I spent 2 years, maybe more, working more than 40-50 hrs a week to build a NN that would give an estimate of the S&P and Bond prices 6 days ahead. The inputs were time series from several markets S&P, US, GC, S, and DM. I would get great results for a while, and then the markets would turn. It was impossible for me to judge a 'trend change' from a 'retracement'. I retrained the nets every weekend but that did not solve the problem. At the top of my efforts I had a small linux cluster of computers working 24/7 testing and training nets. Collecting and validating data was a non-stop job. I was lucky that I stopped trading that stuff before I blew out. I coulda got killed. 
Shortly after I stopped NN development I saw a web site called the 'Casandra Project' which was using an approach similar to the one I was following. I liked the Casandra Project because they seemed to be honest about their approach and posted their results every day including actual broker statements. (They were also trading their own money.) They doubled the account size and then blew out. If you go to www.archive.org, and enter 'www.cassandraproject.net' into the search you can view pages from the project from 1999 when it started until it ended in 2002. Or, http://web.archive.org/web/20020205202136/www.cassandraproject.net/cgi-bin/cgiwrap/cassmk/index.cgi
Some pages on their site were not archived but there is still some interesting content there.
This is not to say NN cannot be helpful or a NN system cannot be profitable over a long period of time. However, NN systems when applied to trading systems have some issues:
1. You cannot monitor the weights of all of the nets and understand 'why' this week's forecast is different from last week. Small changes in inputs can produce large changes in the output. A data error can cost you a lot of time to find, and a lot of money if you do not find it.
2. Preprocessing and cleaning data is a bitch.
3. You can never have anyway near 100% confidence in the forecast so your 'trade management' must be extremely good.
4. The combination of NN training, trade management, forward and backward testing takes 'a lot' of time.
5. You can get 'as good' a return, using 'low tech' approaches with a lot less work (IMHO).
6. The more I learned about NNs the less I trusted them (speaking for myself here).
7. To trade a NN you must have wide stops or have a very short prediction window. You probably also need to trade many markets at the same time as some markets will do much better than others but you cannot select the market to trade before the day begins... You can see cassandraproject did both, wide stops and many markets. When the markets turn, they will usually all turn at the same time resulting in breathtaking drawdowns.
The entire experience was good for me as I started the journey from a programmer point of view and learned enough about the markets to leave NNs behind. It was a failed effort as a trading system, but a great experience as a computer science 'head trip' and learning about the markets.
I am sure others at ET have done better than what I did and that is fine with me....
Good luck,
Shortly after I stopped NN development I saw a web site called the 'Casandra Project' which was using an approach similar to the one I was following. I liked the Casandra Project because they seemed to be honest about their approach and posted their results every day including actual broker statements. (They were also trading their own money.) They doubled the account size and then blew out. If you go to www.archive.org, and enter 'www.cassandraproject.net' into the search you can view pages from the project from 1999 when it started until it ended in 2002. Or, http://web.archive.org/web/20020205202136/www.cassandraproject.net/cgi-bin/cgiwrap/cassmk/index.cgi
Some pages on their site were not archived but there is still some interesting content there.
This is not to say NN cannot be helpful or a NN system cannot be profitable over a long period of time. However, NN systems when applied to trading systems have some issues:
1. You cannot monitor the weights of all of the nets and understand 'why' this week's forecast is different from last week. Small changes in inputs can produce large changes in the output. A data error can cost you a lot of time to find, and a lot of money if you do not find it.
2. Preprocessing and cleaning data is a bitch.
3. You can never have anyway near 100% confidence in the forecast so your 'trade management' must be extremely good.
4. The combination of NN training, trade management, forward and backward testing takes 'a lot' of time.
5. You can get 'as good' a return, using 'low tech' approaches with a lot less work (IMHO).
6. The more I learned about NNs the less I trusted them (speaking for myself here).
7. To trade a NN you must have wide stops or have a very short prediction window. You probably also need to trade many markets at the same time as some markets will do much better than others but you cannot select the market to trade before the day begins... You can see cassandraproject did both, wide stops and many markets. When the markets turn, they will usually all turn at the same time resulting in breathtaking drawdowns.
The entire experience was good for me as I started the journey from a programmer point of view and learned enough about the markets to leave NNs behind. It was a failed effort as a trading system, but a great experience as a computer science 'head trip' and learning about the markets.
I am sure others at ET have done better than what I did and that is fine with me....
Good luck,
