This paper gathers the loss data due to medical insurance fraud from 2006 to 2018,in order to measure the fraud risk of basic medical insurance for urban and rural residents in China. The fraud risk measurement problem has two features: "low loss with high frequency" and "high loss with low frequency", which can separately use normal distribution and Generalized Pareto Distribution (GPD). Under the framework of loss distribution method, the TVaR value of fraud risk loss can be computed from the two-stage (PSD-LDA) model, which is based on the Bayesian Markov Chain Monte Carlo (MCMC). Furthermore, the fraud risk reserve required to be accrued for urban and rural residents' medical insurance can be determined. Research findings indicate the fraud of basic medical insurance has the characteristic of Tail-risks for urban and rural residents in China, the maximum fraud loss within one year being 95% likely to be 72.0 million yuan, as well as 99.9% likely to be 168.5 million yuan, and the government requiring 9.47‰ of the risk reserve to withstand the risk of fraud. The result provides decision basis for scientific calculation of financing standards, fraud risk pricing and early warning of fraud risk in China.