Why machine learning may help stop payment fraud
- 09 February, 2016 11:13
In 2017, the New Payments Platform (NPP) – infrastructure that offers real time payments between financial institutions and their customers’ accounts – will come into effect.
This essentially means that payments will be visible in a customer’s account within around 10 seconds.
What this also means is that banks’ risk systems need to work in real time and make a decision in around 5 to 6 seconds if a transaction should be authorised or not. Frankly, that’s incredibly fast and internal processes will need to be slick.
Under the covers
Banking systems have a set of complex fraud rules that are applied to transactions. These rules have been designed for a different scenario and with only a few seconds to decide, fraud algorithms must be concise.
When the objective is to have fast, secure and reliable payments, ‘design thinking’ is required to overcome the existing paradigm.
A number of other countries such as Singapore and Sweden have already launched faster payments systems. But it is clear that when you have moved from a system where there is a clearing house to transfer money and what I would call ‘old plumbing’, to a real time system – it should be no surprise that fraud is the major concern.
Instant messages, instant cash
As we move to a world where we not only use instant messages instead of email, we also move money at the speed of light with instant cash. At the same time, we are seeing increases in the incidence of money laundering and the funding of terrorist activities.
The National Automated Clearing House Association in the US recently announced it would start to take the first steps towards allowing faster payments and settlement. To enable this change, an agreed 8.2 cents per transaction is to be paid to the receiving institution.
But even with the proposed real time clearing house functionality in-place, back office systems are simply not ready to handle such updates.
Paypal’s fraud systems
The newer players such as Paypal use artificial intelligence (AI) technology with machine learning capabilities to help them identify fraud. These algorithms look for patterns from a user’s past purchasing history and once a new pattern is found, a new rule is created to stop repeated scams.
But how effective are these systems? Paypal has a 0.32 per cent of revenue fraud rate, which is significantly lower than the average of 1.32 per cent.
With this individual profiling, PayPal understand the kinds of items that I would buy and when there’s an anomaly in my purchasing patterns, Paypal is alerted that fraudulent activity might be going on.
In Australia, payment card fraud increased from 46.6 cents to 58.8 cents for every $1000 transacted from 2013 to 2014.
It appears that the only real way to combat this trend, which will only exacerbate with real time payments, is to adopt a deep machine learning (AI) approach.
New machine learning models
We are starting to see new companies that will help us to navigate the change brought about by real time payments. An example is IRIS Analytics, which was acquired by IBM last year.
The objective of this solution is to detect fraud at scale and speed, using a ‘machine learning’ model, which acts to assist the bank as a ‘virtual analyst’.
Without such technologies, it is impossible for companies to combat fraud. A great example is American Express which uses Hadoop and machine learning to manage the volume of more than 1 trillion of transactions per year.
Humans are still needed in this equation and the idea is to use the technology to process the data, and when required, it still makes sense to have a call or text to the cardholder. Such notifications can really help to get the right level of engagement and reduce the level of fraud.
Are we lemmings?
There is no doubt that the world is moving faster and we all want to use instant messages with instant cash. However, it is arguable that the existing payments systems are sufficient for 80 per cent of volume of business.
The business case to have real time payments is not strong but that is a moot point as we are all charging to that cliff. Let’s not be a lemming. The only ‘safety net’ is that we adopt new fraud engines that use advanced machine learning.
David Gee is the former CIO of CUA where he recently completed a core banking transformation. He has more than 18 years' experience as a CIO, and was also previously director at KPMG Consulting. Connect with David on LinkedIn.