IBM Watson

Following progress in the healthcare space made by IBM Watson (Dr Watson), the AI is now turning its gargantuan machine learning intellect towards financial regulation.

In November 2016, IBM acquired Promontory Financial Group, a global compliance consultancy run by former top level regulators, to combine this expertise with Watson's cognitive abilities. To begin with the IBM Watson Financial Services regulation initiative has focused on key compliance pain points – fraud, AML, KYC and corporate surveillance – where manual processes can give way to the lightning recall of all sorts of structured and unstructured data, leading gradually to context specific understanding.

In healthcare, Watson showed its ability to read vast tracts of doctors' notes, medical journals, process images, MRIs, x-rays; and in areas like oncology, the computing system received training from the best medical minds at places like the Mayo Clinic. The aim was to provide doctors with the best possible recommendations.

Alistair Rennie, GM, Watson Financial Services Products, said: "Instead of artificial, it's more augmented intelligence, which is really important when you are dealing with professionals."

Rennie explained the work being done with Promontory is not programming rules or defining the way things are handled in hard code. Rather the system learns by example, by the decisions people are making one by one and thereby gets better and better understanding. In this respect the system is only as valuable as the humans teaching it and Promontory fits that bill (its CEO Eugene Ludwig was the head of the Office of the Comptroller of the Currency). Because as well as the monolithic legalese comprising various financial rule books, it's important to grasp the rationale behind the regulations, not to mention the nuanced ways these can be interpreted and implemented.

"What is interesting is regulators have different styles. Some regulations are meant to be quite explicit, but many are directional and are left to interpretation with the institution. Doctors' notes are similar; they are often not entirely definitive. So rather than generic machine learning, we have built up a really deep ontology to help understand if, say, it's a regulation from this body, then we read it this way and here's how we extract obligation.

"At the end of the day a person inside a bank has to make the decision about which obligation they are going to take. There is no way to outsource that to a computer, and nor should there be. As people make a definitive selection of what obligations they are going to take, we get to learn by the decisions that they are making and it kind of recursively gets better."

News is an important part of what the systems does, made more tricky in these times where Trump might tweet its time to begin dismantling post-2008 regulatory regimes. Watson weighs the importance of something like a speech by Mark Carney or an FCA announcement in terms of the probability it will have of impacting some obligation. Other unstructured data includes examiner notes or policy papers, and lots of work has been done on voice. More broadly, Watson has also been working with large accountancy firms to build up understanding of an audit using machine learning.

Looking ahead, a goal would be to allow clients a horizontal view by line of business or perhaps country. Rennie used the example of a once a day liquidity calculation, to be able to offer that once an hour. "You would be able to ask a simple question of Watson, such as, 'are we optimised for liquidity?' Being able to answer that question requires connecting a lot of dots."

Interpretability is in Watson's digital DNA, as opposed to developing black-box machine learning, even when it comes to the fast paced world of anti-fraud. "We are focused on what we call 'white-box rules': basically being able to explain in a clear way to a fraud analyst what change needs to be made, and how to make it in a way that's clear to the person responsible for supervising or auditing the system," said Rennie.