Re: Optimization

[Bob Fulks <bfulks@xxxxxxxxxxxx>]

At 2:32 PM -0400 8/16/02, CRLeBeau@xxxxxxx wrote:

>This is only one study but I have never seen any study that showed that
>reoptimization actually helped.  However if you give any value to anecdotal
>evidence there are many traders that seem to be fond of this questionable
>technique.

One needs to take most such studies "with a grain of salt" unless you
understand the methodology.

Assume you have developed a lousy system - something like a moving
average crossover system. You can always find some settings of the
parameters that show a profit on some set of data. This is curve
fitting. Try it on any other set of data and the results will be
poorer, illustrating the effects of curve fitting.

Now you try re-optimizing the system on the new data and get another
set of parameters curve-fitted to that data. It also works poorly on
out-of-sample data.

So have you proven that re-optimization doesn't work? No. You have
shown that re-optimizing fails to make a lousy system to work better.

Now consider a second example where you have a great system that is
making a good steady profits trading the Nasdaq in 1998. It has fixed
parameter values specified in points.

The Nasdaq increases from 2000 to 5000 over the next two years and
your system slowly begins to be less profitable. You re-optimize and
find new fixed levels appropriate to the revised values of the
market. Your system is now back to working as it was before.

So have you proven that re-optimization works. Certainly. It would be
expecting a lot to have a system work well with fixed parameter over
such a big change in the market.

You could then modify the system to make the parameters adapt to the
market better by measuring some characteristics, such as the value of
an index or average-true-range, etc., and scale the parameters
automatically as the market changes, making the system more adaptive
to changes in the market.

This can remove the "first-order" effects of major changes in the
market. But it is sometimes hard to make such corrections track each
other without a lot of analytical work, so many people simply go with
the fixed parameters and re-optimize from time to time.

Bob Fulks