This  focus on modeling and estimating tail parameters of bank operation loss severity. Using extreme value theory, we centralized on the Generalized Pareto Distribution (GPD) and compare with standard parametric modeling based on Exponential, Gamma and Weibull distributions. In empirical study, we determine the thresholds of GPD through mean excess plot and Hill plot. Kolmogorv-Smirnov and LR goodness-of-fit test are conducted to assess how good the fit is. VaR and expected shortfall(ES) are also calculated. We also take into account copula functions to calculate the correlation of event-pairs. Bootstrap method also be used to estimate the confidence interval of parameters. Empirical results show that the GPD method is a theoretically well supported technique for fitting a parametric distribution to the tail of an unknown underlying distribution. It can capture the tail behavior of bank operation loss very well.