Artificial intelligence (AI) and machine learning have become buzzwords in financial services, but while this technology can be applied in finance in numerous ways to improve returns, it also has some significant limitations that market participants should be aware of.
This was the message from speakers at the Profit & Loss Forex Network New York conference, on a panel discussion titled “AI: Regular Quants with a Bigger Bazooka?”
“In my mind the biggest problem with machine learning in its application to finance is the problem of non-stationarity. What I mean by that is that you cannot assume that the structure of the data set that you’re using whatever machine learning tools on will remain the same in the future,” said Andrej Rusakov, a partner at Data Capital Management, a hedge fund that uses AI technology to help develop trading strategies.
Rusakov explained that the majority of machine learning tools have been built for firms trying to produce more targeted advertising, and of the data sets that they look at and their relationships to one another are very static. The opposite is true in finance, where the data sets that need to be analysed are constantly changing.
“The time dimension kills the number of observations, so you can’t really tell Federal Reserve to jack the interest rates up and then go down and then up, or China to devalue its currency and not devalue, just for the purpose of your experimentation. So everything changes all the time, therefore the non-stationarity is a problem and therefore you have to make judgement calls on how to patch different times together and make inferences from them,” said Rusakov.
Michael Recce, the chief data scientist at Neuberger Berman, agreed that stationarity issues are one of the main reasons why AI tools fail to yield positive results when applied to finance.
Talking about how AI is being used in advertising, Recce said that the types of data analytics being used by advertising firms are also applicable to finance and, as a result, he expects there to be an influx of human talent from this industry into financial services.
“That exact analysis that they’re doing for showing you the ad is the analysis they have to do to understand who is winning in the marketplace and to, for example, predict an earnings surprise,” he added.
Recce also pointed out that some firms are trying to apply AI and machine learning tools to problems where they are not needed.
“If you can’t solve a problem, don’t ask a machine to solve it, and definitely don’t ask a machine to solve it by throwing tons of data at it,” he said. “Machine learning tools are all too often seen as the go-to direction and the thing that will solve [any problem] no matter what.”
Richard Rothenberg, executive director at Global AI Corp, concurred with this last point, stating: “When we think of AI, it’s a way to augment or enhance what you’re already doing. So part of our process is to do as much as you can with traditional methods and, if you see something there, AI can probably help you do things faster and sometimes more efficiently, but it’s not the ultimate answer. You need a lot of skepticism to deal with all the data mining and over-fitting and all the wrong signals and false positives that may arise.”
The other major challenge for applying AI to finance that was highlighted on the panel was just the sheer investment that firms need to make in order to use this technology in an effective manner.
“It’s not easy, it takes a lot of investment, investment in skill set and technology investment,” said Gaurav Chakravorty, co-founder of qplum and a former partner at Tower Research Capital.
He added: “There are so many ways you can go wrong with the technology, it’s an approach and you can use it incorrectly, that’s a limitation. A lot of people are not going to adopt it because it’s a sea change from their current workflow.”