Unsuperised Anomaly Detection : Methods and Application on Solvency 2 Technical Provisions

Abstract: This thesis work examines anomaly detection methods on large data sets related to insurance funds. Starting from requirements of low time complexity, ease of implementation and thorough definitions of contextual- and collective anomalies, different modelling frameworks are examined. Twelve time series models and a replicator neural network are presented in detail, while other modelling methods that are considered promising are mentioned more briefly. It is shown that a very simple time series-model can be widely applied to detect collective anomalies. It is also shown that a relatively simple replicator neural network can be applied to detect contextual anomalies.