Spliting optimization for uptrend and downtrend

Posted By: nanotir

Spliting optimization for uptrend and downtrend - 01/22/16 07:49

Hi

It is mentioned in the manual that the forex assets are symmetric and therefore the strategy has to pass the detrend = CURVE; option. Since they are simmetric is is also mentioned that to split the optimization of parameteres for uptrend and downtrend should not be done as it introduces overfitting.

However, I have not observed that simmetry myself. And I get a problem by optimizing parameters of the strategy due to it. Because if an asset has been mostly trending upward since 2010, it means that the optimization of a trend indicator will provide better PF if the train favor the long trades. The problem is that as consequence, the strategy does not enter right the short trades and the strategy only provides profit in upward trend.
However if I split the parameter optimization for uptrend and downtrend, then I get short wining trades as well.

I was thinking about spliting the parameters and introducing the detrend=CURVE; in train mode. Will that avoid overfitting?
I have observed in the performance report of the Z12 that it only introduces trades in one direction for certain assets and algos. So that could be a posibility as well. Just avoid entering short in an asset if the backtest only provides good PF results in long trades. But if a downtrend comes in the future, either there is an algo to cover that behaviour as well, or the it can be a problem.

Any ideas?
Posted By: MatPed

Re: Spliting optimization for uptrend and downtrend - 01/22/16 16:24

Nice question Nanitek, detrend is my new nightmare. The finacial hacker is my personal Freddy Kruger laugh

I have intrduced detrend = trades; in trading, but honestly I do not know what I am doing.
Posted By: boatman

Re: Spliting optimization for uptrend and downtrend - 01/24/16 23:36

Note that if your strategy shows poor out of sample performance when the price curve is detrended that it is very likely not suitable for trading. However, the converse is not necessarily true. That is, just because out of sample performance is good on detrended data does not guarantee that your strategy is not overfit or is suitable for trading.

I personally think there is merit in using different strategies for markets trending in different directions. But also consider other market regimes: trending/non-trending, high/low volatility. Of course, how you detect those regimes and how much lag you can tolerate is up to you.
© 2024 lite-C Forums