Neural Programming for Profitable trading | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

Neural Network Trading, Serious people only! 127 replies

Neural Network trading using Matlab and Metatrader 8 replies

Interested in Neural Network Assisted Trading 4 replies

Growing Neural Gas Algo for FX trading 2 replies

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

Neural Programming for Profitable trading

  • #1
  • First Post: Jul 30, 2006 12:58am Jul 30, 2006 12:58am
I know a lot of people do this and don't ever talk about it because everyone would want their program on their computer hooked up to all of the money they had.

If you do know about it, could you please recommened some books, some ideas, some software that is already out there I might get ahold of.

How do I go about mastering neural trading.
I keep a close watch on this chart of mine...
  • #2
  • Jul 30, 2006 7:00am Jul 30, 2006 7:00am
If you mean neural networks, maybe you should buy the "light" version of brainmaker and experiment with it. Look at: http://www.calsci.com/
  • #3
  • Jul 30, 2006 7:08am Jul 30, 2006 7:08am
To integrate a neural network into price data, and screen historical data to develop pattern recognition. You need a very good programmer who has a background in financial markets, and neural networks and pattern recognition. Its very hard to do that, since most institutions would gobble them up. Boutique hedge funds have the resources to do this.

Retail investors not as easy.
Price is the only indicator.
  • #4
  • Jul 30, 2006 9:21am Jul 30, 2006 9:21am
the only thing neural networks would have an edge is lack of emotion. I would have to argue that your own personal neural net(your brain) has much more potential then anything we would be able to design in the next 10 years.

So those trying to develop neural nets for trading are trying to create a artificial trader, and in the end just becomes another system or black box.
Price is the only indicator.
  • #5
  • Jul 30, 2006 12:15pm Jul 30, 2006 12:15pm
JosTheelen Thankyou

Specter Thank you for your insight

In general, myself any many professional traders I have heard would say they do not believe in a "black box" solution. I heard this several times at the forex expo from several speakers. Actually most of them said they traded the same way I learned to which is basically support and resistance, price action and confluence.

But I have heard several amazing stories about neural networks, and I feel strongly that it is worth my time to master so that I can take advantage of the market more than I would trading as normal.

For while I expect to become the best a trader could be, Trading like that all the time, no matter what, I'm not going to sit there 24/7 and have all of my capital working for me every day of the week. I think that the black box is something that will make me a lot of money, but at the same time I'll watch it and still be trading as normal

THanks
I keep a close watch on this chart of mine...
  • #6
  • Jul 30, 2006 1:37pm Jul 30, 2006 1:37pm
Coming from a US equities background, and as an avid computer programmer myself, I read up a lot on neural networks and their application to financial markets. I even started to integrate them into my own options analysis program before putting that on the back burner.

To make a long story short, I was impressed and ended up buying the flagship neural networks software for $1500 USD. Having done that, I would recommend you turn your efforts away from NNs.

What intruiged me about NNs was their supposed ability to "detect non-linear relationships" between inputs (indicators, price, volume, etc). I thought ability would allow me to trade at the very start of a new trend, without getting whipsawed by noise and minor retracements. I thought the NNs would be able to look at the "non-linear" patterns in the inputs and be able to see through these fakeouts, pointing me only to solid trades.

Turns out that even after months of programming them, I'm just now getting them to the point where they can trade about as profitably as I can on my own (note: I'm no super trader, these things just aren't that great). Once you start using them you realize pretty quickly that for a multitude of reasons, they're not yet that great for trading. Sure some people can make them perform decently well, but in the long run, I can't really think of that many serious traders who's NN's would make more money than the trader would by trading without the NN's.
  • #7
  • Jul 30, 2006 1:50pm Jul 30, 2006 1:50pm
thanks for that Loki

I would have to say your description and your recommendation doesn't suprise me because I feel the same way about black box vs human. I certainly do not plan to abandon the true discipline of trading as you suggest, and of course I expect to always trade like that.

Say however, you are going to climb mt. everest, and you don't want to bring your laptop. You want you cash working for you not just sitting in yourh account? what do you do with it? do you enter a long term trade? Any number of other things? Why not a neural network that can trade decently consistantly? Is such a thing impossible.

