Price Analysis with Neural Networks | Page 2 | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

Any success with neural networks in trading 29 replies

Neural Networks: Myths And Reality 19 replies

Extreme scalping using neural networks - EA/Manual 646 replies

Cognitive indicators and neural-networks 19 replies

Neural Networks in Trading (Java, Algodeal, Technical Analysis) 3 replies

Attachments: Price Analysis with Neural Networks
Exit Attachments

Price Analysis with Neural Networks

  • #21
  • Feb 18, 2008 11:44am Feb 18, 2008 11:44am
The only thing closest to GA I've seen is http://www.forexstrategybuilder.com/

It has generation and optimization. However you need to find good amounts of data to feed it. it even takes into acct swap.

Here is the English version: http://forexsb.com/
  • #22
  • Feb 26, 2008 11:40am Feb 26, 2008 11:40am
Charlinks has created a great FANN-MT4 communication layer at the following thread: http://forexfactory.com/showthread.php?p=1870633. Thanks, Char!

I will be moving forward with the development of this NN-based price distribution forecaster using Char's implementation. I'll let y'all know as I make progress.
  • #23
  • Feb 27, 2008 7:24am Feb 27, 2008 7:24am
Good luck with this. Hopefully I can help later on, I've been wanting to implement a lot of ideas for a while. Let's see if we can't at least get some sort of research paper out of this stuff (if not a few million dollars)
Virtue finds and chooses the mean. Aristotle, Ethica Nichomachea
  • #24
  • Feb 27, 2008 10:07am Feb 27, 2008 10:07am
Quoting Plutonite
Let's see if we can't at least get some sort of research paper out of this stuff (if not a few million dollars)
Both would be just fine with me .

So after a few (maybe more ) hours of development, I've implemented an alpha version of the high/low distributions neural network. Its configuration is as follows:

All nodes have a symmetric sigmoidal transfer function.

Inputs
3 inputs for one bar: +upper wick, +- candle body, +lower wick
7 inputs, each corresponding to one day of the week; one is set to 1, others set to 0
1 input for tech indicator: 14-bar MA delta, calculated as (MA[i]-MA[i+1])/(60*Point)

Hidden Layers
1 layer - 30 nodes

Outputs
10 outputs for one bar:
- 5 for the high distribution
- 5 for the low distribution
each output ranges between -1 and 1.

The outputs are plotted as a range of intensity between the background color and some foreground color (purple for highs, green for lows). Each set of outputs is plotted between the last bar and the latest bar. That is, when looking at a bar, the network's predictions are shown to the left of that bar's wick.

The red vertical line on the image shows the first bar of the test set (50 bars). The 10000 bars before that line compose the training set. The network was trained for 1000 iterations over the entire training set with the same weight applied to each bar. The network was then tested on the 50 bars in the test set to produce the colored high/low distributions in the image.

It is interesting to see how well only one bar of input, combined with the day identification and MA delta, can predict the next bar's highs and lows. Of course, each high/low block in the distribution is 10 pips high, so the distribution is not perfectly exact, but the block size is configurable, so this can be adjusted later.

It seems that, if one placed limit orders at the most probable blocks, one could do quite well.

This is actually a much stronger initial result than I expected. I am suspicious that I may have made a coding error somewhere that is somehow giving my learner inside information . I will double-check my code tonight when I have a little more time. For now, I have to go design a circuit for a microcontroller-based apnea monitor.

Happy trading!
Attached Image (click to enlarge)
Click to Enlarge Name: distribution.PNG Size: 22 KB
  • #25
  • Feb 27, 2008 10:15am Feb 27, 2008 10:15am
Quoting drummer486
Both would be just fine with me .

So after a few (maybe more ) hours of development, I've implemented an alpha version of the high/low distributions neural network. Its configuration is as follows:

All nodes have a symmetric sigmoidal transfer function.

