1 (edited by dannnnn_the_man 2018-10-13 10:54:25)

Topic: Significance of timeframe adjustment?

Hello,

While i slowly figure out how to best use EA studio still i have a question.

I am trying to build a portfolio of 20 or so EA's using 100,000 bars of data with a 1M EUR/USD.  I have what looks like a very nice curve but I thought i might as well try it on 200,000 bars to see what happens.  As you can see it heads south for 100,000 bars then right where the all the testing and optimization has been done for this portfolio at 100,000 bars it looks good.

Now I don't know what to think of this porfolio and don't know how significant that is.

Post's attachments

balancechart.PNG
balancechart.PNG 40.96 kb, file has never been downloaded. 

You don't have the permssions to download the attachments of this post.

2 (edited by dannnnn_the_man 2018-10-13 10:55:50)

Re: Significance of timeframe adjustment?

And this is the regular 100k bars chart that was initially created.

Post's attachments

balance100k.PNG
balance100k.PNG 38.47 kb, 2 downloads since 2018-10-13 

You don't have the permssions to download the attachments of this post.

3 (edited by sleytus 2018-10-13 12:09:26)

Re: Significance of timeframe adjustment?

Hi Dan -- it's me again...  smile

So -- this is an easy one.  I've posted about this before and provided charts and examples. but no one really gets it because everyone believes more historical data is better, right?  Wrong...

A strategy combines one or more indicators.  An indicator is a simple algebraic formula -- e.g. y = 2x + 3.  The "2" is analogous to an indicator setting.  And how did the "2" get into this formula?  Because it fit the data better than any other constant.

When you optimize a strategy, it simply computes the best constants for the algebraic formulas.  If you use different data then you get different constants.

Look at your result.  The algebraic formula gives a nice result for the portion of data used to compute its settings, and a crappy result for unrelated data.

Since you trade in the present, then it doesn't make sense to use old data to compute these constants.  All it does is skew the computation in favor of old data.

When you optimize a strategy you'll find you get better results (when trading) if you train strategies using recent data -- i.e. just enough data that results in about 100 - 200 trades.  If you use more than that, then you are simply shooting yourself in the foot.

Sorry -- this is not a case of "more is better".  This is not artificial intelligence -- i.e. the strategy's algebraic formula is not a database that "learns" by providing more input data. But good luck trying to convince anyone -- even when the evidence is staring them in the face.

Related to this -- most people believe that after a while their strategies "go bad".  I'm serious -- they actually believe an algebraic formula stops working.  Of course, that doesn't make sense.  What happens is the market data changes and the settings that were previously computed no longer match well with the current data.  When this happens you do not need to generate new strategies -- rather, "refresh" your current strategies by re-optimizing using recent data.

Re: Significance of timeframe adjustment?

Hey!  You're just helping me a ton recently and I appreciate it!=D

I've been using 6 months of 1m bars.  Maybe i should turn that down to 1-2 months of data.

5 (edited by sleytus 2018-10-13 16:28:52)

Re: Significance of timeframe adjustment?

The amount of data is up to you -- but since we trade in the present then it makes sense to use the minimal amount of data that most closely resembles the current market and which includes enough trades to allow the optimization computation to do its thing.

I think one reason why people resist doing that is because then it implies you have to repeatedly go back and "refresh" strategies with more current data from time-to-time.

We can't predict the future.  However, the probability that next week's data and market conditions will be similar to this week's is greater than it being similar to market conditions 2 years ago.  This gives you an edge.  Using irrelevant data patterns (that will likely never repeat) to compute indicator settings eliminates that edge.