I suspect highly if you had a genie, and you asked him to show you the guy in the world with the best forex nn, you would want it badly.

But thx for your insight I never actually planned to focus all my efforts on I am looking into becoming an engineer and a programmer so I thought it was something that could be good.

Basically aren't NN auto systems good for when you can't trade? Aren't they better than simply throwing on a long term trade or some orders etc? I figure I want my money working for me if its a sure thing my system will make money while I am exploring deep underground caves somewhere

Interesting
I keep a close watch on this chart of mine...
  • #8
  • Jul 30, 2006 2:05pm Jul 30, 2006 2:05pm
this may be a bit off topic so feel free to skip over my comment

having an equities background as well, i gotta say that i agree with the no black box theorey,.. atleast not yet. i suppose it would be great to have an NN as an alerter, or as something to confirm what you see as well.

the one problem with black boxes is that theyre based on the past, and arent quite smart enough to apply the "emotional" quotient into future application.

atleast not until they can program panic fear and greed filters

i've done quite well on equities by exploiting the emotions of traders, and mutual fund black boxes ( which are ultimately programmed to act like an emotional trader ).

the human equation is what makes this business as fun and as harrowing as it is , and emotions play an even larger part of the forex game what with its potentially astronomical profits / losses.

having said that i must admit i wouldnt say no to a reliable, tested and fairly priced NN system
  • #9
  • Jul 30, 2006 3:10pm Jul 30, 2006 3:10pm
Quoting ManInBlack
thanks for that Loki

I would have to say your description and your recommendation doesn't suprise me because I feel the same way about black box vs human. I certainly do not plan to abandon the true discipline of trading as you suggest, and of course I expect to always trade like that.

Say however, you are going to climb mt. everest, and you don't want to bring your laptop. You want you cash working for you not just sitting in yourh account? what do you do with it? do you enter a long term trade? Any number of other things? Why not a neural network that can trade decently consistantly? Is such a thing impossible.

I suspect highly if you had a genie, and you asked him to show you the guy in the world with the best forex nn, you would want it badly.

But thx for your insight I never actually planned to focus all my efforts on I am looking into becoming an engineer and a programmer so I thought it was something that could be good.

Basically aren't NN auto systems good for when you can't trade? Aren't they better than simply throwing on a long term trade or some orders etc? I figure I want my money working for me if its a sure thing my system will make money while I am exploring deep underground caves somewhere

Interesting
I also thought that NNs would be cool when integrated with auto-trading like in Tradestation. However, from all my experience with using NNs, here are my two MAIN complaints:

1) It doesn't seem like NNs intelligently learn "non-linear" patterns in price. I've feed dozens of them almost prebuilt trading systems composed of 2-3 indicators that flutuated between 1 and -1. When the indicators cross 0, it's a position reversal, from long to short, or short to long. The closer the indicators are to 1 or -1, the stronger the uptrend/downtrend. I've backtested the systems without using NNs and if you traded according to the rule "reverse positions every time the indicator crosses 0", you'd make some serious money. The NNs botch even this system and manage to capture only about half the profit that I would have made by hand.

The problem lies with the training of the NNs. The software tells the NNs to keep training until it finds the set of rules that produces the maximum return on your account over the specified period. The NNs then adjust the trading rules until it catches a major move at the bottom and exits right at the top. You now have the maximum return on account. This is great, but now the rules are skewed so that it would have captured that one major move in the past, and you trading system flounders in the future because the NN is still looking for a major move that almost perfectly matches the one in the past.

It's hard to explain online, but hopefully you get the point. Basically it's really hard to train a NN that produces CONSISTENT winners during the walkforward testing, instead of producing a single fluke of a winner and having all the other trades be losers.

2) You can only feed 2-3 inputs into a NN. For reasons that are far beyond the scope of a simple post on the internet, once you start getting more, problem #1 becomes exponentially worse. Your NN might produce returns of 50% during the training period but only produce 3% during the walkforward testing period. Basically it falls apart under the real world because the only thing giving the NN its profit during the training was that one big win it got. The point is, that I'm one of the people who doesn't really feel you can accurately gauge the market with only 2-3 pieces of information.