Inputs
3 inputs for one bar: +upper wick, +- candle body, +lower wick
7 inputs, each corresponding to one day of the week; one is set to 1, others set to 0
1 input for tech indicator: 14-bar MA delta, calculated as (MA[i]-MA[i+1])/(60*Point)

Hidden Layers
1 layer - 30 nodes

Outputs
10 outputs for one bar:
- 5 for the high distribution
- 5 for the low distribution
each output ranges between -1 and 1.

The outputs are plotted as a range of intensity between the background color and some foreground color (purple for highs, green for lows). Each set of outputs is plotted between the last bar and the latest bar. That is, when looking at a bar, the network's predictions are shown to the left of that bar's wick.

The red vertical line on the image shows the first bar of the test set (50 bars). The 10000 bars before that line compose the training set. The network was trained for 1000 iterations over the entire training set with the same weight applied to each bar. The network was then tested on the 50 bars in the test set to produce the colored high/low distributions in the image.

It is interesting to see how well only one bar of input, combined with the day identification and MA delta, can predict the next bar's highs and lows. Of course, each high/low block in the distribution is 10 pips high, so the distribution is not perfectly exact, but the block size is configurable, so this can be adjusted later.

It seems that, if one placed limit orders at the most probable blocks, one could do quite well.

This is actually a much stronger initial result than I expected. I am suspicious that I may have made a coding error somewhere that is somehow giving my learner inside information . I will double-check my code tonight when I have a little more time. For now, I have to go design a circuit for a microcontroller-based apnea monitor.

Happy trading!
Great! One way to know for sure is to make your indicator as an EA and run it through the simulator in visual mode Great stuff can't mess up using future data...
  • #26
  • Feb 27, 2008 1:30pm Feb 27, 2008 1:30pm
Woah, good job! I do not understand your outputs though - what exactly are the 5 values and how are they different from each other? Also, you say the input bars are equally weighted. Have you tried linearly decreasing the weights as you go back in time(or vice versa)?

P.S: you say you are coding in C++.. how come the chart?!! I presume you made a dll and somehow used it as an indicator. Someone will have to teach me how to do that sometime - Windows programming is a no no for me!

P.P.S Those are some rad colors man. MY EYES...the goggles, they do nothing!!


Quoting drummer486
Both would be just fine with me .

So after a few (maybe more ) hours of development, I've implemented an alpha version of the high/low distributions neural network. Its configuration is as follows:

All nodes have a symmetric sigmoidal transfer function.

Inputs
3 inputs for one bar: +upper wick, +- candle body, +lower wick
7 inputs, each corresponding to one day of the week; one is set to 1, others set to 0
1 input for tech indicator: 14-bar MA delta, calculated as (MA[i]-MA[i+1])/(60*Point)

Hidden Layers
1 layer - 30 nodes

Outputs
10 outputs for one bar:
- 5 for the high distribution
- 5 for the low distribution
each output ranges between -1 and 1.

The outputs are plotted as a range of intensity between the background color and some foreground color (purple for highs, green for lows). Each set of outputs is plotted between the last bar and the latest bar. That is, when looking at a bar, the network's predictions are shown to the left of that bar's wick.

The red vertical line on the image shows the first bar of the test set (50 bars). The 10000 bars before that line compose the training set. The network was trained for 1000 iterations over the entire training set with the same weight applied to each bar. The network was then tested on the 50 bars in the test set to produce the colored high/low distributions in the image.

It is interesting to see how well only one bar of input, combined with the day identification and MA delta, can predict the next bar's highs and lows. Of course, each high/low block in the distribution is 10 pips high, so the distribution is not perfectly exact, but the block size is configurable, so this can be adjusted later.

It seems that, if one placed limit orders at the most probable blocks, one could do quite well.

This is actually a much stronger initial result than I expected. I am suspicious that I may have made a coding error somewhere that is somehow giving my learner inside information . I will double-check my code tonight when I have a little more time. For now, I have to go design a circuit for a microcontroller-based apnea monitor.

