2. 9. 2025 Žaneta Dufková
We recently completed the development of a machine learning–based Initial Reserve Predictive Model for a client ranked among the TOP 5 in the Croatian market.
Using XGBoost with the Gamma regression loss and optimized hyperparameters, the model consistently outperforms the client’s existing reserving scheme. While there’s still room to enhance predictive power through additional data sources and features, this initiative marks a step forward in modernizing and automating claims reserving with data-driven methods.
Key highlights:
- Better accuracy than the current business implementation
- Supporting (semi-)automated claim reserving as a replacement for manual processes
- Transparent and structured performance evaluation on training & test data
- Exploration of structured and unstructured data (including claim text descriptions)
This project demonstrates how data-driven reserving can enhance actuarial practice, providing a scalable way to refine reserve estimates with advanced methods.