Systematic trading | Page 24 | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

Trade another day through Systematic Trading 203 replies

Nonparametric Association Measures in Systematic Trading 57 replies

Anyone use Systematic from Alpari? 1 reply

A Systematic Approach to Markets 68 replies

Systematic Trading vs. Discretionary Trading 14 replies

Attachments: Systematic trading
Exit Attachments

Systematic trading

  • #461
  • Edited 7:33am Nov 25, 2009 4:56am | Edited 7:33am
My question is how would you apply this in the real word?

I have tried things like this and came to the conclusion it wasn't very useful.

It basically predicts like a linear regression, it doesnot predict any turns. In other words it will basically just continue the line.

I believe you only can get the edge in a statistical way. Predicting to this level (in a meaningful way) is not possible in my view. (as you said, you are predicting MA anyway which in itself is meaningless as it is delayed and the event has already happened)

I could be wrong though.
  • #462
  • Nov 25, 2009 10:39am Nov 25, 2009 10:39am
Quoting spekkiefx
My question is how would you apply this in the real word?

I have tried things like this and came to the conclusion it wasn't very useful.

It basically predicts like a linear regression, it doesnot predict any turns. In other words it will basically just continue the line.

I believe you only can get the edge in a statistical way. Predicting to this level (in a meaningful way) is not possible in my view. (as you said, you are predicting MA anyway which in itself is meaningless as it is delayed and the event has already happened)

I could...
I think you have hit the nail on the head there, I think everyone who studies AI tries a similar thing at some point. On the positive side challenger78 obviously has some good programming skills.
The breaking of a wave cannot explain the whole sea.
  • #463
  • Nov 25, 2009 3:55pm Nov 25, 2009 3:55pm
Here is some interesting articles :http://www.tradingsolutions.com/news/anand.html

http://www.allbusiness.com/personal-.../554820-1.html

Mikkom ,I hope I am not disrupting the content of your topic.
  • #464
  • Nov 26, 2009 12:38am Nov 26, 2009 12:38am
Quoting purley
challenger78, thanks for the great contribution.When you compare your PNN MA model to your original MA(without NN) model how much difference does PNN MA model make?
Sorry not sure I understand the question. The red line is the real value of MA, the blue is the estimated.
  • #465
  • Nov 26, 2009 12:40am Nov 26, 2009 12:40am
Quoting spekkiefx
My question is how would you apply this in the real word?

I have tried things like this and came to the conclusion it wasn't very useful.

It basically predicts like a linear regression, it doesnot predict any turns. In other words it will basically just continue the line.

I believe you only can get the edge in a statistical way. Predicting to this level (in a meaningful way) is not possible in my view. (as you said, you are predicting MA anyway which in itself is meaningless as it is delayed and the event has already happened)

I could...
The truth is I haven't back tested it thoroughly. By looking at it my guess was that it is possible to win a few pips even though you don't predict the actual price cause you still get the bias significantly enough to make profit. The reason I didn't continue it is that at the time I was quite busy and would have taken quite a lot of time to implement proper backtester; then I got sidetracked as usual...

BTW it is quite different than simple linear regression... the only common they have is that they both fit a function to data, but other than that its entirely different story...
  • #466
  • Edited 12:59am Nov 26, 2009 12:44am | Edited 12:59am
Quoting Craig
I think you have hit the nail on the head there, I think everyone who studies AI tries a similar thing at some point. On the positive side challenger78 obviously has some good programming skills.
Craig this is not AI in the usual form of typical NNs... it is a nonlinear fit on data. Furthermore, the structure creates itself from the data, it is not programmer defined as in usual NNs.

If we assume that the market is non random (which I believe it is not as it is man made), then makes sense to try and fit a highly nonlinear function.

I really hope one day I will get back to this and implement it properly cause I do believe it has some merit, especially in higher TFs where you can integrate macroeconomic data into the model.

