|
PureBytes Links
Trading Reference Links
|
I don't think many will argue that optimization has utility weather applied
to odd ball or anything else. Like any tool, it can be harmful in the wrong
hands and useful in the right hands. Any indiciator that requires lentgh
inputs (MAs, STOCH, ROC, pretty much most out there) uses some number along
the curve of optimized numbers. So any selected number will be an
"optimized number" or some number along the curve. If you plan to use a
system like this, which many people do succesfully, then you should optimize
to at least see where along the curve your default value lies and if there
are better "default' values. As you pointed out some numbers are more
stable than others and optimum numbers can change over time. The only two
ways around this problem I know of are to use a value that has broad
stability or uses an adapative input based on some other market indicator.
The easiest to recognize is one with broad stability. If you're looking for
a number with stability than doesn't it make sense to choose the largest
sample possible? Afterall the larger your sample is, the more accurate your
estmation will be. But just going from experience, I've done it both ways
over and over and I always seem to get the best results using optimized
values over longer terms. I've never been able to get weekly or frequent
optimzations to work for me consistently. That's not to say it can't be
done because others have reported success. I have had better results
optimizing across a "large" set of data (enough data to generate say 100-300
or more trades) and then sticking with that value. Once I choose a number
then I walk it forward and back. If the system holds up then I feel
comfortable trading it.
I try to keep things as simple as possible. I also find that the optimzation
curve can give a clue about how well your system will do. If your system
generates an optimization curve that approximates a bell shape then those
optimum numbers will probably hold up well over time. A system that starts
real profitable with small numbers and then deteriorates from that point
will probably not be as profitable. Not only because it will be harder to
trade with smaller inputs and more frequent trades (more mistakes, slippage,
commission, etc) but also those numbers are inherently unstable with no
support on either side.
Brian.
-----Original Message-----
From: Wes Williams [mailto:softexcl@xxxxxxx]
Sent: Tuesday, February 19, 2002 9:58 PM
To: Brian; List, Omega
Subject: RE: Oddball and the Emperor's New Clothes
Brian,
you wrote:
> I get robust results optimizing over the largest set of data
> possible. I'd
> say back to inception of the S+P is a robust optimization.
I agree with the principle of having a large in-sample period but optimizing
against it all is not good for robustness. Let us apply the generality to
the specific of OddBall with Walk-forward testing and principles of
robustness.
If you go back to the inception of the S&P and optimize the system variable
until **1996**, and THEN run the system from 1997 through the present, I
might consider that to be a *generalized* robustness test. Running
parameters from the inception of the S&P until the *present* is promising
but is not a guarantee of robustness. You will tend to pick the best figures
and the system may fall apart going forward. For instance, when I run
OddBall parameter optimization of 5,3,1, it only returned $3,000 for
2001-present as compared with my gleeful figures for 7,3,1. The difference
between 5-3-1 and 7-3-1 is only two numeric places. Specifically, what
market inefficiency are you trying to capture with the difference of two ( 5
and 7 ) on the length of the rate of change? I do not know. Do you? When I
optimize the rate of change length (the "7" in 7,3,1), I do not see a bell
curve. Instead, I see the 7 as near a steep decline. This is a robustness
issue, the ability to make profits over a wide range of variables. If market
conditions change, how are you going to protect yourself with a cliff only
two numeric places away? I really want to know.
Are you going to periodically reoptimize to solve this "problem"? If so,
then you may wish to walk-forward precisely with THAT kind of testing. In
other words, start with the S&P from inception and optimize on three years
of data and run it against the next year's out-of-sample data. Then divide
your ACTUAL profits with your PROJECTED profits and you will get your
Walk-Forward Efficiency (WFE) ratio. If you have a WFE ratio of 50% by the
time you get to the present, I will more inclined to say that the system is
robust and willing to put money in it. Going forward into the future, I can
roughly expect that the system will return 50% of my optimized results.
However, optimizing against ALL past data and blindly walking forward
exposes the system to breaking down and loss of money.
A question for you: What do you think about the fact that the profits vary
so widely from 5,3,1 to 7,3,1? Perhaps a prudent step to take would be to
move the "7" away from the "5" a little more. Maybe a "9,3,1" is more in the
center of the bell curve and safer, albeit with less profits.
Sincerely,
Wes Williams
> -----Original Message-----
> From: Brian [mailto:blink64@xxxxxxxxx]
> Sent: Tuesday, February 19, 2002 6:51 PM
> To: List, Omega
> Subject: RE: Oddball and the Emperor's New Clothes
>
>
> "may I submit that you do not know the precise parameters to
> use in the future unless your settings are **ROBUST** and yield
> profits (and
> cut losses quickly) in all kinds of market conditions."
>
> I get robust results optimizing over the largest set of data
> possible. I'd
> say back to inception of the S+P is a robust optimization.
>
> -----Original Message-----
> From: Wes Williams [mailto:softexcl@xxxxxxx]
> Sent: Tuesday, February 19, 2002 4:25 PM
> To: omega-list@xxxxxxxxxx
> Subject: Oddball and the Emperor's New Clothes
>
>
> Dear List,
>
> I am one who has tried different variations of Oddball. I also obtain
> wonderful results AFTER optimization. I also see the ecstatic
> results posted
> by others on the list and only wonder if they are fooling themselves into
> thinking that it is a robust system. After optimizing for **tremendous**
> profits, I wonder if we are all not like those in _The Emperor's New
> Clothes_ who stood back and said "Look at the beautiful profits on this
> system" but ignore the Risk of Ruin and apparent lack of robustness.
>
> Can the Oddballer's who marvel at Oddball's beautiful clothes answer this
> elementary question? What happened to the system from 5/21/2000 -
> 9/19/2000?
> Running 100 @SP contracts against $ADV with parameters 7,3,1
> produced a loss
> of $4,500,000 and a debt of $1,100,000. You are ruined. Now before you
> reply and say "ahh.. that is the problem... you are using the wrong
> parameters!", may I submit that you do not know the precise parameters to
> use in the future unless your settings are **ROBUST** and yield
> profits (and
> cut losses quickly) in all kinds of market conditions.
>
> Using the "standard" settings may have done well in the past and done well
> since 9/19/2000, but how do you account for 5/21 - 9/19 ?
>
> I would really like to know how you plan to survive this period if we are
> at, say, 1/1/2000. Does it really have any clothes that are robust?
>
> Sincerely,
> Wes Williams
>
>
>
|