Pattern recognition nets work via reinforcement. Lets say you plugged in 50 characteristics of an Apple, Orange, Grapefruit, Lemon and Banana. Now you present the net with the characterisitcs of a Tomato (yes, the Tomato is a fruit). The net will come back with a guess that what you have is an Apple, with a high (90% probability) assessment.
Now lets say you load only 5 characteristics of 50 fruits and then introduce the basic characteristics of a Tomato. It will come back with a lower probability answer that the fruit is like 2 or 3 closely related fruits, but will be more correct than the former example.
Then you will have to take into account the learning algorithms. Lets say you load in "long nose, thin tail, gray mammal" and you define an elepahnt. If the correct response is mouse, the system needs to learn that, and equally weight the potential responses or be given differentiating characteristic (large/small in this case). However, since nets, like human minds, will give more weight to earlier data, it is important to only add characteristics as necessary and eliminate less useful ones. Multiple occurrance reinforcement (lets say the last ten correct responses were mouse) will also weight the net in favor of that response.
So to the point of using dozens of inputs, I would caution that one is not adding new characteristics, but different ways of looking at OHLC and Volume statistically manipulated. Running 10 years of data will only be helpful if the market is an orderly and repeatable thing. If you are trading GM or GE or AA it might prove quite useful, but if you are trading newer tech stocks, you will not find a correlateable pattern in YAHOO or AOL one yuear ago with last week.
lastshadow |
