Many experts have hailed machine learning (ML) as the next frontier for quantitative investors, with good reason. ML offers strong potential, as it can help uncover exploitable nonlinear patterns and interaction effects that more traditional techniques fail to detect. Yet for all the enthusiasm raised, we also caution: ML is not a magic bullet and faces important implementation challenges.
- Recent studies show promising results when predicting stock returns
- But their reliance on high-turnover signals raises concerns
- Further research is needed to make the most of ML
While ML theory has been around for many years, recent advances in cloud computing power have made it feasible to assess how ML can contribute to investment management. Such increasing popularity is reflected in the soaring number of research papers published in recent years investigating the use of ‘artificial intelligence’ (AI) and ML in quantitative asset management.
Some of the most-cited studies report promising results when predicting one-month stock returns using ML with a large set of traditional predictor variables as input features. Although the models partly pick up known factors, they are able to add value by exploiting nonlinear alpha opportunities and interaction effects.
However, such encouraging findings remain essentially theoretical. Turning the resulting fast alpha signals into profitable investment strategies in practice – once costs and other real-life implementation frictions are taken into account – is easier said than done. For one, the academic literature investigating practical implementation issues remains scarce.
Moreover, most studies suggest the potential for ML models to outperform traditional ones is often reduced by their reliance on high-turnover signals. For this reason, there has recently been an effort to integrate economic structure into loss functions, so the ML model can focus on stocks that are easier to trade. These efforts should increase the likelihood of monetizing the predictive power of ML models.
An evolution more than a revolution
Besides forecasting returns, other promising use cases for ML have been proposed. This includes enhancing traditional factors, creating new variables from unstructured data, and predicting metrics other than return, such as risk or sustainability. So far, ML methods in asset management have therefore been more of an evolution than a revolution.
"Asset managers who will disregard advances in ML may see their performance wane relative to those who embrace ML."
Presumably, asset managers who will disregard advances in ML may see their performance wane relative to those who embrace ML. For example, the ability to automate tasks of traditional analysts, such as reading, seeing or hearing ultimately promises large gains in productivity, provided the asset manager possesses the necessary infrastructure and can investigate different big datasets and signals at scale.
Yet discarding economic theory altogether and turning to a fully data-driven approach can vice versa also set one up for failure. Investors can identify and evaluate the ability of an asset manager to succeed in advancing its investment process accordingly by scrutinizing what research protocol is in place. The latter is key to the success of ML in practice and to navigate the many pitfalls.
Altogether, researchers have just been scratching the surface of the endless possibilities offered by ML, and many exciting new discoveries can be expected in the years ahead. However, human domain knowledge is likely to remain important, because the signal-to-noise ratio in financial data is low, and the risk of overfitting is high.
Read the Full Story on SSRN
