Designing a FOREX Trading System using SOFM and SVM Options

[navyseal]

Hi Peltarion

i am trying to design a Simple 2-Stage Hybrid Intraday FOREX Trading System using SOFM and SVM . In the First Stage , the SOFM is used to cluster the training data sets into several disjointed Clusters (regions) probably 3 to 6 Clusters . In Second Stage, Mutiple SVMs (or SVR) are constructed for each Cluster in order to either Predict the Price or to classify a BUY or SELL

Inputs: O,H,L,C,V and may be 5 other relevant indicators (RSI, MACD, CCI, SMA, STOCH K ) = may be 10 inputs

Output: Option 1: Classify into BUY ( = 1) and SELL ( = 0)
Option 2: Predict the Price of the Pair e.g EURUSD

The System needs to be completely adaptive may be using SWAM or Genetic or Monte carlo Optimization

Please can someone get back to me and i will send you a document describing the system i am trying to adapt to Forex trading system. I have been trying to use Synapse to design the system but i am getting stuck since i am new to Synapse Software. Any help to do this. I have about 20 similar Hybrid Systems that i am planning to design using SYNAPSE software. If this first example proves to be easy to design i will Purchase the Software inorder to complete the remaining Advanced and complex Hybrid Systems.

Thank

Navyseal

[luka]

Both SOMs and SVMs only work on static algorithms where no implicit time dependency exist between samples. They are essentially unsuitable for time series data.

[navyseal]

Both SOMs and SVMs only work on static algorithms where no implicit time dependency exist between samples. They are essentially unsuitable for time series data.

It is kind a strange because my former employer the Largest Investment bank in the world uses exactly that with a system win rate of 95.5% and a drawdown of 2.7%. I was incharge of the trading Systems and i know what was programmed and i can duplicate it. In reality most people think that way .

If you can help i will appreciate it do not worry if it is suitable for time series or not

[johnr]

I think what Luka is trying to say is that what you are asking for is mathematically impossible. Neither SOMs nor SVMs have any form of short-term memories. They have no way of remembering the order of past samples. It's not a question of suitability but a simple fact of the nature of the algorithms.

[navyseal]

I think what Luka is trying to say is that what you are asking for is mathematically impossible. Neither SOMs nor SVMs have any form of short-term memories. They have no way of remembering the order of past samples. It's not a question of suitability but a simple fact of the nature of the algorithms.

i do agree but it has been done. i know for a fact . I have several Private papers or instructions on doing it for Forex Trading which i used at the Commercial bank as a Senior Strategeist and Trader. They still use the same Hybrid structure i described and it is in 3 of their Major (Top) Systems. They really make the money with this system.

I thought Peltarion would make it easy to do it. I guess i will just have a Neural Network Programmer Code it for me. Most of the Programmers at the commercial bank Coded the Neural Network SOMs & SVMs hybrid system so i know for sure because i worked with them to code it.

[luka]

Johnr is right. Self organizing maps and support vector machines don't see time dependencies in data. If you put a time series into a SOM or SVM, you'll get nonsense out as they are static algorithms. If you want to predict FOREX trading signals (there are quite a few Synapse users that use it for that purpose), you should use the dynamic algorithms/system in Synapse that range from Gamma networks to full blown LSTMs.

For an introduction to time series modeling in Synapse, see tutorial 3 (although I would recommend going through the two first tutorials before that).

[navyseal]

Johnr is right. Self organizing maps and support vector machines don't see time dependencies in data. If you put a time series into a SOM or SVM, you'll get nonsense out as they are static algorithms. If you want to predict FOREX trading signals (there are quite a few Synapse users that use it for that purpose), you should use the dynamic algorithms/system in Synapse that range from Gamma networks to full blown LSTMs.

For an introduction to time series modeling in Synapse, see tutorial 3 (although I would recommend going through the two first tutorials before that).

Hello John & Luka,

What are the best Types of Neural Networks in Synapse to use to Build Forex Trading Systems . Just an honest recommendation would be great . Thank you . i am going through the Tutorials to get a good understanding how Synapse works

Thanks

[navyseal]

Hello John or Luka,

Can someone please put up a simple tutorial on how to create a simple FOREX Trading System using Synapse.

Synapse Data Feed --> Synapse ---> Synapse DLL ( Which generates BUY or SELL signals) and then

How to integrate the Synapse DLL into a C# application with Realtime Data Feed

Realtime Data Feed (Data Feed Adapter) ----> Application with Synapse DLL ( e.g Generating Buy or Sell Signals on MT4) ---> Forex Brokers ( Signals are executed on Metatrader 4)

Just a Simple end to end example would be Great and helpful to the forum greatly. You can use Metatrader Data Feed which is Free and Great . I am actually using Metatrader 4 platform and so are millions of Retail professional system traders. This would be Great . Thank you very much

[bansh33]

I think what Luka is trying to say is that what you are asking for is mathematically impossible. Neither SOMs nor SVMs have any form of short-term memories. They have no way of remembering the order of past samples. It's not a question of suitability but a simple fact of the nature of the algorithms.

