A Study of Variable Selection Within A Framework of Real-Coded Genetic Algorithm

Recently variable selection and parameter optimization are getting more and more important due to increase of use of big data. With regard to parameter optimization, Much attention has been paid to Real-coded Genetic Algorithms (RCGA) because of their good searching ability and high flexibility. As for variable selection, traditionally Akaike Information Criterion (AIC) or Bayesian Information Criterion (BIC) are used quite often as selection criteria. These criteria estimate the relative quality of analysis models for a given set of data. This paper proposes a new variable selection method applying RCGA. In this new variable selection, variances of genes in RCGA have an important role. This paper consists of 2 main parts. The first part is for explanation about a developing process of the new variable selection method and the second part is for application of the proposed method to an empirical study in financial field with a structural change model, which is one of discontinuous models. At the beginning of the first part, it is reported that variances of genes are proportional to square value of standard errors of the parameters corresponding to each of genes. Taking advantage of this finding, a new value is introduced. It is named I-value, which is calculated by dividing a squared value of an estimated parameter by a variance of a corresponding gene. I-value is utilized as a measuring indicator of the importance of explanatory variables in second part and it is found of usefulness.