This content is from: Portfolio
Asset Managers Can No Longer Ignore Machine Learning
Researchers at Robeco performed an analysis of the literature to see how managers could implement the innovations.
At first glance, it might appear that machine learning models and asset managers, particularly those that employ quantitative strategies, are a natural fit. But the reality is that many managers are still working to determine how to integrate machine learning into their investment process.
Asset managers may now have reason to speed it up. Researchers at Robeco have found that machine learning significantly improves the investment process.
They performed a sweeping analysis that detailed the existing work on machine learning in asset management and how managers can implement those findings into their work. The analysis is set to appear in an upcoming issue of the Journal of Portfolio Management.
The research showed that machine learning models can improve investor predictions better than other quantitative models. It also showed that these models can be employed in a variety of ways, from predicting corporate bond yields to market betas.
“Machine learning is well equipped to deal with large sets of features and ‘learns’ to give the most weight to the most relevant variables,” according to the paper.
In order to explore the relationship between dividend yields and predicted returns, the researchers used both a machine learning model and an ordinary least squares (OLS) regression model, which is used in some traditional quantitative analyses.
They found that an OLS regression did not immediately reveal a relationship between the two. Instead, the researchers first needed to remove the companies that never pay out dividends from the data sample. Once that change was made, the OLS model revealed a positive relationship between dividend yield and return.
The machine learning model, however, immediately observed this pattern and controlled for non-dividend-paying companies.
In another example, the machine learning model was able to reveal that the relationship between the earnings-to-price ratio and actual returns is stronger during the months in which a company reports earnings than in others.
Asset managers can use these findings in a few ways.
They can start by examining the data related to strategies that are based on predicted short-term stock returns. Three studies referenced by Robeco found that machine learning strategies based on predicted stock returns “substantially outperform” their comparable linear strategies. In each study, the Sharpe ratio of the machine learning strategies was at least two or higher.
Other research has shown that machine learning models can extract more value from analyst predictions, which investors can use to forecast events such as earnings surprises.
Machine learning can also employ natural language processing to turn narrative-driven information, such as Securities and Exchange filings, earnings call transcripts, and news articles, into structured data, which can then be used for predictive analysis. In addition to predicting possible future returns, machine learning models can also forecast market beta, stock volatility, share buyback announcements, dividend announcements, and announcements of mergers and acquisitions.
According to the research, machine learning models can also be used in the trade settlement process to reduce transaction costs by automatically analyzing the best time, size, and place to trade.
“So far, machine learning methods in asset management [have been] more of an evolution than a revolution,” the paper said. “Presumably, asset managers who [choose to] disregard advances in machine learning will see their performance wane relative to those who embrace machine learning.”