On Claims Reserving with Machine Learning Techniques

Machine learning has gained popularity due to advances in computing power and access to large datasets, enabling the training of complex models. This thesis explored its potential in insurance reserve setting by comparing Gradient Boosting Machines (GBM) and neural networks with traditional statistical models. The results showed that machine learning often provided more accurate reserve estimates, with GBM performing best in most cases by balancing over- and underestimations. While all models worked well on simulated data, they faced challenges with real-world data. Overall, the findings suggest that machine learning, offers an interesting approach for improving reserve accuracy in non-life insurance.