Menu
Sister Sites

Options

Bookmark Thread

First Page First Unread Last Page Last Post

Print Thread

Similar Threads

Top 100 best Trading Systems: post good Trading Systems here 51 replies

I Will Learn Coding in 1 Month - Give me the 80/20 of MT4 Coding 14 replies

Reality of Automated Systems Coding 3 replies

Systems, Systems and Systems... 1 reply

Seriously, do forex systems have to be so complex? 11 replies

Attachments: Overly Complex Coding and Trading Systems.
Exit Attachments

Overly Complex Coding and Trading Systems.

  • Post #1
  • Quote
  • First Post: Edited Mar 24, 2006 2:07pm Mar 22, 2006 11:46am | Edited Mar 24, 2006 2:07pm
This thread has been created to discuss Overly Complex Coding and Trading Systems with one goal in mind. Performance.

Everything, from the tricked out custom software to the assistance of spiritual beings can apparantly help you trade. But will it help you trade EVERY TIME?

===============================



my trading system, simple or complex, has to maximize...

Win %
Average win/loss
Position Size

And minimize..

drawdowns
stress
discretion

and take into account anything that can affect it's tested probabilities (news etc...)

If It meets those goals, I have every traders dream. So I am eager to learn everything that is possible in trading
=====================================

I encourage everyone with any inclination to contribute anything they like. I encourage you to contact me personally if you with to work together for the common betterment of each others and the communities trading.

When I am satisfied with the results of the discussion, I will be creating a free ebook and other info on my free website. This is, (like it has been said by another), to help make YOU money. I guess if you didn't want your work in the book, you wouldn't publish it here anyway...

Since I have nothing that I have forward tested (my simple system works, thanks rob) , I open the floor to the rest of my esteemed colleagues here at FF...
I keep a close watch on this chart of mine...
Quoting ManInBlack

my trading system, simple or complex, has to maximize...

Win %
Average win/loss
Position Size

And minimize..

drawdowns
stress
discretion
So how do you measure these things? Is a system with a 15% return with a 1% drawdown better than a system with a 100% return with a 20% drawdown?

What do you do for in sample versus out of sample testing?

How much data do you use?

What program do you use for your testing?

How do you account for spread in your testing?

Do you optimise? What metrics do you use for your optimising?

There is much to consider before you even put your first indicator on the chart.

Simon
Quoting ManInBlack
This thread has been created to discuss Overly Complex Coding and Trading Systems
Oh. Sorry. No contribution to this one......
The framework is there; we just have to work on our frame of mind.
Quoting DrRock
So how do you measure these things? Is a system with a 15% return with a 1% drawdown better than a system with a 100% return with a 20% drawdown?

What do you do for in sample versus out of sample testing?

How much data do you use?

What program do you use for your testing?

How do you account for spread in your testing?

Do you optimise? What metrics do you use for your optimising?

There is much to consider before you even put your first indicator on the chart.

Simon
I hope you don't mind ManinBlack if I answer a few of these.

Hello DrRock

I've enjoyed your posts in the past. So let me know if this makes sense to you.

Drawdown and performance, this would depend on how much someone has in their account. The more you have the more you're willing to risk for higher return. But even with a small account proper money management and diversification will definitely be the key here.

A system that has a large drawdown and a big return could still be traded. If you manage your money properly and you have other systems running simultaneously in a portfolio to minimize the effects of the drawdown.
And let's not fool ourselves that largest drawdown will happen again. Murphy's law works.

Sampling data would also depend on how many trades your system takes I like to see a few hundred trades in the out of sampled data.

The Software that I will be using if I post coding and such is TradeStations but since the title of this discussion is complex coding well. I figure people would be able to translate it to whatever software. They're using.

Optimization, a great tool if you use it properly as far as the matrix that I use. If the system will only be traded on one currency I like using return on account. If the system would be traded on multiple currencies I like using profit factor. These two matrix also tend to minimize the drawdown.

