Forecasting stock returns using variable selections with Genetic Algorithm and Artificial Neural-Networks

Abstract

Modeling stock returns requires selections of appropriate input variables. For an Artificial Neural Network, the appropriate input variables have both linear and nonlinear functional relationship with stock returns as output variables. To capture the non-linear relationships, we propose Weierstrass theorem. However, to estimate the relationships for all possible combinations of input variables, especially for a large set of variables, is too numerous for a simple exhaustive search thus we use a Genetic Algorithm to approximate the non-linear relationships between the prospective input variables and the output variables. The result shows that the Artificial Neural Networks with the selected variables based on both linear and non-linear relationship perform better than the ones with all possible variables for all but one out of the sample of ten US stocks.