Machine Learning and financial crime

When we founded ClearBank, one of the key questions we set ourselves was, “how do we leverage technology to be better”. When I say “better”, we meant better at particular services or functions that a clearing bank have to undertake or offer. This one question really me on the journey of delivering a true Banking as a Service (BaaS) offering to other regulated institutions and FinTechs. It’s also why we wanted to explore Machine Learning to help in the fight regarding financial crime.

I will be honest, I don’t find “RISK” management fun, however, as part of managing risk and trying to mitigate financial crime, things can get quite innovative, and that is where the fun can begin. These areas have some real obvious, yet rather powerful application uses for deep machine learning. I know some will call this AI, but it’s far too narrow to be called AI (I have a real bug bear with people calling things AI when it is Machine Learning – let’s save that for another day though). So, these use cases, what are they:

1. Fraud detection
2. Anti-Money Laundering

Most of us when we read fraud detection think of our experiences with our payment cards, either someone else is seemingly able to purchase “stuff!” with our cards, or we get stopped and cannot use our payment card because the bank thinks something fraudulently is going on. However, fraud is wider than that, think ID theft, actually having your bank account taken over by other individuals, these are two other areas of fraud that impact many of us today. So, how can machine learning help with actual transactional based fraud, and fraud such as account takeover? The answer is to use it to learn about you, you as an individual, and what I like is also learning about you in context of your peers activities. I will come back to that one in a moment.

The second obvious use case is that of AML (Anti-Money Laundering). This is where we can use machine learning to help identify money movements that could indicate a form of money laundering, especially within closed groups. One of the benefits here of being an actual direct clearer (as in connecting to all the payment schemes) is that you can track the money movements across all the channels, helping to gather sufficient data that a machine learning platform can start to identify money laundering techniques.

For the purpose of today’s post though, let’s just focus on fraud detection…

 

What’s normal?

Let’s use machine learning to learn what your transactions look like. Believe it or not, most of us are creatures of habit, we buy coffee typically from the same shops, we purchase our lunch or shopping at similar times in the day from similar locations. We visit restaurants on date night (which for me is always a Friday night with the wife), we drink at the same bars etc etc etc. You get my point. Machine learning can take all that data and start to build a profile of your normal activity. Sure you will have the odd splurge on some big ticket items, a holiday, a sofa, a car etc but across all your activity Machine Learning can build a pretty accurate picture of what looks like “normal” activity for you, as opposed to what looks “strange” for you.

We can apply similar learning to how you access your account, locations when you access it, the devices you use, the time of your access etc. These data points help form a profile again of what “normal” looks like, and therefore “strange” can be identified.

This all sounds great right, however, how many of us have incidents when we find we cannot make a payment, or we get alerts saying, “due to fraudulent activity your card has been suspended”? This can be the result of a “rules” based matrix approach, trying to spot fraud, or a machine learning implementation that could be better. Essentially your provider is identifying “strange” and creating what we call a “false positive”, in other words, thinks it’s fraud when it isn’t.

 

I love context

With Machine Learning, you can add layers of learning, so why not add an additional layer that looks at the context of “normal” or “strange” in relation to your peers. Let me give you an example, because without it I don’t poses the capabilities to articulate what I mean…

You never gamble on horse racing, however, it’s the Grand National here in the UK, and you fancy a flutter. When placing your bet, this looks like “strange” activity for you and your account. It can easily be flagged as attempted fraud and you are stopped from placing that bet. However, if your banking providers machine learning platform understands “context” it can make a better assessment. See, the Machine Learning platform could learn that it is the Grand National, it could also learn that your peers are also all placing bets on that race, this too could look like “strange” for individuals, but as a group of you, all of a sudden this doesn’t look like strange activity. Essentially, your Machine Learning platform has learnt the “context” of that activity, therefore it looks “normal”. The result, well instead of getting stopped from having that flutter, your horse comes in, you make a fair few “quid” and everyone is happy…. The power of Machine Learning with “context”.

 

Compute compute and a little more compute

Machine Learning is highly powerful, and I hope you see just how capable and helpful it can be at protecting your account from fraudulent activities. However, you need data, lots of it, and lots of compute power to crunch those numbers and algorithms to actually provide a decent Machine Learning based platform. The challenge therefore is having enough compute power to learn at an individual level, but also at a group contextual level. Until the Cloud really came along, this made Machine Learning a tool that only the real big players could leverage, simply because of the cost of purchasing enough physical compute power. That’s all changed, the cloud allows us to elastically scale resources associated with Machine Learning up and down, which drastically reduces the cost involved. It also brings far greater flexibility in terms of how these platforms are built and connected into the banking systems.

At ClearBank I always wanted to ensure we had sufficient compute capabilities, that’s why our Machine Learning solutions reside within our Azure cloud, giving us access to all the compute power we need, when we need it. We have partnered with some pretty cool technology companies too, such as FeatureSpace, enabling us to build out deep powerful machine learning solutions to fight financial crime, which do understand “context”.

 

Quick recap…

Essentially a good Machine Learning based solution can protect your account from fraudsters. It can learn what normal looks like for you, and when understanding “context”, it can even spot when activities are yours that typically don’t fall into your normal activity. The keys to unlocking this level of capability is harvesting enough data, and having the compute power to process it all. The cloud here is an enabler, helping financial service providers take advantage of the endless scale of the Cloud in the fight against financial crime.

It would be great to hear your comments and thoughts on this, but also any ideas or applications where you can see the use of Machine Learning really having an impact in financial services.