I wanted to share this question from Michael Ebal that was emailed to me directly.
Hello Shawn. My name is Mike Ebel. I read your article in December's Journal of Financial Planning titled, "Beyond Monte Carlo Analysis: An Algorithmic Replacement for a Misunderstood Practice". Very interesting and enlightening.
But I have a question. I notice that you made mention of Gobind Daryanani and his work on Sensitivity Simulations. I read that work. Also, I use MoneyGuidePro planning software that makes use of such sensitivity simulations through a tool they coin, "Beyond Monte Carlo". I am not strong mathmatically or statistically speaking, but I have enough of an understanding to muddle my way through.
My question to you is: Isn't Daryanani's sensitivity analysis different from the typical monte carlo engines to which you refer in the article? My understanding is that he feels this to be somewhat more accurate and certainly far less time consuming to come up with a result. Or are you saying that we can group sensitivity analysis in with the problems/concerns that you outlined with the mainstream monte carlo engines?
Thank you so much for your time and consideration.
Hi Michael
I am on the road at the moment and would like to review the article from Daryanani before responding, but based on my recollection you are quite right that he outlined a methodology he called Sensitivity Simulations that gave similar results as a Monte Carlo (or better when compared to only a small number of simulations 1000 or 2000) with only a subset of the calculations (60+/-). It was not an algorithmic solution but was definitely superior to a brute force Monte Carlo and I believed addressed issues like correlations between asset classes.
So, my understanding is yes it is different, better in that in allows for correlations (not handled in MCS), faster and more sensitive when compared to small MCS runs.
In respect to the last part of your question, the concerns I expressed were that any statistical methodology that applies a single Capital Market Assumption (mean and standard deviation) and "rolls up the results", is not in fact providing any true test of outlying returns - and I do not believe the methodology would matter - MCS, Sensitivity Simulations or my own Reliability Forecast. MCS can generate these results (albeit inefficiently) is if simulations were subgrouped and reported upon. If someone has implemented the Sensitivity Simulations, I cannot speak to what results they are either able to or are presenting.
I hope this helps.
Warm regards
Shawn