Neural Networks - Genetic Algorithms - Boltzmann Machines - FOREX crystal balls?

gtatix

Junior member
Messages
40
Likes
3
Comments and discussion related specifically to the usability for forex trading - or the hypothetical usage of such in Forex trading - of the titled topics would be appreciated.
Further discussions as per the titled items regarding:

Dr. Jeffrey Owen Katz and N-train along with a concise conclusion to his findings in "The Encyclopedia of Trading Strategies" are welcomed.
Dr. Jeffrey Katz's lamish website - [I don't believe that HE is lame in any way however!] here:
Scientific Consultant Services -- Master Index
a more specific link to my query here:
Software please read beyond the price quotes regarding the products.

Boltzmann Machines and Dr. Geoffrey E. Hinton of the U of T information can be found here:
Boltzmann machine - Scholarpedia
and here:
http://www.cs.utoronto.ca/~hinton/absps/netflixICML.pdf
WTF is wrong with this discussion group software? I can't upload a small PDF file due to ?????? Now the following youtube video is alledgedly not available - which is a LOAD...it is available. Go to youtube [afraid to type the URL in case it doesn't work!!!!!!!!!!!!!!] and type in "Boltzmann machine" there are only TWO videos that come up. Pick the first one it is a Google engEDU video so it is SOUND! I hate having my time wasted with error-filled days. Sorry for the anger but....what a waste!
and here:
YouTube - The Next Generation of Neural Networks

I think that advanced neural networking and proper implementation of basic to quasi-complex forex technical strategies are the answer to successful trading. What is your experience?
 
Last edited:
Hi

Can knowledge base articial intellegince be applied to forex?Could it be applied with probabilistic knowledge?

OILFXPRO
 
Hi

Can knowledge base articial intellegince be applied to forex?Could it be applied with probabilistic knowledge?

OILFXPRO
Well it has basically been done. Much of the testing that is up front and clearly available are very basic tests of varying moving average crossovers. I am thinking more on the lines of being able to input data/history with formulae but incorporating a lot more.
I think an awesome concept would be to program a neural net with all the results of market reaction to - perhaps - the 10 most influential fundamental announcements over the past ??? years. Once done, the net would have the "memory" of these reactions and be easily able to predict - with accuracy - the true human reaction. I think this would be possible because the "human reaction" is very predictable. Just look at history in general. Clearly "bad" economic news should result in market downs. "Good" news in market ups. Is this always true? Is it usually true? Is it true - but only with a certain amount of information or is it true when the news from a fundamental announcement is virtually devestating or amazing? This could be quite easily done with the right data and the right - fairly basic - software.
 
Here's a a few tips when using neural networks.

1) standardise the inputs by using the natural log of lagged returns (eg take the log of the difference in closing prices over x bars)
2) look for correlations between a target variable of AT LEAST 50 bars in the future, and inputs using lagged returns covering the past 50-250 bars
3) use intermarkets as the primary source of inputs
4) keep the number of inputs below 7-8 (as few as possible)
5) do not use more than 1 hidden layer
6) build several networks and use emsembles
7) do not try and predict price as the target variable - rather predict the log return of price over the next x days, or something like the rsi.
8) stay away from predicting moving averages - they are easy to predict and utterly useless in real trading (one particular software vendor uses predictions of moving averages and i know from bitter experience its a pointless exercise)
9) before building a network do seperate analysis on correlations

I have reserached neural nets for some time and have had some success - particularly with forex. (e.g over 90% directional accuracy at predicting direction over the next 50 bars on out of sample and live trading). Its not easy, and every market behaves differently, so no holy grail.

happy trading.
 
Here's a a few tips when using neural networks.

1) standardise the inputs by using the natural log of lagged returns (eg take the log of the difference in closing prices over x bars)
2) look for correlations between a target variable of AT LEAST 50 bars in the future, and inputs using lagged returns covering the past 50-250 bars
3) use intermarkets as the primary source of inputs
4) keep the number of inputs below 7-8 (as few as possible)
5) do not use more than 1 hidden layer
6) build several networks and use emsembles
7) do not try and predict price as the target variable - rather predict the log return of price over the next x days, or something like the rsi.
8) stay away from predicting moving averages - they are easy to predict and utterly useless in real trading (one particular software vendor uses predictions of moving averages and i know from bitter experience its a pointless exercise)
9) before building a network do seperate analysis on correlations

I have reserached neural nets for some time and have had some success - particularly with forex. (e.g over 90% directional accuracy at predicting direction over the next 50 bars on out of sample and live trading). Its not easy, and every market behaves differently, so no holy grail.

happy trading.
 