Happy trading!
Virtue finds and chooses the mean. Aristotle, Ethica Nichomachea
  • #27
  • Feb 27, 2008 2:20pm Feb 27, 2008 2:20pm
Quoting Plutonite
Woah, good job! I do not understand your outputs though - what exactly are the 5 values and how are they different from each other? Also, you say the input bars are equally weighted. Have you tried linearly decreasing the weights as you go back in time(or vice versa)?

P.S: you say you are coding in C++.. how come the chart?!! I presume you made a dll and somehow used it as an indicator. Someone will have to teach me how to do that sometime - Windows programming is a no no for me!

P.P.S Those are some rad colors man. MY EYES...the goggles, they do nothing!!
Use the MT4 plugin I made and posted!
  • #28
  • Feb 27, 2008 4:56pm Feb 27, 2008 4:56pm
Quoting charlinks
Use the MT4 plugin I made and posted!
Doh! Just saw it now! Drummer: did you post your code too? I dont want to write another classifier!!! *tears*

PS: About books, the most complete I have studied was Pattern Classification by Duda. Excellent reference and very heavy on multivariate bayesian stuff, which could help if you want to use more than just a NN. Ok, Ok stop throwing things at me - just food for thought...
Virtue finds and chooses the mean. Aristotle, Ethica Nichomachea
  • #29
  • Feb 27, 2008 6:26pm Feb 27, 2008 6:26pm
Quoting Plutonite
Doh! Just saw it now! Drummer: did you post your code too? I dont want to write another classifier!!! *tears*

PS: About books, the most complete I have studied was Pattern Classification by Duda. Excellent reference and very heavy on multivariate bayesian stuff, which could help if you want to use more than just a NN. Ok, Ok stop throwing things at me - just food for thought...
Did not post my code. It still needs quite a bit of tweaking. In fact, my suspicions were confirmed: there was a bug in the output calculation. The results in my first development update are thus mostly invalid. I'll be testing out a bunch of ideas over the next week or so, and I'll be sure to post any major successes.

As for pattern classification... I was indeed thinking about using a Bayesian learner. Either that, or possibly a SVM. Too many possibilities!

Divide and conquer sounds like the best strategy to me... you figure out how to use Char's code? Seems like the three of us, and maybe a few others, could add some serious meat to this discussion.
  • #30
  • Feb 29, 2008 7:08pm Feb 29, 2008 7:08pm
Quoting drummer486
Did not post my code. It still needs quite a bit of tweaking. In fact, my suspicions were confirmed: there was a bug in the output calculation. The results in my first development update are thus mostly invalid. I'll be testing out a bunch of ideas over the next week or so, and I'll be sure to post any major successes.

As for pattern classification... I was indeed thinking about using a Bayesian learner. Either that, or possibly a SVM. Too many possibilities!

Divide and conquer sounds like the best strategy to me... you figure out how to use Char's code? Seems like the three of us, and maybe a few others, could add some serious meat to this discussion.
Quite possible that a linear classifier wont work. God knows theres enough possibilities with the NN alone. I think we should stick with those, actually. I haven't looked at theory in a while so I've forgotten what kind of functions can be modeled by which classifier, but I think multilayer NN are the best bet for now. The range of inputs/outputs we can design it for is quite daunting, and that's without even talking about a trading strategy(s) that uses them. Good luck for now, I'll pop back once I've cleared up my workload a little and can actually program something on the side. grad school is not fun!
Virtue finds and chooses the mean. Aristotle, Ethica Nichomachea
  • #31
  • Mar 3, 2008 10:56pm Mar 3, 2008 10:56pm
  •  Trader#0000
  • | Status: Unregistered Trader
Quoting charlinks
The only thing closest to GA I've seen is http://www.forexstrategybuilder.com/

It has generation and optimization. However you need to find good amounts of data to feed it. it even takes into acct swap.

