Skip to Content

The Changing Landscape of Application Fraud Detection

Nick Sime, Head of Modelling

The author

Nick Sime

Head of Modelling

It's been 25 years since Nick Sime, Jaywing's Head of Modelling, built his first application fraud model. Reflecting on the evolution of these models and the rapid adoption of AI/ML-based approaches over the past couple of years, Nick outlines fraud’s changing landscape and highlights some key areas of focus for financial institutions.

The fight against fraud

Undoubtedly, all lenders want to minimise the ‘cost of fraud’. This cost has multiple aspects, most obviously the actual fraud losses incurred, the operational cost of referral activity, 3rd party data and system expenses, and the opportunity cost associated with delaying applications or inconveniencing applicants with requests for information that reduce conversion.

Linear scorecards, standard credit reference, and traditional rule-based systems, all of which have been around for decades, have previously dominated the fraud landscape. Often working in silos, these methods of fraud detection proved challenging to many financial institutions, particularly due to their time-consuming nature. So, while the fight against fraud is not new, the ways in which firms are mitigating this risk has evolved significantly.

UK Finance reports that fraudsters stole a total of £753.9 million in the first half of 2021, an increase of +30% from the previous year. During the pandemic, fraudsters evolved their strategies to take advantage of growing financial insecurities and increasing internet usage. Indeed, today, almost all scams contain an online element.

Covid-19 also had a profound and immediate impact on many lenders as they adopted a more prudent credit strategy and, in some cases, even ceased lending for a brief period. Changing behaviours, particularly that of increased internet usage, triggered higher potential for 3rd party fraud to occur and forced financial institutions to revisit their detection and prevention strategies. Finally, one of the most obvious changes we have seen over the last couple of years is that lenders are rapidly moving from regression-based models to the use of AI/ML-based approaches, creating far more powerful models than ever before.

Out with the old, in with the new: the mass adoption of AI in banking

With the mass adoption of AI/ML-based fraud models, particularly from challenger banks and fintechs, it’s become clear that AI-based scoring is largely replacing the traditional scorecard. Rule-based systems might allocate points to suspicious events, but it’s only by developing a statistically-based score that this data comes into its own, helping you to make sense of the noise. The latest techniques can generate the most accurate outcome, helping your fraud investigation team to focus on what’s important from a host of data sources.

AI-based approaches allow financial institutions to keep up with ever-evolving fraudster techniques. While these models are robust, it is good practice to ensure that they can be updated and refreshed to ensure firms can stay ahead of criminals' strategies. Frequent refreshes are far simpler with AI-based models than traditional rule-based/linear approaches, allowing you to update your system as soon as a threat emerges or even on a reactive basis, if needed.

An ongoing challenge for many financial institutions involves explaining the results of these AI tools - understanding exactly how they do what they do. Jaywing’s AI predictive modelling software, Archetype, overcomes this “black box” problem. As the 2021 winner of the Credit & Collections Technology Awards for best 'Anti-Fraud Solution', Archetype allows lenders to view all the model's processes and inner workings, as well as inputs and outputs. Archetype is explainable on both a global and local level. Global explainability ensures that model inputs can be constrained to have a logical impact on model predictions. Local explainability provides transparency on an individual case level, allowing underwriters to understand why an application has been referred for review. This explainability is becoming vital to an effective fraud model as it allows the underwriter to investigate suspicious applications in the most appropriate and efficient way.

Prevent fraud or satisfy your customer? Do both

As the fraud modelling landscape evolves, new challenges arise. The emergence of numerous fintech companies and challenger banks have made the credit landscape far more competitive. Organisations are heavily focused on minimising friction within their application process so that automated lending decisions can be made quickly, without frustrating the customer and negatively impacting the customer journey. This adds to the challenge of making approval models and processes as slick and powerful as possible.

Finding a balance between mitigating risk and satisfying customers is imperative - and it will involve your whole business. Lenders must strengthen their fraud model and automate where possible whilst also working closely with their marketing teams, and potentially external CRO & UX experts, to ensure that all communications are clear and transparent and that the customer journey is seamless.

The fight continues

Following an increase in fraud this year, some lenders may be looking to further develop their fraud detection systems. It must be recognised that adopting too many systems could create an unwieldy process comprised of disconnected systems, each generating its own set of rules and referrals that potentially choke the onboarding process.

Before licensing a new system or source of data, it is highly recommended that a proof of concept is performed so that their incremental value is fully understood. One must assess the potential impact on multiple KPIs including; fraud losses, refer-rates, and false-positive rates etc. Investment in new data systems need to be cost-justified through detailed analysis and ROI of existing data/systems should be reassessed as strategies change.

Organisations should continue to harness their predictive models from relevant available sources, this often includes data primarily used for credit purposes. The model development process should reveal which sources of data/systems are giving the greatest ROI. Weaker or more expensive systems should be challenged, particularly as new competitor products come to market.

In an increasingly competitive market that sees a growing share of business originate through brokers, the battle for business is largely based on price and product features. To be able to price and design attractive products and gain market share, lenders must strengthen their underlying decisioning models and ensure they are utilising the most advanced and innovative technology.

If you’d like to learn more about Jaywing’s award-winning fraud solutions and find out how we helped Virgin Money identify 86.6% of fraud incidents by reviewing just the top 10% of cases, click here.