Of course proper optimization is a big issue so I will start with some sophisticated software

The picture below is a 3-D graph of an optimization run the system looked mighty promising when you looked at the report after optimizing but a quick look at the graph shows different.Keep in mind a robust system would have a flat area where parameters could be changed without much change to the net profitor or end results.
Attached Image
A truly wise man, always has more questions than answers.
Quoting Claude

The picture [edit: above end edit:]is a 3-D graph of an optimization run the system looked mighty promising when you looked at the report after optimizing but a quick look at the graph shows different.
Clearly my eyes are getting old. I had to take a long look to see it.

Can I ask what software you are using to generate the graph?

Thanks Claude. I am really hoping you get your teeth into this thread.
How does your garph work? Claude?
Quoting Walker
How does your garph work? Claude?
Hello Bemac and Wallker

The graft is produced by a software called 3-D view by rina systems.
TradeStations puts out a report of all the optimization runs. Then I make a text file from this and feed it to the software. It takes the thousand best runs and creates the graph.

Tomorrow if I get a chance I'd like to talk a little about neural networks since it seems to have become more acceptable and mainstream,especially since some members here are using vantage point.
A truly wise man, always has more questions than answers.
very interesting stuff claude, look forward to more

Dr. Rock, I know basically nothing about what we discussing here. My trading is all about "traditional" and im new to trading in general.

Claude... what would you recommend to a trader who wants to master the kind of stuff you do? Is there books, websites. gurus, mantras that can help me become a wizard?

For now, I will be googling some of those topics

rock on
I keep a close watch on this chart of mine...
Cool thread MIB, I think I'll share some of the stuff I've been doing. Maybe this should be more about "cutting edge" and less about "overly complex." I dunno

I'm convinced the markets are wavelike, and if you use any kind of oscillator, you agree. One thing I mentioned before is square wave patterns. When the market is "flat" or "range bound", it's really building up energy in the form of a widening frequency spectrum.

To start, export a chart's data from MetaTrader. Just have a chart open, the do file->save as, and it'll save it as a .csv file. You can open this file in Excel or OpenOffice.

This file has a bunch of columns. Delete all but the closing price to get one column of data that looks exactly like what a signal analysis tool is looking for.

I tried this with some shareware which you can get at SigView.com. After you install and run SigView, you can open a text document under the File menu. It will ask you for a sample rate. I like 1000 but it doesn't matter.

Now you'll see your chart, and it should look the same as a line chart of the data you exported from MetaTrader. From there, you can manipulate the data in some really cool ways.

I like looking at:

The frequency spectrum, which shows the peak frequencies that contribute to the sample. This helps expose some of the minor "hidden" oscillations in the data.

The Time FFT, which does the above in little slices over time. This can show the expansion and contraction of the spectrum. As a rule of thumb, when the spectrum is wide, there is a lot of pent-up energy in the system.

Filters. You can filter out the overall trend, as well as a lot of the random noise. If done correctly (and I can't begin to tell you how to do this correctly), it should be possible to get a feel for how many of the "important" frequencies are pointing up or down.

So if you're into this kind of stuff, give it a try. Maybe you'll spot some patterns that help explain some mysterious market move. If you're looking for a system that already works, this ain't it
  • Post #10
  • Quote
  • Edited Mar 25, 2006 3:43pm Mar 24, 2006 11:19pm | Edited Mar 25, 2006 3:43pm
"Trading systems range in complexity from the simple to the inscrutable." Joe Krutsinger

A long time has passed since the first neural model by McCulloch and Pitts (1943). But now, with so many home having a computer. These artificial neural networks. can be used by just about everyone.

They are are nonlinear models, loosely based on how the human brain is structured.

I will try to skip a lot of details because I do not want to over complicate things. Just the basics.

Our simple artificial neural network will have tow inputs these inputs can be anything. They will first go into two neurons each small bit of data will be giving a weight{kind of like a level of importance} from their they will go to three other neurons, which are interconnected that information will be processed . This information will then be sent to the output neuron the information will be processed and more weights will be added. Then all these weights{ these levels of importance}. Will be added up.

Based on this information. We will either buy or sell, the British pound since this will be running in a brute force optimizer. It will return that it has just lost or made x amount of dollars , the neural network will have learnt that the solution that it just came up with was garbage or good and from there. It will try another solution.

The process will repeat till it comes up with the best solution to this problem.

This neural network is one of the simplest. It is only a foreword learning, meaning the information only goes one way from input to output and then it tries again, to see if it can come up with a better solution.

