Experimental feature

Listen to this article

00:00

00:00

Experimental feature

or

Wall Streeters have been coming out ahead in their struggles with demanding Washington regulators and liberal activists. Investment management companies in particular no longer face the threat of being regulated as systemically important.

The financial industry’s leaders should enjoy their moment of triumphin Washington. They may not have much time before facing the consequences of their under-investment in research and development.

The big companies are beginning to recognise the competitive challenges posed by advances in machine learning. Their licensed oligopolies give them strong market positions and data libraries, but also rigidities and slow reaction times. Wall Street should be spending more on R&D, but budgets are being cut due to cost controls and regulatory disputes over who should pick up the bill.

Portfolio managers and market strategists are only starting to consider which groups of machine-learning algorithms are best adapted for financial markets analysis.

According to researchers at JPMorgan, the US bank, an ensemble machine-learning technique called random forest shows the most promise for trading the 10-year Treasury market. Société Générale, the French bank, on the other hand, says support-vector machines are better, over time, for making long/short equity investment decisions. Academic researchers sniff that both methods are dated.

Financiers should consider what has happened to car manufacturers. They are at risk of becoming mere component manufacturers for the tech companies that could dominate automated driving and vehicle sharing. Now some of the software and hardware developers who have worked on automated driving have turned to the financial markets.

That will lead to serious competition for banks and investment managers. So far, their response is spotty. Top management seems to think that some strategic investments in fintech start-ups and intensively PowerPointed consultancy reports will do the trick.

Not really. Banks and institutional investors have enormous IT budgets, but much of those go to maintaining legacy systems or meeting compliance deadlines. Hedge funds have sharp managers, but their limited partners have little patience for R&D projects that may not pay off for years.

Still, some bank researchers are trying to keep up. The JPMorgan study, “Do androids dream of electric bonds?”, tested a variety of widely available, off-the-shelf machine-learning programs to determine whether they could outperform buy-and-hold or simplistic, naive trading strategies.

The researchers measured the algos’ predictive power in the markets for 10-year Treasuries, swap spreads and 10-year swaptions volatility. They tested how the algos would have performed in 2017 when they were trained against two sets of data, one from 2000-16, and another more detailed set from 2008-16.

Basically they found the standard machine-learning programs did well for Treasuries, OK for swap spreads and failed to outperform a simplistic strategy for options volatility. That is pretty significant, since you would think that it was hard to find a new edge in such a deep and heavily arbitraged market as the 10-year.

Also, the researchers apparently did not have a bunch of giant supercomputers at their disposal, just a couple of desktops with graphics processor units stitched together to run in parallel. You could probably buy the same equipment and programs with a five-figure budget. The team did get to use JPMorgan’s market data, but so could any of the bank’s institutional customers.

It would be implausible for such a research project to create a box that would end up owning the world, and that is not what the researchers claimed. Instead, they noted that their favourite strategy, random forest, had “two distinct personalities”. These created predictors that “appeared to be decently ‘self-aware’, with realised test-sample hit rates trending much higher on days when they had higher conviction (where they traded in large size). This suggests that one of the most compelling applications of ML is in timing execution.”

Distinct personalities, eh? I hope that is not discomforting for you.

Others who have been working on machine learning for longer believe the JPMorgan study might have used more recent and fully developed work on deep-learning programs, such as convolutional neural networks. Marek Bardonski, managing partner and head of AI at Sigmoidal, a Warsaw machine-learning consultancy, has been developing such programs since he won a Nasa competition when he was 17. Before joining Sigmoidal, he was a senior deep-learning research engineer at Nvidia, where he worked on developing programs for autonomous vehicles. He is now 25.

From his perspective, Mr Bardonski says financial time-series forecasting (what many FT readers are paid to do each day), is now a really interesting problem. “Most of the theoretical work for autonomous driving has been done. It just needs more testing, so it can handle all situations, and regulation.” In contrast, he says, “in financial markets we have more opportunities to find interesting ideas.

“With driving, the observations [analysing images of the road] are difficult, but the problem is easy. With financial markets the observations [price and volume histories] are easy, but the problem is difficult.”

Not impossible, though.

Leave a Reply

Time limit is exhausted. Please reload the CAPTCHA.