i am effectively using forcasting models based on NNP for my long-term positions with more success than a simulated modified OLS regression forecast that I run simultaneosly. i am curious if anyone has been able to "train" NN for day trades? mk
does "effectivly" and "more success" mean making money - i personally dont believe it is possible to forecast - but it may mean you are caused to take a 50/50 bet that has been working out i do believe if enough money is put behind a trade by a big bank and the trade was based on mulitiple technical factors - that trade may cause the event to happen - and if you have a model that replicated that result - you will feel that you system is working - but to me - making money in the markets is about controlling losses - and nn for a private person is not a low risk high return method the banks are quite happy to trade to just provide liquidity themselves for their client trades - at least thats the positive they fall back on when they are losing money trading i would genuinly appreciate your comments on your experiences
IMHO, any NN that is applied to trading has to be of the of the kind that takes temporal dimensional considerations, e.g., "Back Propogation Through Time." Check out http://www.nd.com for neat stuff if you are interested. Just be ware that, again IMHO, you are better off spending your time learning to trade with your Neural Net than with an artificial one _FIRST_. nitro
I also curious about what your interpretation of "effectively using" means. a) Works great with paper trades and real money b) Works great only with paper trades and not with real money c) Profits are enough to make a living d) Profits are enough to supplement my income e) Profits are enough to pay for all the tools and effort put into mechanical trading system f) some of the above...if so...please provide further details ________. Note: I'm not saying it doesn't work because I know some that actually do make money from their artificial neural systems (fuzzy logic, pattern matching, storage, recall and so on). Simply, what's your story? Also, which financial instruments are your long-term positions in? Nihaba Ashi
Early on people thought that NN's would be magic. You would throw a NN at a problem and they'd form themselves into a connect-o-plasm and solve all problems without effort. There is one thing NN's are really good at- over fitting to historical data perfectly with NO forward predictive ability at all. Did you ever check out a NN S&P website and see the crap it comes up with? "The S&P close for today has a 53% chance of being between 1.5 and 3 points higher than yesterday." Can anyone make money on junk like that? NNs take some time to shift from one pattern to a new one and markets love to hurt people who stay to long with a pattern. NNs are the perfect sitting duck who gets squeezed.
I agree with the above comment, since I would never try to use anything like I have now for intraday trading, NN are good at data mining and "fitting" to historical data, and, if set-up wrong have no use to try to play out different scenarios of the future. The key is to split the historical data to calibrate the model (historic interval from 1995 to 2002 on any given market), and then use 1995 - 2001 for calibrating and 2001 - 2002 to check what is the "batting average" and relevant statistics. One can also infuse some "white noise", or garble the data to see if you have a super duper fitting machine or actually can add value to your LONGER term trades (it your NN can fit the garbled data with R^2 of 25% or more then all you have is a fitting machine, rather than anything of value).
Neural Networks can work and I have a couple of setups that are pretty interesting (in monte carlo analysis) on paper trade (slippage+commission). The advantage is that it finds many correlations that a human eye cannot see, the downfall is that the human has no understanding what triggered such and such event. The KISS philosophy is not really part of the Neural Net world.
Traditional NN approaches add very little value to trading systems development: they are worthy - although in many cases inferior - pattern recognition systems. There is a lot of hype in the field and the published info is not very useful. There are however ideas related to NN that are more valuable. None of the useful information appears to be published. If you really want to evolve your approach you will have to approach the subject through your own resources. ......
I agree with Caltrader that nothing of substance realting to capital markets is available YET to the best of my knowledge, other than some PhD papers that usually target different applications of NN's, but the basic algorythms can be inferred from the literature, since getting the basic algorythms in place wasn't an easy task for me. I've been working on and off trying to shorten the timeframes of historical data without having the computer calculate for days.