Have anyone worked with Neural Networks and what is your opinion? Is there a good product that works well with Tradestation? Or should I program my own?
Thanks.
Thanks.
Neural Network
- Thread starter knifecatcher
- Start date
yes, and near useless. Every experiences I"ve had with them (both in terms of finance, option modelling, and swing price modelling in the electricity industry) has shown me that they require insane amounts of resources to construct/maintain, and give no advantage to simpler methodologies.
Hi 99,
Do you use a canned package? What specific problems were your NNs trying to solve?
Thank you for your warning. I got a feeling there is a lot of pitfalls such as curve fitting.
Thanks,
KC
Murray Ruggiero
Sponsor
I have done a lot of work with neural networks and have written many articles on the subject. They are part of my Technology and Trading book.
http://www.tradersstudio.com/Produc...ID/5/psnavcmd/CatalogItemDetails/Default.aspx
In addition I just finished a three part articles series for Futures , the last installment will be printed in August.
Here the key to using neural networks. They can't be used as the core of the system. The way I have been successful using them is to use them to improve a existing strategy that makes money, for example if you developed a system which uses RSI you might create a predictive RSI indicator using a neural network and use that instead. You don't want to develop a system in which if the the neural network degrading will cause catastrophic failure, this is why you start with a good system which we use Neural Networks to improve. Other strategies to reduce this problem are using multiple neural networks and for example averaging the outputs.
Another fact is that you should retrain the networks on a regular basis, for example Train with 1500 bars and then retrain an 200 bar walk forward window. This walk forward process will show you how well your network is designed because if it does not retrain in each window that shows that the set of inputs you are using are truly not predictive of the desired output and good training during previous windows might be the results of short term effects in the data during previous windows. An example of this is during the period 1998-2001 in the Nasdaq. The Nasdaq when straight up and straight down. During this period something like Close-Close[3] could predict future direction 70% of the time. A network might have trained well during this period and then starting in 2002 stopped training because this short term artifact is gone.
http://www.tradersstudio.com/Produc...ID/5/psnavcmd/CatalogItemDetails/Default.aspx
In addition I just finished a three part articles series for Futures , the last installment will be printed in August.
Here the key to using neural networks. They can't be used as the core of the system. The way I have been successful using them is to use them to improve a existing strategy that makes money, for example if you developed a system which uses RSI you might create a predictive RSI indicator using a neural network and use that instead. You don't want to develop a system in which if the the neural network degrading will cause catastrophic failure, this is why you start with a good system which we use Neural Networks to improve. Other strategies to reduce this problem are using multiple neural networks and for example averaging the outputs.
Another fact is that you should retrain the networks on a regular basis, for example Train with 1500 bars and then retrain an 200 bar walk forward window. This walk forward process will show you how well your network is designed because if it does not retrain in each window that shows that the set of inputs you are using are truly not predictive of the desired output and good training during previous windows might be the results of short term effects in the data during previous windows. An example of this is during the period 1998-2001 in the Nasdaq. The Nasdaq when straight up and straight down. During this period something like Close-Close[3] could predict future direction 70% of the time. A network might have trained well during this period and then starting in 2002 stopped training because this short term artifact is gone.
I am curious - do you actually trade? The reason I ask is if you would be willing to show audited returns of an automated system that you trade that is based on NNs in any way shape or form?
You see, 99% of this site are people who don't trade for a living and give lots of advice. I am just wondering which side of the 99% you are?
nitro
Hi Murray,
Thanks for your input. But should one just keep feeding every bars like that into the NNs?
Regards,
KC
Thanks for your input. But should one just keep feeding every bars like that into the NNs?
Regards,
KC
Hi nitro,
Can you give some pointers as to how one should get started with NN? Should I try program NN myself or buy a professional package?
Thanks,
KC
Yes I have worked with NN. No I have not worked with them as the centerpiece of any trading strategy.
Look, folks, NN is just a fancy way of doing interpolation. That's it. That's all it is. If you expect it to be something more than that, you are going to be disappointed. If you realize what a NN is and use it for something appropriate, there's nothing wrong with it. Just don't expect it to do something that any other number of interpolation methods can't do.
Fletch
Look, folks, NN is just a fancy way of doing interpolation. That's it. That's all it is. If you expect it to be something more than that, you are going to be disappointed. If you realize what a NN is and use it for something appropriate, there's nothing wrong with it. Just don't expect it to do something that any other number of interpolation methods can't do.
Fletch
I purchased your ORB video several months ago and I see on the back cover that you were written up in Business Week for your early work in using neural networks in financial applications. Do you remember what issue that was in?