Potential new machine learning style software. | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

Machine Learning with algoTraderJo 941 replies

Algorithmic Quant Trading (Machine Learning + Stat-Arb) 25 replies

Forex with Machine Learning Software Project 1 reply

Machine Learning + Retail Forex = Profitable? (Quant) 1 reply

My most recent advancements into machine learning 16 replies

Attachments: Potential new machine learning style software.
Exit Attachments

Potential new machine learning style software.

  • #1
  • First Post: Dec 4, 2009 2:05pm Dec 4, 2009 2:05pm
I'm actually traveling right now, but I saw this in one of my RSS feeds.

http://ccsl.mae.cornell.edu/eureqa

In short:

Quote:
----------------------------------------------------------------------------
Eureqa is a software tool for detecting equations and hidden mathematical relationships in your data. Its primary goal is to identify the simplest mathematical formulas which could describe the underlying mechanisms that produced the data. Eureqa is free to download and use. Below you will find the program download, video tutorial, user forum, and other and reference materials.
----------------------------------------------------------------------------






Reference article: http://www.wired.com/wiredscience/20...bot-scientist/

Has anyone had any experience with this sort of program?


Once I get back from my visit I'll be playing with it myself.
  • #2
  • Edited 3:31pm Dec 4, 2009 3:03pm | Edited 3:31pm
Quoting Ronald Raygun
I'm actually traveling right now, but I saw this in one of my RSS feeds.

http://ccsl.mae.cornell.edu/eureqa

In short:

Quote:
----------------------------------------------------------------------------...
OMG I just read "Once I get back from my visit I'll be playing with myself."

Seems to be similiar to RapidMiner which is an also free DataMining Programm. You should check it out if you're not familiar with it.

Thanks for the Article.

Edit: Compared to RapidMiner this one is much easier to use. Like it alot.
  • #3
  • Dec 4, 2009 3:46pm Dec 4, 2009 3:46pm
Quoting Ronald Raygun
I'm actually traveling right now, but I saw this in one of my RSS feeds.

http://ccsl.mae.cornell.edu/eureqa
This piece of software rocks!

Finally a tool for perfect curve fitting! This may be of some use for eurusd H1, since this pair is showing extremely strong and stable cycles and autocorrelations since last September.
  • #4
  • Dec 4, 2009 6:53pm Dec 4, 2009 6:53pm
Quoting 7bit
This piece of software rocks!
After playing around a bit it seems a bit buggy, it immediately finds a solution for the example problem but not for almost anything else i tried, not even for simple things with only 20 points and the curves it plots for individual solutions sometimes don't have anything to do with the formula it displays:

Attached Image (click to enlarge)
Click to Enlarge Name: wtf.png Size: 38 KB


Nice idea, but definitely still some work needed.
  • #5
  • Dec 4, 2009 7:31pm Dec 4, 2009 7:31pm
what's wrong with gauss' least squares method?
  • #6
  • Dec 4, 2009 8:44pm Dec 4, 2009 8:44pm
Quoting 7bit
Nice idea, but definitely still some work needed.
Meant to be a proof of concept not a commercially ready piece of software. Besides, bugproofing is just grunt work.
  • #7
  • Dec 4, 2009 8:51pm Dec 4, 2009 8:51pm
Thanks for sharing
The greatest force on earth is compound interest.
  • #8
  • Dec 5, 2009 6:20am Dec 5, 2009 6:20am
Has anyone tried this with eurousd vs usdchf ?

I am trying and not having any luck
  • #9
  • Dec 5, 2009 7:37am Dec 5, 2009 7:37am
Quoting toddanderson
Has anyone tried this with eurousd vs usdchf ?

I am trying and not having any luck

Depends on what you want to calculate. If you want to know the correlation between both just use Excel there is a predefined fuction which just take 2 datarows as parameters to calculate the correlation.
  • #10
  • Dec 5, 2009 9:26am Dec 5, 2009 9:26am
Quoting 7bit
After playing around a bit it seems a bit buggy, it immediately finds a solution for the example problem but not for almost anything else i tried, not even for simple things with only 20 points and the curves it plots for individual solutions sometimes don't have anything to do with the formula it displays:

Attachment

Nice idea, but definitely still some work needed.

I suspect more data is better...
  • #11
  • Dec 5, 2009 9:31am Dec 5, 2009 9:31am
I saw the video and it looks like I can use it to try and model some of the following:

 

  1. Distance between Zigzags
  2. Distance between fractals
  3. Number of bars until a reversal
  4. Moving average direction....


I figure if I use it to model future zigzag turning points across multiple timeframes, I can more or less accurately predict the future.

(Holy grail here I come! haha)

  • #12
  • Dec 5, 2009 9:46am Dec 5, 2009 9:46am
In dealing with human behavior and human reactions, there ain't any "hidden mathematical relationships". It's just emotions, especially greed, and there's just no accounting for it...

But of course if the tool works, do let us know...
  • #13
  • Dec 5, 2009 9:52am Dec 5, 2009 9:52am
Quoting blogzr3
In dealing with human behavior and human reactions, there ain't any "hidden mathematical relationships". It's just emotions, especially greed, and there's just no accounting for it...

