I love the MultiNormalValue function for generating correlated values, but the data I'm simulating is anything but normal. Could you please comment on the validity of the following approach:

1) Transform raw data to closer to normal (e.g., Box-Cox transform in Minitab: let's say the transformation is to take the square-root of the original data).

2) Develop the correlation matrix of the transformed data (e.g., in Minitab)

3) Calculate mean and standard deviation of each of the transformed variables.

4) Simulate using RiskAmp's MultiNormalValue function using the transformed correlation table and transformed means/stdDev for each variable.

5) For each itteration of the simulation, calculate the inverse transform (square the MultiNormalValue results, in this example) for each variable.

6) Use the value derived in (5) as the simulation value for each variable.

Thanks

Hi,

Let me say tentatively that this sounds reasonable, although it will tend to increase error so you likely will need to increase the number of trials to account for this.

Box-Cox is particularly suited to this as it's designed to preserve rank. I need to consult with some of our statistics gurus but theoretically it sounds solid and I've done some very simple testing that shows fairly good results.

Alex
alex@riskamp.com

Did the statistics guru's buy it?
Thanks

Yes, although they say there are better solutions (involving a good bit more math). I'm still waiting on an alternative proposal. The upshot is this is a reasonable approach.

Alex
alex@riskamp.com