Neural Networks - page 12
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Implement a simple feed forward optimisation algo, stimulate winning trades with some form of asymmetrical MM . Access your real trading results with standardized statistical measurements, so you can know if the whole model or part of it is broken when start under performing. Go back to optimization/drawing board.
All this only if your network captured, some underlying cyclical exploitable behaviour. Hint that you are on path of a success is if your pattern is of fractal nature.
Looking on your previous posts, you probably suffer of over-training/fitting syndrome.
I would go for big number of lower quality trades, so I could wed them out with simpler procedures...
You've made some basic mistakes:
- using typical feed forward neural nets for predicting time series like currencies exchange rates is a very poor idea
- you're trying to predict values at H1 time frame - it can't be done for Forex with good results. Use D1 or H4 (for currencies that have low volatility)
- you use to much data as an input - neural net gets "to used" to training data and it will perform very poor in live trading
- you're getting to excited looking at training data
- it is impossible to train a neural net that will work effectively for a long time. Typical neural net trained to predict time series gives about 20-100 good predictions and then it has to be retrained to fit recent changes
If you want to create useful neural nets for time series predictions read about evolutionary neural networks (feed forward neural networks coded as flexible neural trees; their architecture is optimized using PIPE or GEP; flexible activation function parameters are optimized using PSO, EPSO or simulated annealing etc.)Yeah I thought that was a little too good to be true
It took less than an hour to build that network. Would these evolutionary neural networks have anything to do with UTHONNs? I've got an article that suggests this type of ANN works much better than traditional FF back prop models. I'm working with MATLAB, would it be possible to do what you said with this? Also what would you suggest for software or is MATLAB ok?
@Kazam
Leave some money on the table man
Implement a simple feed forward optimisation algo, stimulate winning trades with some form of asymmetrical MM . Access your real trading results with standardized statistical measurements, so you can know if the whole model or part of it is broken when start under performing. Go back to optimization/drawing board.
All this only if your network captured, some underlying cyclical exploitable behaviour. Hint that you are on path of a success is if your pattern is of fractal nature.
Looking on your previous posts, you probably suffer of over-training/fitting syndrome.
I would go for big number of lower quality trades, so I could wed them out with simpler procedures...I've read about over-fitting and I'm assuming this is what has happened here. Would I be better splitting that ten year sample into 10 1 year samples and training on one year and testing on the next? It appears to me that the GBPJPY goes through a 5-10 year cycle. I'd chosen a 10 year sample because I'd hoped to train the network to be able to identify that cycle. The hope being that a network trained over that cycle would be able to identify what part of the cycle it was in and respond appropriately.
My previous strategies have involved waiting for big moves and then cashing in but I would hope to use a neural net as part of a scalping strategy. I'd be looking to trade nearly every tick based on whether the forecasted close was above or below the current.
Evolutionary neural networks are feed forward neural networks build and optimized using genetic computation algorithms and optimization methods. It's something else than UHHONN's (btw UHHONN's look very promising and are next on my testing list
).
I think those two papers should be a good start
http://www.softcomputing.net/insci.pdf
http://www.softcomputing.net/chen-neucom2.pdfYou could use Matlab for creating ENN but You'd have to find or write m-scripts for GEP, PIPE etc.
Evolutionary neural networks are feed forward neural networks build and optimized using genetic computation algorithms and optimization methods. It's something else than UHHONN's (btw UHHONN's look very promising and are next on my testing list
I think those two papers should be a good start
http://www.softcomputing.net/insci.pdf
http://www.softcomputing.net/chen-neucom2.pdfCool, I'll have a look at those papers and go from there. I've read about genetic optimization and it seems the most logical approach. I did have a trial of Alyuda NeuroIntelligence that had a genetic optimization of neurons and weights but its something like $600 and if I can do it in Matlab that would be preferable.
Have you noticed that a lot of papers on neural net forecasting seem to come out of Jinan University? I've got a document on a multitude of different HONNS, if you want I'll email it to you, just PM me with your address. Tried zipping it up and attaching it but I was missing a token apparently
Evolutionary neural networks are feed forward neural networks build and optimized using genetic computation algorithms and optimization methods. It's something else than UHHONN's (btw UHHONN's look very promising and are next on my testing list
I think those two papers should be a good start
http://www.softcomputing.net/insci.pdf
http://www.softcomputing.net/chen-neucom2.pdfHi,
How do you implement the ideas in those papers?I mean is there any Matlab or C++ code?
biddick
I write my own versions of all algorithms I use because I'm using a server that's based on CUDA technology from Nvidia (it makes computations with a speed of 1000 times grater than the newest Intel Quad processor
).
But mostly I use C language.
If you want a good source of information about HONN's get the book "Artificial Higher Order Neural Networks for Economics and Business". It's a new publication so it has very good info. The price is a killer (180$) but you can find it in pdf
.
Some other books I recommend:
- "A field guide to genetic programming" - you can download it for free at
- "Introduction to genetic algorithms" - published by Springer in 2008.
- "Biologically Inspired Algorithms for Financial Modeling" - from Springer, published in 2006. Very good publication.
- "Network Models and Optimization Multiobjective Genetic Algorithm Approach" - from Springer, published at 2008. Also a very good publication.
- "Gen Expression Programming" - by Candida Ferreira.
All of them can be found in pdf. I usually don't encourage to download copyrighted materials but most of the recent books about ENN cost way over 200$.
Those are just few books that I think are really worth reading. I have over a hundred other so if you'll be interested in something more I'll recommend you another books.
About PIPE you can read here:
[CODE]http://edocs.tu-berlin.de/diss/2003/salustowicz_rafal.pdfBy the end of the year I'll finish my PhD thesis about "Evolutionary Neural Networks for financial time series forecasting" and I'll probably translate it to English so i might upload parts of it.
Kazam,
You have a good result in your trading with this stuff ?