Machine Learning & Artificial Intelligence - Financial Crime

It seems that everyone is talking about Artificial Intelligence (AI) at the moment: whether it’s Elon Musk and Mark Zuckerberg disagreeing publicly on the doomsday type scenarios that AI might bring [1], or banks predicting AI to be the primary way in which they interact with customers in the future [2], there’s wide-ranging interest in what AI can do for society as a whole, companies and individuals. But, to be clear, and before going further, what exactly is the difference between Machine Learning and AI, or is there indeed a difference?  The clearest explanation we’ve seen goes something like this:

·      Artificial Intelligence – this is the high level concept that machines can do something in a way that we, as humans, would consider “smart”

·      Machine Learning - is a current application of AI based around the idea that we should really just be able to give machines access to data and let them learn for themselves [3]. (Thanks Forbes!)

Similarly, in the financial crime space, numerous articles exist about how AI and Machine Learning can help to combat illegal activity in banking and beyond [4].

At FINTRAIL, we believe that AI and Machine Learning have huge potential to deliver great results in the financial crime space.  Whether it’s AI helping investigators to detect previously unknown connections between entities and typologies, or Machine Learning helping refine transaction-monitoring rules by different customer sets and behaviour, the benefits for companies and their customers are huge. Imagine for a moment that your bank could tell whether purchases made at a high-end online retailer at midnight just after you received a bonus cheque were genuine or fraudulent, based on your previous behaviour in a similar scenario.  Great, right?  No annoying text messages, or blocked transactions if it were genuine, and peace of mind that that kind of transaction would be blocked if it were fraudulent, and you didn’t actually have a compulsive online shopping habit (ahem).

But, as with anything new and relatively untested, there are pitfalls.  One of the key ones is making sure that any Machine Learning models start off with relevant data, such that they can begin the learning process appropriately, and you don’t program in algorithmic bias.  Typically – and let’s take the case of a Machine Learning engine for transaction monitoring -  this is relatively easy to build: you have a known scenario, which is fed into the engine for it to learn and refine over time as the transactional data is processed and fed into it.  However, this can be tricky in financial crime situations, as ideally you don’t want any money laundering or bribery (for example) to go through your system before you work out what the scenario or relevant data for the Machine Learning engine is. 

So, how do we address this?  Well, something we are passionate about at FINTRAIL is making sure that firms have a thorough risk assessment; truly understanding your business model and the ways in which criminals might seek to exploit it will help to build the best scenarios for any future financial crime Machine Learning engine. These can then be used to create the baseline relevant data that goes into the Machine Learning engine, such that it can start to learn behaviours. Examples here might include understanding your typical customer profile, such that you can build a Machine Learning model to automatically categorise them by risk profile, or Machine Learning models that take into account transactional behaviour and a range identifying particulars to reduce sanction re-screening hits.

Another tactic we’ve seen is to combine more traditional models with Machine Learning.  Again, in the transaction monitoring space, combining a rules-based approach with Machine Learning is a great way of teaching the engine to learn, and giving good baseline scenarios that it can work from.

So, all in all, we’re on Mark Zuckerberg’s side of this particular argument – we think AI has great potential, but that it, and Machine Learning in particular, needs strong data to support it, and as with humans, the right conditions to succeed.






Image Courtesy: Saad Faruque, Flickr (Creative Commons)