Quote:
Originally Posted by Egomancer
The ones I have worked with either did not have those features, or if they had them, I didn't see them. If they do it, and if it is done properly, indeed that is some heavy stuff ;).
|
Most don't have it, but I can tell you, when it works it's magic.
The biggest problem with features like that is that they have to work with a changing market.
So, for example, you train the learning algorithm with data from the past 10 years excluding the very last year, then test it on the data of the very last year to see if it works. If it works, you can include data from the past year, and implement it. The more the market has changed, though, the less effective it typically is.
That problem isn't too big - the real problem comes up when changes in the market speed up. The software has to learn to interpret both changes in the market, as well as the speed of those changes.
In a market that's static or changing slowly, you typically want it to use as much data as possible. In a rapidly changing market, it might be better to use only the data of the past few months, rather than the past few decades.
The challenge is to have the software learn about the data, as well as about the metadata, and make predictions based on the combination of both.
It's seriously interesting stuff, once you get into it.