Caution: Math Geek at Work | Page 12 | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

Murrey Math indicator is very useful in Forex 69 replies

Murray Math Levels 30 replies

Murray Math 13 replies

Turbo's Murry Math 23 replies

Attachments: Caution: Math Geek at Work
Exit Attachments

Caution: Math Geek at Work

  • #221
  • Jun 11, 2008 4:26am Jun 11, 2008 4:26am
Quoting redpine
Scott,

I read in one of your posts, that you'd need a genetic program to further solve a problem.

Dan
Hey, I missed that. Though until now I have not achieved success researching the stock market using a neural net, I have not given up on automated trading yet. But the programming language in this field now I think must be MQL4. I don't know if Scott is still interested. He said recently that an excel spreadsheet is enough for him. As time allows, I wil do something. As the terminator said: "I'll be back!". Good trading.
  • #222
  • Jun 11, 2008 8:51am Jun 11, 2008 8:51am
Quoting redpine
Scott,

I really find your analysis insightful. I started at your web site which led me to this forum. I was running out of ideas and now I'm full of them again.

Anyway, your approach to pattern analysis is very good. I read in one of your posts, that you'd need a genetic program to further solve a problem. Well, I have a full blown simulation of 1 years worth of data covering 21 pairs, and it just happens, I've been running a genetic program looking for patterns in candle sticks.

Have any ideas you want analyzed with my GP simulation?

Dan
Welcome Redpine!

At this point I've settled on a simple "position management" trading approach (as opposed to a predictive one) which is pretty comfortable for me, so I've actually stopped doing research into price behavior for now.

However, I still have a longer term interest in GP, and hope to one day try the technique in a full project. Let me clarify what I mean by Genetic Programming though, as it may be different from what you're doing.

GP was originally conceived by a guy named Koza, and is a more generalized form of the idea of Genetic Algorithms. The idea is that you have an initial "population" of programs, each composed of modules that can be connected together in various ways (like genes). At first, these programs wouldn't be very effective. Then these programs compete with each other, and the most successful ones get to combine their structures and "reproduce" a set of new programs for the next generation. So after many generations, the theory is that you get a highly evolved population of programs which are very effective. We humans might not even understand the structure of the most successful programs or how and why they work.

But anyway, the difference I see is that the trading programs which are created through this evolution process would take known recent price behavior as inputs and specific trades (direction, target and stop) as outputs. They wouldn't be limited to just looking for patterns; they would be actual "traders."

Is this the kind of thing you're thinking of? If so, I know a person on FF who's doing this kind of work (at least he was). I'll see if I can find the thread again.

EDIT: Sorry, the person I was thinking of is doing work with neural nets, which is a completely different approach from GP.
  • #223
  • Jun 11, 2008 9:19am Jun 11, 2008 9:19am
Here's another example of "filling in the gap" when one of my orders gets filled, and then price turns around and continues in the direction it was going causing me to place even more orders behind it and leaving a gap.

In this case, I had 7 sell orders on EUR/JPY lined up behind the price as it moved up. Then the 7-unit order got filled. But price then continued upward, allowing me room to place an 8-unit and 9-unit order. This created a gap in the open order sequence: 1,2,3,4,5,6,missing,8,9. So given the fact that I've committed to the risk of selling this many new units, and in the spirit of always trying to get a better price for my trades, I got rid of the first three orders and replaced them with a new 7-unit order to fill the gap. This gets me a better price for those units when the order fills.

