Thursday, May 31, 2018

Why don't Neural Nets Work?

There have been a few high visibility funds that have proclaimed they are going to trade on signals from the neural network type of artificial intelligence. Several of those have not done well, and I hear that there will be some other disappointments as well. So why isn't this technique working? After all, it has had great success in image and speech recognition, and is making some progress in language recognition.

There's been so much hype in the popular press about AI that it seems journalists have suspended their critical thinking. Here's the story. Virtually all machine learning, include the neural nets, are basically automated statistics. The algorithms range from simple curve fitting to extremely complicated, but they are all about finding regular patterns in data. For example, suppose you want to identify an image of a cat as a cat. The algorithm takes the pixel information and performs a very complex optimization that essentially relates every pixel to every other pixel. It gets fed in data for possibly millions of images of cats (and non-cats) and finds the optimal relationship that makes the images "cat-like." This can work pretty well because cats don't change much. They may get bigger or smaller or change color, but there is something about their look that remains cat-like. In stat terms, the basic look of a cat is a stable system.

If the system thats being analyzed is not stable, this will not work. Liquid assets returns are not stable. There's been a lot of work on this. They were somewhat statistically stable years ago, and that's why the systematic guys were able to print money for so long. But these have been arbitraged out. Not only are asset returns not stable, but they are like a game in which the players are intentionally working to destroy whatever stability there is.

OK, enough of my being Debbie Downer. There are regularities in liquid assets returns that can be traded. The key is to look for regularities that are hard to arbitrage with leverage. People who have read this blog for awhile know that I believe long-term trades are included in this. There are some others as well, which I'll go into as time goes on.

1 comment:

  1. Maruti Suzuki gains over 2.5% as May 2018 volume packs a punch, up 26% yoy
    mcx tips

    ReplyDelete

Comments are welcome, although I can't promise to answer every question.