System evaluation

Hi,
I trade using eod data. I am interested in how you evaluate your systems.
Any alternatives to walk forward optimization?

 

I dont know much about walk forward optimization, can you tell me a bit about it and which backtesting platform you use?

Thanks,

Eric

 

- split the whole period into equal pieces (e.g. years)
- optimize for the first 2
- backtest for the 3rd
- optimize for the first 3
- backtest for the 4th
...
the measure is gained by adding the profit values at the end

platform: AmiBroker

 

Ok, I use Amibroker too. I heard I should use IO.zip. Im pretty happy with my current system, Im not trying to optimize, just see how it performs in "future" type environments, and Im told that Walk Forward is the way to go to get an idea as to whether the system performs well in the future.

Im using 25yr data right now

 

what does this IO.zip do?
where can I get it?

thanks in advance

 

yahoo groups amibroker and amibroker-ts have it just become a member.

 

this is very nice
a lot of functionality...
thanks again

till now i have been using my own optimizer/backtester written in javascript

I would not trust a system that has not been optimized for the particular instrument
it sounds quite strange that you do

and where do you get the data?
at the moment, I have back-adjusted futures data for several kinds of futures but not for such a long period, only 22 yrs max

data seems to be a big investment, unless you find a company with a better offer, (which I havent yet)

 

I use only stock data, free EOD from yahoo or msn downloaded via Amiquote.

I read that optimizing a system is a fundamentally dangerous thing, so I dont do it, and I have never found any solutions in using my optimizing function. I manually adjust parameters myself and 99% of the time, I got a better result than one of my selections in the optimize function.

In the future, I may get into trading futures and commodities, because Its easier to afford stock trading then futures trading.

 

as I understand, walk forward is based on the following idea:

- in the past, optimizing for the history that was already known, we checked whether the system was successful in the next year, or not

- we did this year-by-year

- if it was mostly successful, why not continue in the future?

- if over-optimization was not a problem, we assume that it will not be a problem for this system in the future, either

- if it was not successful, we need to find another system

 

ps:
hope you can understand it
it is not complicated at all
but as i read it back, it looks a bit fuzzy