A slightly more complex neural network uses back propagation meaning information goes forward and backwards the neurons would try to come up with the solution. The error rate would be calculated. The smaller the error, the closer it is to solving the problem. This information would be passed back to the neurons. So they understand that they are headed in the wrong direction or the correct one. ((We might look at this one as well))

Weight can be positive or negative. So that it doesn't take too much time. We only use -1 ,0 or 1 since we have nine areas where weight can be added and three different weights can be put in at those points. This gives us 19, 683 Different combinations. For an even more refined solution. We could allow it to go from -10 to 10, and so on, but this would be very time-consuming in TradeStations.

Keep this as simple as possible. The only two inputs, we will give it is a simple moving average of three as compared to itself. Three periods ago.

The second input will be the simple moving average of three compared to itself five periods ago.

Remember these inputs could be anything, examples: the closing price of gold. The RSI, ADX a different currency, and so on.

The only difference here is that as a system designer. We do not have to explain to it buy if the moving average is greater than one period ago and lower five periods ago at slightly higher than two periods ago. So on and so on. This will all be figured out by the neural network. And the level of importance for each one will also be figured out.

Function for TradeStations name: Htangent

Input: x (NumericSimple), { input to function }

NTerms (NumericSimple); { # terms in series }

Var: pi (3.1415926536),

Sum (0),

ii (0);

Sum = 0.;

For ii = 0 to NTerms Begin

Sum = Sum + 1./(Power(((ii + 0.5)*pi), 2) + Power(x, 2));

End;

Htangent = Sum * 2 * x;

_____________________________________________________________________________

Signal for TradeStations name: Neural net simple

Inputs: synapse1 (0), synapse2 (0),synapse3(0),synapseA1 (0),

synapseA2 (0),synapseA3 (0),synapseB1(0),synapseB2 (0),

synapseB3 (0);

Var: inputneuron1 (0),

inputneuron2 (0),

Hiddenneuron1 (0),

Hiddenneuron2 (0),

Hiddenneuron3 (0),

neuronOut (0);



{ inputs }

if (average(C,3) - average(C,3)[2]) >0 Then inputneuron1= 1 Else inputneuron1 = -1;

if (average(C,3) - average(C,3)[5]) >0 Then inputneuron2= 1 Else inputneuron2 = -1;



{neural net }

Hiddenneuron1 = Htangent(synapse1*inputneuron1 + synapseA1*inputneuron2, 50);

Hiddenneuron2 = Htangent(synapse2*inputneuron1 + synapseA2*inputneuron2, 50);

Hiddenneuron3 = Htangent(synapse3*inputneuron1 + synapseA3*inputneuron2, 50);

neuronOut = Htangent(synapseB1*Hiddenneuron1 + synapseB2*Hiddenneuron2 + synapseB3*Hiddenneuron3, 50);



{ buy or sell }

If neuronOut >= 0.5 then

Buy next bar at High stop;

If neuronOut <= -0.5 then

Sell next bar at low stop;

_________________________________________________________

Neural net are no longer expensive you can have them running in Excel for about 70 to 150 bucks simply enter all your data in Excel sheet and let it go to town. I just thought it be neat if you could see the coding process that lays behind.

Perhaps next time, we can take a look at the more complex neural net, where the error rate is calculated with back propagation. And instead of it figuring out how to use something in real-time, we will get it to forecast the closing price for the next few days.
Attached Image
A truly wise man, always has more questions than answers.
all I can say about this thread is wtf........
Quoting Mike Jolley
Cool thread MIB, I think I'll share some of the stuff I've been doing. Maybe this should be more about "cutting edge" and less about "overly complex." I dunno

I'm convinced the markets are wavelike, and if you use any kind of oscillator, you agree. One thing I mentioned before is square wave patterns. When the market is "flat" or "range bound", it's really building up energy in the form of a widening frequency spectrum.

tried this with some shareware which you can get at SigView.com. After you install and run SigView, you can open a text document under the File menu. It will ask you for a sample rate. I like 1000 but it doesn't matter.

Now you'll see your chart, and it should look the same as a line chart of the data you exported from MetaTrader. From there, you can manipulate the data in some really cool ways.

