Neural network system

[Mediator]

Hello People,

i'm very interested in this field of artificial intelligence. I'm researching and coding a long time and can tell you that this is not so easy like some will think so.

NN in its pure existance will not made more or better profits. it depents on how you train them (supervised or unsupervides) and for what task you use a NN. There exist many task like classification, generalization, clustering, pattern recognition and so on to name a few. And there exists hybrid systems also.

In the near future i will start a new (own) thread because i need help from good coders here. My plan is to create an adaptive multi-agent system (AMAS). This is evolutionary computing including genetic algorithms/programming, evolving strategies coded with simple encoding structures (which could easily done with GPU/CUDA power and cluster networking for shared development).

It sounds difficult but it is very simple because this is strategy evolution with darwian logic.

So be prepared. i could imagine that i could help here with NN and later we could implement this ideas into the AMAS.

Cheers

Mediator

[TraderDesk]

A company that I have been looking into using there software to develop an AI for my trading are located at: http://www.ai-one.com/

I don't know if it would be applicable here but they have some cool white papers...

The Stanford course is pretty good and is a good primer on AI....

[guandi]

Hello all,

i do not know how many of you read this thread in trade2win.com, but one of the posters Arryex has done enormous amount of work using NN with Empty4, his thread is here, http://www.trade2win.com/boards/metatra ... shell.html
,maybe we can invite him here to share his experience.

best regards

[Galen88]

Hi Gamma and friends,

Many thanks to Gamma for introducing this system and sharing the information he has on this as I too am very interested.

I found out about Louis Mendelsohn's system recently and his commercial system claimed 86% accuracy. Another independent report gave a mean accuracy of 78%. They have 2 documents giving some details.

The first gives a glimpse into how their system works:
VantagePoint software uses five neural networks, in a two-level hierarchy, to forecast five different indicators for that market.
1. The first forecasts the next-day high to help set stops for entry and exit points.
2. The second forecasts the next-day low to help set stops for entry and exit points.
3. The third forecasts a five-day moving average of closes two days into the future to indicate the expected short-term trend direction within the next two days.
4. The fourth forecasts a ten-day moving average of closes four days into the future to indicate the expected medium-term trend direction within the next four days.
5. The fifth indicates whether a market is expected to change trend direction, making a top or a bottom, within the next two days.

The second document is an accuracy assessment conducted by Dr Warnock on 26 randomly selected instruments over a 2 year period (including some forex pairs such as USDCAD.) His report showed a mean combined accuracy of 78%, forecast for longs were slightly better.

For example, accuracy for longs for USDCAD was 83.3% versus shorts at 80.6%

As there is a 256kb limit to uploads, I could not upload them here (each pdf is more than 1 mb in size). If anyone is interested, pm your email and I will pass them to you that way.

Finally, I encourage everyone to take the free machine learning course from Stanford University as mentioned earlier @ http://www.ml-class.org Professor Andrew Ng does a great job of simplifying the learning algorithms as much as possible and even providing visual examples or what he calls intuitions which help us to see how these "regression" algorithms work.

I look forward to participating actively in this thread.

[magft]

Hi Gamma and friends,

Many thanks to Gamma for introducing this system and sharing the information he has on this as I too am very interested.

I found out about Louis Mendelsohn's system recently and his commercial system claimed 86% accuracy. Another independent report gave a mean accuracy of 78%. They have 2 documents giving some details.

The first gives a glimpse into how their system works:
VantagePoint software uses five neural networks, in a two-level hierarchy, to forecast five different indicators for that market.
1. The first forecasts the next-day high to help set stops for entry and exit points.
2. The second forecasts the next-day low to help set stops for entry and exit points.
3. The third forecasts a five-day moving average of closes two days into the future to indicate the expected short-term trend direction within the next two days.
4. The fourth forecasts a ten-day moving average of closes four days into the future to indicate the expected medium-term trend direction within the next four days.
5. The fifth indicates whether a market is expected to change trend direction, making a top or a bottom, within the next two days.

The second document is an accuracy assessment conducted by Dr Warnock on 26 randomly selected instruments over a 2 year period (including some forex pairs such as USDCAD.) His report showed a mean combined accuracy of 78%, forecast for longs were slightly better.

For example, accuracy for longs for USDCAD was 83.3% versus shorts at 80.6%

As there is a 256kb limit to uploads, I could not upload them here (each pdf is more than 1 mb in size). If anyone is interested, pm your email and I will pass them to you that way.

