Training Financial Decision Support Systems with Thousands of Decision Rules Using Differential Evolution with Embedded Dimensionality Reduction

Abstract

This paper proposes an improvement of the training process of financial decision support systems, where evolutionary algorithms are used to integrate a large number of decision rules. It especially concerns the new computational intelligence approaches that try to replace the expert knowledge with their own artificial knowledge discovered using very large models from very large training datasets, where the large number of decision rules is crucial, because it defines the degree of freedom for the further learning algorithm. The proposed approach focuses on enhancing Differential Evolution by embedding dimensionality reduction to process objective functions with thousands of possibly correlated variables. Experiments performed on a financial decision support system with $5000$ decision rules tested on $20$ datasets from the Euronext Paris confirm that the proposed approach may significantly improve the training process.