Filters. You can filter out the overall trend, as well as a lot of the random noise. If done correctly (and I can't begin to tell you how to do this correctly), it should be possible to get a feel for how many of the "important" frequencies are pointing up or down.

hello Mike Jolley

Once the data has been manipulated can you produce a XL file to export the new filtered data into another application.
A truly wise man, always has more questions than answers.
Quoting Claude

Function for TradeStations name: Htangent

Input: x (NumericSimple), { input to function }
NTerms (NumericSimple); { # terms in series }
Var: pi (3.1415926536),
Sum (0),
ii (0);
Sum = 0.;
For ii = 0 to NTerms Begin
Sum = Sum + 1./(Power(((ii + 0.5)*pi), 2) + Power(x, 2));
End;

Htangent = Sum * 2 * x;

_____________________________________________________________________________

Signal for TradeStations name: Neural net simple

Inputs: synapse1 (0), synapse2 (0),synapse3(0),synapseA1 (0),
synapseA2 (0),synapseA3 (0),synapseB1(0),synapseB2 (0),
synapseB3 (0);

Var: inputneuron1 (0),
inputneuron2 (0),
Hiddenneuron1 (0),
Hiddenneuron2 (0),
Hiddenneuron3 (0),

neuronOut (0);

{ inputs }
if (average(C,3) - average(C,3)[2]) >0 Then inputneuron1= 1 Else inputneuron1 = -1;
if (average(C,3) - average(C,3)[5]) >0 Then inputneuron2= 1 Else inputneuron2 = -1;

{neural net }
Hiddenneuron1 = Htangent(synapse1*inputneuron1 + synapseA1*inputneuron2, 50);
Hiddenneuron2 = Htangent(synapse2*inputneuron1 + synapseA2*inputneuron2, 50);
Hiddenneuron3 = Htangent(synapse3*inputneuron1 + synapseA3*inputneuron2, 50);

neuronOut = Htangent(synapseB1*Hiddenneuron1 + synapseB2*Hiddenneuron2 + synapseB3*Hiddenneuron3, 50);


{ buy or sell }
If neuronOut >= 0.5 then
Buy next bar at High stop;

If neuronOut <= -0.5 then
Sell next bar at low stop;

Hi Claude,
In an effort to put some mental images on this {I don't mean for the rest of the thread but just for now} If we think of VWB;

1) Would the above Function and Signal return an answer telling us more accurately where the VWB Fib "Line of Interest" should be no matter what Timeframe we applied it to?


2) How would you {could you} solve for (C3)? [ie VWB 24 EMA]
Would you use another set like...

Input: synapse4 (0), synapseA4 (0), synapseB4 (0);
Var: inputneuron3(0);
Var: Hiddenneuron4(0);


Thanks Claude, I was previously a bit overwhelmed by neural nets, but now I think you have given me a much better foundation to build on.
Looking forward to your next post.

B
Quoting Bemac
Hi Claude,
In an effort to put some mental images on this {I don't mean for the rest of the thread but just for now} If we think of VWB;

1) Would the above Function and Signal return an answer telling us more accurately where the VWB Fib "Line of Interest" should be no matter what Timeframe we applied it to?


2) How would you {could you} solve for (C3)? [ie VWB 24 EMA]
Would you use another set like...

Input: synapse4 (0), synapseA4 (0), synapseB4 (0);
Var: inputneuron3(0);
Var: Hiddenneuron4(0);


Thanks Claude, I was previously a bit overwhelmed by neural nets, but now I think you have given me a much better foundation to build on.
Looking forward to your next post.

B
Bemac ,Claude and MIB,

I have been interested in neural networks for some time. I was just wondering if you could recommend a good book on them as they pertain to trading.
I have several different platforms; AmiBroker, Tradestation, Metastock, MT4, Visual Basic and of course I could use C++.

I have been writing programs for a few years now. I am not the best programmer, but also not the worst... I have been fascinated by your discussion. Any help would be very much appreciated.

Thanks SMJ
Quoting Bemac
Hi Claude,
In an effort to put some mental images on this {I don't mean for the rest of the thread but just for now} If we think of VWB;

1) Would the above Function and Signal return an answer telling us more accurately where the VWB Fib "Line of Interest" should be no matter what Timeframe we applied it to?


