We invite you to a backstage look at our tech solutions. Read how Nethone secured the global traffic of large online luxury marketplace with Machine Learning and used its proprietary Profiler to increase conversion rates while lowering both overall chargeback rate and the amount of manually-reviewed traffic.
The largest online luxury marketplace in the world reached out to us with three issues they had to confront on a global scale:
- The growing amount of manual review traffic generated by other anti-fraud solutions providers.
- High chargeback ratios, which had to be below 0,5% for all of our client’s acquirers.
- Insufficient chargeback detection, which resulted in financial burdens.
Within 4 months of full integration, our Machine Learning models — backed by our proprietary Nethone Profiler, which gathers over 5000 attributes and fingerprints about each user or device — increased conversion rates while lowering both overall chargeback rate and the amount of manually-reviewed traffic.
Technical details - how does our solution actually work?
Our fraud prevention solution developed for this e-commerce client was based on custom technologies, including in-depth user profiling, device fingerprinting, and advanced Machine Learning. Our actions resulted in a substantial number of slashed chargebacks all the while ensuring a seamless user experience for desktop and mobile customers. The accuracy of all recommendations was boosted additionally by proprietary Nethone Profiler.
Our Data Science (DS) team implemented two supervised machine learning models to detect chargebacks. We built the first to detect chargebacks directly, while we built the second with the intention of replicating the results of the previous fraud prevention solution's decision-making process. As a result, we managed to slash the manual review rate by nearly 60% while increasing the percentage of accepted traffic in comparison to competitors’ rates of accepted traffic.
Using an in-house Machine Learning pipeline, the DS team trained and tested multiple models using various cutting-edge statistical techniques. Ultimately, of the tested models, two acutely-tuned XGBoost models achieved the highest performance. The Data Science team deployed two fraud detection models in October as our company became the primary fraud-recommendation system for the client.
As the DS team monitored our models’ performance in production, the team made monthly adjustments to match the client’s internal benchmarks. In anticipation of large transaction volume during Black Friday and the following holiday season, the team scaled our data processing resources to ensure our model’s predictions would not experience any downtime.
After excellent performance during the holiday season — a time in which transaction volume is highest — our team deployed minor optimizations to our in-production models. These optimizations helped expand our fraud detection capabilities for risky geographies and targeted our previous models’ false positive rate to ensure we were only rejecting transactions that were most likely to be fraudulent.
The improvements had a large impact on the client’s bottom line, effectively lowering the overall chargeback rate by over 12% and lowering the chargeback rate for some of its most valuable brands by as much as 89%.
Results for the client
Thanks to the above-described actions, in under 4 months of full integration we cut down the client’s chargeback and manual review rates.
The growing amount of manual review traffic generated by other anti-fraud solutions providers resulted in large financial costs and unnecessarily burdened the client’s fraud management team. With our proprietary Machine Learning solutions, we were able to cut these costs significantly.
Ultimately, the client met its internal KPIs and kept its overall chargeback rate below 0.5% for all of its acquirers.