I'm not one of those people who believe in swiss-army knife type charts, with dozens of indicators and fancy lines, but I do use more than 2-3 pieces of informatin (price included) expecially in Forex, in which fundamental data plays such a big role. For example, if I wanted to trade GBP/USD, I would probably look at the dollar index to gauge the overall strength of the dollar. I might look at the yield difference between various fixed income securites to determine market sentiment and potential cash flows. I'll look to see if there is any important economic data coming up for either of the countries. I might check out the price of gold since the USD seems to have a very negative correlation with it. After all this is done, then I'll look at the indicators such as MACD, CCI, Stochastics, etc.



I guess this post has kind of degenerated into a rambling criticism of NNs in general, so I'll stop it now. To get back to your original post, if you wanted/needed an automated trading system, then yes NNs will probably work. You REALLY need to combine them with some sort of money management system if you're going to let the NNs trade unattended though, employing stop orders, trailing stops, etc.

My point in this post isn't that NNs are worthless or unprofitable. I'm sure there are many people out there with NNs that will successfully out trade me 6 ways to sunday (not that it's saying that much). It is that they are by no means a black box solution, and they in NO way, shape, or form, even an approximation to "artificial intelligence". They are a slightly better method of pattern recognition, so if you feel that you see enough reliable patterns in the market to trade automatedly, sick NNs on them. If not, stay away from them because they are not an experienced trader in a box.
  • #10
  • Jul 30, 2006 3:17pm Jul 30, 2006 3:17pm
Loki, have you experimented with genetic algorithms? A artificial construct exchange with 1000 traders, with genetic algorithms, give each trader simple math correlation tools. Then feed in a live data feed to this artifcial exchange, and reproduce the survivors over generations. A form of evolution would occur.

The evolution of the math system for profitable trades might be discerned to some extant from this.
Price is the only indicator.
  • #11
  • Jul 30, 2006 3:26pm Jul 30, 2006 3:26pm
Another system would be to hook up two traders to physiological signs monitors, and have them take opposing positions in a given derivative, the true direction of price movement could be discerned from this also.

The long and short instability in position would become evident. This would be a good PhD topic or research project funded by a trading house.
Price is the only indicator.
  • #12
  • Jul 30, 2006 3:26pm Jul 30, 2006 3:26pm
loki you sound like you've had an emotional relationship with your NNs. Basically you seem to say "don't get your hopes up!" For sure we too often place our expectations on the results we will get from doing something.

So I feel like saying, hey man you're right, it is all based on your understanding of what is fact and illusion. I feel if we concentrate on what we can gleam of the truth about possibilities and the best way to take the opportunities that present themselves, we will be successful.

The problem we face is seriously that there are virtualy no limits to how you can trade. In any other job, there are only so many ways to hammer a nail. In forex there no limits to how you can do your job and it becomes more of an art than a job you do.

So for sure it's how you are looking at it. If you are trying to make the perfect system auto trader, you will probably not be happy with what it can do compared to what you could do trading normally, unless you really are a master of it, but it can be very useful.

For instance my plans for the future make trading something I can't do as much as i'd like to. So thus an autotrader will be far more profitable than not trading at all...

THanks again...

Are there really any NNs for forex for sale? not saying id buy one right now...
I keep a close watch on this chart of mine...
  • #13
  • Jul 30, 2006 3:30pm Jul 30, 2006 3:30pm
Hi Loki,

Sounds like you've done a bunch of work on this before.

Say I am looking at a 30 min Chart and would like to have 4 Trades [2 Long & 2 Short] over a Midnight to Midnight 24 Hour Period.

What happens if you just ask it to find, say, the Maximum Number of Pips that the Price Deviated from a 28 EMA before turning back toward that EMA over the Previous 24 hour Session? And added to that, that it done this Twice for the Long Side and Twice for the Short Side?

So the Short Logic would be something like...
If(Cum(Cross(28EMA+"n",Close))-2 =0,-1,0);

Is that feasible?
  • #14
  • Jul 30, 2006 7:48pm Jul 30, 2006 7:48pm
Quoting Spectre2006
Loki, have you experimented with genetic algorithms? A artificial construct exchange with 1000 traders, with genetic algorithms, give each trader simple math correlation tools. Then feed in a live data feed to this artifcial exchange, and reproduce the survivors over generations. A form of evolution would occur.

