This is the accepted version of the paper.This version of the publication may differ from the final published version. Permanent repository link:

We describe an expert system to monitor the stability of insurance markets. It consists of two components: an agent-based simulation component and a temporal data mining component. Like other financial markets, insurance markets experience destabilizing cycles and suffer episodic crises. The expert system assists market regulators by monitoring the financial position of individual insurers and of the overall market, and by forecasting cycles and impending insolvencies. The agent-based simulation component runs a forward simulation allowing for interaction among insurers in a competitive market, and between insurers and customers. The temporal data mining component extracts useful information for market regulators from the simulations. A prototype of the system is applied to the automobile insurance market. We show how the system may be used to forecast cycles, investigate stability, and analyze insurers' herding behavior on the market. A practical policy conclusion is that regulators should monitor individual insurers' pricing pattern because aggressive price undercutting creates a "winner's curse", with subsequent losses and market instability.

Insurance is critical to the fabric of modern societies and economies, but the insurance industry continues to su er deep cycles and periodic crises. These have a great socioeconomic cost as insurance cover can become prohibitively expensive or unavailable, damaging livelihoods, property, belongings and employment. These phenomena are poorly understood. A set of socio-anthropological and behavioural hypotheses have recently been posited. We investigate these explanations by means of an agent-based simulation model. The model is parameterized on actual property insurance industry data and is carefully validated. Our main result is that simple behaviour and interaction at the individual level can result in complex cyclical industry-wide behaviour. Heterogeneity and interaction at a micro level must therefore be understood if cycles and crises in the insurance industry are to be managed and prevented.

No abstract

Underwriting cycles are believed to pose a risk management challenge to propertycasualty insurers. The classical statistical methods that are used to model these cycles and to estimate their length assume linearity and give inconclusive results. Instead, we propose to use novel Time Series Data Mining algorithms to detect and estimate periodicity on U.S. property-casualty insurance markets. These algorithms are in increasing use in Data Science and are applied to Big Data. We describe several such algorithms and focus on two periodicity detection schemes. Estimates of cycle periods on industry-wide loss ratios, for all lines combined and for four specific lines, are provided. One of the methods appears to be robust to trends and to outliers.