Cyber risks are rapidly evolving threats for organizations and cyber insurance serves as an integral part to transfer risk. The insurance market for cyber is relatively young and standard approaches for actuarial risk assessment are still to be established. Apart from an adequate model specification, access to relevant data is limited. Based on a policy and claims database from a leading cyber reinsurer, we build an extensive dataset for commercial line cyber insurance with more than 400,000 policies. For this, we establish a data pipeline to identify a subset of high-quality data and conduct in-depth analyses to assess the approximately 100 potential risk factors. Following the approach of a collective risk model, we focus on the frequency modeling, where a binary target is used to indicate whether a policy has suffered no or at least one loss. During feature engineering, we add insurance-specific concepts, e.g., controlling for incurred but not reported (IBNR) effects. We test multiple configurations for generalized linear models (GLM), regression trees, and random forests, where the best performing one is chosen to compete against the other model classes. Overall, we observe that the random forest performs best. The GLM shows a comparable performance and has the advantage of a good interpretability of the model parameters. The results show that inference is possible using the available covariates. We also discuss potential extensions to the presented modeling approach that could provide additional insights or improve the analysis.     «

Cyber risks are rapidly evolving threats for organizations and cyber insurance serves as an integral part to transfer risk. The insurance market for cyber is relatively young and standard approaches for actuarial risk assessment are still to be established. Apart from an adequate model specification, access to relevant data is limited. Based on a policy and claims database from a leading cyber reinsurer, we build an extensive dataset for commercial line cyber insurance with more than 400,000 po...     »