Here is the English version: http://forexsb.com/
Where can one find a free data source for currency swap rates?
  • #32
  • Mar 16, 2008 4:57pm Mar 16, 2008 4:57pm
Quoting Plutonite
Good luck for now, I'll pop back once I've cleared up my workload a little and can actually program something on the side. grad school is not fun!
I'm in the same boat . I've been unable to do much work on the NN stuff lately, as I've been going through finals week at school. Thankfully I'm done with most of my schoolwork, save a couple projects and old problem sets I need to finish off this weekend. Then I'll be off to sunny FL for a much-needed vacation. I'm really interested in getting back up to speed on this project, however. I'll be graduating this June, and it sure would be nice to have either some good job interview fodder or some nice trading profits based upon our developments.

I think I mentioned this elsewhere, but I may be taking a course called "Human Level Artificial Intelligence" this coming quarter. If I can get in, it should be very interesting, and hopefully quite useful.

As for further development, while many on this forum believe that trading based upon "naked" charts is best, it seems that indicators may prove quite useful for a machine-learning algorithm. While it's easy for traders to see volatility, trends, and support/resistance on a chart, it is more difficult for a neural network to pick out these features with limited training. It might be best to instead calculate these factors in the form of indicators (on various timeframes). Neural networks are great at finding patterns among data, but their job becomes exponentially (in the actual mathematical sense) easier if the inputs are preprocessed a little to select important variable features. In light of this, I may update my design specs to include some selected indicators along with or maybe even instead of pure price action.

The reason price action can work better for human trading is that we can calculate our own "indicators" internally that are much more complex than a simple moving average. On the other hand, it can take us years of staring at price action to gain a workable understanding of market movements, while we'd like the neural networks to learn everything they need in just a few weeks (or hours, or minutes!).

This also brings me to one more point: I've played around with a free adaptive (genetic) population modeler called Altreva, and it seems to have some features that might be very useful in our development efforts. One key principle it uses to prevent curve-fitting is that its models evolve only in a time-incremental fashion. That is, training and/or population selection is performed in real-time, as the quote series is fed through from beginning to end. Thus, what results is not only a system that performed well, but one that learned well over time. This suggests that such a system would be able to properly adapt to future shifts in market behavior, adapting as circumstances change to maintain profitability. I thus suggest that the learner resulting from this thread only learn in a time-incremental fashion, building knowledge as it progresses from the beginning of time in the market.

Has anyone ever used this type of approach in the application of neural networks to time series?
  • #33
  • Mar 16, 2008 8:25pm Mar 16, 2008 8:25pm
Hi drummer,

I didn't want to impose too much on your thread, but I find it interesting how similar a lot of your initial thoughts are to the thoughts I had when I began developing my artificial neural network a year ago. I wish you luck on your journey. Get ready for revision after revision and optimization after optimization... you have a lot of work ahead of you. It will be worth it in the end!
  • #34
  • Mar 16, 2008 8:45pm Mar 16, 2008 8:45pm
Quoting Advacast
Hi drummer,

I didn't want to impose too much on your thread, but I find it interesting how similar a lot of your initial thoughts are to the thoughts I had when I began developing my artificial neural network a year ago. I wish you luck on your journey. Get ready for revision after revision and optimization after optimization... you have a lot of work ahead of you. It will be worth it in the end!
Thanks for the encouragement, Advacast. It sounds like you've got some good experience under your belt. I'm sure you're somewhat tight-lipped about what actually works for you, but if you can contribute any suggestions, they'd be more than welcome. What sort of inputs do you use to your network(s)? And what sort of outputs?

I'm interested in any and all opinions, contrary to mine or not.
  • #35
  • Mar 16, 2008 11:59pm Mar 16, 2008 11:59pm
With my experiences, each step has been a hurdle to overcome the computational limitations with each layer (input) I've added. I've developed a number of preprocessors to speed things up drastically. I treat my inputs like senses data, then stack layers of various processing techniques to those inputs. The reason I compare to senses is because I store and refer to the data using the same type of correlation you might experience when you see something and it reminds you of a song or a smell. I correlate pairs that don't even technically correlate because I think there is something to gain from having all options available. I've written very broad algorithms to allow the software to invent its own methods of extrapolation. It controls the parameters of those functions using GA to produce the best results within live trading, historical simulations and (will be completed shortly) theoretical simulations.

