[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Suggestion for better backtesting/optimization



PureBytes Links

Trading Reference Links

When doing this, I'm not yet in the model validation phase - where I 
test with out-of-sample data. I'm still in the model selection 
phase, so I'm trying to pick the best model. In order to do this, I 
want to accurately simulate how the parameter selection will change 
over time.

Larry Lewis

--- In amibroker@xxxx, "wavemechanic" <wd78@xxxx> wrote:
> 
> ----- Original Message ----- 
> From: lel4866 
> To: amibroker@xxxx 
> Sent: Sunday, September 29, 2002 12:29 PM
> Subject: [amibroker] Suggestion for better 
backtesting/optimization
> 
> 
> My suggestion is for support for walk forward optimization. This 
is 
> what we all actually do in real life. Since we can't look into 
the 
> future, we test/optimize based on the past, then apply the 
results 
> to the following days trading. We then look at the results and 
make 
> additional changes.
> 
> By "walk forward" do you mean, as in real life, to test as 
future data comes in? Or do you mean to test a system with out-of-
sample data? The second is considered to be the strategy of choice 
by the system experts.
> 
> There are 2 simple variables: 1) The length of time we look 
back, 
> and the length of time (or other conditions) we wait until we re-
> optimize.
> 
> There's a 3rd, more complicated variable - how we choose the 
> optimium parameters from the last optimization run. My 
suggestion is 
> for a "score" formula that takes into account things like: 
maximum % 
> drawdown, best compound rate of return, highest minimum of the 
> running 1 year returns (or whatever period you like).
> 
> I particularly like the last one - I always look at a graph of 
the 1 
> year returns (Equity() - Ref(Equity(), -253)). Ideally, it 
should be 
> as flat as possible (or maybe rising). I've been treating this a 
> little like analysis of variance - it should look like white 
noise 
> with a given mean and standard deviation - any periods that 
don't 
> bear investingation.
> 
> Right now I use AmiBroker for experiments, but when I find an 
idea I 
> like, I must write my own program to do walk-forward testing.
> 
> Larry Lewis
> 
> 
> 
> 
> Post AmiQuote-related messages ONLY to: amiquote@xxxx 
> (Web page: http://groups.yahoo.com/group/amiquote/messages/)
> 
> Check group FAQ at: 
http://groups.yahoo.com/group/amibroker/files/groupfaq.html 
> 
> Your use of Yahoo! Groups is subject to 
http://docs.yahoo.com/info/terms/