When a company becomes large enough to have a large customer base, a key concern is the development of a reliable credit scoring model to determine which customers should be granted or denied credit. Though adequate credit scoring systems are available from third parties, larger firms can potentially save millions of dollars by creating in-house solutions that are fine-tuned to their specific needs. Creating an in-house system is extremely time-consuming, requiring labor for data preparation, the selection of variables, searches for data correlations, and model fitting.
A much more cost-efficient approach is the use of genetic algorithms to construct a credit scoring model. Genetic algorithms follow the Darwinian approach of initially creating a random set of models that compete to find the best solution; the best solutions are then “bred” together by combining elements of each other’s models, or mutating through the insertion of random elements. This sparks another round of competition amongst a new generation of models, with the surviving models going through the same process of breeding and mutating. After several thousand generations and potentially millions of models, an optimum solution is created that can then be used for credit scoring. Given the benefits of high-speed computers, this massive modeling approach becomes an overnight process.
Genetic algorithms are capable of using hundreds of variables, so they are more likely to locate subtle predictors of credit scoring than a manually-created model that may only use a few dozen variables (at most). There is also little need for the labor of a statistician in constructing a scoring model, since most of the work is automated. Further, they can be updated regularly with minimal effort, so there is less risk of having the results of the scoring model decay over time.
The genetic algorithm concept is mostly confined to university research. However, the “Model” analytics tool is available from Massachusetts-based Genalytics (www.genalytics.com ). Genalytics has an excellent white paper on the subject, which is available at http://www.genalytics.com/products/model.htm