The inputs are simple pricing data: open, high, low, close, volume in every standard time frame. I used to output to a colorful grid but have recently decided to switch to a numerical output. Easier to make decisions.

Have you run into any computation limits yet? Are you performing any simulations, or only learning as the market moves live?
  • #36
  • Mar 17, 2008 2:02am Mar 17, 2008 2:02am
Hi there i did some work on smoething like you. I did not use neural network. I used clustering & machine learning...

WHat make my head big is what parameters to use. I had my own profitable trading systen .. long term trend trading.

doing the machine learning thing is for self fulfillment.
No really useful yet... but improving while i can.
Keep it up.
Just here to cheer you
  • #37
  • Mar 17, 2008 11:14am Mar 17, 2008 11:14am
Quoting drummer486
This also brings me to one more point: I've played around with a free adaptive (genetic) population modeler called Altreva, and it seems to have some features that might be very useful in our development efforts. One key principle it uses to prevent curve-fitting is that its models evolve only in a time-incremental fashion. That is, training and/or population selection is performed in real-time, as the quote series is fed through from beginning to end. Thus, what results is not only a system that performed well, but one that learned well over time. This suggests that such a system would be able to properly adapt to future shifts in market behavior, adapting as circumstances change to maintain profitability. I thus suggest that the learner resulting from this thread only learn in a time-incremental fashion, building knowledge as it progresses from the beginning of time in the market.

Has anyone ever used this type of approach in the application of neural networks to time series?
I discovered Altreva a while ago, but I didn't find the models to be all that impressive. The equity curve did look nice, and it appeared that the systems it produced were not curve-fit, but the profits were minuscule. The example model that they show for the S&P does not beat buy and hold.

Do you have any other example models you can share?
  • #38
  • Mar 24, 2008 11:34am Mar 24, 2008 11:34am
Quoting pipsqueaks
I discovered Altreva a while ago, but I didn't find the models to be all that impressive. The equity curve did look nice, and it appeared that the systems it produced were not curve-fit, but the profits were minuscule. The example model that they show for the S&P does not beat buy and hold.
You are probaby referring to the return over the last 10 years (which is shown by default in the sample model on the Performance tab) of the sample model that came with version 0.98. The compounded average return over all the 46 years (including the last 10) are close to 20%.

The new sample model that is on the website since the release of version 0.99 (on 12 March) shows a compounded avg annual return of 8.8% over the last 10 years (which is an excess annual return of 6.7%) and a 20.8% compounded avg annual return over the last 47 years. Also notice the smooth equity curve since 1999.
  • #39
  • Mar 25, 2008 6:34am Mar 25, 2008 6:34am
Quoting Altreva
You are probaby referring to the return over the last 10 years (which is shown by default in the sample model on the Performance tab) of the sample model that came with version 0.98. The compounded average return over all the 46 years (including the last 10) are close to 20%.

The new sample model that is on the website since the release of version 0.99 (on 12 March) shows a compounded avg annual return of 8.8% over the last 10 years (which is an excess annual return of 6.7%) and a 20.8% compounded avg annual return over the last 47 years. Also notice the smooth equity curve since 1999.
Altreva,

Perhaps you can show us a few more example models? I believe GP has the potential to produce models even more profitable than that.
  • #40
  • Mar 25, 2008 10:39am Mar 25, 2008 10:39am
Quoting pipsqueaks
Altreva,

Perhaps you can show us a few more example models? I believe GP has the potential to produce models even more profitable than that.
Altreva, I too would be interested in seeing profitable models that you've developed.

Could you post more in a separate thread? I'd like to keep this thread focused on NN development. Thanks
0 traders viewing now