
In summary, evidence indicates that a dynamic factor model derived from index-tracking ETF returns outperforms conventional factor models in explaining and predicting hedge fund returns.
Cautions regarding findings include:
- The ETF-based process may be more realistic (investable) than the conventional factor specification process, but it is complex in terms of setup and computation. Investors would bear a fee for delegating the ETF-based process to experts.
- Large portfolios of hedge funds are not realistic for most investors. Applying the methdology to select one or a few hedge funds may be unreliable.
- Calculations do not account for any costs of/constraints on rebalancing portfolios of hedge fund or clone portfolios.