Well it has basically been done. Much of the testing that is up front and clearly available are very basic tests of varying moving average crossovers. I am thinking more on the lines of being able to input data/history with formulae but incorporating a lot more.
I think an awesome concept would be to program a neural net with all the results of market reaction to - perhaps - the 10 most influential fundamental announcements over the past ??? years. Once done, the net would have the "memory" of these reactions and be easily able to predict - with accuracy - the true human reaction. I think this would be possible because the "human reaction" is very predictable. Just look at history in general. Clearly "bad" economic news should result in market downs. "Good" news in market ups. Is this always true? Is it usually true? Is it true - but only with a certain amount of information or is it true when the news from a fundamental announcement is virtually devestating or amazing? This could be quite easily done with the right data and the right - fairly basic - software.
why isn't this thread being continued???

you're talking about vantage point software, aren't you? :rolleyes:

besides that obvious example, neunets have been employed financially
i found this one interesting and i'll start experimenting some things
if you know your way around this stuff, i'd like some help
Neural network software, stock prediction, market forecasting, trading, Excel
Backtesting Software, Stock Technical Analysis, Forex Neural Nets

i know others but this one is the one i remember now (y)
 
Trying to revive this interesting thread.

I'm not a big fan of neural nets because they are a black box and you can't see the rules and calculations learned. I much prefer genetic algorithms (or genetic programming) because you get the same benefits i.e. a solution learnt from the problem data but the solution is a program that is human readable so can be reimplemented in other systems e.g. MetaTrader or even a set of written rules that a manual trader could use.

I've built a spread betting simulator that coupled with minute by minute data acts as a fitness function. The algorithms create functions that are used to populate an order i.e. order direction (buy or sell), order type (stop or limit), order level (market value to enter at), stop loss, limit (profit target). The algorithms have to create none, one or more orders per day based on the past daily candles and various indicators. The orders are then executed with minute by minute data to determine profit / loss.

So far I've had some interesting results that can be extremely profitable on the trained data, the trick is to produce algorithms that remain profitable on unseen data which is proving harder to achieve.

Is anyone else trying this approach ?

- Andrew
 
I always thought a patriot act style program that picks up on key words in new news article releases could be incorporated in an ai based trading system.
 
Yes I agree - a bayesian categoriser like a spam filter would work well for that e.g. if the word 'challenging' featured heavily in the company report then you'd sell.

So basically you'd be looking to sell as soon as possible after the news appeared in realtime along with every other daytrader ?
-
I tend to be looking for daily trends by which time any bad news will already be reflected in the price movement so would be picked up in the technical analysis. But the magnitude of the bad news might invalidate any existing support levels so it might be useful from that point of view.
 
Trying to revive this interesting thread.

I'm not a big fan of neural nets because they are a black box and you can't see the rules and calculations learned. I much prefer genetic algorithms (or genetic programming) because you get the same benefits i.e. a solution learnt from the problem data but the solution is a program that is human readable so can be reimplemented in other systems e.g. MetaTrader or even a set of written rules that a manual trader could use.

I've built a spread betting simulator that coupled with minute by minute data acts as a fitness function. The algorithms create functions that are used to populate an order i.e. order direction (buy or sell), order type (stop or limit), order level (market value to enter at), stop loss, limit (profit target). The algorithms have to create none, one or more orders per day based on the past daily candles and various indicators. The orders are then executed with minute by minute data to determine profit / loss.

So far I've had some interesting results that can be extremely profitable on the trained data, the trick is to produce algorithms that remain profitable on unseen data which is proving harder to achieve.

Is anyone else trying this approach ?

- Andrew

Hi Andrew,

Neural networks and AIs have been and, are being used, in the various financial markets. There are a number of trading robots which, allegedly, will produce profits while you sleep. Imho, making and executing your own trading decisons is best.
 
Imho, making and executing your own trading decisons is best.

There's a lot of sense in that. But I think genetic algorithms can discover patterns and ideas that are counter-intuitive but profitable which could then be incorporated into manual trading.

As an example, when I checked on my latest run this morning before leaving for work, one of the algorithms was setting the stop loss to the value of stochastic indicator (0 - 100) on a sell order. I wouldn't have thought to do that but it makes perfect sense as the higher the stoch is, the more oversold the market is likely to be and therefore the more confident you are of a sell so the more money you would be prepared to risk on it. You could never have gained that nugget of information from a NN as it's just a collection of weights and connections that is meaningless.

