Comments on: Eating Your Own Tail: HPC in 2009

By: matt.reilly

matt.reilly — Wed, 30 Nov -0001 00:00:00 +0000

One might suggest that the derivative valuation task is computationally intractable, and that it is the crux of the failure in derivative markets.

I\'d propose an alternate interpretation: The true goal of all derivative valuation programs is to predict the valuation attached to the derivative by OTHER valuation programs. That is, the entire industry is running models of each other\'s models. Derivatives are the reductio ad absurdum of the notion that the \"worth\" of a thing is what someone else is willing to pay for it. But in the case of derivatives, it seems that there has always been a moment when somebody asked the dangerous question: \"Are the models really even close to reality?\" In the case of the latest collapse, the answer eventually was \"no.\"

This wasn\'t a matter of computer systems creating their own disaster. The hedge fund industry would have gotten to the same place if they\'d had to run the Black-Scholes model on a soroban. If you run the model predicting model for long enough, it will wander away from the underlying reality. If, at the same time, the consequences of absurd risks are made irrelevant to the behavior of the people operating the machine, while the upside is apparent, then the system will fail. Over and over again.

(Has everybody forgotten LTCM? The Savings and Loan
collapse? This isn\'t ancient history.)

By: tpierce

tpierce — Wed, 30 Nov -0001 00:00:00 +0000

RE \"high-tech irony of selling machines so advanced that they stupidly contribute to their own demise,...\"

The Dune saga, with Omnius and the Butlerian Jihad?

speaking out on the great issues of our times.....

By: ak3la

ak3la — Wed, 30 Nov -0001 00:00:00 +0000

I hate to sound like someone\’s High School English teacher, but I had trouble paying attention to the content of this article \”do to\” all of the grammatical errors. Is there an editor on staff? With all due respect to the letters after Mr. Eadline\’s name, writing is not his strong suit.

By: jenniw

jenniw — Wed, 30 Nov -0001 00:00:00 +0000

The Models ARE actually to predict the true valuations. The exact algorithms _are_ intractable. There are quite a few problems in the real world the computation of which are Intractable, generally when you are searching for the optimal in a search space lager than 3 dimensions. However, these problems are solved by one of a _huge_ family of Machine Learning algorithms that provide an approximation of the optimal result. For example, one family of algorithms \"prunes\" the search space in each direction beyond which point it is improbable for the optimal solution to occur; the resulting limited space in tractably searchable. Another family randomly samples the search space, which process, after some time, gives a clear indication of where the optimal solutions are likely to occur in the search space; and then you search these localized areas exhaustively. An example is the Monte Carlo simulations often used for Financial Models. They give a \"probably approximately correct\" result. For example, VaR is the amount which, with 99% probability, you are likely to lose on the particular financial instrument the next day. This corresponds to an AA rating. if you use the value with 95% probability, which will take less time to estimate, the organization gets a B rating.

These huge families of algorithms are tractable, but take a long time to compute. These are amoung some of the problems for which HPC servers are used in Finance. The problems with the credit crunch is not a Computational or Computability problem. It is a problem in the underlying Maths and Statistical Models; the underlying assumptions of the models are inaccurate; those Models that are accurate predict values under \"normal\" conditions well, and they do not predict correctly under once in a lifetime \"extreme\" conditions.

So the Credit Crunch has nothing to do with HPC Computing.