Here's the chart:

Attached Image (click to enlarge)
Click to Enlarge Name: position.JPG Size: 85 KB
  • #224
  • Jun 11, 2008 2:31pm Jun 11, 2008 2:31pm
Quoting capitalist88
At this point I've settled on a simple "position management" trading approach (as opposed to a predictive one)
You should search the net for 'DrZogg' and/or 'GrimFD', he had the same idea.
The breaking of a wave cannot explain the whole sea.
  • #225
  • Jun 11, 2008 4:49pm Jun 11, 2008 4:49pm
Quoting Craig
You should search the net for 'DrZogg' and/or 'GrimFD', he had the same idea.
I found a thread where people were trying to guess what he was doing, so it looks like he never disclosed it. Then I found a site where he says all his accounts got margin called.

http://www.houseofforex.com/index.htm

Sounds pretty dramatic and mysterious, lol. I'm more boring and hopefully more clear (or not).
  • #226
  • Jun 11, 2008 5:23pm Jun 11, 2008 5:23pm
Let's hope so, he made some fairly bold statements and then it all fell flat, it was at least entertaining.
The breaking of a wave cannot explain the whole sea.
  • #227
  • Jun 11, 2008 5:36pm Jun 11, 2008 5:36pm
Quoting capitalist88
Welcome Redpine!

At this point I've settled on a simple "position management" trading approach (as opposed to a predictive one) which is pretty comfortable for me, so I've actually stopped doing research into price behavior for now.

However, I still have a longer term interest in GP, and hope to one day try the technique in a full project. Let me clarify what I mean by Genetic Programming though, as it may be different from what you're doing.

GP was originally conceived by a guy named Koza, and is a more generalized form of the idea of Genetic Algorithms. The idea is that you have an initial "population" of programs, each composed of modules that can be connected together in various ways (like genes). At first, these programs wouldn't be very effective. Then these programs compete with each other, and the most successful ones get to combine their structures and "reproduce" a set of new programs for the next generation. So after many generations, the theory is that you get a highly evolved population of programs which are very effective. We humans might not even understand the structure of the most successful programs or how and why they work.

But anyway, the difference I see is that the trading programs which are created through this evolution process would take known recent price behavior as inputs and specific trades (direction, target and stop) as outputs. They wouldn't be limited to just looking for patterns; they would be actual "traders."

Is this the kind of thing you're thinking of? If so, I know a person on FF who's doing this kind of work (at least he was). I'll see if I can find the thread again.

EDIT: Sorry, the person I was thinking of is doing work with neural nets, which is a completely different approach from GP.
That is the kind of GP I'm doing. I've tested with the GP evolving tech indicators and candle patterns. The Candle patterns are based currently on the last 5 bars, include high, low, close, open and a charaterization of each, such as upper shadow, real body, etc.. I ran this across 100,000 bars of 5 minute data for 21 pairs and got pretty good results (60%+). However, I haven't really pursued it much more. I read your last paper on predicting Zig Zags (real out of the box thinking). It got me thinking that maybe if I structured my tests a little different, I might do better. Rather than using a fitness sample of profit, maybe accuracy of picking zig zag reversals...

A trader approach sounds interesting. I'm not sure how to use it, wonder...
I was looking at atomic TA indicators, in my latest testing, but had early worked with more interpretive type indicators, like X MACD up, or Close above EMA. These were used as part of thing with the product I use (Beagle), called CoEvolution, which uses Genetic Algorithms and Genitic programs. So you might use the GA to optimize the number of bars and the GP, to find the algorithm.

Dan
.
  • #228
  • Jun 11, 2008 10:54pm Jun 11, 2008 10:54pm
Quoting redpine
That is the kind of GP I'm doing. I've tested with the GP evolving tech indicators and candle patterns. The Candle patterns are based currently on the last 5 bars, include high, low, close, open and a charaterization of each, such as upper shadow, real body, etc.. I ran this across 100,000 bars of 5 minute data for 21 pairs and got pretty good results (60%+). However, I haven't really pursued it much more. I read your last paper on predicting Zig Zags (real out of the box thinking). It got me thinking that maybe if I structured my tests a little different, I might do better. Rather than using a fitness sample of profit, maybe accuracy of picking zig zag reversals...

A trader approach sounds interesting. I'm not sure how to use it, wonder...
I was looking at atomic TA indicators, in my latest testing, but had early worked with more interpretive type indicators, like X MACD up, or Close above EMA. These were used as part of thing with the product I use (Beagle), called CoEvolution, which uses Genetic Algorithms and Genitic programs. So you might use the GA to optimize the number of bars and the GP, to find the algorithm.

