I was wondering if there is someone on this forum that have had any experience with commercial neural network softwares for financial markets.
I'm referring to softwares like NeuroShell, TradingsSlutions and the like, and not to end-user systems like VantagePoint.
Specifically I'm interested in the possibility to:
1) define the topology of the net on my own, or having that determined genetically
2) build committees of networks
3) THE MOST IMPORTANT ONE: speed of training and genetic evolution
Time ago I worked extensively on neural networks with an NN framework completely written on my own. I ended up with a very good profitable system. Unfortunately I was never able to use it because in order to work I had to create a new committee of 40 networks every week, genetically determining their inputs and training all of the 40 nets on 1 minute bars of some previous weeks market data. On every new week I had to throw away the old committee and create a new one.
The problem was that the whole building process required at least 10 days of computation despite I was able to work in parallel with 8 speed PC.
I made all of my best to optimize my code in order to have the committees be built and trained in the lasting of a week-end, but I was not able and had to give up.
Is there someone that have any experience with commercial neural network softwares and nets with, say, 40-80 inputs, being trained in big quantities (40 or more) in less then 48 hours?
I'm referring to softwares like NeuroShell, TradingsSlutions and the like, and not to end-user systems like VantagePoint.
Specifically I'm interested in the possibility to:
1) define the topology of the net on my own, or having that determined genetically
2) build committees of networks
3) THE MOST IMPORTANT ONE: speed of training and genetic evolution
Time ago I worked extensively on neural networks with an NN framework completely written on my own. I ended up with a very good profitable system. Unfortunately I was never able to use it because in order to work I had to create a new committee of 40 networks every week, genetically determining their inputs and training all of the 40 nets on 1 minute bars of some previous weeks market data. On every new week I had to throw away the old committee and create a new one.
The problem was that the whole building process required at least 10 days of computation despite I was able to work in parallel with 8 speed PC.
I made all of my best to optimize my code in order to have the committees be built and trained in the lasting of a week-end, but I was not able and had to give up.
Is there someone that have any experience with commercial neural network softwares and nets with, say, 40-80 inputs, being trained in big quantities (40 or more) in less then 48 hours?
.Ubeee.