The evolution of the math system for profitable trades might be discerned to some extant from this.
I've actually programmed my own Genetic Algorithms before, but it sounds like you and I might be talking about different types. The GA's I'm used to hearing about are for optimizing trading systems and rules.

Basically the problem faced by any technical analyst, especially one using a mechanical trading system, is deciding what parameters to use in their indicators and system. For example, if you're using a mechanical trading system using a moving average crossover, the question becomes what periods do you use for the moving averages? If you use shorter periods, you will get into a profitable trade more quickly, exit closer to the optimum point, catch smaller movements a "longer" system might miss, etc. On the other hand, you'll also get whiplashed into losing trades more often, exit out of profitable ones due to slight retracements, be killed by the commisions/spread, etc. The opposite is true for longer term moving averages.

The traditional method for "solving" this dilema was brute-force optimization. You define a range of say 1-10 periods for the first moving average, 5-25 for the second moving average, 15-50 for the third, etc. and then go through every single iteration. You would back test the system with periods of 1-5-15 and see how much it made. Then you would do 1-5-16. Then 1-5-17 and on and on and on and on.

The problem with this method is that your search space grows exponentially instead of linearly with an increase number of factors (i.e. how many moving averages you use) or an increase range for a facor (i.e. search from 1-10 periods for MA #1). The hypothetical system above is simple as dirt and it requires over 7000 backtests to optimize. When you're dealing with neural networks, the systems aren't as simple, and you can literally get to a point where it would take a Cray supercomputer thousands of years to optimize the brute force way. It's basically like trying to call your friend, and dialing 111-111-1111, seeing if it's him that picks up, and dialing 111-111-1112 if it's not, then 111-111-1113.....

This is where genetic algorithms step in. They use mock evolution to combat this problem. Basically it will create a world full of random "systems". For example, it might create a 2-11-27 period crossover system, an 8-16-48 period system, etc. It then evaluates the "fitness" (net profit, percentage winning trades, largest drawdown, or other such things) of each system using a "fitness function". Once all the random systems have a fitness score, the ones with the highest score are allowed to "breed", meaning that the two of them produce another system who is theoretically more "fit" than the two parents. This process then repeats. In reality is significantly more complicated then that, as it involves artificial "mutations", repopulations, and more, but that's the basic idea. The end result is almost the same exact system that you get from the brute force method, but you arrived there in a period of time that's orders of magnitude less than if you had done the brute force method.


The bottom lines is that GA's won't produce a trading system for you like NN's will. The only function that it serves is to take an existing system, and attempt to optimize it so that it extracts the maximum possible profit from the market.
  • #15
  • Jul 30, 2006 7:57pm Jul 30, 2006 7:57pm
Quoting Loki

The bottom lines is that GA's won't produce a trading system for you like NN's will. The only function that it serves is to take an existing system, and attempt to optimize it so that it extracts the maximum possible profit from the market.
How do you compensate for a random set that catches an anomaly Spike that leapfrogs it to the top of the heap for Profit?
  • #16
  • Jul 30, 2006 8:11pm Jul 30, 2006 8:11pm
Quoting ManInBlack
loki you sound like you've had an emotional relationship with your NNs. Basically you seem to say "don't get your hopes up!" For sure we too often place our expectations on the results we will get from doing something.

So I feel like saying, hey man you're right, it is all based on your understanding of what is fact and illusion. I feel if we concentrate on what we can gleam of the truth about possibilities and the best way to take the opportunities that present themselves, we will be successful.

The problem we face is seriously that there are virtualy no limits to how you can trade. In any other job, there are only so many ways to hammer a nail. In forex there no limits to how you can do your job and it becomes more of an art than a job you do.

So for sure it's how you are looking at it. If you are trying to make the perfect system auto trader, you will probably not be happy with what it can do compared to what you could do trading normally, unless you really are a master of it, but it can be very useful.

For instance my plans for the future make trading something I can't do as much as i'd like to. So thus an autotrader will be far more profitable than not trading at all...

THanks again...

