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

Re: Update on the NeuralFuzzy Project



PureBytes Links

Trading Reference Links


To those interested in applying Neural Nets to the market:
> 
> Hi Tom:
> 
> Years ago, I tried 1. Market Mirror (pattern recognition software); 2.
> BrainMaker (neural network); 3. ModelWare (based on some advanced modelling
> algorithm, purported to be of great predictive value).  I spent thousands
> of hours on each, and never got anything out of any of the three programs.
> 
> My conclusion:  they may be great for describing today's prices, but very
> bad for predicting future prices; or maybe I'm very stupid and don't know
> how to use them to my advantage.
>
Apart from NN, I was astounded to find that the set of indicators I
had created (from NN methods) had predictive strength of 1 to 2 hours to
over the next morning.  I THOUGHT I would be getting 1 - 3 day predictors.
Check your time-frame.
 
> Maybe because my background is not math, programming or engineering
> oriented.  Or maybe I have limited IQ.  Anyways, it was a very
> disheartening experience.
> 
> A few years ago, I talked to Mark Jurik over the phone, and we discussed
> neural network application to stocks and futures.  I told him I no longer
> believed in nn.  He said he had good results.  (If you visit his website,
> some of his products are nn-related.)
>
This is not to brag, but to reflect ...  I started studying Neural
Net technologies in 1962 (Perceptrons, they were then) with K.S. Narendra
at Harvard (He's now at Yale).  A couple of years later, I developed
an algorithm to compile arrays of some hardware devices developed by
Analog Devices in Cambridge into what-worked-like NN circuits.  In my
post-doc with D. Gann at CWRU, we developed Boolean Algebraic models
which work someone like NNs in order to model unsee-able physiologic
systems.  I taught AI and NN technologies to my grad students.

So why DIDN'T I use NNs when I decided to see if I could model the
markets?  (1) Dynamics are difficult to capture in an understandable
way using NNs.  Not that NNs can't DO dynamic models, but it is difficult
to dissect a dynamic NN model to grasp what it is doing.  It is somewhat
easier (for us Boolean Modelers  :=}}  ) to extract the LOGIC of an NN
that is static.  (2)  NNs have this certain problem of producing an
output even when they have no prior experience ("training").  I was NOT
going to put my money on the line with a predictive model that couldn't
tell me what is was doing, why it was doing it, nor assure me that it
wasn't a momentary wild twitch of its neurons that drove its decision.

 
> I never pressed on, and he never pressed me either.
> 
> I'd rather stick to the old traditional indicators that I'm familiar with,
> like moving average, RSI, Stochastic etc.
> 
> Anyways, thanks for sharing your experience with us!
> 
> 
> Regards,
> 
> Wong

Having set the NN table straight above, let me say that I HAVE found
certain steps in my market modeling process where NNs could be VERY 
helpful.  These one ones where a mathematical constraint limits the
scope/range/nature of the inputs to the NN.  Why not NNs here?  Simply
that I am re-engineering my system to run in real-time (every tick, perhaps)
instead of once every 15 minutes.  When that's done, then I'll look at
the literally hundreds of improvements/questions/issues that arise from
a totally new approach to market modeling. (Well, if anyone HAS done it
before, they haven't been talking about it, eh?).

Cheers,
Robin Lake
Environmental Modeling Inc.
rbl@xxxxxxxxxxx