A Comparative Study On The Stochastic Loss Reserving Models

Outstanding claims reserve is the main source of debts for P/C insurers, which almost accounts for half of the liability items. Besides, it directly determines the reasonableness of the company's solvency ratio, which is the main focus of CIRC.However, the traditional reserving techniques can only provide a point estimate and can't measure the extent of volatility. In light of this, the stochastic loss reserving model has been a hot topic in non-life actuarial field and numerous methods have been raised in recent years.Unfortunately, the literature on the comparative study is very few. This thesis starts from the theoretical analysis of stochastic models, and then tries to show the difference between them by quantitative analysis through three aspects.For each aspect, the author builds a unique framework and individual American company's data are employed to run the test. First, the risk measure perspective. To reflect the difference between stochastic models at different scales, the 60%, 75% and90% quantiles are selected. Test result shows that non-parametric Bootstrap method always corresponds to the minimal, and GLM model is the highest at 90%quantile.Second, the accuracy of prediction perspective. The author designs evaluation criterias for two main user groups: managers of insurance companies and regulatory agencies. After quantitative tests of four business lines, GLM and lognormal model's performance are generally better than Mack and nonparametric Bootstrap method. But for long-tail business, the prediction accuracy of the four methods all significantly decreases. Third, reserve risk perspective. This thesis sets RBC reserve risk measurement as a benchmark. The author defines risk factors and absorbs idea of balancing the industry standard and company's feature. At last, it shows risk factors of different stochastic models are quite close to the output of RBC,which means the stochastic model can be used to measure the reserve risk under any line of business. However, due to limited data, the stability of the resulting risk factors remains to be investigated.