« A Chat with PRM Candidate of the Year 2007, Mr. Vinayak Kamat, Pune, India |
Main
| Damages caused by Risk Management Practices »
November 13, 2007
My Frustrations with Stress Testing
This blog is concerned with my general dissatisfaction with Stress Testing. While enormous theoretical and empirical work of diverse qualities have been done in many spheres of risk modelling and quantification, Stress testing happily lives in a world which is pathetically slow in which one variable moves at a time. This is surely not our world, a world in which subprime losses get transmitted to all corners of the globe even before the most bragged about sophisticated risk management systems wake up from slumber. Whom are we fooling? In a world of cointegrated movements in macro variables and risk factors, the standalone stress testing models are bound to be deficient. The defensive response citing absence of good alternatives may seem reasonable to those who live in the present but such line of argument badly fails to satisfy those who dream of a better tomorrow.
There is no doubt we need good techniques for multivariate stress testing in a situation where credit risks and market risks are intertwined. We essentially need a good VaR estimate (my experience tells me this is not as easy as it may seem), coupled with a good model where risk factors can be efficiently shocked to predict future scenarios. Probing into financial econometrics, I have a feeling that the technique of Vector Autoregression holds promise. The technique of VAR ( not VaR) has several useful characteristics- the vector of variables move together ( thats real),and you have a fairly advanced technique of providing shocks( of your chosen intensity) with scientific choice of lag lengths that leads you to attractive impulse response functions. Used in inflation predictions, it has outperformed structural models routinely. It has been useful in judging the interrelations between segments of financial markets. Time to put this to good use in risk management and stress testing? I am trying out models using the following course of action a)Calculate VaR for various instruments in your portfolio b) conduct risk aggregation, if you wish to c) put possible variables that can shock the VaR values (do not put too many variables as things may go out of hand) d) instead of conducting standalone stress test, map the impulse response paths- those curves can throw much light on the shape of the things to come. Thats the broad idea.
( to my pleasant surprise, I find a presentation of Stress Testing from Monica Mars uploaded from Greece ! )
Posted by sunandoroy at November 13, 2007 11:25 PM
Dear Mr. Roy,
Thank you for maintaining this blog.
I agree fully with your stressing of the stress test. But the models I know still share a basic inadequacy with eg. VaR. They base on curves, correlations, volatilities, but there is no place for cyclic dependencies.
Now that we know about the trust-pitfall causing the gap between interbank and prima rates causing a dry-up of credits causing mistrust etc, we understand. But do you know of a statistical tool that is able to predict this.
Now I'm not the one to put the thinking of John Sterman into the Risk Management realm. (I don't have the time nor the background.) I think that there is awarenes among traders about some of these cyclic dependencies, bit I did not find a RM-model that included them.
Maybe you know of one?
Kind Regards,
Johan Steunenberg
Posted by: Johan Steunenberg at November 26, 2007 09:14 AM
Is it possible that the tool you are looking for here is a widget that will convert spreadsheet models into C/C++ algorithms, which will make multi-variate analysis much easier, and a whole lot faster, to perform? I believe even Mr. Steunenberg's concern about cyclic references could be solved therein.
Posted by: Nitish Shukla at December 12, 2007 05:27 AM
Post a comment