2) How would you {could you} solve for (C3)? [ie VWB 24 EMA]
Would you use another set like...

Input: synapse4 (0), synapseA4 (0), synapseB4 (0);
Var: inputneuron3(0);
Var: Hiddenneuron4(0);


Thanks Claude, I was previously a bit overwhelmed by neural nets, but now I think you have given me a much better foundation to build on.
Looking forward to your next post.

B
The last book I read on this was "Understanding Neural Networks, Computer explotarations" by: Muareen Caudill and Charles Butler. but that was publishied in 1992 and had nothing to do with trading or markets.

Thanks again, SMJ
Claude,

Which Neural Net software do you use and recommend.
I did read someone offering signal produced by Neural Net for FX and e-Mini for next day price movement. They post the actual day movement and the forecast from Neural Net and all I can say its amazing.

I think Neural Net would brings us closer to the search of the grail why because we cant figure out how to trade trending and nontrending market at the same time.

Either we set the system for trending or nontrending but not both. If you open a plain chart every price is related, every single bar/candlestick is significant but cant explain how and applying it to forex would be ideal since forex is 24/7.

I hope discussion in this area would prosper and show how to think outside the box or on another dimension.
  • Post #18
  • Quote
  • Edited 3:32pm Mar 25, 2006 3:22pm | Edited 3:32pm
WoW!!!

Didn't expect this much response after all, this is breaking with tradition technical analysis.

Bemac:

Quote
1) Would the above Function and Signal return an answer telling us more accurately where the VWB Fib "Line of Interest" should be no matter what Timeframe we applied it to?


2) How would you {could you} solve for (C3)? [ie VWB 24 EMA]
Would you use another set like...

Input: synapse4 (0), synapseA4 (0), synapseB4 (0);
Var: inputneuron3(0);
Var: Hiddenneuron4(0);

the above system is an extremely simple example. It only returns of 1 or a -1 for a buy or sell if I understand you correctly, you are asking at what level does the market become critically overbought or oversold we can build a neural network to define this. Not sure if a simple one like this can be used.

I know you're coding is pretty creative and you could probably do it. But by simply adding 1 neuron. You must remember it's to the power of three, because it will add three new weights. So instead of just having 19683 combinations. It would be in the millions of different combinations. This simply becomes too time-consuming to do TradeStations it is not set up for that. You could do it in XL program that has a neural network built in and simply hard code the different weights. that the net comes up with.

I will write another entry that has a neural network in TradeStations that uses{ nearest neighbor} look at X. brs back to determine the future perhaps this one will be more functional for what you are asking as it uses internal loops and is less intensive on the optimizer.

Quote
Sajones

I have been interested in neural networks for some time. I was just wondering if you could recommend a good book on them as they pertain to trading

from the sounds of it. You've already read more than most. I've read two or three of them, but must admit, they often go into the intricate structures of a neural network. Today, software. This is not necessary. I have a friend, who is a great driver, but knows nothing about the internal combustion engine.

Of course, if you're like me and wanted to know. A quick search on Amazon, turned up three pages worth of neural networks for trading books.

I'm in the middle of Cybernetic Analysis for Stocks and Futures : Cutting-Edge DSP Technology. But I must admit, I usually go by the reviews that other people have written..

Quote
Azmi

Which Neural Net software do you use and recommend

I use a rather expensive one but I use it for other purposes as well. This is not necessary to study market data. Market data can be considered relatively easy to neural network. There are several of them, that work with an Excel spreadsheet. I would probably first look at these. They range in price from $70 to $600. Just pay more attention to ease-of-use. I personally would not pay into the thousands of dollars for a neural network to study the market it is just not necessary.
A truly wise man, always has more questions than answers.
  • Post #19
  • Quote
  • Edited 10:41pm Mar 25, 2006 10:35pm | Edited 10:41pm
Hello

Something that Bemac said. Got me thinking.

This overly complicated equation is called nearest neighbor. Ask yourself when you optimize let's say three months worth of bars does it represent conditions that are occurring currently . Of course not.



Enter neighbor theory what if we told the system last time my system took a trade and my indicator looked like this. What happened. Well, we can do this. By calculating the error rate every time a trade is taken. and comparing.

