Banking recommendation system

About the client

The client is a leading commercial bank with a large branch network and a multi-million customer base. The bank is actively investing in digitalization and automation of customer services to improve personalization and user engagement.

How it all started

In 2024, XPG Factor began working with the client on an initiative to improve the bank’s customer experience on digital platforms. The bank’s team approached us with a request to develop an intelligent recommender system capable of analyzing customer behavior and predicting interest in financial products.

Challenges we faced

The project involved several technically challenging issues. One of them was the wide variety of behavioral patterns among clients. Their financial activity varied significantly by segment, region, income level and consumer habits, making it difficult to build a universal algorithm.

Additionally, the model worked with data where the majority of values were zero, reflecting very limited customer interest in certain products. In some categories, only 2% of users showed any activity. Our priority was to make sure the model learned to recognize and incorporate even such rare cases.

Implemented solution

As a result of the project, we delivered a scalable system that can grow with data volume and adapt to evolving customer behavior patterns.

To enhance the precision of predictions, we used a range of models, from traditional gradient boosting to deep neural networks (DNN, LSTM), as well as hybrid models. Thus, the Transformer-based model demonstrated outstanding performance with new customers.

Solution’s key attributes

• Accurate predictions in cold-start scenarios
  Thanks to its architecture and hybrid models that analyze similar behavioral patterns, the system provides accurate predictions, even under cold-start conditions.
• Processing of sparse financial data
  The system is designed to handle large financial data where many fields are empty or have non-zero values, especially in some product and transaction categories.
• Instant recommendation delivery
  With real-time data processing, the system enables bank employees to respond promptly to user actions.
• Auto-validated ML lifecycle
  The system is built around continuous learning, with automated quality control in place. This ensures that only validated updates are released into the live environment.

Security and privacy

During the project, we implemented a set of strict security measures that fully complied with the bank’s internal requirements. Thus, data transfer between system components was carried out via protected internal channels, and the entire system was hosted in the client’s isolated cloud infrastructure. 

To provide more control, we have also implemented role-based access, which provides full transparency and reliable supervision of how the system is used.

Testing

To check the model’s ability to handle new customers, we formed a test sample including users whose data were not used in training. This allowed us to test how accurately the system predicts interest in products without any user history.

The model’s performance was evaluated using key metrics such as AUC-ROC, Precision@K, Recall@K and F1-score, which provided insights into the accuracy, completeness and balance of the predictions. Additionally, after system deployment, we conducted numerous A/B tests.