Which Factors Drive the Hedge Fund Returns: A Machine Learning Approach

Arbitrage is a central concept in finance. It is defined as simultaneous long and short positions in similar assets to exploit mispricing. Hedge funds experienced fast growth over the past three decades, as real-world arbitrageurs as a group. As they increasingly influence the financial market, it is important to understand the economic drivers of hedge fund returns. Therefore we would like to present a paper dealing with the development of a parsimonious factor model, based on anomalies, to explain hedge fund returns.

Identifying hedge fund factors from anomalies is a very demanding task, regardless of the intuitive appeal. The literature detects hundreds of anomalies, but very important for the authors of the paper is to form a parsimonious set of factors, in the spirit of the APT, that can explain a large fraction of the time-series variation in hedge fund returns. Next problem is that hedge funds use sophisticated strategies, involving both long and short positions as well as derivative securities, that are often highly nonlinear and dynamic. What should we do with such problems? The authors implemented a machine learning approach to identify anomaly factors.

In this novel approach is proposed a new hedge fund factor model, using adaptive-Lasso shrinkage techniques. This model combines anomaly factors with conventional market and macro factors. The factor model outperforms existing models in explaining hedge fund returns. The adaptive LASSO estimator prescreens candidate factors and penalizes nuisance factors with smaller coefficients based on simple OLS. The authors show that the statistical variable selection successfully identifies a small set of nine factors, and the selected factors are robust across testing assets.

Paper contributions to the hedge fund literature:

1. A new factor model for hedge fund returns is very important because hedge funds have evolved at a fast pace in recent decades. This paper proposed a nine-factor model to uncover the risk factors that are most crucial for the hedge fund industry.

2. A significant decrease in alpha in the model provided fresh evidence on hedge fund performance and added to the debate about the value of active asset management.

3. The significant fraction of underperforming hedge funds detected by the proposed model calls for greater scrutiny from hedge fund investors when making investment decisions.

4. The proposed factor selection framework allows to update risk factors to accommodate style shifts in hedge fund strategies and, more generally, highlight the value of factor selection via machine learning.

The authors proposed a new factor model for evaluating hedge fund performance based on shrinkage techniques. They named their factor selection procedure – “Post-Adaptive-LASSO”. It systematically selects the best risk factors from a pool of 44 factors. They used a comprehensive sample of 7,314 hedge funds from TASS and HFR databases over the period from 1997 to 2019 to select factors.

The newly proposed model consists of nine risk factors:

– an asset growth factor,

– a betting-against-beta factor,

– a low-risk factor,

– an equity market factor,

– a return-on-asset factor,

– a time-series momentum factor,

– the monthly change in the 10-year treasury yield factor,

– the monthly change in credit yield spread factor,

– the term spread factor.

The model revealed substantial performance heterogeneity across funds, which is unlikely attributed to pure luck based on a bootstrap analysis.

Authors: Chen, Young and Li Zhengzi, Sophia and Tang, Yushan and Zhou Guofu

Title: Anomalies as New Hedge Fund Factors: A Machine Learning Approach

Link: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4320496


We identify factors from a large set of anomalies for explaining hedge fund returns using machine learning methods. Our new model combines anomaly factors with market and macro factors and outperforms existing models both in-sample and out-of-sample. Moreover, the model leads to a significant reduction in hedge fund alphas compared with other models, while revealing substantial cross-sectional performance heterogeneity. Further subsample analysis provides evidence of style shifting in the hedge fund industry. Overall, the anomaly factors help quantify hedge fund strategies and risk exposures and are useful for fund performance evaluation.