So instead of optimizing over a period of the past. We optimize over a sample of trades. When the indicator looked as close as what it does now.

So we choose nearest neighbor, 20 which means it will find the last 20 trades that look as close as the trade you are about to take. Then calculates what happened afterwards so that the computer can predict if price actually will go up.



Then, you can simply say, if you predict that price will go up then take the trade. If not forget about it.

I know, I know it sounds like I'm talking to the computer as if it has feelings but when we spend so much time with it :-)



The system that I am posting uses only Bollinger bands you can easily see where this is and change it to whatever you would like RSI CCI Bollinger bands whatever your heart desires.



It will buy, if the prediction of the next three bars is up and it will sell if the prediction of the next three bars is down. When you run the system. It will also write an XL file to your C drive so that you can see the prediction for the next six bars. This will be a prediction of the closing price.



You can optimize the period of the indicator being used, and you can also optimize how many trades that it looks at to make its prediction { how many neighbor}



I also like to thank someone that goes by the name of sexy Susie. That's the only name I know her by.

I did modify the code. Quite a bit and adapted it to the forex market. But a lot of the coding is hers.


Now this is not a system I would trade as there is no stop loss or any money management. Included. This is only a basic looked at what is possible.

_____________________________________________
_______________________________________________________

System name: neighbor



Input: NNN (50), { Number of nearest neighbors used in prediction }

NPredict (5), { Number of days of prediction }

FName ("C:\predict.csv"), { wright to C drive}

indicatorPeriod (8); {indicator Period}



Array: Errors[500, 2](0), { Error relative to current bar }

NNBars [70](0), { Bar numbers of nearest neighbors }

Predict [20](0); { Prediction of indicator 1 to 20 days in future }







Var: indicator1 (0),

indicator2 (0),

indicator3 (0),

indicator4 (0),

I1Max (0.0001), { Max value of indicator 1 }

I2Max (0.0001), { Max value of indicator 2 }

I3Max (0.0001), { Max value of indicator 3 }

I4Max (0.0001), { Max value of indicator 4 }

Sum (0), { Sum of predicted }

MinErr (0), { Minimum error }

MAXBARS (500), { Max dimension of Error array }

MAXNNN (70), { Max dimension of NNBars array }

MAXPRED (20), { Max dimension of Predict array }

NBars (0), { Number of bars to look back }

NoMatch (True), { flagged }

ii (0), { counter }

jj (0), { counter }

kk (0), { counter }

StrOut (""); { output character string }







{ Initialize file on first bar }

If BarNumber = 1 then Begin

FileDelete(FName);

StrOut = "Date, Close, PredBar1, PredBar2, PredBar3, PredBar4, PredBar5, PredBar6," + NewLine;

FileAppend(FName, StrOut);

End;







NBars = MinList(MaxBarsBack - 20, MAXBARS);





{ indicators }

indicator1 = BollingerBand(C,indicatorPeriod, 2) - BollingerBand(C,indicatorPeriod, 2)[1];

indicator2 = BollingerBand(C,indicatorPeriod, 2) - BollingerBand(C,indicatorPeriod, 2)[2];

indicator3 = BollingerBand(C,indicatorPeriod, 2) - BollingerBand(C,indicatorPeriod, 2)[3];

indicator4 = BollingerBand(C,indicatorPeriod, 2) - BollingerBand(C,indicatorPeriod, 2)[5];







{ Loop over prior bars to find max indicator values }

for ii = 1 to NBars Begin

if indicator1[ii] > I1Max then

I1Max = indicator1[ii];

if indicator2[ii] > I2Max then

I2Max = indicator2[ii];

if indicator3[ii] > I3Max then

I3Max = indicator3[ii];

if indicator4[ii] > I4Max then

I4Max = indicator4[ii];

End;





{ Calculate errors }

for ii = 1 to NBars Begin

Errors[ii - 1, 0] = Square((indicator1 - indicator1[ii])/I1Max) + Square((indicator2 - indicator2[ii])/I2Max) +

Square((indicator3 - indicator3[ii])/I3Max) + Square((indicator4 - indicator4[ii])/I4Max);

Errors[ii - 1, 1] = ii - 1;

End;





