|
PureBytes Links
Trading Reference Links
|
Neural nets/ fuzzy logic = an illusional composite of statistically
insignificant periods, all stuffed inside a shiny package.
>> Has anybody managed to come up with tradable intra-day system using neural
>> networks or fuzzy logic (for E-min S&P maybe).
TBR> I have read about 3 NN projects for index price prediction in the
TBR> literature, 2 of the 3 claim to have achieved profitability.
TBR> There is a person on the realtraders list that is using NNs.
TBR> Seems to me that he is always switching from one net to the other
TBR> based on what works this week. Not very confidence inspiring.
TBR> Many claim NNs are nothing but overoptimisation engines. IMHO
TBR> there is much truth in this, but it may not be such a bad thing.
TBR> I would be quite comfortable with an NN that required weekly
TBR> retraining (optimization) given sufficiently long term profitable
TBR> backtesting.
TBR> I can think of at least 10 others that have bought such software,
TBR> none of whom use it today. I know of no one using a successful
TBR> intraday system based on a NN.
TBR> Some time ago I came up with approx 8 data series that I was
TBR> pretty sure held important factors that influence prices
TBR> intraday. It seemed to me that tweezing the significant
TBR> relationships out of this data would take me months and was just
TBR> the kind of job neural nets should be good at. I bought BrainCel,
TBR> spent quite a bit of time reading up on neural nets, got
TBR> everything working, found out that the initial nets gave REALLY
TBR> bad signals and it would take several months of CPU time to
TBR> search for the best net (assuming I got the data representation
TBR> perfect first time, not bloody likely!). Some might claim genetic
TBR> optimisation can avoid this time constraint, but this approach
TBR> does not fully explore the solution space.
TBR> IMHO...
TBR> Garbage in - garbage out is a top level design issue for such
TBR> systems. Randomly applying indicators and expecting the software
TBR> to generate a profitable net is unlikely to work, no matter what
TBR> the vendors say. There is no free lunch in this business.
TBR> Data representation is a key issue. Be ready to put significant
TBR> time and effort into learning about this.
TBR> NN vendor's claims should be viewed the same as claims made by
TBR> trading system vendors. Make sure you get a money back trial
TBR> period for any software you buy. If you haven't developed NNs
TBR> before a 1 month trial period will not be long enough for you to
TBR> even get your feet on the ground.
TBR> regards,
TBR> tbr
|