Dan
I think using the profit as a fitness criterion as you've been doing makes the most sense. After all, that's the bottom line. I'm not familiar with any commercial GP products; if I ever do a GP project it will be from scratch using a recursive programming language like LisP.

What I mean by each program in the population being a trader is that it would receive the raw data on the candles as input first. So in your case it would receive 20 values; the open, high, low and close for 5 candles. Then it would run through its genetically evolved algorithm on those values (using some string of operations on the variables; addition, subtraction, multiplication, divison, logical comparisons between higher/lower, larger/smaller, rising/falling etc.). The output would consist of 3 values; an action (buy sell or hold), a stop distance and a TP distance.

In essence, each program in the population would do what a trader does. They would compete, using profit as a fitness criterion, and the winners would produce genetic offspring for the next generation of "traders."
  • #229
  • Jun 12, 2008 4:39am Jun 12, 2008 4:39am
Quoting capitalist88
I think using the profit as a fitness criterion as you've been doing makes the most sense. After all, that's the bottom line. I'm not familiar with any commercial GP products; if I ever do a GP project it will be from scratch using a recursive programming language like LisP.

What I mean by each program in the population being a trader is that it would receive the raw data on the candles as input first. So in your case it would receive 20 values; the open, high, low and close for 5 candles. Then it would run through its genetically evolved algorithm on those values (using some string of operations on the variables; addition, subtraction, multiplication, divison, logical comparisons between higher/lower, larger/smaller, rising/falling etc.). The output would consist of 3 values; an action (buy sell or hold), a stop distance and a TP distance.

In essence, each program in the population would do what a trader does. They would compete, using profit as a fitness criterion, and the winners would produce genetic offspring for the next generation of "traders."
Been there done that. Raw data doesn't seem to do much good. There are an infinate number of combination with math operators and 20 numbers. If I give it a little help it does better. Below is the output for a candle pattern test on EUR/USD with a 50 pip fixed SL/TP. It produces 2462 pips over a one year period with a 63% win/loss. HighLow is simply the candles High - Low values.

if(AND(EQ(HighLow5(), HighLow3()), EQ(HighLow2(), HighLow3()))

I've mucked around with trying to add multiple choices in it's logic, such as long/short and usually it short circuits the algorithm to only ever make one decision. I could probably force it to do that though. Another approach is to use Coevolution, where the different decisions are evolved in seperate threads. This takes the best of one thread and uses it to evolve the other thread, and vice versa. I've only played with that option and gotten mixed results.

Beagle is free source and all written in c++. The GA is the best I've seen, but unfortunately, GA is an optimization technique and this GA curve fits real well, so forward testing results are usually bad. I often see it optimize to 80% win/loss with forward testing reduced to a coin flip.

Dan
.
  • #230
  • Jun 12, 2008 8:44pm Jun 12, 2008 8:44pm
Quoting redpine
Been there done that. Raw data doesn't seem to do much good. There are an infinate number of combination with math operators and 20 numbers. If I give it a little help it does better. Below is the output for a candle pattern test on EUR/USD with a 50 pip fixed SL/TP. It produces 2462 pips over a one year period with a 63% win/loss. HighLow is simply the candles High - Low values.

if(AND(EQ(HighLow5(), HighLow3()), EQ(HighLow2(), HighLow3()))

I've mucked around with trying to add multiple choices in it's logic, such as long/short and usually it short circuits the algorithm to only ever make one decision. I could probably force it to do that though. Another approach is to use Coevolution, where the different decisions are evolved in seperate threads. This takes the best of one thread and uses it to evolve the other thread, and vice versa. I've only played with that option and gotten mixed results.

Beagle is free source and all written in c++. The GA is the best I've seen, but unfortunately, GA is an optimization technique and this GA curve fits real well, so forward testing results are usually bad. I often see it optimize to 80% win/loss with forward testing reduced to a coin flip.

