On Day Two of DATAx New York's Big Data & AI in Banking track at the Midtown Hilton, speakers covered a variety of topics and challenges facing firms attempting to leverage AI in the finance space.
However, the trials firms who have already begun to incorporate ML functionalities into their operation confront every day is dwarfed by those problems facing companies trying to enter the AI realm for the first time.
These first-time challenges are those Meninder Purewal, data scientist at the Bank of America, tried to address in his presentation. Following the completion of his first year as an adjunct ML professor at NYU, he tried to consolidate the information he had garnered over his time teaching himself and his students ML from the ground up, his interactions with the "quants" – quantitive analysts – who use ML models daily and the technology currently available. Here are the five lessons he picked up over that year.
Visit DATAx Singapore on March 5–6, 2019
Five introductory lessons learned from teaching ML in finance
ML is extremely hard: Because ML is such a new science, finding the right talent is nearly impossible. Purewal stated that the three qualities you should look for when trying to hire someone to work on or create ML models at your firm should be finance, math and computer science. "If you find someone who has two out three of those skills, you're way ahead," claimed Purewal. "And it is almost impossible to find someone with three out of three of those qualities."
ML in finance is extremely easy: Because it is still such a vague term, you can be a "data scientist" with no more than a cursory knowledge of ML algorithms. "You can go onto the internet, copy four lines of code and just like that, you can have a highly effective neural network," explained Purewal. "This is the danger and appeal of ML; people can incorporate these models into their processes, they work and all of a sudden, consider themselves data scientist." As Purewal put it, "some "data scientists" who work models on excel deal in megabytes, whereas those who work on the cloud deal in petabytes. That's a difference of a magnitude of a factor of nine". That is a huge difference in depth of knowledge.
Expectation versus reality: Purewal compared the current state of ML to the very early stages of the internet as, while he genuinely believes it will eventually change the world, the expectations placed on data scientists should be tempered at the moment. "It's a grind to make these things work. Everything is extremely bespoke and not generalizable."
There are not many great examples to teach with in ML: "The finance industry is instinctively secretive. They hoard secrets," admitted Purewal. This makes it inherently difficult to teach yourself how to build finance ML models. To get around this he suggested reading blogs, going on LinkedIn and reading ML posts and just trying to piecemeal information together because "there simply isn't a single source of information and that should give you an idea of where the technology currently is," he added.
Deep and few versus shallow and many: According Purewal, the art of ML is all about your approach: "Do you want to know a few algorithms very well or have cursory knowledge of as many algorithms as possible?" In his experience, he found researchers and specialists tend to gravitate toward deeper understandings while consultants and management tend to seek out cursory. "But it's practically impossible to have deep knowledge about many kinds of models," Purewal stressed. "So, advertise your skills appropriately as well as your learning goals."