With around two million active shoppers and over 90 million transactions per year, Tinka’s challenge was to process as many legitimate payments as possible without causing user friction. To achieve that goal, the company needed help from Eraneos in leveraging machine learning (ML) and artificial intelligence (AI) capabilities to better know its customers, reduce risk, and ensure compliance with strict financial and regulatory standards – all while keeping to the client’s values of transparency and honesty. To adjust quickly to changes in fraud and/or pay behavior, these models also needed to be easily retrainable and explainable, which would help reduce bias and improve the prediction process. With this as our main focus, we got to work.
Together with Tinka, we set out to improve data activation capabilities by building an end-to-end machine learning platform using AWS Sagemaker, which would allow the company to prepare, build, train, and deploy high-quality ML models quickly and efficiently.
On top of that, we supported Tinka through the development of several easily-retrainable models for invoicing alongside a fraud detection model. We also planned to implement explainability models that were built using XGBoost algorithms to help them avoid black-box situations.
Tinka’s ambition is to be the preferred financial services partner for thousands of (web) stores across the Netherlands. Its portfolio includes financial products such as pay-later options, pay-in installments, and credits. Through its (web) store partners, the company serves millions of shoppers in a responsible and secure way. This is accomplished thanks to a Credit Risk Management system, which allows it to closely monitor the spending behavior of its customers so that products are used responsibly.
We built and deployed an innovative end-to-end Machine Learning platform using AWS Sagemaker, which allows the client to test and develop advanced AI and ML use cases. Next to that we built two cloud-native and retrainable invoice models and helped Tinka implement their own fraud detection model. The invoice models predict the probability of defaulting, while the fraud detection model predicts if a transaction is fraudulent.
Next to that, explainability was integrated into all the models both on local level (each transaction) and global level (which features are predictive of fraud according to the model). This way customers can also get an explanation if their pay-later option is declined and what drove the algorithm’s decision. The purpose of adding explainability models is to continuously improve the algorithm, avoid bias, and keep high level of accuracy.