04.10.2016 Views

PCM Vol.2 - Issue 6

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Thought Leaders Corner<br />

Fighting fraud with Bayesian analytics<br />

Effective fraud detection has many forms,<br />

from traditional rules-based approaches to<br />

sophisticated analytical models and machine<br />

learning.<br />

out from the crowd in its machine learning<br />

capabilities – the system can learn rapidly as<br />

frauds are tagged, constantly learning and selfcalibrating<br />

as fraud trends change. This means<br />

that fraud teams know that they are in the best<br />

possible position to identify and stop fraud when<br />

it happens.<br />

Within the industry, the most widely used<br />

analytical approach is to build Neural<br />

Network models using historic<br />

consortium data to build up a<br />

set of patterns to later use in<br />

detection of live card fraud.<br />

This approach has been used<br />

for many years with varying<br />

success. However, recently<br />

new approaches have become<br />

increasingly popular and have<br />

begun to overtake the results<br />

being achieved with traditional<br />

modelling techniques.<br />

NCR has pioneered a modern<br />

approach of fraud analytics,<br />

choosing to base its models<br />

on a more flexible Bayesian<br />

analytical technique. This<br />

methodology enables fraud<br />

teams to build models faster<br />

than it is traditionally possible,<br />

while achieving extremely<br />

good detection rates at low<br />

false positives.<br />

About NCR<br />

easier.<br />

The models form parts of the<br />

Fractals Adaptive Classification<br />

Engine, which provides automated, intelligent<br />

fraud detection using a combination of Bayesian<br />

statistical analysis and proprietary inference<br />

techniques. By applying mathematical models to<br />

each incoming transaction, the system identifies<br />

suspicious activity, calculates a probability-based<br />

fraud score that indicates the likelihood of fraud<br />

and then triggers an action.<br />

NCR Corporation (NYSE: NCR) is the global<br />

leader in consumer transaction technologies,<br />

turning everyday interactions with<br />

businesses into exceptional experiences.<br />

With its software, hardware, and portfolio of<br />

services, NCR enables more than 550 million<br />

transactions daily across retail, financial,<br />

travel, hospitality, telecom and technology,<br />

and small business. NCR solutions run the<br />

everyday transactions that make your life<br />

Unlike a Neural Network, which is nothing more<br />

than a “black box” that produces a score,<br />

Bayesian models produce scores that<br />

come with a clear set of reasons<br />

and probabilities. This “white box”<br />

approach gives the end user much<br />

more granular information to work<br />

with when assessing a transaction<br />

or speaking with a customer and, as<br />

we know, more information means<br />

better customer service - the key to<br />

successful fraud operations.<br />

Based on the risk score, pre-defined<br />

actions are triggered and the<br />

transactions are approved, rejected<br />

or referred. Custom rules can then<br />

be set for referred transactions. If<br />

an unusually large transaction is<br />

detected for a customer profile, the<br />

bank can inform the customer via<br />

SMS or e-mail about the suspicious<br />

transaction and get confirmation<br />

that the payment is authorized.<br />

Comparative tests have shown that<br />

the Bayesian models outperform<br />

industry standard solutions<br />

in detection rates and cause<br />

significantly fewer false positives. Enhanced<br />

with self-learning functions such models that<br />

can automatically adjust to new fraud scenarios.<br />

Fractals scores transactions in real time and can<br />

decline a transaction during the authorization<br />

request.<br />

The future of fraud<br />

The model is initially set and tuned using financial<br />

institution’s recent historical data and fraud tags<br />

so that it can identify the unique fraud patterns<br />

facing each unique organisation. The result is<br />

a bespoke and highly accurate fraud detection<br />

model that can be set up in just a few short<br />

weeks.<br />

Each model in the Adaptive Classification Engine<br />

is based on a series of mathematical algorithms<br />

that are targeted to identify specific fraud<br />

patterns or irregularities in the account’s holder<br />

behaviour. However, the system really stands<br />

The truth is that fraud isn’t going to disappear any<br />

time soon – criminals have got too much to lose.<br />

There will always be new breaches or techniques<br />

to get past even the strongest security and fraud<br />

systems. And, unfortunately, there is no solution<br />

which is the answer to every problem.<br />

However, a fraud system that can assess a<br />

transaction in real time, based on a comprehensive<br />

understanding of customers’ historic behaviour<br />

and intelligent analysis compared to the latest<br />

fraud trends, will arm your fraud team to stop<br />

the highest possible levels of fraud.<br />

005

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!