|
PureBytes Links
Trading Reference Links
|
>TJ >wrote
>bob, you've touched on something that's bothered me for a while about
system backtesting. mark brown and others insist the one must use
>clean data to test systems. i understand the reasons. but when you apply
the system on dirty real time data. and bang, the system often
>falls apart in the real world. so what's the point then?
This is going to get real complicated: You need to test a system to death
on as many data points as you can. I test on at least 17,500 or more data
points before I would ever even consider a system.
Of that data that I test on I only pick what really amounts to as a random
point. That being one of the following:
a point in time
mean of the data
an average over time
average range over time
an average over a number of reported trades
average range over a number of reported trades
a point in time using a synthetic perpetual data series
I try and ignore specifics like the high of the bar and the low of the bar
ect. I much prefer using the close price at some interval of time. I have
come sort of full circle to make this decision. I believe that just as Bob
states below it is a matter of building a model loosely that will help it
hold up over time. If the data is slightly off then the system should
perform just about as well. I know this seems impossible at first, but all
this has to do more with finding the magic cycle in the data and
capitalizing on that even when that cycle takes a deviation from the norm.
Your data doesn't have to be perfect, it needs to be consistent as you can
get it. That consistency has a direct correlation to how you collect it and
or build it. So I believe you must trade on the same data that you test on
else your system will be off in real time trading. This is exactly why when
you test it looks good and when you trade it falls apart. Mark Brown
>Bob Fulks wrote:
>Make me wonder why you think you need clean data. If the data you get
real-time is "dirty" (and it is), then what good is a trading system
>that only works on clean data.
>
>Wouldn't it be more appropriate to have one that works on dirty data since
that is what is available in real-time?
|