Dan
Hmmm, a good illustration of why I try not to mess with ready made software, "wizards", etc. It seems like you're being constrained by the features they put in the program.

When you tried to use just the 20 values with a set of arithmetical/logical operators, what was the best performing program that evolved? Was it a better performer than the best program in the initial population? And how many generations did it take to achieve that improvement?

Remember, it took millions of generations for US to evolve. Maybe the number of generations is simply too low.
  • #231
  • Jun 13, 2008 4:50am Jun 13, 2008 4:50am
Quoting capitalist88
Hmmm, a good illustration of why I try not to mess with ready made software, "wizards", etc. It seems like you're being constrained by the features they put in the program.

When you tried to use just the 20 values with a set of arithmetical/logical operators, what was the best performing program that evolved? Was it a better performer than the best program in the initial population? And how many generations did it take to achieve that improvement?

Remember, it took millions of generations for US to evolve. Maybe the number of generations is simply too low.
The above mentioned algorithm was produced for EUR/USD. I can post the rest for each pair if you like. As with any testing the results are only as good as the experiment. With the genetic program, I'm using it has settable limits on the number of nodes and tree depth of the alogrithm. It also has a capablity (I think) to kill off part of the population, which is an approach to try when it get stuck in hole in the solution space. Neural Nets have a similar problem where they just can't improve any more on the fitness of the solution... down another rat hole. Anyway the point is that designing a good experiment to address limitation of the tool and to produce a good measure of success is important. Setting the fitness just on pure profit can be a problem. Quite often, if raw profit is used, you'll get a system that will make 5,000 pips a month, but trade 10,000 times. Obviously half a pip trade is not gonna be profitable. So in that instance, you need to include the cost of trade in your calculaiton of profit. I usually penalize each trade 3 pips. Other measures can produce even stranger results, like max profit per trade, usually will end up with one trade. This list of issues for an experiment are endless.

The evolution works real good, always the results in the end are better than the start, with one exception, CoEvolution. With CoEvolution, if you try to evolve entries and exits in seperate threads, the results bounce all over the place. Not sure, why, I've pretty much dropped that strategy for now, and focus on a single thread of evolution for an entry.

I've found the best results are to evole 4 populations of 100. The cross over between poplutions, really moves things along. Most evolution stops improving around 50 generations.

Dan
.
  • #232
  • Jun 13, 2008 6:22am Jun 13, 2008 6:22am
One other idea I wanted to run by you is the correlation of pairs. I've been looking into this for some time and have a technique that seems to work, but could probably be improved on.

When I get a signal, that it is time to trade a pair based upon a regression slope, I run a correlation against all pairs and take a vote, if the majority of highly correlated pairs vote with their regression lines in the same direction, I take the trade. Correlations are taken on 50 one hour bar close data. Highly correlated pairs are defined as > .8. I also look for highly negative correlations and an oppisite direction. This gives me a 67% win rate and in demo testing I've had positive results.

1. I was surprised to find that there was that much correlation on hourly data.
2. Why is my win rate only 67%? I would expect it to be better, but I guess it is really only a confirmation that pairs are in the same state at that time. Would be nice if this was more a predictor.
3. As was suggest by a collaborator on this system, I shifted the hour bars for the other pairs by 5 hours and took their next three hours as a predictor.
a. I was surprise that there were any highly correlated pairs, but even
more that there were about as many.
b. The win/loss was around 55%.

Any opinions or ideas?

