Hi Howard, Markus,
Please pardon the intrusion, but both questions relate to my question.
It may be naive, but I would think that a trend following system would
perform best when there is a reasonably strong trend, and that a mean
reversion system would perform best when the trend is weak, if there is one
at all. So, then, suppose one has written a set of 8 functions that compute
the buy, sell, short and cover signals: half for the trend following and
half for the mean reversion parts of a composite system, and a 9th that uses
a reasonably responsive indicator of trend strength to generate a boolean
controlling whether or not one wants trend following or mean reversion at a
given point in time for a given security (so if oil is showing a strong
trend but telcos are relatively flat, one can use trend following in the one
case and mean reversion in the other: if you don't like that pair, pick a
pair that you'd expect to be uncorrelated and the point remains the same).
I may be a heretic, but I normally assume there are no invariants, whether
through time or through space: in this case the space is defined by the
universe of securities.
I can see it getting complicated as I would expect any open positions at
those times when one switches between trend following and mean reversion
would have to be assessed as to whether they ought to remain open (would it
be most sensible to close them regardless or just leave them unless the
signal from the current system is either buy or neutral). Obviously, to be
viable, one must take steps to prevent the trend following functions from
becoming mean reversion functions. In this case, and assuming the duration
of a trend is typically shorter than the in sample period, one needs to be
able to tell the optimization function/system that it is to be constrained
so that a given relation must always exist between two or more of the
parameters, and that there may be more than one such constraint.
I can see how I'd code this in C++ (that is almost trivial), but how would
one code this in AFL?
Could one use something like the following and get the expected result:
MA1Length = Optimize("MA1Length",50,50,200,1);
MA2Length = Optimize("MA2Length",201,MA1Length,300,1);
MA3Length = Optimize("MA3Length",100,100,300,1);
MA4Length = Optimize("MA4Length",50,50,MA3Length,1);
Cheers
Ted
On Wed, Jan 6, 2010 at 8:19 AM, Howard B <[email protected]> wrote:
>
>
> Hi Markus --
>
> When the relationship between the two moving average lengths changes, the
> system changes from being a trend following system to being a mean reversion
> system. It remains a "long" system when the signals generated are Buy =
> cross(MA1,MA2); It becomes a "short" system when the signals generated are
> Short = cross(MA1,MA2);
>
> Why exclude them? You may very well find that the system works best as a
> mean reversion system.
>
> But if you insist on limiting the variables, try this:
>
> //////////////////////////////////
>
> MA1Length = Optimize("MA1Length",50,50,200,1);
> MA2Length = Optimize("MA2Length",100,100,300,1);
>
> MA1 = MA(C,MA1Length);
> MA2 = MA(C,MA2Length);
>
> // Generate a Buy signal when MA1 crosses up through MA2,
> // but only when the length of MA1 is less than the length of MA2
> Buy = (MA1Length<MA2Length) AND Cross(MA1,MA2);
> Sell = Cross(MA2,MA1);
>
> ///////////////////////////////////
>
>
>
> Thanks,
> Howard
>
>
>
> On Wed, Jan 6, 2010 at 5:44 AM, Markus Witzler <[email protected]> wrote:
>
>>
>>
>> Hello,
>>
>> let´s say I intend to optimize a 2 MA crossover system with MA1 and MA2.
>>
>> Possible range for MA1: 50-200
>> Possible range for MA2: 100-300
>>
>> Now, there are some instances in which periodicity of MA1 is higher than
>> the one of MA2, thereby creating a "short" system.
>>
>> How does one exclude these "redundant" combinations from optimization?
>>
>> Thanks
>>
>> Markus
>>
>
>
>
--
R.E.(Ted) Byers, Ph.D.,Ed.D.
[email protected]
CTO
Merchant Services Corp.
350 Harry Walker Parkway North, Suite 8
Newmarket, Ontario
L3Y 8L3