But of course if the tool works, do let us know...
which came first -- the chicken or the egg? which came first -- the price action or the human behaviour?
  • #14
  • Dec 5, 2009 10:04am Dec 5, 2009 10:04am
Quoting ABoulfoul
which came first -- the chicken or the egg? which came first -- the price action or the human behaviour?
All of the above. For some, it's the chicken, others it's the egg. Some just follow the price action and go with it, some try to anticipate and place their orders accordingly. And the rest have no ideas, so they hope and pray and then place their orders.
  • #15
  • Dec 5, 2009 10:54am Dec 5, 2009 10:54am
Quoting Ronald Raygun
(Holy grail here I come! haha)
This whole topic is somehow related to something i am currently trying.

Consider the following code:

Inserted Code
double getValue(){
   return(
      iAC(NULL, 0, 0) * a +
      iAC(NULL, 0, 1) * b +
      iAC(NULL, 0, 2) * c +
      iAC(NULL, 0, 3) * d +
      iAC(NULL, 0, 5) * e +
      iAC(NULL, 0, 8) * f +
      iAC(NULL, 0, 13) * g +
      iAC(NULL, 0, 21) * h +
      iAC(NULL, 0, 34) * i +
      iAC(NULL, 0, 55) * j +
      iAC(NULL, 0, 89) * k
   );
}
This is a simple linear model of something, using values of the accelerator oscillator as its regressors. It could also be seen as a FIR filter of 11th order or as a single layer perceptron. This something that is modeled here is the trade direction at any time (always in, no stops, only reverse) that would have given the highest profit over a given period of time. It is used like this:

Inserted Code
void onOpen(){
   double v = getValue();
      
   if (v<0){
      if (getNumOpenOrders(OP_SELL, magic) == 0){
         closeOpenOrders(OP_BUY, magic);
         sell(lots, 0, 0, magic, comment);
      }
   }

   if (v>0){
      if (getNumOpenOrders(OP_BUY, magic) == 0){
         closeOpenOrders(OP_SELL, magic);
         buy(lots, 0, 0, magic, comment);
      }
   }
}
Here is a screenshot from a forward test (i fit it to the last 2 months of H1 data every day), the small window is the equity curve:
Attached Image (click to enlarge)
Click to Enlarge Name: perc.png Size: 48 KB


Currently i fit it using the backtester optimizer, optimizing it for maximum profit, the range for each of the 11 coefficients is [-20 ... 20]. I could also fit them by ordinary least squares or something similar instead of bruteforcing them if i *knew* what exactly the regressand (the above mentioned something) for maximum profit would be, I suspect it is not just the body of the next candle, i'm also not sure what kind of error metric would be the best in this case to maximize profits for this kind of trading (maybe absolute error if i use the next candle body as regressand?).

I tried (and keep trying) to play around with eureqa to find (and fit) similar things, but haven't had any success yet.

PS: it is possible to connect MQL to R for the purpose of fitting and predicting all models R has to offer and do other funny things by using the mql4-python bindings and rpy, but it is difficult to set up, someone needs to write a simple and direct mql4 - R interface.
  • #16
  • Dec 5, 2009 10:59am Dec 5, 2009 10:59am
  •  sloof
  • | Joined Apr 2009 | Status: Birejji (ビレッジ) ►FOREX◄ | 141 Posts
hahaha I've been looking for something like this for a while now! Thanks for the find RR

If there are any better ones out there please let me know =P
  • #17
  • Dec 5, 2009 11:26am Dec 5, 2009 11:26am
  •  sloof
  • | Joined Apr 2009 | Status: Birejji (ビレッジ) ►FOREX◄ | 141 Posts
I'm currently testing the program on 1M data from GU (Only about 1500min worth of data... I tried putting in over 3 Months of data and it crashed lol)

The pic is of the program running so far...
Attached Image (click to enlarge)
Click to Enlarge Name: EureqaTest1.png Size: 98 KB
  • #18
  • Dec 5, 2009 11:33am Dec 5, 2009 11:33am
Quoting sloof
I'm currently testing the program on 1M data from GU (Only about 1500min worth of data... I tried putting in over 3 Months of data and it crashed lol)

The pic is of the program running so far...
Maybe instead of trying to model price directly as a function of time you could try to model price as a function of past prices, indicator values, indicator values of n bars in the past, prices of other pairs, etc.
  • #19
  • Dec 5, 2009 11:37am Dec 5, 2009 11:37am
Quoting 7bit
Maybe instead of trying to model price directly as a function of time you could try to model price as a function of past prices, indicator values, indicator values of n bars in the past, prices of other pairs, etc.
Modelling the price as a function of past price won't work.
Why? Plott a Graph with Close Prices. Now plott graph with Close-1 Prices.
Do they fit? Yes -> Solution found.

Price as a function of Time will probaly just give you a linear regression which you can calculate yourself in a fraction of the time with Excel.
  • #20
  • Dec 5, 2009 11:38am Dec 5, 2009 11:38am
Quoting Tjpld
Plott a Graph with Close Prices. Now plott graph with Close-1 Prices.
Do they fit? Yes -> Solution found.
Then maybe price differences instead of prices.
0 traders viewing now