Thanks,

Dan
.
  • #233
  • Jun 13, 2008 2:32pm Jun 13, 2008 2:32pm
Quoting redpine
(...)which is an approach to try when it get stuck in hole in the solution space. Neural Nets have a similar problem where they just can't improve any more on the fitness of the solution... down another rat hole. (...)
The solution space is a canyon the size of a galaxy... . I suspended my research on neural nets due to lack of theoretical basis (you and the thread author have that in excess ) and a thought invaded me: if I had not devised a profitable system the "random walk" of the NN in the canyon could not do that for me. There is always a layer of error that cannot be reduced without forcing the NN into "madness" (you can reduce error arbitrarily to zero - yes, zero! But then the system becomes extremely dependent on initialization values and the forecasts become random, different at each experiment and an absolute nonsense). Sorry for my bad expression. English is not my natural language. I must have left some BS on this and previous posts but I can't see where they are... my bad. Good trading to all.
  • #234
  • Jun 13, 2008 7:05pm Jun 13, 2008 7:05pm
Quoting Jairo
The solution space is a canyon the size of a galaxy... . I suspended my research on neural nets due to lack of theoretical basis (you and the thread author have that in excess ) and a thought invaded me: if I had not devised a profitable system the "random walk" of the NN in the canyon could not do that for me. There is always a layer of error that cannot be reduced without forcing the NN into "madness" (you can reduce error arbitrarily to zero - yes, zero! But then the system becomes extremely dependent on initialization values and the forecasts become random, different at each experiment and an absolute nonsense). Sorry for my bad expression. English is not my natural language. I must have left some BS on this and previous posts but I can't see where they are... my bad. Good trading to all.
There are a number of traders who claim the NN gives them an edge. Working in the Software industry, all my AI savy coherts say they can't see how. And I can tell you I found no success. There are a couple of approaches highly recommended by these same people, Genetic Programming and NN with reinforcement learning. I haven't tried the reinforcement learning yet, cause the code is very difficult to follow.

If anyone want's to look around at NN and NN with reinforcement learning, I highly recommend FANN, just cause it's free and there are some really smart helpful people there including the author.

Dan
.
  • #235
  • Jun 13, 2008 7:46pm Jun 13, 2008 7:46pm
Quoting redpine
There are a number of traders who claim the NN gives them an edge. Working in the Software industry, all my AI savy coherts say they can't see how. And I can tell you I found no success. There are a couple of approaches highly recommended by these same people, Genetic Programming and NN with reinforcement learning. I haven't tried the reinforcement learning yet, cause the code is very difficult to follow.

If anyone want's to look around at NN and NN with reinforcement learning, I highly recommend FANN, just cause it's free and there are some really smart helpful people there including the author.

Dan
Great stuff there (FANN)! Thank you for pointing it. I believe now that a neural net needs a "master" to be effectively trained to trade. Like a child, you must point the right direction. The universe of possibilites in the market is immense and always changing. The NN will not be able to find the way, unless an experienced human trader with a profitable strategy gives a hand. After getting there, the net will have to "rethink the market" from time do time to account for the changes. I am planning to restart my research by a simple scalping strategy with a known track record. Then, if necessary, build gradually around, step by step.
  • #236
  • Jun 13, 2008 8:45pm Jun 13, 2008 8:45pm
Quoting redpine
One other idea I wanted to run by you is the correlation of pairs. I've been looking into this for some time and have a technique that seems to work, but could probably be improved on.

When I get a signal, that it is time to trade a pair based upon a regression slope, I run a correlation against all pairs and take a vote, if the majority of highly correlated pairs vote with their regression lines in the same direction, I take the trade. Correlations are taken on 50 one hour bar close data. Highly correlated pairs are defined as > .8. I also look for highly negative correlations and an oppisite direction. This gives me a 67% win rate and in demo testing I've had positive results.

1. I was surprised to find that there was that much correlation on hourly data.
2. Why is my win rate only 67%? I would expect it to be better, but I guess it is really only a confirmation that pairs are in the same state at that time. Would be nice if this was more a predictor.
3. As was suggest by a collaborator on this system, I shifted the hour bars for the other pairs by 5 hours and took their next three hours as a predictor.
a. I was surprise that there were any highly correlated pairs, but even
more that there were about as many.
b. The win/loss was around 55%.

Any opinions or ideas?

