James Rundle: Rise of the Machines

Artificial intelligence (AI), as a topic in computing, has been around for a long time. In science fiction, it’s been around for even longer. The genre is filled with tales of murderous machines, as seen in The Terminator, or a single, callous overlord, as in Harlan Ellison’s I Have No Mouth, and I Must Scream. In trading, machine learning and AI have been discussed for some time, but the topic has picked up a lot of steam in the past few years.

In discussions about front-office technology, AI has typically been proposed in relation to algorithms. This is the dream of many a quant, of course—building the ultimate algo, self-aware enough to adapt to market changes and rewrite itself in new and fantastic ways, able to oscillate between venues and asset classes as if they were mere lines on a rugby pitch, rather than imposing hurdles on a number of interlocking race tracks. Some machine learning is indeed used within algorithms, but none has yet achieved a level where they can truly be called “intelligent,” as a scientist would understand the phrase.

Wider Uses
However, people are starting to look at machine learning in other contexts that could be equally as valuable to sell-side firms as algos with an appropriately aggressive moniker that they could use on their own desks. One head of a technology arm at an exchange that I spoke with last week related a number of these to me. First, in terms of infrastructure and development, was the use of machine learning in the testing phase. This is by far one of the most onerous tasks in modern software for high-level entities that hold crucial positions in market structures—everything must work perfectly, which means thousands of hours spent running through the code and the alpha builds. These tasks are time-consuming, repetitive, and yet entirely necessary.

What they posited was using an intelligent form of machine learning to automate most of these testing tasks, layering in technology that has the ability to correct errors it finds on an iterative basis—not just setting a pre-programmed subroutine to clean up bugs, but recognizing issues outside of standard templates, and employing non-standard solutions to fix them in the process. The stigma that science fiction has given AI shows through in this conversation, though: The exchange exec is firm on calling it machine learning, rather than artificial intelligence, which he says in a sheepish, brisk way.

Constant Monitoring
There are numerous other applications. Again relating to testing, AI algorithms have the ability to constantly monitor software performance in a technology stack, alerting engineers to problems that aren’t necessarily related to code. Humans will always be needed for physical boxes after all, as anyone who has wondered how the Terminators actually managed to build the factories where the metal monsters were created in the first place can attest. Machine learning also has another use in front-office operations, too­—if we live in a world where rogue algorithms can pump thousands of orders into a market before humans can react to pull the plug, perhaps the only solution is to combat them on their own terms, by using super-reactive, intelligence watchdog algorithms to do it for us.

Likewise, in market surveillance, the computational power of AI can also be used to alert compliance officers to new, potentially abusive scenarios that don’t fit the standard template. Many compliance engines are rules- and scenario-based, with only the most advanced being able to mutate new procedures in a cross-market, cross-asset environment, which arguably needs that capability the most.

The point is that the fiction needs to be separated from the science. We’re unlikely to have a future any time soon where traders will have sentient AI assistants, but there might well be a point where algorithms with elements of sapience help to manage a riskier, faster trading environment. And that point may be closer than we think.