|
PureBytes Links
Trading Reference Links
|
It seems you are conscious of one of the tests to see if your parameters are
curve fitted: parameter sensitivity. That is when 111 works, but for whatever
reason, 110 or 112 give significantly different results. It sounds like you
have taken care of that problem.
Another thing you may want to consider is running your test over just a portion
of your entire data. Optimize your parameters over the middle third (or
half--different people use different periods) of the data. For example if your
data is from 8/1986 to the present, that is 12 years of data or 3 4-year
periods. Optimize over data from 8/90 to 8/94, then test the optimized results
over the remaining "tails" of data (from 8/86 to 8/90 & 8/94 to 8/98). If you
get similar results in the tails as you did during the optimized period, you
have a robust set of parameters that have not been curve fitted.
When creating a system, be honest with yourself. Be pessimistic in your
assumptions regarding slippage and commissions. Good luck!
Regards,
Bill Bancroft
MORIMOTO, Kozo (FinAssets) wrote:
> Like everyone else, I'm striving for a robust trading system which can cope
> with as much adverse conditions as possible. So my question is when
> searching out specific values for your trading system parameters, when does
> is become curve fitted? For example, I need 2 values for my system. I run
> it thru the system tester in Metastock and it spits out the 100 most
> profitable sets of parameters. I normally don't pick the no. 1 parameters
> but I usually look thru the list to see which values crop up a lot. In this
> particular case, the optimal was [14,111]. When I look thru the list, the
> numbers 13,14,15 comes up time and time again for the first value and
> anything in the range of 100~120 for the second value. So to pick out a
> round set of numbers, I chose [15,110]. Does this constitute as a curve
> fit?
>
> This question must get pretty complicated when we start to talk about neural
> nets wouldn't it? When would the nets be curve fitted and when would it be
> generalised?
>
> ------------------------------------------------------------
>
> The contents of this message are the views of the Author and
> do not necessarily reflect the views of SUNCORP-METWAY LTD.
>
|