Are there really any NNs for forex for sale? not saying id buy one right now...
I think you hit the nail square on the head (pardon the pun). If you're looking for NNs as a means to have the best autotrading system you can, keeping in mind that it will not be as profitable as if you were there, you're right on target. For what they are, NNs are great. NNs are a TOOL. They're like a really good hammer. With that really good hammer you can drive nails better, drive more nails, miss the head less, etc. It would be foolish however to to look at a hammer and say "build me a great house". The problem is that NNs are often marketed in this way, (as though they are master carpenters, as opposed to just a good hammer) and so many people put their entire faith in them, and end up either a) hating NNs b) broke, or more likely c) both.

Since I knew the workings of NNs before I bought the software, I knew that they weren't a miracle system. I did underestimate how much work they are to PROFITABLY program (they are easy as hell to program, it's doing it so they make money that's the b*tch). If you spent maybe 2-3 months tinkering with them, and then integrated the NNs into a money management scheme with trailing stops, stop losses, etc, you would definatley have a profitable system. You shouldn't think that NNs are terrible or will lose you money. They just aren't as good as a human trader. In your case though it sounds like they're just what you're looking for.

As far as NN Programs for sale for forex, yes there are. I don't want to recommend them specifically because I'm still under the 50 post minimum for recomendations, but there are two major ones out there: both do Forex. Speaking of, 90% of my experience with NNs comes from my experience using them in the equities market. I have no doubt the NNs will be more profitable in the Forex market. I don't have the daytrading version of my particular software, so it only allows me to analyze on a daily time frame, hence I don't really use it.
  • #17
  • Jul 30, 2006 8:23pm Jul 30, 2006 8:23pm
Quoting Bemac
Hi Loki,

Sounds like you've done a bunch of work on this before.

Say I am looking at a 30 min Chart and would like to have 4 Trades [2 Long & 2 Short] over a Midnight to Midnight 24 Hour Period.

What happens if you just ask it to find, say, the Maximum Number of Pips that the Price Deviated from a 28 EMA before turning back toward that EMA over the Previous 24 hour Session? And added to that, that it done this Twice for the Long Side and Twice for the Short Side?

So the Short Logic would be something like...
If(Cum(Cross(28EMA+"n",Close))-2 =0,-1,0);

Is that feasible?
The first part of your system is really easy to implement. What would happen is that you would feed one an input akin to this:

Sub(Close, 28EMA)

Which is basically the difference between the close and the ema. You can then also feed it a linear regression of that indicator, or a MACD of it should you wish.

I'll get back to you on the 2 long/2 short per day thing because to tell you the truth I'd have to tinker with my software first.
  • #18
  • Jul 30, 2006 8:52pm Jul 30, 2006 8:52pm
OK, I think I figured out the 2 long, 2 short trade thing somewhat. You can tinker with the settings and basically force the maximum trade span (the time the position is held) to X number of bars. You could set it to 6 hours (one fourth of 24 hours), but this obviously isn't optimal because it means that you would only have a 6 hour window from the time you bought or shorted the pair, until when the position had to be closed out, regardless of if its still in the uptrend/downtrend.
  • #19
  • Jul 30, 2006 10:05pm Jul 30, 2006 10:05pm
a bit off topic but can anyone describe confluence to me, or point me to some articles or discussions about it? thanks.

Quoting ManInBlack
In general, myself any many professional traders I have heard would say they do not believe in a "black box" solution. I heard this several times at the forex expo from several speakers. Actually most of them said they traded the same way I learned to which is basically support and resistance, price action and confluence.
THanks
  • #20
  • Jul 30, 2006 10:15pm Jul 30, 2006 10:15pm
Quoting ManInBlack
I know a lot of people do this and don't ever talk about it because everyone would want their program on their computer hooked up to all of the money they had.

If you do know about it, could you please recommened some books, some ideas, some software that is already out there I might get ahold of.

How do I go about mastering neural trading.
MIB, you should see this post on another forum.

http://www.forex-tsd.com/general-dis...algorithm.html

I am user silverpike, I posted several replies.

If you have any other questions about my experience with NNs, ask away. Most users there didn't have the knowledge to deal with my posts.
0 traders viewing now