I think one has to be careful not to go towards optimisation or
curve fitting. I think the basic crossover system with no
additional filters is one simple option and then any other options
with partial TPs need to be forward tested rather than backtested.
Thats my opinion any way
vimal- curve fitting is not an issue here because
we are not changing the way the GDX system takes trades, we are
only taking some profits after a certain % gain such as after 10%
or 20%, therefore the worst that could happen is that say the BPGDM
system had a monster trade, the modified version would not catch
the whole trade % but otherwise the dynamics of the system are
unchanged. Also on purpose I chose to lock in gains after a
10% and 20% gain, very wide. Now if I would have taken 10%,
11%, 12%, 13%, 14%, 15% and so on, then yes I would have been
tweaking it too much, but by choosing wide numbers, I leave it very
open.
I think too many people have heard the word 'curve fitting' and
overuse the word and just assume that it applies to anything you do
with a system, not true
curve fitting is when you look at a strategy and its results,
see that its not quite where you want it to be, add in some new
variables that you know improve it and hey presto the results
improve. Thats curve fitting
Vimal -- yes, agreed, that's what curve-fitting is -- but in
this instance what new variable is matt adding to the system?
He is not adding any new variable. He's just taking profits
at a predetermined level. If anything, he is simplifying the
system not making it more complex.
....the only other point I will make having coded and tested
many many strategies (and also from research done by collegues) is
that overall, dynamic exits work far better than fixed TPs or TP
%s.
For example, the simple MA Crossover systems which in the main
use a crossover to enter short or enter long is far better
than using the crossover to enter short and then setting a x%
TP or $TP
So the question Matt is whether you have tested using a dynamic
indicator driven exit? This will be far more robust in dynamic and
differing market conditions than one that uses a fixed TP whether
it be % or $
no problem with the TP level being set. Good idea given that
profit moves to 10% in many cases but as seen in recent trades,
this can be given back
If we now have a series of 6 trades which only make it to a
9% profit before giving this back then I would not expect the
dynamics of the strategy to change such as to move this TP to 9%
with a rerun of historical results to show what 9% would have
delivered
Otherwise, yes good system and no problem with a TP provided its
not regularly adjusted as per the above example
Your 9% example would be true if the series of 9% profit trades
were purely random. If, however, the market had changed and a
9% TP consistently (say over several months) provided better
returns, then it would be a wise choice to change the system to a
9% TP. My assumption is that some market changes aren't day
to day random, but persist over several months. One way to
test this hypothesis with with walk forward optimization.
In my opinion, the problem isn't with optimization per se but
with using exotic criteria to form fit a system. Clearly, any
system can be made to appear incredibly productive if one were to
choose exotic criteria (for example, buy the market as of March of
2009) that are clearly not reproducible.
However, criteria such as price targets can work very well.
Again, in my opinion, the only problem isn't over
optimization but underoptimization. Markets change, so a set
of conditions that works well in one market may begin to
underperform in other markets. The best way that I know to take
care of that is to optimize frequently, and for relatively recent
periods, and to use profit curves to make sure that a system is
continuing to be profitable. A downsloping profit curve that
can't be corrected by reoptimization is a sure sign that the system
is no longer working.
Newsletter
Subscribe to our email list for regular free market updates
as well as a chance to get coupons!
I think one has to
Improved GDX swing system and statistics
Posted by vimal on 28th of Jul 2010 at 05:00 pm
I think one has to be careful not to go towards optimisation or curve fitting. I think the basic crossover system with no additional filters is one simple option and then any other options with partial TPs need to be forward tested rather than backtested. Thats my opinion any way
vimal- curve fitting is not
Posted by matt on 28th of Jul 2010 at 09:03 pm
vimal- curve fitting is not an issue here because we are not changing the way the GDX system takes trades, we are only taking some profits after a certain % gain such as after 10% or 20%, therefore the worst that could happen is that say the BPGDM system had a monster trade, the modified version would not catch the whole trade % but otherwise the dynamics of the system are unchanged. Also on purpose I chose to lock in gains after a 10% and 20% gain, very wide. Now if I would have taken 10%, 11%, 12%, 13%, 14%, 15% and so on, then yes I would have been tweaking it too much, but by choosing wide numbers, I leave it very open.
I think too many people have heard the word 'curve fitting' and overuse the word and just assume that it applies to anything you do with a system, not true
curve fitting is when you
Posted by vimal on 29th of Jul 2010 at 03:05 am
curve fitting is when you look at a strategy and its results, see that its not quite where you want it to be, add in some new variables that you know improve it and hey presto the results improve. Thats curve fitting
curve fitting
Posted by Michael on 29th of Jul 2010 at 05:52 am
Vimal -- yes, agreed, that's what curve-fitting is -- but in this instance what new variable is matt adding to the system? He is not adding any new variable. He's just taking profits at a predetermined level. If anything, he is simplifying the system not making it more complex.
....the only other point I
Posted by vimal on 29th of Jul 2010 at 08:06 am
....the only other point I will make having coded and tested many many strategies (and also from research done by collegues) is that overall, dynamic exits work far better than fixed TPs or TP %s.
For example, the simple MA Crossover systems which in the main use a crossover to enter short or enter long is far better than using the crossover to enter short and then setting a x% TP or $TP
So the question Matt is whether you have tested using a dynamic indicator driven exit? This will be far more robust in dynamic and differing market conditions than one that uses a fixed TP whether it be % or $
no problem with the TP
Posted by vimal on 29th of Jul 2010 at 07:58 am
no problem with the TP level being set. Good idea given that profit moves to 10% in many cases but as seen in recent trades, this can be given back
If we now have a series of 6 trades which only make it to a 9% profit before giving this back then I would not expect the dynamics of the strategy to change such as to move this TP to 9% with a rerun of historical results to show what 9% would have delivered
Otherwise, yes good system and no problem with a TP provided its not regularly adjusted as per the above example
Vimal, Your 9% example would be
Posted by algyros on 29th of Jul 2010 at 08:46 am
Vimal,
Your 9% example would be true if the series of 9% profit trades were purely random. If, however, the market had changed and a 9% TP consistently (say over several months) provided better returns, then it would be a wise choice to change the system to a 9% TP. My assumption is that some market changes aren't day to day random, but persist over several months. One way to test this hypothesis with with walk forward optimization.
vimal -- I see what
Posted by Michael on 29th of Jul 2010 at 08:10 am
vimal -- I see what you're saying.
In my opinion, the problem
Posted by algyros on 28th of Jul 2010 at 06:00 pm
In my opinion, the problem isn't with optimization per se but with using exotic criteria to form fit a system. Clearly, any system can be made to appear incredibly productive if one were to choose exotic criteria (for example, buy the market as of March of 2009) that are clearly not reproducible.
However, criteria such as price targets can work very well. Again, in my opinion, the only problem isn't over optimization but underoptimization. Markets change, so a set of conditions that works well in one market may begin to underperform in other markets. The best way that I know to take care of that is to optimize frequently, and for relatively recent periods, and to use profit curves to make sure that a system is continuing to be profitable. A downsloping profit curve that can't be corrected by reoptimization is a sure sign that the system is no longer working.