Genetic Algorithms |

Genetic Algorithms

I had an interesting set of meetings and conversations this past week. I was presented with a tool to mine for useful factors and combinations that may have predictive value. The general idea is to borrow from evolutionary research into natural selection and survival of the fittest. In this case, instead of “genes” you have factors, instead of mutations you insert different factors. I’m happy to explore the idea further, but I want to highlight a few issues that kept coming up for me:

1. Locally optimal vs. globally optimal: for those of you coming from the sciences you’ll recognize this problem. In essence, there is a path dependence for mutations. Once you start down a specific path, you can improve on that mutation, but it’s extremely difficult to mutate a completely different structure from that original path. An common example of a locally optimized structure is the eye. A locally optimized trait can be traced back more easily and usually you’ll be able to find other structures that would compete. A globally optimal structure is more challenging. It is often used by those challenging evolution, since it implies predetermined design and there can be no improvement based on the function; a common example used is the structure of a beehive.

Testing factors, there is an issue in recognizing globally optimal vs. locally optimal factors a priori.

2. Predictive value: Genetic algorithms are based on successive “generations” of factors with only the strongest surviving. It is usually tested with a stable set of conditions, i.e. within a certain time period. The obvious issue develops of the predictive value of success. In biology, a good example is the evolution of specialized dinasaurs that were optimally evolved for their environment. Unfortunately, the ones most specialized had the least chance of survival. Thus, being strong in one period may actually increase your risk in successive periods. So the algorithm needs to factor in a) predictive value in successive periods and b) some value of robustness and ability to deal with change (extremely difficult).

3. Data mining: As with all models that look for successful factors, the danger is always in data mining, but this is a common danger and if nothing else, the tool can be used to uncover relationships that are then examined separately.

4. Tradeoffs in the model: As with all models, there are tradeoffs that need to be made. Finding a more optimal solution often sacrifices robustness, for example. Another instance is the data availability and ability to keep data virgin for testing.

In all, I think there is a lot of potential to use scientific modeling techniques and lessons when looking at financial problems. Let’s discuss.

Last 5 posts by Yaron Sadan

Calling all themes - February 10th, 2012
Portugal, and Yields, and Iran . . . - January 30th, 2012
The Hunt for Yield - January 26th, 2012
All eyes should be on Iran - January 23rd, 2012
Quick note - January 19th, 2012