Thanks LastShadow.
Great analogies.
I agree that there only so many ways to look at OHLC; however I think this would apply more to continuous indicators which are just variations on averaging and ratios.
But what about non-continous indicators such as candlesticks? I liked your fruit and animal analogies. But don't they apply more to more specific types of pattern recognition where there is a large universe of objects that you are trying to find, such as handwriting, voice recognition, etc. In trading aren't you really trying to find only two types of object, Buy, Sell. To use your analogy wouldn't that be more like trying to separate or distinguish two different classes in the animal kingdom, Mammals or Reptiles. If mammals are buys then you're not just looking for elephants, any tiger, lion, rhino, will do. Some may be better than others. In otherwords sometimes you might buy a stock because it is oversold, sometimes because it is breaking out. The difficulty lies in that they all have the similar traits, the mammals as well as the reptiles. Crocks have legs eyes and heart just as any mammal would. Same thing with buy or sell decisions. Superficially two stock charts can seem radically different, a mouse and an elephant, but they are both mammals.
Sometimes a mammal might be confused for a reptile, a rhinoceros is not so different than a dinosaur, a reptile-like class (now I'm really mixing up the analogy), but that's the point.
I think that to build successful nets you really have to define what you're looking for and hope that the inputs you put in really are distinguishing traits and not something meaningless ,such as having legs although that in itself might be useful in a very limited way because there are no two-legged reptiles that I know of off hand but there are a few mammals that do. Let's say having two eyes. That input might not help at all. The problem with using nets to trade stock is that the relationships are hidden otherwise why would we want to use a neural net if we know the right relationships to begin with.
There is no way to know which inputs are required unless you chance upon a net that works which brings me back to my original question, how do you deal with noncontinous indicators and their lags, when the universe is potentially so large?
I don't think a balance of power will work in situations where you can't reduce the types of patterns down to a handful. |
