Nerds on Wall Street: Math, Machines, and Wired Markets

8.7. Chromosomes for Forecasting Models

The preceding list suggests the design of a chromosome for a forecasting model to be developed using genetic and evolutionary techniques. It will consist of three segments, corresponding to the three main design components of the model: what it forecasts, how it forecasts, and a large and complex segment specifying the variables and transformations with which it forecasts.

Not all of the basic GA manipulations on these chromosomes will correspond to reasonable models. For example, a model that looked only at data from five years ago would not make much financial sense (even though it might have great statistical properties, it would be a fluke of data mining). These nonsensical chromosomes, such as those that don't look at the most recent data, can be weeded out in the mechanics of the GA itself by rejecting some products of crossover and mutation, or they can be penalized severely in their fitness function. It is a better use of computational resources to get rid of them early.

Variants on the chromosomes^[] used for the forecasting models, as seen in Figure 8.4, allowed for higher levels of flexibility. The simplest chromosomes assumed that the standard predictor variables used in the existing models were utilized, and only the transforms were adjusted. A more complex variant allowed the predictor variables to change as well as the transform. The most sophisticated chromosome allowed for variation in functional form, introducing ideas from ...