Finally, I encourage everyone to take the free machine learning course from Stanford University as mentioned earlier @ http://www.ml-class.org Professor Andrew Ng does a great job of simplifying the learning algorithms as much as possible and even providing visual examples or what he calls intuitions which help us to see how these "regression" algorithms work.

I look forward to participating actively in this thread.

This system doesn't sound too tricky to implement using FANN2MQL4 so i'll have a look into this and will post an update probably towards the weekend. I'll pm you my email so i can have a read. I have tried something a lot more simpler than this and it didn't work too well. I think we would need to look at developing a self optimising system that retrains each day (user specifiable) to refresh the NNs using the new data. Do you think it would work for intraday say using the same ideas on a smaller timeframe?

[magft]

I'v had a read through the VantagePoint PDF and the Networks are as follows:

• 1. The first forecasts the next-day high to help set stops for entry and exit points.
  2. The second forecasts the next-day low to help set stops for entry and exit points.
  3. The third forecasts a five-day moving average of closes two days into the future to indicate the expected short-term trend direction within the next two days.
  4. The fourth forecasts a ten-day moving average of closes four days into the future to indicate the expected medium-term trend direction within the next four days.
  5. The fifth indicates whether a market is expected to change trend direction, making a top or a bottom, within the next two days.

The first four neural networks at the primary level of the hierarchy make independent market
forecasts of the high, low, short- and medium-term trend. These predictions are then input into
the fifth network, along with other intermarket data input at the secondary level of the network hierarchy, to predict market turning points.

I need ideas as to what the text in bold can be, i can start putting together the rest for testing. I plan to initially create 4 indicators for points 1-4 and then finally another indicator for point 5 using the other 4 and this extra data. Once we have the indicators showing hopefully good signals we can build the EA round it, maybe even build all into a combined EA.

Any thoughts?

Mike

[Galen88]

This system doesn't sound too tricky to implement using FANN2MQL4 so i'll have a look into this and will post an update probably towards the weekend. I'll pm you my email so i can have a read. I have tried something a lot more simpler than this and it didn't work too well. I think we would need to look at developing a self optimising system that retrains each day (user specifiable) to refresh the NNs using the new data. Do you think it would work for intraday say using the same ideas on a smaller timeframe?

Hi Mike,

I have attached the pdfs to you, happy reading and coding!

Yes, I believe you are right, a self-optimizing system that retrains at specific periods (could be 1 min if the the computing power is adequate) is the way to go. It should work on any time series as long as it has enough time to learn the sequence - i.e. reach the desired mean square error.

My understanding is that the artificial neural network or ANN learns in a very similar way to how we humans learn - i.e. by supervision and/or experience or deduction. Therefore the ideal system should have elements of both types of learning.

Other than the fann2mql4 framework, we could also look at other open source frameworks such as Encog.

Here is a link showing how to add the Encog's framework to Ninjatrader:
http://www.heatonresearch.com/wiki/Enco ... ng_Started

Here is a link to a video from Jeff Heaton, the developer:
http://www.youtube.com/watch?v=52IK5wHSkrA

We will discover very quickly that they are a lot of different types of ANN including SVM, another algorithm with a lot of promise.

One way to explore this quickly is to study the NextClose-V13 code from Neuro and go from there. I find his V14 version pretty fascinating, works well for EURUSD and GPBUSD 5m, not so well for USDJPY 5m though.

I will report on my study of NextClose-V13 once I have any significant to report. As I am still learning this ANN stuff, it may take a while.

Happy Trading!

[magft]

I'v had a read through the VantagePoint PDF and the Networks are as follows:

• 1. The first forecasts the next-day high to help set stops for entry and exit points.
  2. The second forecasts the next-day low to help set stops for entry and exit points.
  3. The third forecasts a five-day moving average of closes two days into the future to indicate the expected short-term trend direction within the next two days.
  4. The fourth forecasts a ten-day moving average of closes four days into the future to indicate the expected medium-term trend direction within the next four days.
  5. The fifth indicates whether a market is expected to change trend direction, making a top or a bottom, within the next two days.

The first four neural networks at the primary level of the hierarchy make independent market
forecasts of the high, low, short- and medium-term trend. These predictions are then input into
the fifth network, along with other intermarket data input at the secondary level of the network hierarchy, to predict market turning points.

I have created NNets 1 and 2 as shown in attached screenshot. I'll work on the next 2 later on today. I have used fann2mt4 and adapted an example. Currently it just solves when first dropped on chart so the next phase is to get them to recalc as new bars are added. The NNet outputs are scaled to between 0 to 100.

Mike

[DING]

Mike,

Nice work, I will follow up.

[magft]

Another update, i think i have the MA nets working now