Yes that is true. But I think you can still use SOFM/SVM for trading, if the features you pass into your model hold enough temporal information. Basically, you need to convert the dynamic prediction problem to kind of static pattern recognition problem. Which is VERY difficult.
There is a lot of material on how SVM is used to predict time series/ forex rate changes available on the internet - for instance, in the book I have they describe a GASVM model for predicting monthly exchange rate change and the success ratio is above 80% and this method is presented as the best one in the book (compared to other neural network based methods.)

So basically, there are two rules:
1) Features you select needs to be relevant to the market movements and hold enough temporal info to be noticed by the SVM...this is the hardest part, because if you pass too much non relevant features to SVM you get bull***t as an output...you could use some GA optimizer to select just relevant features
2) The task you use SVM for must be simple enough....you may need to use multiple SVMs to get enough info if to make BUY or SELL decision
3) Noise will be your worst enemy...you need to define you features and classes carefully to avoid noise as much as possible

Also, there are modifications to both SVM and SOFM to gain better performance for this type of operation (DSVM, c-ascending SVM, temporal SOMs), which might help you to gain some accuracy in addition (basically, all these methods works with dynamic cost factor - recent samples get higher C)

Good luck!

[business-investm...]

i tried to do in that way but couldn't succeed

[bansh33]

I think this is good reading for start:

"A two-stage architecture for stock price forecasting by integrating self-organizing map and support vector regression"
http://linkinghub.elsevier.com/retrieve/pi...957417408007860

I want to try this later too (now I stick with single/multiple SVR without using SOM)...

Edit: Also this one is interesting: http://www.svms.org/finance/Cao2003.pdf

[Steve]

i do agree but it has been done. i know for a fact . I have several Private papers or instructions on doing it for Forex Trading which i used at the Commercial bank as a Senior Strategeist and Trader. They still use the same Hybrid structure i described and it is in 3 of their Major (Top) Systems. They really make the money with this system.

I thought Peltarion would make it easy to do it. I guess i will just have a Neural Network Programmer Code it for me. Most of the Programmers at the commercial bank Coded the Neural Network SOMs & SVMs hybrid system so i know for sure because i worked with them to code it.

Hi NavySeal - I note that this is an old thread - but did you ever get a resolution to your question?
I would like to explore your ideas with you, with a view to collaboration toward a working trading system.

I have been developing trading ideas using neural networks for over a year now - based on a time series analysis.
I've had extremely poor results & conclude that this is NOT THE WAY TO GO.

To predict FUTURE price or buy/sell points using a time series the network must be able to analyze a repeating pattern, and the pattern should repeat consistently into the future.
It does not work this way!
It is simply too random. (Certainly in today's markets)

My conclusion is that the best that one can do is to be reactive to changes in patterns - with some statistical analysis behind it - without unreliable future prediction on price action.
The more work I do leads me towards static pattern recognition / vector analysis with some statistical input.

[svm]

Hi NavySeal - I note that this is an old thread - but did you ever get a resolution to your question?
I would like to explore your ideas with you, with a view to collaboration toward a working trading system.

I have been developing trading ideas using neural networks for over a year now - based on a time series analysis.
I've had extremely poor results & conclude that this is NOT THE WAY TO GO.

To predict FUTURE price or buy/sell points using a time series the network must be able to analyze a repeating pattern, and the pattern should repeat consistently into the future.
It does not work this way!
It is simply too random. (Certainly in today's markets)

My conclusion is that the best that one can do is to be reactive to changes in patterns - with some statistical analysis behind it - without unreliable future prediction on price action.
The more work I do leads me towards static pattern recognition / vector analysis with some statistical input.

I've tried everything, som, LVQ, PNN and the results were poor, I have a week testing with SVM and I'm getting pretty good results, I think is the solution ..
Regards

[bansh33]

I've tried everything, som, LVQ, PNN and the results were poor, I have a week testing with SVM and I'm getting pretty good results, I think is the solution ..
Regards

I am also experimenting with predicting time series behaviour using SVM (in regression mode), and preliminary results are interesting - it seems to be able to "see" certain price moves few bars ahead, also in the full out-of-sample data segment. The question is, if this performance is good enough to build a trading system of top of it - I don't know yet.

Look at the images - I try to predict percentual change of EMAH3 (exponential average of highest price, period 3) 5 bars ahead. Time series is EURUSD, timeframe 4H. SVM inputs and paramets are chosen by a genetic optimizer. There are three images attached for demonstration:

- training - used for SVM training, 1000 bars
- validation - used for validating the model (out-of-sample for SVM, but in-sample for genetic optimizer), 200 bars - used to select imputs and SVM parameters
- outofsample - full out-of sample data (not seen by any component of the system), 200 bars

Attached File(s)

 eurusd_training.png ( 77.51K ) Number of downloads: 22
 eurusd_validation.png ( 45.82K ) Number of downloads: 21
 eurusd_outofsample.png ( 44.01K ) Number of downloads: 22