NNs is a curve-fitting method. Very useful where curve-fitting is required, completely useless where curve-fitting is to be avoided, like in trading.
Tradecision is most likely what you're after. I don't think anything else is turnkey and still good. Most NN's are designed for analyzing people (banks issuing credit) Wave59 appears to be 10+ years old and run by narcissists (frankly) . Note-Forecasting price is generally less effective than error management on indicators. If you forecast price, you will need to write specific indicators and have plenty of computing horsepower. You will also need to spend lots of time preprocessing the data before you train the net. Most likely software will do this for you. Cheers.
Below is a post of mine from 2006... I follow this technology and have not seen anything since that would change my perspective from the comments below..... You are not going to find anything in the commercial world that can be used without a lot of very serious software and database experience. Anyone who tells you differently is selling you some expensive POS that has better alternatives in open source world used in universities. Are there great applications using NN? Yes Are people making money with NN? Yes Are you likely to either apply NN to a useful application or make money.... Probably not. Focus on Price Action, longer term trading. What is the market doing rather than some overfitted forecast. If you look at what is available, get open source software and get it to work in real time with CUDA. Maybe. I liked GAs a lot better but at the end of the day I had this insight, "Hell, I can make as much money trading x, y, or z without spending nights and weekends programming like a demon..." Good luck... Slacker =============== Interesting thread... I wasted 2 years on Neural Nets. I purchased a lot of software including packages from BioComp, Ward and Braincell. The best package that I used was SNNS. It was best because it had the best documentation, user interface and examples, runs on Linux and Windows, and was free. The commercial packages have better UIs for newbes but at the end of the day you will probably be using scripts anyway so the UI does not help. I trained the system daily using many different markets and was looking for an intermarket edge. Some thoughts from that time: Coordinating data is a bitch. One data-fart can cost a lot of money and be very hard to find. Often I would not know that a data-fart exists until after I would lose a lot of money and go looking for the cause. There are tools that can help but not much. Unless you have to time to write real data validation routines that can do more than identify when a time series has shifted 1 or more bars or the data is way outside the range of the last 20 bars.... Maybe validating data against multiple sources would work. I was able to get the same or better results using less and less data as I learned more about the problem. Throwing tons of data at a NN and expecting it to sort everything out does not work (IMHO)... More on data. What is the best way to handle 'market shocks' with a NN? I don't know. Market shocks in historical data messes up your results. If the shock was a scheduled report maybe you can plan to handle that day differently. You could clean your data to remove market shock spikes, but what will your system do when it faces that unscheduled market shock tomorrow??? One way many improve the NN results is to 'train every day'. Fine. The system signals a LONG, the next day the market goes south, the system says BUY MORE, market goes faster south BUY LOTS MORE, you retrain the nets that night and the system says, 'SOLD 3 DAYS ago, SELL MORE!' WTF!!!! What about 'walk forward testing'. Some vendors claim 'walk forward testing' is not 'curve fitting' because even newbees know curve fitting can be dangerous. Curve fitting can give the user false confidence that their curve fitting results during the sample data period did not destroy their account and then find that even routine moderate volatility days can completely eliminate profit goals reached during testing. The vendors claiming walk forward testing is not curve fitting are being dishonest to the point of committing fraud in my opinion. The curve fitting and walk forward testing assume that tomorrow there will be a buyer for every seller and a seller for every buyer. Not always the case, 9/11, Oct 1987, and LTCM are good examples. The NN is a trend following system and needs wider stops than most trend systems because of the noise in the forecast. Even a good system in backtesting will have 5 losing trades in a row. If you are trying to forecast a price 2 days in the future or 6 days in the future you will probably need a 1.2 or a 1.5 * AvgTrueRange of 2 or 6 days, this plus 5 losing trades in a row starts to get expensive. Managing risk is very difficult to do as the nets are always changing. Every time you re-train the system; the system you will use tomorrow is new. Think about that statement, tomorrow's NN is NEW. Your results from the past 3 months have nothing to do with tomorrow's NET retrained last night. With a state machine or a GeneticAlgo you can track down the specific trigger for an entry or an exit. With a multilayer NN you can look at the layer weights and inputs but it is hard (impossible for anything except the most simple net) to find the trigger that gets into into/out of a trade. Just how much do you trust your net? One guy on the web, around 2000, started a site and published his real time results for 20 markets. He did great for 6 months, people were subscribing, sending money. Looked great. Then the market turned. All of his NNs were on the wrong side of the market and stayed wrong for about 3 weeks. It was ugly. He held fast to his NN until he reached 50% drawdown and then closed the site. Each day he would publish a 'holding fast WE believe our technology' letter to the web site. As long as the behavior of the market says either volatile or within a range the nets are fine. However, when the markets turns from a trend to a range period or from range to volatile, look out baby!!!! How can you be confident when you have had 4 losing trades in a row and not wonder if you are moving to a market fundamentally different than the one used to train data??? You cannot. IMHO I started to scale back my NN project and just use NN to replace indicators such as RSI. In effect making the NN a better filter. Not predicting the future as much as reducing noise. This is a good application of NNs. However, there are better filters that will do the same without the overhead such as T3Average and the Jurik Research products. (You can find alternatives to Jurik's stuff on Tradestations forum. The Tradestation stuff is free. However, if you want to play with NNs you probably have a fat software budget so spend away. Jurik's indicators are good and cheap for what they do.) Can a system be profitable and not re-trained 'often'? The fact that all NN vendors retrain and now have tools to choose from 1 out of many alternative systems is proof that they do not work. "They can demo, but they cannot trade." Big difference. The good news is that computers are cheaper now and much faster. Databases are much better. And best of all, if you spend 2 years developing a NN you can learn enough about the market to reach a point where you can say, "Hey, I can do this good with a price action or KISS system and enjoy my evenings and weekends!" Only then will you start to focus on money management, position size, and price action and maybe make some money. Maybe. Lastly, the biggest advocates of NNs have something to sell, or are in denial (halfway into the woods). Don't believe any 'historical testing' of NN as it is fake. I would only believe system results that were gained trading real dollars. It doesn't have to be large size, 100 shares of anything would be fine. Trading micro-pips at oanda is fine but it has to be something real. You won't find any vendor providing it or names of anyone who you can talk to not vested with the software provider. Somebody prove me wrong and post the result here please or where they can be found. I still have all of my work on the cold linux boxes over in the corner. Give me a reason to attach them to a monitor!!! Sorry for the Hershey length post!!!! Good luck!
Looks like people here use NN and know lots about it. In my opinion spot a trade first and then use TA, and not the other way around. So, I have a problem, i know the solution but i am missing the tool. Wonder if NN can be used. Do you 'train' NN on subset of past data in hope that tricks learned will predict movements ? Say simple example, on 11 mondays there is a gap down and you get uptrend 8 weeks, therefore on next gap down NN will spit out a buy on monday open ? Or is there more to it ?
Erroneous prices happen in real time, too. Without a clear method for dealing with them in real time, I would be hesitant to scrub anything.
To understand why NN might not work, look at this first http://www.elitetrader.com/vb/showthread.php?s=&threadid=240936