Machine learning has increasingly attracted the attention of actuarial researches in the recent past. Particularly neural networks - maybe the most popular representative in this field - have repeatedly shown to detect complex mechanisms amongst data even under a low level of model specification. In this master thesis, we investigate the usage of feedforward neural networks in a regression framework for the purposes of estimating actuarial reserves, particularly in a combination with the established chain-ladder method, in the legal expenses insurance, a non-life branch uniting short- and long-tail business. By our experience in the area of neural networks, the wide range of predefined model configurations, which vigorously influence their predictive perfomance, can only be handled by a large amount of tests. With the help of real claims data, we investigate these effects and furthermore demonstrate the improvement through additional model extensions. We discover the necessity to a comprehensive experimental setup and finally determine a collection of models, which fulfill our requirement of a performance improvement over the conventional chain-ladder method, which serves as a benchmark. In that process, we pragmatically counter an instability issue of neural networks, namely their sensitivity to their initial conditions. We test different empirical methods to determine the most promising network architectures as well as optimizer specifications. In order to eventually make sound statements, we simulate various initial states and build our final predictions upon the arithmetic mean. Our evaluation will be mainly based on three different performance measures applied to a one year and a two year forecast.     «

Machine learning has increasingly attracted the attention of actuarial researches in the recent past. Particularly neural networks - maybe the most popular representative in this field - have repeatedly shown to detect complex mechanisms amongst data even under a low level of model specification. In this master thesis, we investigate the usage of feedforward neural networks in a regression framework for the purposes of estimating actuarial reserves, particularly in a combination with the establi...     »