Iman–Conover Rank Correlation for Aggregating Simulated Risks

Summary

The document describes a simulated operational risk model in which individual losses are assigned distributions, calibrated from selected percentiles, and sampled using event frequencies. It then asks whether the Iman–Conover method is the standard way to introduce dependence among simulated risks before summing their losses. The method referenced in the discussion rearranges simulated values to induce a target rank correlation while retaining the marginal samples.

The text provides the modeling setup and identifies the central dependence question, but it does not compare Iman–Conover with alternatives or establish that it is a gold standard. Its stated calibration relies on optimistic and pessimistic loss estimates because historical data are unavailable, and the choice of frequency distribution depends on whether an event can occur once or repeatedly in a year. The discussion is therefore a useful framing of dependence modeling in aggregate loss simulation, rather than a complete evaluation of aggregation techniques or their accuracy.

Key ideas

• The model samples severity and event frequency separately for each operational risk.
• Iman–Conover is used to induce rank correlation while preserving sampled marginal values.
• Aggregate simulated loss is computed by summing losses across risks after dependence is imposed.
• The document raises, but does not answer, whether other methods are equally suitable or more standard.

Tags

# Is Iman Conover the standard approach for risk aggregation?


# Is Iman Conover the standard approach for risk aggregation?












I am genuinely interested in understanding a little more about the risk aggregation approaches that are out there.

I have been recently working on building an operational risk model under Basel 3.1, where the following approach applies. Risks are being modelled via LogN distributions (with the option to use Weibull or Pareto). Given there is no data history, an optimistic and a pessimistic loss is being used (e.g. the 50th and the 90th percentile of the distribution) in order to back solve for the mean and the std. Then a frequency distribution decides the number of draws (a Bernoulli or a Poisson, depending on whether the event can fire only once per year or more than once), then the sampling from the LogN.

Once all risks have been calibrated and sampling is complete, the losses per risk are shuffled and ranked via a method called Iman Conover (ref. Iman RL, Conover WJ: "A distribution-free approach to inducing rank correlation among input variables", CommStats 1982). Then next stage is the aggregation e.g. the sum of each loss per correlated risk.

My question is, are there any other equally sophisticated aggregation techniques or is the IC method the gold standard when it comes to aggregating simulated losses? What other methods are out there?

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.