How Machine Learning Will Change the Way Banks Fight Fraud

There is a lot of excitement in the air about the potential of machine learning when applied to many walks of life, from banking to insurance, from cancer research to pharmaceuticals. Our imagination is captivated by the notions of accelerated learning, greater knowledge, and productivity gains.

While the concept of machine learning has been around since the late 1950s, it’s only in recent years that the required software infrastructure matured to the level where it is good enough and affordable enough to be adopted by large organizations.

It is expected that by the year 2020, there will be 5,200 gigabytes of data for every person on Earth, according to a Digital Universe study. The downside of this abundance of personal data is that nothing is “personal” anymore. If it’s stored on a device or on the cloud, there’s a chance that it will get breached one day. Over 15 million consumers were victims of identity theft last year, up 16 percent from the previous year, according to a Javelin report. $16 billion was lost to this fraud in 2016, up $1 billion from 2015.

The result of this stolen information is a steady increase of fraud across the globe. Global card losses have nearly quadrupled since 2010, from $7 billion in 2010 to $27 billion in 2017, according to The Nilson Report. These card losses are projected to keep rising to over $30 billion in 2019.

Keeping pace with the fraudsters

In this high-stakes environment, banks without big data contextualization capabilities will remain one step behind the fraudsters, exposed to new vulnerabilities and loopholes. A single breach can cost tens of millions, like the Bangladesh Bank heist of last year, where hackers withdrew $101 million from a Bangladesh Bank account at the Federal Reserve Bank of New York.

In the never-ending race to satisfy today’s real-time consumption culture, we are witnessing faster payments moving in at a staggering pace. From PSD2 to Same Day ACH (SDACH) to Zelle’s P2P offering, payments are becoming more and more exposed to fraudsters’ last-minute manipulations.

With a shorter payment integrity audit time, fraudsters can take advantage of the limitations of technology to screen large volumes of payments in real time. We need technology that can handle those large payment volumes and mountains of individual payment data while still making sensible decisions at an acceptably low False Positive Rate (FPR), all in real-time.

Consider that online banking fraud more than doubled the year after the UK launched the Faster Payments Service, in 2008, to reduce payment times between different banks’ customer accounts. Today’s real-time payments trend puts a larger burden on banks to prepare and protect against the rising risk of fraud.

Reaching limits, and transcending them

While in the old days, anticipating fraud through a set of pre-defined business rules was commonly sufficient, in today’s big data world, it’s practically impossible to anticipate how the next fraud breach will shape up. We need a machine that can generate thousands of small profiles for devices, terminals, users, payment types, originators, beneficiaries, geographies, etc., so an aggregation of anomalous indicators can point to account takeover or other criminal intent, all in real-time. At Feedzai, we call this capability the “segment of one”.

I joined Feedzai out of the belief that the magnitude and complexity of today’s fraud-fighting challenges require the latest and greatest machine learning (ML/AI) technologies. We will need to harness and learn to trust machine learning to perform multi-channel big data aggregation and analytics contextualization. This will enable banks and large Financial Institutions to identify complex fraud patterns that are too evasive for the human eye.

Machine learning can serve as the orchestration layer for many fraud prevention and authentication systems already in place and provide a fraud shielding umbrella that intelligently links the dots.  Machine learning will become the most critical layer in detecting and fighting financial crime.

As machine learning becomes more predictive and accurate, banks may let it gather and quantify multiple aspects of client risk (i.e. monetary, reputational, loyalty and churn and CLV)   to help make better long-term relationship decisions. Machine learning is here to stay, and I venture to guess that our wildest imaginations are only starting to scratch the surface in terms of what the technology could do for us ten and twenty years from now.

Subscribe to stay infomed