{ Take the smallest NNN values as the nearest neighbors }

for ii = 0 to MinList(NNN - 1, MAXNNN - 1) Begin

MinErr = 100 * Errors[0, 0];

for jj = 0 to NBars - 1 Begin

if Errors[jj, 0] < MinErr then Begin

NoMatch = true;

for kk = 0 to ii Begin

if Errors[jj, 1] = NNBars[kk] then

NoMatch = false;

End;

if NoMatch then Begin

MinErr = Errors[jj, 0];

NNBars[ii]= Errors[jj, 1];

End;

End;

End;

End;







{ Calculate prediction for next NPredict bars }

for ii = 0 to MinList(NPredict - 1, MAXPRED - 1) Begin

Sum = 0;

for jj = 0 to MinList(NNN - 1, MAXNNN - 1) Begin

if (NNBars[jj] - ii - 1) > 0 then

Sum = Sum + (C[NNBars[jj] - ii - 1] - C[NNBars[jj]]);

End;

Predict[ii] = C + (Sum/NNN);

End;







{ Write out predictions }

StrOut = NumtoStr(Date, 0) + "," + NumToStr(C, 4) + "," +

NumtoStr(Predict[0], 4) + "," + NumToStr(Predict[1], 4) + "," +

NumtoStr(Predict[2], 4) + "," + NumToStr(Predict[3], 4) + "," +

NumToStr(Predict[4], 4) + "," + NumToStr(Predict[5], 4) + NewLine;

FileAppend(FName, StrOut);



{ Buy if predictions are up; sell if down }

if Predict[0] > C and Predict[1] > C and Predict[2] > C then

Buy next bar at C limit;



if Predict[0] < C and Predict[1] < C and Predict[2] < C then

Sell next bar at C limit;
Attached Image
A truly wise man, always has more questions than answers.
NNs might look like the holy grail, but in practice they need quite clever (definitely not trivial) data preprocessing for producing anything useful. You can't just make a bunch of interconnected "neurons", feed it with raw price (probably a window of past prices) and expect anything useful to come out. Feeding it with multiple inputs (such as various indicators, probably derived from price) will probably make things worse.

One problem is that typical feed-forward NNs (most commonly used) accept a limited input range, which creates a problem for signals that contain trends (such as price). One could overcome this with detrending by using 1st-order differences, or by subtracting a moving average of the signal, or some more involved method. After that the detrended signal should be normalized (i.e.: put into some well defined range, like -1..+1). All that preprocessing requires a lot of prior knowledge about the data to work as expected.

Another problem is overwhelming the NN with information. Time series prediction with NNs is normally done by using a window of past values (e.g.: prices). Using, for example, 5 historical price changes for the inputs creates a 5 dimensional input space for the net to classify into appropriate predictions. The problem with high dimensional spaces is that they're often too sparse to work with (it's a similar problem as having insufficient samples for making any significant conclusions/generalizations in statistics). On the other hand, using a window that is too narrow results in the net not properly capturing the underlying dynamics (connecting patterns that aren't really connected). So, thinking about feeding all sorts of inputs to a NN simultaneously is a bad idea. Noisyness combined with sparse data (because of high dimensionality) doesn't not help at all. (Type "the curse of dimensionality" in Google to find out more about this, if interested.)

There's also always the danger for the net to overfit the data. The net would literally memorize the training set, while totally failing at predictions on new data.

I could go on, but ... the point is, there's a lot of voodoo required to make anything useful with NNs. A lot of people think a NN is some black box that magically learns to map history to future, you just need to feed it with tons of information. That's actually a brute force approach that's doomed to failure. I think people should know at least the basics about how NNs work before paying for NN software. And I'm almost certain, that those who really master the stuff, would find available software insufficient for implementing their ideas (except, maybe, some general purpose mathematical packages out there).

But, it is fun, nevertheless .

Quote
I did read someone offering signal produced by Neural Net for FX and e-Mini for next day price movement. They post the actual day movement and the forecast from Neural Net and all I can say its amazing.

Another thing about those charts that show actual price along with predicted price by NN. These are actually quite misleading. The "predicted" chart is just price at [t-1] + prediction. Predictions could be totally wrong, but the chart would still seem "amazingly accurate". I don't know about that particular signal provider, though. Just speaking in general.
0 traders viewing now