SOA Monograph- A New Approach to Managing Operational Risk - Ch.8



Reading Source:

Topics Covered in this Reading:

  • Goals of Measuring/Assessing Operational Risk
  • Actuarial Approach
  • Data Requirements
  • Frequency and Severity Distriubtions
  • How to Model Frequency
  • How to Model Severity
  • Combining Internal and External Loss Data
  • Risk Assessment, Scenario Analysis and Stress Testing
  • Calculating Value at Risk
  • Modeling the Operational Risk Component of Other Risks

SOA Monograph- A New Approach to Managing Operational Risk Ch. 8

Hi all,

I have a question about “Exhibit 8.7 – Risk Assessment Under Modern ORM”. Can you confirm me that in the table, the last column “1 in N years Output” is the result after the LEC was fitted to a normal distribution?




Good question. The notes indicate that they are fitting a log-normal distribution to the LEC. I can’t seem to reproduce the same percentiles with the given mu and sigma though. Was anyone else able to get the correct output results given the parameters indicated?


That’s correct. The output column is produced by the model. They don’t explain how to actually fit it, and it’s probably proprietary since there is a “patent pending.”


Hi there, I had a question about how to model frequency.

It says that you should use a:

  • Poisson distribution when there is no variability in the frequency relative to the mean
  • Negative Binomial distribution when there is excessive variability in the frequency relative to the mean
  • Binomial when there is very low variability in the frequency relative to the mean

I think I can make sense of the Poisson and Binomial examples, but I am struggling with the Negative Binomial. Can we talk through some examples of operational risks whose frequency follow the three different distributions?


I’ll give binomial distribution a shot

One type of operational risk is process risk. So let’s say we have a conveyor belt that puts coke caps on coke bottles. Let’s say it it does it 30 bottles per second.

Let’s say that on average, every 1000th bottle has an issue. It may be the 998th bottle, or the 1013th bottle, but it’s always around the 1000th bottle.

I would say that this follows a binomial distribution because the risk occurs not on every 1000th bottle on the dot, but AROUND the 1000th bottle (low variability).

Does this sounds reasonable?

I can’t really think of an example where there’s no frequency relative to the mean - If that’s the case wouldn’t we know exactly when the risk will occur (ex/ every 1000th bottle on the dot), meaning we can just exercise an action to stop the risk from occurring?