Machine learning – no silver bullet?
Machine learning has evolved rapidly over the past decade, with huge consequences across industries. But does the hype exceed its potential impact? In this article, we discuss with Portfolio Manager Mark Richardson the value of machine learning for the world of quantitative finance.
- There is a great deal of hype around machine learning (ML), given the potential ramifications for the world of quantitative finance.
- It is too early to say with any certainty that machine learning represents a potential ‘silver bullet’ in the world of quantitative finance. But there are areas of interest that could be worth investigation.
- One of the more potentially credible applications of machine learning is in modelling the evolution of classical models in a time series.
Machine learning is a nascent paradigm in modern quantitative investment management, and like many industry participants, Janus Henderson has been testing the water with this developing technology. This includes looking at applying a machine-learning approach to augment existing models the team operates in the equity derivatives space. In the views of Portfolio Manager Mark Richardson, “We felt it was important to begin a systematic and dispassionate investigation of the potential scope of the emergent Machine Learning toolkit in order to assess the implications for certain aspects of our investment process”.
At this point, it is too early to say with any certainty that machine learning represents a potential ‘silver bullet’, but there appear to be several areas of potential research interest. For example, one of the key strategies the team operates involves dynamic trading around a persistent supply and demand imbalance in Euro Stoxx 50 forward contracts, which manifests in significant, if ephemeral, price distortions. A key component of the trade involves modelling of price changes in the Euro Stoxx 50 dividend term structure relative to a given move in the underlying index.
While the Janus Henderson Alternatives team has an existing model for this that works very well, as interest in machine learning grows, they have been keen to see if there is something to be gained by applying techniques from the machine-learning toolkit. As Mark says, “Initial investigations suggest that it is difficult to beat our existing models”.
Elsewhere, the team has looked at applying machine-learning techniques to help forecast volatility surface moves, one of the more potentially credible applications of the machine-learning toolkit. The question is, given a parametric specification of the previous day’s implied volatility surface, is it possible to utilise machine-learning techniques to describe its time-series evolution, given the observable intraday futures move? As Mark comments, “Increasing the accuracy of implied volatility forecasts has the potential to be significantly impactful to the extent that it smooths portfolio volatility, removes subtle directional exposures, and ultimately produces more stable hedges”.
The team continues to progress its research agenda in this area. “Clients expect us to have an opinion on the efficacy of machine learning in general and the only way of developing such a view is to carry out extensive and rigorous experiments. All the while we remain open to the possibility that there could be better ways of doing things”.
Machine learning has been touted as a potentially transformative technology for the investment industry. But the measure of all progress is whether that potential can be harnessed for aggregate gains on productivity or performance. While the team is approaching the idea of machine learning as a potentially useful development, they are doing so from a starting position of measured scepticism.
A necessary condition for determining whether machine-learning tools are suitable for prime-time use is getting comfortable with the trade-off that (potentially) superior solutions to trading optimisation problems come at the cost of substantially reduced model interpretability. While the team understands exactly what is going on inside their current models, it would be a significant departure from this intellectual framework to move across to what would be an essentially ‘opaque’ system. In Mark’s view: “At this point, we do not believe that clients would be comfortable with such an abdication of understanding. It would require truly exceptional machine-learning model performance before we could ‘trust the black box’ and seriously consider their use in real-world trading activity”.
“[Using Machine Learning for] increasing the accuracy of implied volatility forecasts has the potential to be significantly impactful to the extent that it smooths portfolio volatility, removes subtle directional exposures, and ultimately produces more stable hedges. ” Portfolio Manager Mark Richardson
More Alternative Perspectives
Can investors follow the pack and hope to outperform? In this article, Portfolio Managers Andrew Kaleel, Mathew Kaleel and Maya Perone investigate the characteristics of trend-following strategies and consider their potential value as part of a strategically positioned portfolio.
In this article, Portfolio Manager Aneet Chachra and Investment Director Alistair Sayer consider the cost of trading when volatility spikes, and ask if liquid alternatives provide an alternative for investors to manage their portfolios during periods of uncertainty?