Thanks for the compliment

PS: if you google you will find many papers, but the mother site (IMO) is this: http://www.gmdh.net/articles/index.html

When they started researching them in eastern Europe they were called Polynomial NNs and the method Group Method of Data Handling, in the west they were called statistical learning networks.
  • #467
  • Nov 26, 2009 2:29am Nov 26, 2009 2:29am
Quoting challenger78
If we assume that the market is non random (which I believe it is not as it is man made), then makes sense to try and fit a highly nonlinear function. .
There are a lot of fairly trivial ways to prove that the market is mostly random, so I'm not sure I agree on that one, however that subject has been beaten to death elsewhere. But cheers for the link, aways interested in learning more!
The breaking of a wave cannot explain the whole sea.
  • #468
  • Nov 26, 2009 3:32pm Nov 26, 2009 3:32pm
Quoting Craig
There are a lot of fairly trivial ways to prove that the market is mostly random, so I'm not sure I agree on that one, however that subject has been beaten to death elsewhere. But cheers for the link, aways interested in learning more!
Quite right, and I'm not interested in starting yet another debate on market randomness. But I will add some thoughts here, for what is worth it:

1. If market was random, non of us would ever stand a chance to make consistently profits (we would be making profits for short amount of time due to luck, but lose in the long run due to transaction costs).

2. It is intuitive to perceive something highly complex as random just because we can't see with our human eyes a pattern. A voice signal/waveform for example appears random, but can be analysed in components with Fourier transform that can then be utilised to do things such as voice recognition. This is a proven fact and people actually do it. The waveform is just very complex with some added noise due to instruments used (measurement noise) and some other forms of noise. The market is much more complex and dynamic than speech because it's the sum of all participants all over the world, which in turn get affected by many other factors, so if speech waveform appears random, imagine how the price movement appears...

Imagine this idealised scenario: you are God, and you can know everything that happens everywhere at any time in the world. I mean everything. And because you are God, you have the ability to actually combine and interpret all this information into a real deterministic model. Then you would be able to predict with perfect accuracy every single market move.

But because non of us is God and non of us can do all this, we just call it random, which in effect means "we can't really fit a model to it"... but just because we can't doesn't mean a model doesn't exist And all we are doing, all of us, when we create a trading system, is trying to find a model that fits a close as possible to the ideal one.

and I leave it to this point.

All the best
  • #469
  • Dec 2, 2009 2:32pm Dec 2, 2009 2:32pm
hello mikkom and all the other experts in this thread

first, id like to thank you mikkom for starting this thread in the first place. since you started posting your progress i was looking into this stuff myself.
im currently working on my own first implementation, and its fun

the gp-core framework is done in its initial form (rather simple, but working, making it more complex is for the future)
im now starting on the backtest framework

theres so much free educational stuff to read and learn, its overwhelming...

hows your own project going mikkom? it seems you have quite a huge success with it... congrats on that

thanks for all the valuable info in here (including links )
Patience is key.
  • #470
  • Dec 2, 2009 3:10pm Dec 2, 2009 3:10pm
Quoting challenger78
The truth is I haven't back tested it thoroughly. By looking at it my guess was that it is possible to win a few pips even though you don't predict the actual price cause you still get the bias significantly enough to make profit.
The thing with predicting MA's is that the higher the ma, the lower the one bar means so it's quite easy to predict them with simple maths as well.

If I understood this correctly then your nn will try to estimate the turning speed of ma (because ma length-1 elements are already known) - not probably the most important thing you could predict so basically you are still predicting the last bar even if you don't think so So the results that you get for trying to predict next close will be as good as the results for predicting MA.
  • #471
  • Edited Dec 3, 2009 1:46am Dec 2, 2009 4:55pm | Edited Dec 3, 2009 1:46am
I have read the book. I agree, the argument is silly.
Vincit qui se vincit.
  • #472
  • Dec 3, 2009 12:48am Dec 3, 2009 12:48am
Quoting mikkom
The thing with predicting MA's is that the higher the ma, the lower the one bar means so it's quite easy to predict them with simple maths as well.

