Empirical Research, Domestic Commercial Banks Operational Risk Measurement

One of the most important risks of financial organization is operational risk. Comparing with the measurement research of marketing risk and credit risk where people already had mature measurement model, the measurement study of operational risk is still on its basic period.For the purpose of the measurement study of domestic commercial banks, this thesis mainly applies Monte Carlo simulation to 464 operational loss events of domestic commercial banks which collected from some public media. The thesis does a few try of different distribution functions in the simulation of loss severity and makes some comparing analyses of them to find a better fitness. The main procedure is as below:1. At first, the thesis analyzes the statistics performance of the samples, to make sure that the samples have the character of operational risk loss distribution like high peak, thick tail.2. Monte Carlo simulation is applied to study the operational risk in details later. The Negative Binomial distribution is selected to simulate the loss frequency distribution since it has a better fit which proved by Kolmogorov-Smirmoff (KS) testing, while the Weibull distribution is selected to simulate the distribution of log data of loss severity. The thesis gets an operational risk yearly loss distribution, the VaR on 99.9% and an operational risk capital requirement finally. The simulation result is verified. Furthermore, the thesis analyzes the advantage if we fit the log data of loss severity with Weibull distribution.3. The thesis is continued to look for a better fit of the loss severity, the POT (the Peaks-over-Threshold) Model is utilized this time. The Generalized Pareto Distribution (GPD) is selected to fit the tail of loss severity distribution, and then the thesis makes a comparison of the simulation result between GPD and log-Weibull distribution, and finds out that we get a better fit with GPD to the tail of loss severity data, while we get a better fit with log-Weibull distribution to the head of loss severity data which lower than the threshold. According to this conclusion, we choose GPD to fit the tail of loss severity distribution, and select log-Weibull distribution to fit the head of loss severity distribution. After that, the thesis gets an operational risk yearly loss distribution through Monte Carlo simulation and verifies the simulation result again.4. In the end, the thesis makes an integrated comparison of all the simulation results, sums up the feature and suitability of log-Normal distribution,log-Weibull distribution and GPD. Next, the thesis gets a conclusion that the performance of the Monte Carlo simulation depends on the character of samples strongly. The applicability of Monte Carlo simulation is discussed also.