One guy I knew of spent a lot of time with neural nets looking for a longer term solution to the markets and the one thing he learned was that he never got anything that worked consistently beyond about a week. He had a buy the dips strategy going. I was able to develope a better one as a newbie using free charting services!!
Take a look at SNNS at http://www-ra.informatik.uni-tuebingen.de/SNNS/ This project runs on Windows or Linux, has excellent documentation, and has more features than most commercial products. Although it looks like this technology is not 'new' it is well supported and there is really not much new developed in this space in the last 15 years anyway. However, SNNS requires that you read the documentation and have some technical skills. After purchasing several NN packages I ended up using SNNS. (I no longer develop NN trading systems.) The problem with NNs is not the package being using but the work that must go into developing the inputs and managing the data necessary to feed the inputs. Bad data, or bad indicator selection can ruin months of work. (You only know its bad when something costs you a lot of time and money, not before.) Getting something that can be used in trading will probably take much longer than you can imagine. Some thoughts: 1. At the end of the day any 'prediction' using GAs or NNs will win or lose depending on money management. 2. Any NN or GA system will still need a trade/money management approach that assumes the system will not work. 3. Any NNs will usually need wide stops, drawdowns can be exciting. and greater than discretionary or indicator based system. Does your blind faith in the system exceed your margin requirements? 4. The system can 'learn', profits increase, and then the markets make a fundamental change in structure executing all of your wide stops. Your system becomes worthless every n number of months and you don't know when the market changed until after the fact. 5. Early results that are curve fitted always look promising. Spend more time on 'out of sample' testing before trading real money. NNs and GAs are no 'short cut' to trading success. Good books to read for high tech trading approaches include: Fooled by Randomness The Predictors When Genius Failed (on LTCM, a great book) NNs and GAs greatest benefit (IMVHO) is that if you work hard enough to get them to work well over a short time you have learned enough about the market and trading to no longer need them.... Time spent on money management is time well spent. bona fortuna! ramora
My experince with the up-to-date mathematical methods (anything you number it from artifical intelligence to nonlinear dynamics) is that it is not worth of time for individual investors. The market is too complicated, it takes a vast energy for a develop and time for a computer to find good pattern hopefully. I am not discourage anyone, just saying from my experience. I spent a couple of yrs on that, and at the end, I found that the well-known classic Dow theory serves my purpose just right.
I felt the same way - that was why I was interested in seeing if any of the developed systems that are used by big names are available to the public - or something similar to them. For some reason, and I don't really have any good reason why other than a gut feeling, I don't know that I trust the stuff on my own systems. But then again, the web site things (Stensland Systems and Trendax) look like they do all the computing on powerful systems and they have already developed the technology. The Trendax site looks pretty impressive, but they have no statistics up at all - and like I said, I'm familiar with Kim Schmitz, so I'm wary of it. I will give him the benefit of the doubt and maybe this is his company that is legit and proper. The Stensland Systems site has some statistics up, and I'm not sure if they are amazing, but just the fact that it has some data up makes it more impressive to me - at least willing to show what it has. I will do the free period with them once they resolve their bank connection system thing that they mention, and then if it sucks, so be it. If I remember, I can try to post up my experience with it.
Bear uses predict and professional II + for in house development. http://www.neuralware.com/products_pro2.jsp
They made quite a bit of money. The methods used are nothing astounding. Standard fare stuff with a few twists. The way they put it together was what was primarily responsible for the value. There is value in approaches related to NN but most of the stuff that you will see out in the publicised commerical space is of limited value. The stuff that Bear uses is IMHO not very useful and if you are truly interested in this stuff my advice is not to listen to anything pitched by a particular firm or group: Truly useful approaches will be trade secret and not discussed anywhere.
Your question is similar to "is anyone here successful using excel or a trading software ?". The answer would be yes and no . ANN is just a tool like Excel except that it is not a piece of software but a tool for modeling. But for modelling something you must know the parameters and the law of the world you want to model. And so there can be many hypes behind ANN sometimes. If you want to have an idea of ANN applied to stock market I just quoted a book in another thread "Neural Networks for Economic and Financial Modelling" Thomson Publishing (the edition I got is 1996 perhaps they have updated I didn't check) which deals also with genetic algorithm since these algorithms are used in the ANN for optimising the objective function. This book deals essentially with Feed-Forward with Back propagation model as they said that it is in their opinion the best suited general architecture for financial modelling. Personally ANN has not been useful for creating my model. It is only after my model has been created that I could use ANNs profitably but I prefer another approach which is similar to what is called classifier (but at the moment there is a practical problem of databases). Although it has the capability of being statistically an Universal Approximator for non-linear model - and my model is non-linear - notwithstanding that the structure of the network is also correct - this is essentially a black box modelling approach and I prefer white box modelling approach because you model the causes and know what you are doing by design whereas with a black box approach there are much uncertainty about the robustness of the model.
With ANNs he has excellent mean to catch pigeons and so do others . From the marketing point of view he is an intelligent guy : it is better to make people think that you have found something with ANN than make real research.