Optimizing Non-Life Insurance Technical Reserves: The Contribution of Neural Networks Under the Solvency II Directive

Abstract

Insurance companies today face strong demands from regulations such as Solvency II. These rules require more accurate and reliable ways to estimate claims reserves. Both changes in the market and updates in supervision add to this pressure. As a result, many insurers continue to use the Chain Ladder method. However, this approach does not always capture the complexity of current insurance data. To keep up with these new demands, many companies are now exploring more advanced ways to estimate reserves. One of the most promising options is the use of neural networks, such as SplineNet. These models are well suited to today's complex data and can adapt when new trends appear. Because they learn directly from past data, they often produce more accurate results than traditional methods. This paper examines how neural networks perform compared to traditional reserving methods. We start by describing the main criteria for reserving under Solvency II. Then, we look at earlier studies that compare the two approaches. After that, we use simulated SPLICE data to test both the Chain Ladder and SplineNet models, focusing on accuracy, capital optimization, and compliance. The results show that neural networks bring real improvements. We find higher accuracy and lower capital requirements, which suggests these methods could shape the future of insurance reserving.