So I'm not saying that NN's are not effective, I'm sure that they are, but I think you can learn more from genetic algorithms even if you don't use them to autotrade. Personally since I have to work for a living and I can't scalp during the day, then autotrading is quite attractive to me so long as it has controls in place to stop losing before your account is cleared out !
 
There's a lot of sense in that. But I think genetic algorithms can discover patterns and ideas that are counter-intuitive but profitable which could then be incorporated into manual trading.

As an example, when I checked on my latest run this morning before leaving for work, one of the algorithms was setting the stop loss to the value of stochastic indicator (0 - 100) on a sell order. I wouldn't have thought to do that but it makes perfect sense as the higher the stoch is, the more oversold the market is likely to be and therefore the more confident you are of a sell so the more money you would be prepared to risk on it. You could never have gained that nugget of information from a NN as it's just a collection of weights and connections that is meaningless.

So I'm not saying that NN's are not effective, I'm sure that they are, but I think you can learn more from genetic algorithms even if you don't use them to autotrade. Personally since I have to work for a living and I can't scalp during the day, then autotrading is quite attractive to me so long as it has controls in place to stop losing before your account is cleared out !



Hi Andrew,

As a new and eager learner to this forum, a network engineer of 30 years experience and a research student in the field of neural networks, this thread intrigues me. What program are you using to model your GA on and how have you linked it accross to carry out your autotrades? (If you dont mind me asking);).

Paula.
 
Hi Paula,

l've written my own trading simulator that works with historic data. The genetic algorithms are then evaluated in the simulator as their fitness function. This blog post has more details.

regards

Andrew
 
Also forgot to say, although I haven't got this for yet I will autotrade by taking the winning algorithms and recoding these as a Metatrader EA - I have an MQL framework ready to plug them in.

This will allow me to then confirm the results by backtesting in MT and then forward test in a demo account to ensure continued profitability before trying real money.
 
Thanks Andrew,

Very interesting. Im currently looking at neural networks and also comparing with GA's looking at crossover and mutation to see if the No free lunch theory holds for market trading or not. Am currently using a java neural net simulator based on the SNNS from TU Stuttgart. Was quite intrigued when I found the link accross to trading as that is more of a long time interest that Ive decided to try and learn a bit more about. Gotta pay for that conservatory somehow ;-)

Paula
 
Here's a a few tips when using neural networks.

1) standardise the inputs by using the natural log of lagged returns (eg take the log of the difference in closing prices over x bars)
2) look for correlations between a target variable of AT LEAST 50 bars in the future, and inputs using lagged returns covering the past 50-250 bars
3) use intermarkets as the primary source of inputs
4) keep the number of inputs below 7-8 (as few as possible)
5) do not use more than 1 hidden layer
6) build several networks and use emsembles
7) do not try and predict price as the target variable - rather predict the log return of price over the next x days, or something like the rsi.
8) stay away from predicting moving averages - they are easy to predict and utterly useless in real trading (one particular software vendor uses predictions of moving averages and i know from bitter experience its a pointless exercise)
9) before building a network do seperate analysis on correlations

I have reserached neural nets for some time and have had some success - particularly with forex. (e.g over 90% directional accuracy at predicting direction over the next 50 bars on out of sample and live trading). Its not easy, and every market behaves differently, so no holy grail.

happy trading.

Hi, Jayram

Thanks for that valuable info.

I'll appreciate any help you can give us.

Number of bars that you use to train the neural network?

Do you use FANN library, or another package?

What kind of inputs do you use?

Now, I'm testing the NN with some inputs of log/diferences between price and the high and low of x bars ago, bolinger bands, etc. (EUR5 or EUR15). The output/s are the log returns. Until now, the prediction are not successful.

Thanks.
 
Neural nets are not as great performers as first appears. It is possible to make them work but the markets change and stationary learning will not produce consistent profits.

I'm working on some strategies in this area applied to the spot forex markets. My main tool is Matlab but I code the decent strategies into .dlls which can be used in metatrader, etc.

I'd be interested in a discussion on this if anyone is still looking at this thread.
 
I'd recommend that you ditch NN's and look at GP instead. I've found several consistently profitable strategies using GP although it certainly wasn't easy.
 
I'm not clear about Genetic Alogarithm. What is it? What is the difference with Neural Network?
 
Top