River Island transforms fraud strategy with automated decisioning

About River Island

River Island is one of the UK’s leading fashion retailers with a strong high street and ecommerce presence nationally and internationally. They operate 300 stores and six websites, selling men’s, women’s and children’s clothing and accessories.

River Island is an icon of the British high street and has been serving style and leading trends for over sixty years.

Background - before Ravelin

River Island’s fraud strategy was both rules and labor intensive. Human decisioning had been effective in keeping fraud rates low for the fashion retailer. But this meant that analysts were working 16-hour days. This was not sustainable or an effective use of resources – especially as order volumes picked up during peak periods.

To maintain sustainable and secure growth, they needed to get smarter with how they detected, reviewed and prevented suspicious orders. They needed a solution that could keep up with their growing and evolving online channel with minimal intervention.

Nick Kirby, River Island Fraud Operations Manager, expands on the situation:

“The impact of running a 7 days per week/ 16 hours per day fraud prevention operation was significant for a small team. Coverage was a constant stretch and it impacted morale and motivation. At a time when I could see that the industry was getting smarter with resource deployment, it felt like River Island had got bogged down without a clear sight of a suitable solution.”

Goal

River Island’s priorities were two-fold. They wanted to ensure that genuine customers had a frictionless shopping experience. But also protect themselves against fraud and its associated losses. River Island needed an intelligent and bespoke solution that reduced the need for manual intervention. And took full advantage of their extensive customer data.

They were looking for:

• Real-time machine learning decisioning to improve efficiency and accuracy

• Full ownership of the fraud detection management solution

• Collaborative engagement and iterative modeling

The process

The move from rules to machine learning was an iterative process over an extended period of time. Rules were added, removed and amended to find which ones were actually useful. The aim was not to remove rules completely, but to optimize their use alongside the machine learning model.

Grant Shipway, Global Fraud Manager at River Island, describes the collaborative relationship between the teams at River Island and Ravelin over this period:

“Our Islanders and the Ravelinos worked in harmony. We met every week and, once the pandemic allowed it, we would frequently meet in person for workshops and to discuss the longer-term strategy. This built trust on both sides as for us we want Ravelin to enjoy working with us as much as we enjoy working with them.”

Challenges

River Island had to overcome hurdles to understanding Machine Learning both conceptually and in practice. It was important to demonstrate how the models were comparable to but more accurate than human decisioning. And clearly outline its impact on fraud and or customer experience.

There were a lot of legacy issues that needed to be addressed when Ravelin replaced River Island’s previous fraud provider. This included analyzing the effectiveness of each existing rule to reduce redundant rule sets.

All of this was necessary to begin the work of setting the correct thresholds and establishing Prevent rule sets.

Other practical challenges included changes in personnel on both sides and business priorities that had to take precedence over this project. This was a long term project requiring commitment and perseverance over a sustained period in order to deliver the goals. But throughout the process, there was clear evidence of continuous improvement.

The results

• Reduction from 30,000 manual reviews in 2021 to zero manual review requirement by peak 2022

• More than 2,500 analyst hours freed up to focus on new challenges, such as first-party fraud prevention

• Chargeback and fraud rate maintained at a consistent level. The iterative nature of the process meant that there was no adverse impact on these metrics.