The real value of using Machine Learning in Financial Markets

Co-authored with Dan Liebau

In his recent book, “The Man Who Solved the Market: How Jim Simons Launched the Quant Revolution”, Gregory Zuckerman describes how the hedge fund, Renaissance Technologies, identified market patterns through the use of machine learning. The most crucial detail we took away from the book: Your prediction algorithm only has to be right 50.75% of the time. Jim Simons’ net wealth as of Dec 2019 is US$21.6bn.

The use of technology in financial markets is by no means new. Many other firms including the likes of Millenium, AQR Capital Management (with its total AUM of US$186bn in Q4 2019) on the buy-side and brokers like Bank of America Merrill Lynch, UBS and others, have invested millions in algorithmic trading capabilities. Exchange operators like ICE and Deutsche Boerse run colocation data centres to reduce latency for participants. Dark pools like ITG PosIT and BATS aim to enable the crossing of more significant positions with minimal market impact.

The technological arsenal in the pursuit of alpha not only continues to expand in possibilities but is becoming more commonly available to many market participants. Anyone with access to commodity machine learning platforms and alternative data sources can make use of these methodologies that once had been the proprietary skill of a select few. More recently, affordable access to powerful computers, coupled with the abundant availability of feature-rich, big data has encouraged some investors to expand and exploit alternative information sources. Such alternative data include, for example, satellite imagery of oil tankers. Fund managers may use such information to estimate oil supply to gain an information advantage compared to other market participants. All done with a single drive - the extraction of alpha, the portion of the return above the market.

Alpha is primarily captured by exploiting information asymmetries, capitalising on opportunities that others have yet to uncover. We question whether this exploitation will still be possible when the majority of market participants are machine learning and big-data-dependent? The algorithms and data might be approximately similar. Such a scenario quickly leads to an arm’s race: Who can be the fastest to identify and execute? It reminds us all too much of what we have already experienced in the high-frequency trading space.

Algorithmic and data convergence combined with in-human speeds is what makes the use of machine learning to chase alpha potentially cataclysmic. It presents a systemic risk to the financial markets - eventually even the entire system.

Machine learning algorithms may unintentionally converge, and such behaviour may result in a swarm-like impact.

What was meant to execute orders without market impact may result in driving down prices at scale. Equally, we already know of adversarial algorithms with the sole purpose of learning other algorithms’ behaviour for subsequent manipulation. Such considerations signify the importance of the ethical use of machine learning in financial markets. They include operationalisation of principles such as fairness, privacy, transparency, explainability and accountability. Just because you can do it, does not mean you should! We believe regulators need to be able to run large scale simulations of adverse human and algorithmic behaviour. After all, there are many ways in which machine learning can fail.

We believe the focus should instead be on the exploitation of operational-alpha. Highlighted in “Rethinking Alternative Data in Institutional Investment” the authors Monk, Prins and Rook suggest that understanding risk must be the priority rather than chasing absolute returns. They suggest the use of alternative data should focus on the building of defensive strategies (with a focus on capital preservation and prudent risk-taking, for example, ESG). They continue to emphasise the value of defensible strategies (for instance, based on datasets that relate to internal operations of the hedge fund). Such strategies create more value than engaging in rivalry (the extent to which one entity’s use of a resource diminishes its value for another entity). They are also superior to exclusion (the degree to which one entity can prevent another from using a resource). Hong Kong based BlueFireAi supports investment managers with their machine learning and natural language processing capability, also in mandarin, to extract insights about listed companies that are struggling and create early warning signals for portfolio managers to consider.

The primary objective of investment managers and speculators is and will remain, the generation of attractive risk-adjusted returns. Machine learning can propel trading success. However, given the current day data and algorithmic maturity, the scale and availability of alternative data - can the same success be achieved today and tomorrow by many? And perhaps more importantly, do we understand the consequences and implications of our actions and the actions of our algorithmic creations? We find ourselves in a counter-intuitive situation. Maybe machine learning should be used to “humanise” the financial markets. Perhaps it should focus on inclusion of “systemic algorithmic risk” into asset selection and trading processes.

These questions highlight that the fiduciary duties of the portfolio manager have expanded to include the need for deeper appreciation and skill sets in data and machine learning. In this context, the spurring of activities by institutes such as the Financial Data Professional Institute (FDP) to address the growing need in finance for a workforce that has the skills to perform in a world with an increasing number of data driven decisions, are truly welcomed.

References

1. “The Man Who Solved the Market: How Jim Simons Launched the Quant Revolution”, by Zuckerman
2. “Rethinking Alternative Data in Institutional Investment” by Monk, Prins and Rook
3. https://en.wikipedia.org/wiki/Jim_Simons_(mathematician)
4. https://en.wikipedia.org/wiki/AQR_Capital
5. “Machine Learning in Finance: From Theory to Practice” by Dixon, Halperin and Bilokon