Neural Networks - page 15
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
Simba what do you mean by out of sample? I tried building another network with about 6 more inputs and that reduced the accuracy. I think I've got a good set of inputs. It's based on my trading strategy which works well in trending markets and poorly in ranging markets. This makes sense with the predictions as well. The two long ranges at the top and bottom of the chart are where the predictions start to falter. I think some work on ranging strategies could turn up some useful inputs.
Kazam, this is the code used to create the network. To me that says create a new feed-forward back prop network. I'm not sure why it's churning out good results either it seems a little too good to be true. I summed the values of correct pips minus incorrect pips and over that period (about 30 months) it would have profited 294873 pips... I'm scared
numHiddenNeurons = 13;
net = newff(p,t,numHiddenNeurons);
mrwobbles
Out of the sample means checking how the NN works with data different from the one it wast trained on. In the sample means opposite.
The code is ok. Are you sure you don't use as input and target data in the same time step /input[x1(t0), x2(t0), ..., xn(t0)], output[y(t0)]/?
Some more papers I found in my bookmarks /the first one should be really interesting for you/:
http://www.chaos2008.net/zzProceedings/CHAOS2008%20(D)/PAPERS_PDF/Atsalakis_Nezis_Skiadas-Forecasting_Chaotic_time_series_by_a_Neural_Network.pdf
I also decided that in 3-4 weeks time I'll try to build a ENN capable to predict exchange rates at the D1 and maybe H4 time frames. I think I'll tart with combination of FNT, GEP and PSO and then try to improve the ENN using PIPE instead of GEP, EPSO, Simulated Annealing etc
Thats what I figured but no it was trained on a 7 year chunk from 1999 to 2006 and then simulated on data from then till now. So it was out of sample.
But that might explain it. It could well be trying to predict the current high/low and close rather than the next hour
. I'll have a look at the date stamps and see. Thanks for the links will have a look see.
Yeah turns out I was feeding it input and target data from the same hour rather than input from current hour and target from the next. I trained the network on the same inputs but with the targets from the next hour and the results are more what I was suspecting. It was accurate on direction about 51% of the time and recorded about 10,000 pips over 30 months. I normalized the inputs to [-1,1] and this improved the performance marginally. Gonna try a few different network model. Have been looking at time-delayed networks, thou the flexible neural tree approach does seem more and more appealing.
mrwobbles
Skip TDN's.
You should be only interested in 3 types of networks:
- Flexible Neural Trees - designed with the use of either GP, GEP, PIPE or ECGP and fine tuned with the use of Simulated Annealing, PSO, EPSO, Ant Algorithms, Artificial Immune System Algorithm etc.
- Bayesian Networks - either for time series prediction or for building a trading system
- HONN's
For building a trading system you may also find useful:
- RIPPER algorithm - someone used it an EA made for Automated Trading Championship if I remember correctly
- C4.5/C5.0
You don't need anything more and anything less /as for now/
Thanks for the advice Kazam, I think I'm fairly set on looking at some form of genetic algorithm, ECGP looks the most promising. It's just my coding skills aren't that great and ANNs are fairly new to me. I am having some trouble with choosing the function set. Correct me if I am wrong but this a set of all possible functions for mapping nodes to each other? So I would specify a random distribution of input weights and the terminal set would contain all the node transfer & activation functions? Assign them uniform probabilities with the sum of the probabilities of sets F and T being 1. The optimization of the input weights would happen using SA or PSO thou? This resultant optimized best fit tree would be passed back to build a new PPT?
On a seperate note, the Uni library has all the editions of the European Conference on Genetic Programming from 1998 to 2004. Think they might be worth a read? They've also got a copy of 'Genetic programming and evolvable machines' by Dordrecht : Kluwer Academic Publishers, c2000 that I'm thinking about taking out.
mrwobbles
Don't start from ECGP. Start from something simple like designing NN using Genetic Programming or Gen Expression Programming. There are ready to use classes which implement those methods.
For example:
http://69.10.233.10/KB/recipes/aforge_genetic.aspx
http://www.codeproject.com/KB/recipes/aforge_neuro.aspx
[/CODE]
I don't know if there is a freely available implementation of ECGP, but there's one of ECGA /ECGP is based on ECGA/:
[CODE]
http://www.kumarasastry.com/2006/03/26/extended-compact-genetic-algorithm-in-c-version-11/
But like I said, skip ECGP/ECGA for now.
EDIT
As for the books - leave the conference materials for later /don't try to learn everything about NN's and Evolutionary Algorithms, focus on tools that might be useful for trading - time series prediction and decision making/.
Start from the "Field guide to genetic programming" that is free to download /I posted a link some pages ago/.
let's talk about how to choose samples for neural network
any neural network depends on how you choose the samples. i believe this question bother you too.
i build some eas use neural networks. but i havn't found a good sample prepare way.
here , i tell you my sample choose way, i 'm waiting for yours.
use three bars as the sample, the next bar as the result.
use few bars as the basewindow, and few bars as the feature window, set some feature value like ma value close price openprcie etc. any indicator you can use as the feature.
Neuro Net is a SCAM!!!
I tested this EA and found it doesn't produce the results that he is claiming. Very small profits, and huge losses. After several days of testing and working with the company, this EA couldn't produce results for more than a week or two at a time.
They claim to have a neural net in the MQ4 language, no dll. I find this very unlikely.
I asked for a refund and he would not give it to me. SCAMMER!!!!!
I'm not sure if this will be of any use to anyone but I found a GA toolkit for MATLAB called SGALAB. Anyway here's the link.