Voice of the Industry

Best practices in fighting fraud and reducing false declines

Friday 25 February 2022 13:45 CET | Editor: Simona Negru | Voice of the industry

Katarina Garai, EMEA Retail Loss Prevention Lead at SAS Institute, suggests why identification of genuine transactions is vital to a successful business nowadays and how to reduce false declines

Identification of genuine transactions and minimising false positives is a vital part of a successful business nowadays. One could even argue that this delicate balance can represent a difference between a market leader and an average, less successful player. First of all, why do false declines happen? Simply put, they happen whenever the transaction is flagged as potentially fraudulent, but in fact, it isn’t. The term false positive or false decline is a broad term that describes a situation in which a company’s fraud detection system triggers an alert, yet there’s actually no fraudulent activity taking place. This is very dangerous, as triggering unjustified alerts diminishes the value of real alerts. If you have too many false positives to investigate, it becomes an operational nightmare, and you most definitely will overlook real fraudulent attacks. There are also false negatives, which represent a fraudulent activity that is not being detected by the fraud detection system. Expectedly, the last category is formed by true positives, which is the category that represents successfully detected wrongdoing. The chart below provides a simple visual representation of these concepts.

Distinction between true positives, false positives, true negatives, and false negatives. Graphic inspired by blog post: Click flooding detection and the false-positive challenge

False declines versus customer experience, satisfaction, and retention

Among younger demographics, in particular, digital channels are increasingly becoming the default shopping platform – anything that can’t be effortlessly done online is seen as a major inconvenience. False positives represent the bane of fraud investigators’ existence, the same way as increased volume of falsely declined transactions can represent the difference between a happy customer and a steadily rising slope of customer complaints, abandoned shopping carts, and decreasing ability of customer retention. 

So, let’s further see what the most effective strategies in designing an optimally functional fraud management system are, while ensuring the right balance between great customer experience and preventing fraud losses.

Introducing AI and analytics 

By using effective machine learning strategies, companies can manage fraud and risk across the entire customer lifecycle without alienating good customers. AI/ML techniques can be utilised to build models that can analyse previous cases and separate out the behaviour patterns that are truly suspicious from the purely superficial anomalies. 

Traditional rules-based approaches to fraud detection are still useful but unequal to challenges of today’s highly digital world. They capture the patterns known to indicate possible fraud – and they’re relatively easy to understand. The other challenge is that fraudsters are constantly probing, and they are – in some cases – able to ‘guess’ the rules and avoid triggering alerts.

For example, in my personal experience, whenever we have introduced a monetary threshold into a fraud detection rule, such as ‘If the amount is EUR 5.000 or more, flag this transaction’, the fraudulent transaction was captured at EUR 5.000, but the fraudster then tries and succeeds at EUR 4.700, and has learned how to increase their odds of success. 

Through machine learning, we can implement large numbers of conditions, variables, and events into analytical models and detect patterns that would slip by the business rules and could never be noticed by a human investigator. The combination of machine learning models and business rules has proven to be much more effective at finding fraud than business rules alone. When we use combinations of machine learning methods – including some of the newer ones like random forest, gradient boosting, or deep learning – the models become extraordinarily accurate.

Challenges with machine learning models

Since machine learning models can look into so many things in so many ways, their outputs can be very hard to interpret. Yet the fraud investigator needs to understand the rationale behind a fraud alert and make an informed decision about the necessary investigation path. This can be resolved by building a surrogate model to present and explain the results (white box companion) in the form of a scorecard or a set of visuals or an auto-generated narrative of the key conditions indicating fraud. The objective is to provide the investigator with the data and insight to explore the case – and not be wasting time by wondering how the model works.

Another common point of failure when implementing advanced analytics, is relying too heavily (or even exclusively) on a single approach, a single type of model, for detecting fraud. The variety of techniques that fraudsters are using today requires a combination of approaches to spotting them.

Achieving the right balance

In my opinion, machine learning and detection rules are always going to work side by side. We can see increasing efforts being applied to the combination of machine learning models and business rule sets, often referred to as the hybrid approach

To understand how these concepts can be transformed into reality, let’s imagine an organisation that has previously achieved a certain maturity in its fraud detection systems using a sophisticated combination of business rules. We can observe that in order to detect 64% of the fraudulent transactions, the fraud investigators needed to review 30% of the transactions referred by their existing detection system. After the implementation of a hybrid detection model, combining the right portion of business rules and advanced analytical modelling, the fraud investigators were able to review just 9,6% of referred transactions to detect this percentage of fraudulent transactions.

Advanced Analytical Model Implementation results

Customers are without doubt looking for convenience throughout their shopping experience. Constantly inventing and re-inventing ways how to continuously improve the convenience of the customer experience whilst maintaining the right fraud detection and prevention strategies is the uneasy quest of today’s merchants. AI and advanced analytics are here to help.

This editorial is part of The Fraud Prevention in Ecommerce Report 2021/2022, the ultimate source of knowledge that delves into the evolutionary trail of the payments fraud ecosystem, revealing the most effective security methods for businesses to win the battle against bad actors.

About Katarina Garai

Katarina is the EMEA Retail Loss Prevention Lead at SAS Institute and is helping companies understand how advanced analytics can help solve their struggles with fraud, waste, and abuse. Before joining SAS, Katarina worked as Head of Fraud, leading a team of fraud fighters, and cooperating with merchants across Europe on their fraud management strategies. Her focus has always been Fraud management, Internal Controls, Internal Audit and Compliance.

About SAS Institute

SAS is the leader in analytics. Through innovative software and services, SAS empowers and inspires customers around the world to transform data into intelligence. SAS gives you THE POWER TO KNOW.

Free Headlines in your E-mail

Every day we send out a free e-mail with the most important headlines of the last 24 hours.

Subscribe now

Keywords: false declines, false positive, identity verification, fraud management, artificial intelligence, machine learning
Categories: Fraud & Financial Crime
Countries: World
This article is part of category

Fraud & Financial Crime