|
PureBytes Links
Trading Reference Links
|
> If for example I optimise a set of parameters over 2000 days of
> trading and get a near a near linear equity curve - then the same
> parameters that work over the first 1000 days should work equally well
> over the next 1000 days as well.
Yes. If your system shows very consistent performance over
varying market conditions, then you have a much better chance of
it continuing to perform well in the future.
However, be wary of curve-fitting. When you say "2000 days" are
you talking about 2000 bars of EOD data? That's not a lot of
history to test on. How many trades are you taking in this
period?
I have systems that have performed extremely well for 5 years of
30min bars -- that's about 17000 bars of data, much of it out-of-
sample -- and still they have rolled over and died in the market
conditions we've seen in the last 9-12 months. Even very robust
systems can have troubles if the underlying market changes enough
to violate the systems' assumptions. You always have to keep an
eye on changing conditions.
> Therefore is it true to say that most important thing to look for in
> optimizing in sample data would be a linear equity curve rather than
> net profit or profit factors and then apply these optimized
> parameters to the out of sample data.
A smooth equity curve is the primary thing I look for in my
systems. Net profit is interesting, win% is interesting, profit
factor is interesting, but none of them are as important in my
mind as consistent performance. I wish Tradestation supported
optimization on Sharpe ratio. Failing that, when optimizing
systems I use the ROA% as a first-order selection criterion, and
then I eyeball the optimization results to make sure other
measures (like average trade size, etc.) also peak smoothly in
the same area. Then I go check the most promising areas for best
Sharpe.
I also use the HeatMap spreadsheet I recently posted to the list,
found at http://www.frii.com/~fritz/trading. It helps me see
where the best-performing areas are in a 2-D parameter space, and
lets me easily compare the Sharpe and other measures in that
space.
Gary
|