The Promise of Machine Learning

Machine learning shows great promise for empirical asset pricing and has the potential to improve our understanding of expected asset returns.

Due to increased computing power and data availability, decreased data storage costs and algorithmic innovations, machine learning methods are increasing in popularity in financial research. A number of aspects of empirical asset pricing make it an attractive field for analysis with machine learning methods. There is now a zoo of predictors that various researchers have argued possess forecasting power for returns, many of which are highly correlated. With its emphasis on variable selection and dimension reduction techniques, machine learning is well suited for such challenging prediction problems by reducing degrees of freedom and condensing redundant variation among predictors. There is also the question of whether relationships are linear or nonlinear. Machine learning is explicitly designed to approximate complex nonlinear associations.

Doron Avramov, Si Cheng and Lior Metzker, authors of the October 2021 study “Machine Learning versus Economic Restrictions: Evidence from Stock Return Predictability,” examined whether investors can harvest extra profits generated by various machine learning signals given plausible restrictions on the investment universe. They considered a comprehensive set of both linear and nonlinear models (a diverse collection of high-dimensional models for statistical prediction). In doing so they imposed several economic restrictions:

They limited the universe of stocks to those that were relatively cheap to trade by excluding microcaps or distressed firms.

In the time series, they examined whether investment profitability was more pronounced during high limits-to-arbitrage market states, such as high volatility and low liquidity.

They assessed the turnover and the corresponding transaction costs associated with implementing machine learning-based strategies.

They explored the economic foundations of trading strategies advocated by seemingly opaque machine learning methods.

Their full sample covered U.S. stocks over the period 1957-2017 divided into three subperiods: a training sample and a validation sample (both of which varied in length depending on the machine learning method), and the remaining 31 years (1987 to 2017) for out-of-sample testing. They trained the model every year so that the training sample expanded every year. Following is a summary of their findings:

The value-weighted long-short portfolio generated returns of between 0.95% and 2.18% per month, with corresponding Fama-French six-factor (FF6) adjusted returns (beta, size, value, momentum, profitability and investment) of between 0.62% and 1.87% per month: “Such large and significant figures reflect the impressive success of machine learning techniques in generating outstanding performance relative to traditional methods such as nonregulated regressions and portfolio sorts based on individual anomalies.”

Imposing economic restrictions revealed that the predictability of deep learning methods weakened considerably. Relative to the full sample evidence across all stocks, the value-weighted FF6 adjusted return was between 48% and 71% lower after excluding microcaps, between 53% and 77% lower after excluding firms that did not have credit rating coverage, and between 69% and 94% lower after excluding financially distressed firms that faced further deteriorating credit conditions. None of the deep learning methods generated significant value-weighted FF6 adjusted returns at the 5% level after excluding distressed firms. They also found that applying the same economic restrictions to traditional methods resulted in a similar proportional magnitude of performance deterioration.

Accounting for nonlinearities was especially useful for difficult-to-value and difficult-to-arbitrage stocks, helping to explain anomalies to asset pricing models based on linear regressions.

All the machine learning signals generated portfolio turnover that was considerably higher than most individual anomalies – the monthly turnover in the long-short portfolio ranged from 87% to as much as 163%. Thus, in the presence of reasonable trading costs, the machine learning-based investments would struggle to achieve a statistically and economically meaningful risk-adjusted payoff.

Machine learning signals often required taking extreme long-short positions – impractical from a risk management and regulatory perspective.

All four machine learning signals continued to predict cross-sectional stock returns in the post-2001 period across all stocks, with the anomalous return patterns confined to difficult-to-arbitrage stocks.

All four machine learning methods successfully identified mispriced stocks consistent with most anomaly-based trading strategies. Stocks in the long positions of machine learning-based investments were typically small, value, illiquid and old stocks with a low price, low beta, high 11-month return (medium-term winners), low asset growth, low equity issuance, low credit rating coverage and low analyst coverage.

Machine learning methods were able to mitigate downside risk and protect investors from extreme crashes. The maximum drawdown for machine learning portfolios ranged between 20% and 35% for the full sample, while the market portfolio experienced a larger drawdown, at 49%. All four machine learning methods generated positive returns across four major episodes of market downturns (e.g., the 1987 market crash, the Russian default, the bursting of the tech bubble and the Great Recession).

Intra-industry strategies delivered substantially higher returns than interindustry strategies – machine learning added value in terms of stock selection more than industry rotation.

Their findings led Avramov, Cheng and Metzker to conclude: “Despite their opaque nature, machine learning signals successfully identify mispriced stocks consistent with well-established empirical facts, without preselection of truly useful characteristics and models. Our findings highlight the merits of employing machine learning methods to avoid the data snooping problem in the anomaly literature and suggest that black-box-like machine learning models are reasonably interpretable, which is essential for a robust and credible assessment of out-of-sample predictability.” They also concluded that “machine learning routines possess superior ability to detect complex features in the data that otherwise remain unnoticed.” They added: “Our findings further support the concept that machine learning-based investments could hold considerable promise for asset management.”

Investor takeaways

In their study, Avramov, Cheng and Metzker found that while machine learning substantially improved the investment payoff compared to traditional methods (such as ordinary least squares (OLS) regression and portfolio sorts based on individual anomalies), they also found that both machine learning and traditional methods delivered reduced payoffs in the presence of realistic economic restrictions, and their performance deteriorated by a similar proportional magnitude. Despite that, they found that even considering economic restrictions, deep learning signals were profitable in long positions while also reducing downside risk. Thus, they appear to hold promise for traditional long-only investors. Their work also highlighted the importance of considering trading costs and economic restrictions when reviewing the results of studies using machine learning.

Finally, some words of caution. A major benefit of artificial intelligence tools (such as machine learning) is that they have great capacity to deal with massive amounts of data. But that also creates the risk that findings can be the result of “torturing the data until it confesses.” Correlation doesn’t necessarily mean causation, as machine learning methods on their own do not identify fundamental associations among asset prices and conditioning variables. Thus, it is critical that any findings should be supported by either risk- or behavioral-based explanations, and those findings should be persistent across economic regimes, pervasive across asset classes and regions, and survive transactions costs.

Larry Swedroe is head of financial and economic research for Buckingham Wealth Partners.

For informational and educational purposes only and should not be construed as specific investment, accounting, legal, or tax advice. Certain information is deemed to be reliable, but its accuracy and completeness cannot be guaranteed. Third party information may become outdated or otherwise superseded without notice. By clicking on any of the links above, you acknowledge that they are solely for your convenience, and do not necessarily imply any affiliations, sponsorships, endorsements or representations whatsoever by us regarding third-party websites. We are not responsible for the content, availability or privacy policies of these sites, and shall not be responsible or liable for any information, opinions, advice, products or services available on or through them. The opinions expressed by featured authors are their own and may not accurately reflect those of Buckingham Strategic Wealth® or Buckingham Strategic Partners®, collectively Buckingham Wealth Partners. Neither the Securities and Exchange Commission (SEC) nor any other federal or state agency have approved, determined the accuracy, or confirmed the adequacy of this document. LSR-22-430

The Promise of Machine Learning

Sponsored Content

Trending Topics View All

Upcoming Webinars View All