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

RE: Neural Networks



PureBytes Links

Trading Reference Links

>>  Building nets isn't easy and it's no holy grail, if you can define your model 
in a way that the net can analyse it you can get results but it's not magic.
The problem is there is just too much noise in most markets and that throws
of the net.  Any good net will have to have a good noise filter or it will
never come close to giving you  an edge.<<

Very true.

As markets become increasingly efficient, the persistence of 
information diminishes, shortening the forecast horizon.  
However, short forecasts are degraded by market noise, and 
so I see a squeeze play in action.  

To alleviate some of the squeeze, good noise filtration is essential.
The problem is that classical frequency low-pass filters are not
optimal for filtering market noise.  This is because filters 
designed for a Gaussian distribution of price changes do not
properly handle large price gaps, a real aspect of price action. So, 
more advanced filters are needed.

Even so, I tell clients not to train nets to forecast a target as
unstable as price.  Rather it's better to find some quantitative 
analysis in your current trading system that could use improvement, 
for example, a better estimate of support and resistance.  Then 
train a neural net to produce this better estimate and use it 
in their existing system.

As for Linda Rashke's assessment of neural nets ... did she really 
think that anyone with a killer net is going to spill their guts to 
her, and expose their competitive edge?  

- Mark Jurik










Attachment Converted: "f:\eudora\attach\RE Neural Networks"