Why Correlated Variables Matter in Monte Carlo Credit Loss Simulations
Summary
The document clarifies the distinction between independence across Monte Carlo trials and dependence among variables within a trial. In the example, each of 1,000 simulation draws contains 100 credit-related components. The draws can be independent of one another while the 100 components inside each draw follow a joint distribution with a specified pairwise correlation. Transforming independent normal variables into correlated ones represents assumptions about how the underlying credits may default together.
That dependence matters for portfolio loss analysis: defaults that cluster can produce a different loss distribution and affect tail-risk estimates than independent defaults. The answers mention varying assumed correlations and using dependence models such as copulas in stress testing. They do not provide a calibration procedure or demonstrate empirical accuracy. Correlation assumptions are model inputs, and the brief discussion does not address the limitations of normal dependence models or how relationships may change during market stress.
Key ideas
- Independent Monte Carlo trials can each contain dependent variables.
- Within-trial correlation represents modeled dependence among the underlying credits.
- Changing dependence assumptions can alter simulated portfolio loss outcomes and tail risk.
- Copulas can be used to impose or stress dependence in portfolio simulations.
- The document does not explain calibration or validate a particular correlation model.
Tags
Full text
# Why do we need correlated random variables in a Monte Carlo simulation? # Why do we need correlated random variables in a Monte Carlo simulation? Question: I don't understand why a Monte Carlo simulation needs correlated random variables. Isn't each simulation thread independent? Background: Specifically, I'm referring to the below example on pg 319 in the Malz text (http://ca.wiley.com/WileyCDA/WileyTitle/productCd-0470481803.html). He describes a Monte Carlo simulation with 1,000 simulation threads to calculate credit losses on a CDO with 100 underlying credits. #### #1. In the simulation, we setup a matrix of 1,000 draws from a 100 dimension joint normal distribution. #### #2. We posit 4 separate assumptions for pairwise correlation 0, 0.3, 0.6, 0.9 #### #3. For each correlation assumption, the matrix of 1,000 random normals is transformed into matrix of 1,000 correlated random normals (which of course are still 100 dimensional normals). I don't understand why we need to transform the matrix of uncorrelated random normals into correlated ones? Isn't each simulation thread independent of the previous ? ## Answer by dm63 (score 3, accepted) https://quant.stackexchange.com/a/25962 Surely what is meant is that the 100 components are pairwise correlated but the 1000 draws are independent. ## Answer by user6430 (score 1) https://quant.stackexchange.com/a/25975 Correlated (simulated) features or variables are used when the source data has correlated features. If the original features in the source data are not correlated and are orthogonal, then there is no reason to use correlation when simulating. Many assets are correlated, mostly through volatility clustering and sentiment. Secular bull and bear markets can also cause many assets to correlate. If there was no correlation, then you could merely build a portfolio by randomly selecting the assets, and then weighting the amount owned by historical returns -- which is never a good idea. Bank portfolio stress testing involves use of copulas (simulations) to artificially introduce greater levels of correlation between portfolio assets to determine Variance at Risk, Expected Tail Loss, etc. Basically, when times are bad (e.g., subprime mortgage crisis), many assets begin to become strongly correlated. Thus, if there are large market corrections (loss), when many assets start correlating, the chances of greater loss are larger -- because you are not only losing net worth via one asset, but losing a lot more via loss among many correlated assets. (a particular type of portfolio which can safeguard against high correlation during increased volatility with correlation is called a collar portfolio -- which is a form of hedging). When correlation is also greater, there is less diversity -- which is less sustainable.
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.