|
PureBytes Links
Trading Reference Links
|
Dear friends,
I am very interested in the idea of using Weibull distributions for modeling the
results of a sistem. In fact I know a little about of this type of distribution
because I used it in the past for modelling the fracture behaviour of ceramic
materials and it is quite usefull (although sometimes it has also some
drawbacks).
I would like just to point out a couple of things, and I hope sometime I will
be able to give some other ideas:
1) To have a reliable Weibull fitting it is neccesary to have al least 30
data,... that means if you broke down the results in wins and losses, you will
require a longer period of time to obtain a good fit
2) "The Median Rank numbers" in the spreadsheet (that uses the strange numbers
0.3 and 0.4 to make the calcutations) is just the probability of "survival" at
this level of the x function. You can use these numbers (but be carefull because
they assume equal espacing in the B column) or any other technique (maybe a
possibility will be the s-shaped curve of the cumulative earnings distribution
histogram)
3) The more important paremeter of Weibull distribution usually is the beta. I
think that for our porpuses, the main point is to be sure that the beta value
does not change to much with time (or to control the change with time). The
other paremeter is also important but there is a fixed relationship between it
and the mean value of the distribution (and so it is aesier to calculate).
4) Did you check if adding a different constant number to the earnings give up
different fitting values? This is a typical problem that we have in the past.
I hope these comments will help and sorry for my English but sometimes it is not
easy to explain just in a few words rather complex ideas. If any of you does not
understand me, please let me know and I will try to explain more in detail.
Regards
Angel
rudolf stricker wrote:
> Glen,
>
> On Sun, 15 Aug 1999 22:34:21 +0200, you wrote:
>
> >Thanks for the Weibull Analysis info. It looks really interesting. I have
> >attached an extract from the .pdf article. Could you please explain the
> >rationale for the author's calculation of "Median Ranks"? Specifically, why
> >does he subtract 0.3 from the numerator and add 0.4 to the denominator?
>
> Imo, this "median rank" calculation is *a part* of an implicit
> integration formula (with different possible formulas for different
> integration methods, like right-angled, trapezoidal, etc.) This way a
> kind of "weighting" is done for the given discrete values to represent
> them appropriately in the context of all other points. Because the
> integration formula is "broken up" in different steps, it is not easy
> to describe the "meaning" of every single constant.
>
> I have used the numerical recipe as described for several win & loss
> distributions, and error calculations showed sufficient accuracy, if
> we have an appropriate number of trades. So I'm satisfied with the
> procedure without analyzing explicitly every single step ...B-)...
>
> >Also, how do you adapt the calculations to deal with negative results
> >(losses)?
>
> You may transform (and re-transform afterwards) your wins & losses by
> adding an appropriate number to all trading results.
> A better way is to set-up separate Weibull distributions for wins and
> losses (which gives a *very nice* representation of the unsymmetrical
> shapes). For this, you have only to multiply the values of the loss
> distribution by -1 ...B-)).... This way I got the best results and
> the opportunity to *concentrate on the losses* during system
> optimization, and Weibull's "beta" then represents the "characteristic
> win" and " characteristic loss", respectively.
>
> mfg rudolf stricker
> | Disclaimer: The views of this user are strictly his own.
|