Measuring Ronbustness for a Complex Model

D

dima777

Hi!
I understand that a system is robust if input values form flat areas on the 3-D chart where the third axis is the Profit or Drawdown. I attached the example of this chart. But what to do with multi-input models. Say, I have a system with 5 inputs - so there will be 10 charts with different input couples to consider against the corresponding PROFIT surface. I wonder if there is a systematic way to identify system robustness without having to use the 3D charts. Rina's SmartView uses the robustness function which is very interesting.
Any opinion is highly appreciated!
Dima
PS: given that each of the above inputs has 20 different values - the 3 million possible combinations can be very hard to represent on a chart - much less if there are billions of possible combinations.
 

Attachments

ASSUME that you have chosen a preliminary design (which is a point in N-dimensional space, one coordinate for each of the N system parameters).

One thing you could do, is simulate other designs that only move each coordinate, at most, +/- 15 pecent away from the preliminary design. (or 20 percent, or whatever value you prefer). In effect you are exploring the nearby neighborhood. And ignoring faraway parameter sets.

For each of your performance measures, you could construct frequency distributions. Then you could summarize the data by saying: The 50th percentile (median) value of the distribution of Compound Annual Growth Rate, in the neighborhood of my preliminary design, is +41% per year. The 95th percentile value of the distribution of Compound Annual Growth Rate, in the neighborhood of my preliminary design, is +36% per year. Compute the 50% point and 95% point for each of your performance measures. Now look at them. Do they make you feel uneasy? That's non-robustness, as measured by your own gut reaction.
 
yep...i am thinking to get a preliminary solution using genetic algorithm and then use risk software to perform Monte-Carlo simulation of the input values around the best chosen by the genetic algorithm to see if they hold the performance of the best solution - do you think that is a way to do this?
thanks a lot for your input!)
 
You must log in or register to reply here.
Back
Top