Thanks,

Dan

Yes, this is just a shot in the dark, but have you looked at a plot of the correlations over time? I'm thinking that they must vary quite a bit. So a fifty-bar correlation of say .92 between two pairs might be a correlation of .60 during a different 50-bar period. I don't know if this is the case, but if it is, then it suggests that the correlation itself is partially a result of randomness. This may be why you got lower win rates than you thought. There may be less explanatory power in the correlation; i.e. "correlation does not imply causation."
  • #237
  • Jun 14, 2008 9:37am Jun 14, 2008 9:37am
Quoting capitalist88
Yes, this is just a shot in the dark, but have you looked at a plot of the correlations over time? I'm thinking that they must vary quite a bit. So a fifty-bar correlation of say .92 between two pairs might be a correlation of .60 during a different 50-bar period. I don't know if this is the case, but if it is, then it suggests that the correlation itself is partially a result of randomness. This may be why you got lower win rates than you thought. There may be less explanatory power in the correlation; i.e. "correlation does not imply causation."

Interesting comments. My assumption going into this was that the correlations between pairs were transitory on the scale I chose. For example if you look at USD/JPY and GBP/JPY over the one year period the trends are very similar, with divergence at various points, but the general form of the trend is the same. My assumption is that during these non correlated times, is when systems, base upon indicators/patterns break down, mainly because the to correlated markets are out of balance and they are not well behaved.

In my system I compare a target pair with all other pairs looking for short term correlations. Maybe I should only compair pairs on the short term that have long term correlations????

Dan
.
  • #238
  • Jun 14, 2008 9:41am Jun 14, 2008 9:41am
Quoting Jairo
Great stuff there (FANN)! Thank you for pointing it. I believe now that a neural net needs a "master" to be effectively trained to trade. Like a child, you must point the right direction. The universe of possibilites in the market is immense and always changing. The NN will not be able to find the way, unless an experienced human trader with a profitable strategy gives a hand. After getting there, the net will have to "rethink the market" from time do time to account for the changes. I am planning to restart my research by a simple scalping strategy with a known track record. Then, if necessary, build gradually around, step by step.
If I was working this problem full time, I'd be looking into an AI predicting direction based upon parsing news feeds for predictors of price movements. There was a very promising paper written on this approach from the Penn/Lehman project. But, I have to earn a living, so just scalping is all I have time for.

Dan
.
  • #239
  • Jun 14, 2008 9:25pm Jun 14, 2008 9:25pm
Quoting redpine
In my system I compare a target pair with all other pairs looking for short term correlations. Maybe I should only compair pairs on the short term that have long term correlations????

Dan
No, I think mixing the two time frames just confuses the issue, so let's keep it as simple as possible. The question is essentially whether the short term 50-hour correlations are meaningful or just the result of randomness. In order to do this, you need to compare several independent periods to each other.

The tool I would pick is the coefficient of variance (COV). This is just the standard deviation of a set of data divided by its mean, so it gives you a dimensionless measure of how "consistent" the data points are. You could take correlations from 10 periods.

The ideal case is that all the correlations are high, and tend to hover right around the same values. So suppose you get results where the mean is very high, say around 90% and the STDEV is low, like around 10%. This would give you a COV of 0.11. I would say any COV of 0.10 or less would indicate that there is SOMETHING (no idea what) non-random which is causing a CONSISTENTLY high correlation.

Now I don't know which would work better for you; to take several periods of 50 hours each, or to break up one 50 hour period into several smaller periods.

Either way, you can use COV to test to see if the correlations you're seeing are just the result of randomness or not. This is a very basic "back of the envelope" kind of test, and a real statistician would probably laugh at me, but it's quick. One thing to remember is that in order for COV to give meaningful results, ALL OF THE DATA POINTS MUST HAVE THE SAME SIGN.

Hope this helps a bit.
  • #240
  • Jun 14, 2008 9:27pm Jun 14, 2008 9:27pm
Bleh. These are the periods that really test a trader's confidence in their systems. Down another 4.7% last week.
2 traders viewing now