curve fitting, question for the mathematician

Hi,

I have generated some strategies with the following process.

1) Random Generation with Eurusd–fhdb-data

ISS:  1.1.2008-27.09.10

Oss: 27.09.10-1.1.2012

parameter:

File: settings.jpg

2) I Found following good strategy.

File: equity_fhdb_1step.jpg

3) The robustnesstest was good an the equitycurve was good too. So I decided to make an forward test with fhdb and dukadata

Forward: 1.1.2012-29.03.2013

Iss+Oss: 1.1.2010-1.1.2012

Forward: 1.1.2012-29.03.2013

forward with fhdb-data(graphic below)

File: equity_fhdb_2forward.jpg

forward with duka-data(graphic below)

File: equity_duka_2forward.jpg

4) I am happy, because I found a strategy who is good in Iss, Oss and this strategy is good on unseen data (1.1.2012-29.03.2013) too.

I think If I made a test with unseen data the EA can´t be curvefitted ????

5) In the next step I have done the following.

I checked this EA with data from 29.03.2013 till 10.10.2013(dukadata)

Iss+Oss+Forward:1.1.2010-29.03.2013

OSS: 29.03.2013-10.10.2013

The result was the following.

File: equity_duka_3forward_bad.jpg

6) question: who can explain this?

I have done a forwardtest of a long period 29.03.2013 till 10.10.2013, I have done a robustness test. I have checked this EA with data from

dukacopy and fhdb.

All tests show me that this strategy is good.

And in the next step of unseen data this strategy is only loosing. I don´t understand this.

thomas

it is hard to say what could have been done better. I think you did almost everything from statistical point of view – the only thing more that can be done is WF Matrix.

We just have to realize that even multiple repeated tests on unknown data don’t guarantee anything,

I think multiple OOS testing they greatly reduces chance that strategy will be unprofitable, but that possibility is still there.

Mark

Hi bill,

I want to be independend from the market.

I generate every year a new portfolio with the actual market conditions.

The GB tool is very good to do this.

thomas