If I understood this correctly then your nn will try to estimate the turning speed of ma (because ma length-1 elements are already known) - not probably the most important thing you could predict so basically you are still predicting the last bar even if you don't think so So the results that you get for trying to predict next close will be as good as the results for predicting MA.
Well, in that example the objective was to predict the change in MA value between the current and next bar. But because MA always lags price, even if we did get the MA prediction 100% correct, we would still have error between our predicted value and the close price of the next bar. The idea was that if the MA was low lag (which it was), the error between the predicted value and the actual close price of next bar would be a small enough, compared to the price move between the 2 bars, to allow some profit.
  • #473
  • Edited 5:52am Dec 3, 2009 1:16am | Edited 5:52am
Quoting challenger78
Well, in that example the objective was to predict the change in MA value between the current and next bar. But because MA always lags price, even if we did get the MA prediction 100% correct, we would still have error between our predicted value and the close price of the next bar. The idea was that if the MA was low lag (which it was), the error between the predicted value and the actual close price of next bar would be a small enough, compared to the price move between the 2 bars, to allow some profit.
What I was trying to say is that for example ma(5), 4 parts of the ma calculation are already known so you are actually just predicting the last one (1 of the 5 elements) - therefore the prediction of MA should be exactly as accurate as trying to predict next close.

(edit: you can easily calculate the next close by based on the MA you predict because it's always (tomorrow+h1+h2+h3+h4..)/masize where hx is history at x and tomorrow is the value you are trying to predict - every other element than tomorrow is known)
  • #474
  • Dec 3, 2009 1:23am Dec 3, 2009 1:23am
Quoting Pippopotamus
May the market be random from one perspective, and not from another? challenger78, have you just thrown Mr. Heiseb=nberg out the window??
As this is my thread, I officially PROBIHIT silly arguments about whether the market is random or not. There are already about million threads that have degregated into arguments about market randomity and none of them brings any new into discussion - search is your friend.

This book should answer the question in more scientific manner
http://press.princeton.edu/books/lo/
  • #475
  • Dec 4, 2009 12:38am Dec 4, 2009 12:38am
Yes this is correct, but the price close is more noisy than the MAs, this is why it's not as accurate predicting the price directly.

Your 2nd paragraph is a smart thing If I ever come back to this project I'll make sure I implement it

Quoting mikkom
What I was trying to say is that for example ma(5), 4 parts of the ma calculation are already known so you are actually just predicting the last one (1 of the 5 elements) - therefore the prediction of MA should be exactly as accurate as trying to predict next close.

(edit: you can easily calculate the next close by based on the MA you predict because it's always (tomorrow+h1+h2+h3+h4..)/masize where hx is history at x and tomorrow is the value you are trying to predict - every other element than tomorrow is known)
  • #476
  • Dec 4, 2009 12:39am Dec 4, 2009 12:39am
Quoting mikkom
As this is my thread, I officially PROBIHIT silly arguments about whether the market is random or not. There are already about million threads that have degregated into arguments about market randomity and none of them brings any new into discussion - search is your friend
This is why I didn't continue the discussion. Respect to the thread. Pippopotamus no offence to you, nothing personal. BTW I like your nick, I think it's very cute You need to get yourself a cute picture though for avatar!
  • #477
  • Dec 4, 2009 4:20am Dec 4, 2009 4:20am
Quoting challenger78
Yes this is correct, but the price close is more noisy than the MAs, this is why it's not as accurate predicting the price directly.

Your 2nd paragraph is a smart thing If I ever come back to this project I'll make sure I implement it

I think the decision to give up on the project was the wise one, that is exactly what I did when I started to understand the dynamics.
  • #478
  • Dec 5, 2009 12:41am Dec 5, 2009 12:41am
Quoting spekkiefx
I think the decision to give up on the project was the wise one, that is exactly what I did when I started to understand the dynamics.
Well, that project is a very interesting one to say the least and I definitely want to look into it again one day. It's just that it's too time consuming, and at this stage there are other ways to tackle trading and make some money, that are more rewarding per unit time spend
  • #479
  • Edited 4:27pm Dec 5, 2009 3:03pm | Edited 4:27pm
I might have found something really important today.

I have been looking for bug from my framework for the past something like 6 hours and found nothing - I still can't believe that this is real - too good results usually aren't.

All I can say that if this is not a bug then some of my market beliefs have changed. Please don't ask more because at this point I can't say much more because I still think this is a bug and I don't want to embarrass myself

Anyway, one pic tells quite lot more than words.
Attached Image (click to enlarge)
Click to Enlarge Name: Screenshot-AlphaFactory-76.png Size: 23 KB
  • #480
  • Dec 6, 2009 12:57am Dec 6, 2009 12:57am
Quoting mikkom
I might have found something really important today...
Mikkom, is any part of this picture out of sample?
0 traders viewing now