Robustness in forecasting future liabilities in insurance

Abstract

The Gaussian distribution has been widely used in statistical modelling.Being susceptible to outliers, the distribution hampers the robustness of statistical inference. In this paper, we propose two heavy-tailed distributions in the normal location-scale family and show that they are superior to the Gaussian distribution in the modelling of claim amount data from multiple lines of insurance business. Moreover, they also enable better forecasts of future liabilities and risk assessment and management. Implications on risk management practices are also discussed.