[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: GEN Neural Networks are not magic



PureBytes Links

Trading Reference Links

Lawrence,
Thank you for an excellent summary on Neural Nets.
I first worked with these in 1991, and my opinion was,
and still is, that you need a lot of non-correlated inputs
to predict time series data.  Perhaps a better use would
be working from chart patterns, even P&F.
Recently, however, there was a post to the Omega-list
claiming to have found the Holy Grail trading the S&P
with NNs.
I would now like to do some testing with genetic algorithms.
If you have any suggestions for software, let me know.
Neal 

----------
> From: Lawrence Lewis <lel@xxxxxxxxxxxxxxxx>
> To: realtraders@xxxxxxxxxxxxxx
> Subject: GEN Neural Networks are not magic
> Date: Friday, April 24, 1998 10:00 AM
> 
> I have spent the last several years studying neural networks and their   
> use in predicting prices and developing trading systems. This includes a   
> couple of graduate courses at University of Washington, purchase of a   
> commercial neural network package, several hand written programs, and   
> subscription to a neural network trading service. For anyone interested,   
> I thought I'd give you a quick summary of what I've found so far. The   
> following is not mathematically rigorous.
> 
> 1.  The typical neural network used in trading systems has nothing to do   
> with simulating the way humans think. It doesn't adapt or learn on the   
> fly. It simply computes a non-linear mathematical function of its inputs   
> that produces a predicted output. The neat part is you don't have to know   
> what function you need to compute, There are algorithms which find a   
> function that fits a set of training examples that are given to it during   
> a training phase. For those of you who hate curve fitting, neural   
> networks aren't for you. That's exactly what they're intended to do.
> 2.  Neural network fitting algorithms can approximate non-linear and   
> chaotic functions, and are tolerant to noise in the training and test   
> data. This makes them suited to price series, since I think most traders   
> believe that prices are chaotic functions with a noise component. Chaotic   
> doesn't mean random - chaotic functions are completely deterministic. It   
> typically means a function that is dependent on prior values of itself   
> and is extremely sensitive to its initial value. I believe SP500 prices   
> are chaotic with noise because I believe they are dependent on some of   
> their prior prices and have noise. Typical chaotic functions can look   
> random.
> 3.  Training a neural net involves optimizing a set of parameters (called   
> weights) in order to generate a function which closely matches the   
> training data. For those of you who shrink at the dreaded thought of   
> over-optimization, neural nets may be the ultimate evil. A neural net   
> that takes 5 inputs (say price, yesterdays price, price 2 days ago, price   
> 10 days ago, MACD), will typically have a minimum of 20 weights to   
> optimize (maybe many more). How many of you like trading systems with 20   
> optimized parameters? If you have 10 inputs, you might have 50 or more   
> weights. A great deal of research in neural networks tries to address   
> this problem. In the end, you train on a set of training data, which   
> calculates the weights, then test on a different set of testing data   
> using the previously calculated weights. The theory is, if it does well   
> on the test data, that what you learned by "fitting" was the underlying   
> mechanism of the market.
> 4.  One way to feel better about all this curve fitting is to have a good   
> understanding of some reasonably advanced statistics. This is probably   
> not a bad idea for trading systems developers of any kind. You need to   
> know whether your results could have been achieved by chance or not. You   
> need to know whether your system is still working if it has 5 straight   
> losing trades. Remember, completely random coin tossing will usually get   
> a run of 5 consecutive heads or tails 3 times during 100 tosses.
> 5.  Neural networks are not self designing! To create a neural network   
> trading system you, the human, have to a) select a set of inputs, b)   
> select a network architecture, c) decide how to initialize the weights,   
> d) select a set of activation functions, e) select a learning algorithm,   
> f) select an error function, g) decide how to pre-process the data, and   
> h) decide how to use the output to generate buy/sell signals. Oh...first   
> you have to spend the time to learn how to do all that. A number of   
> people have tried to use genetic algorithms that randomly generate a   
> whole bunch of different combinations of these things and then select out   
> the ones which work bes
> t. Talk about more optimization! It also takes a   
> ton of time.
> 6.  Neural networks clearly work for some applications. They are used   
> everyday in radar processing, chemical processing, optical character   
> recognition, physics labs, etc. That is, in places where there is no   
> question that they were able to learn an underlying mechanism from a set   
> of examples. In some cases, neural networks have solved problems that had   
> not been solved (at least in a reasonable amount of time) previously.   
> They are also use very successfully in a variety of financial   
> applications like credit scoring.
> 7.  On the other hand, most kids are pretty good at character   
> recognition. I'm not sure whether financial forecasting is more like   
> character recognition or more like radar processing. Humans are awfully   
> good at certain types of pattern recognition and awfully bad at others.
> 8.  I bought Biocomp Systems Neurogenetic Optimizer (NGO) and their   
> Profit add-on package (www.biocompsystems.com). NGO is a general purpose   
> neural network development system. Profit is really still a work in   
> progress, although it seems to demonstrate that it is possible to develop   
> profitable neural network based trading systems. Good quality software,   
> but I found that it was just a stop along my own personal journey. I   
> wanted to try things that were not yet supported in the software  and I   
> didn't want to wait. On the other hand, it's always nice to have a known   
> good test system to verify some ideas. I think most traders would benefit   
> from software like this if they're interested and willing to put in the   
> work to understand how it works (not just its operation, but some theory   
> behind it).
> 9.  I subscribed to Joseph Prewitt's TrendySystems   
> (www.trendysystems.com) neural net based trading system. I've watched it   
> since December 1997. It appears to just plain work as long as you are a   
> short term trader (1-2 days per trade), can accept 15 point stops (SP),   
> can accept being out of the market 2/3 of the time, and can accept not   
> knowing why it's doing what it's doing. But, it's still a short track   
> record in the best bull market of all time. I don't think I can stick to   
> a system when I don't understand the details of how it works.
> 10.  I subscribe to the Journal of Computational Intelligence in Finance   
> (formerly Neurovest Journal). Great journal, but not for the   
> mathematically challenged. Need at least undergraduate level math and   
> statistics. I have a math degree, but not a lot of formal statistics, and   
> it's tough for me. Try on Generalized Autoregressive Conditional   
> Heteroskedasticity for size.
> 
> Good Trading.
> 
> Lawrence E. Lewis (lel@xxxxxxxxxxxxxxxx)
> Director of Engineering
> Thrustmaster, Inc. (http://www.thrustmaster.com)
> 
> 
>  ----------
>