As always, we present several interesting figures and tables:

Notable quotations from the academic research paper:

“Our hedge fund sample is from the Lipper TASS (TASS) and Hedge Fund Research (HFR) databases. Following the literature, we focus on funds that report net-of-fee returns in USD on a monthly basis.6 Although these databases contain records from as early as 1977, the coverage of both live and defunct hedge funds started only in 1994. To mitigate survivorship bias (see Brown et al. (1992), Fung and Hsieh (2000), and Liang (2000)), we exclude all monthly returns prior to 1994. To address the concern of backfill bias, for each fund, we only keep monthly records starting from the month the fund is first added to the database.7 Besides monthly returns, TASS and HFR also provide information on categories of hedge fund strategies. To maintain high statistical powers on category-level inferences, we include ten categories with sufficient numbers of funds, including Convertible Arbitrage (CA), Dedicated Short Bias (DSB), Event Driven (ED), Emerging Markets (EM), Equity Market Neutral (EMN), Fixed Income Arbitrage (FIA), Fund of Funds (FOF), Global Macro (GM), Long-Short Equity Hedge (LSEH), and Multi-Strategy (MS). Finally, we require each fund to have at least 48 monthly return observations. Our final sample consists of 7,314 unique hedge funds over the period from January 1997 to August 2019. In section 3, we also collect monthly net-of-fee returns for a different set of hedge funds from BarclayHedge to examine the performance of our factor model on this alternative hedge fund universe.

We first apply adaptive LASSO to each set of testing assets at fund level, category level (EW and VW), and industry level (EW and VW), and record the factors with non-zero coefficients. Panel A of Table 3 reports the number of testing assets at each level, and the average number of factors selected by adaptive LASSO. On average, adaptive LASSO selects 7 factors for individual funds, 11 factors for the EW category portfolio, and 12 factors for the VW category portfolio. For the EW and VW industry portfolios, adaptive LASSO selects 12 and 7 factors, respectively. The resulting average number of selected factors across testing asset levels is around 10 ((7+11+12+12+7)/5=9.8). The last two columns in Panel B directly present the selected factors for the two industry portfolio series.

To justify the number of factors in our new hedge fund model, we rely on Principal Component Analysis (PCA) to uncover the number of orthogonal latent factors required to capture the vast majority of hedge fund return variations. PCA requires a balanced sample, yet different hedge funds enter and exit our sample over time. To utilize as many funds as possible, we adopt a fixed-window PCA approach to retain more funds. Specifically, we divide our full sample into 11 two-year consecutive windows. For each window, we perform PCA on the excess returns of all hedge funds with two-year history. Then we calculate the average variance explained ratios by each principal component. Figure 1 succinctly summarizes the average variance explained ratios across principal components. Detailed PCA results for each two-year window are deferred to Table A.4 in the Appendix. The first principal component alone explains more than 40% of the return variation, whereas the 10th principal component explains merely 2% of the return variation. The cumulative variance explained ratio by the first nine components exceeds 80%. Altogether, the evidence strongly corroborates the choice of nine factors in our model.

Panel A of Table 5 presents the time-series regression results by regressing EW or VW industry portfolio returns on factors from various models. The three benchmark models CAPM, FF5, and FH7 generate sizable alphas ranging from 0.17% to 0.28% per month, all of which are statistically significant as well. In addition, the adjusted R2 ’s of these models are between 53.02% and 72.63%, indicating that a large proportion of the industry portfolio return variation is still not explained by those commonly used factors. In sharp contrast, our new HF9 model produces insignificant alphas close to zero, and much higher adjusted R2 ’s of 85.34% (EW) and 79.16% (VW), highlighting the strong exploratory power of our new factor model.”

Are you looking for more strategies to read about? Sign up for our newsletter or visit our Blog or Screener.

Do you want to learn more about Quantpedia Premium service? Check how Quantpedia works, our mission and Premium pricing offer.

Do you want to learn more about Quantpedia Pro service? Check its description, watch videos, review reporting capabilities and visit our pricing offer.

Are you looking for historical data or backtesting platforms? Check our list of Algo Trading Discounts.

Or follow us on:

Facebook Group, Facebook Page, Twitter, Linkedin, Medium or Youtube

Share onRefer to a friend

Subscribe for Newsletter

Be first to know, when we publish new content

    The Encyclopedia of Quantitative Trading Strategies

    Log in

    We boasts a total prize pool of $15,000
    Gain a Share of a Total Prize Pool of $15.000
    Gain a Share of a Total Prize Pool
    - bi-weekly research insights -
    - tips on new trading strategies -
    - notifications about offers & promos -