A semantic-based, distance-proportional mutation for stock classification

Abstract

Decision making in the field of financial management is a very complicated and dynamic process. How to incorporate domain knowledge into evolutionary computing to solve financial decision-making problems has long been an interest of researchers. This paper investigates the use of domain knowledge in an evolutionary process, especially in the mutation process. A semantic network of financial attributes is created and used to measure the variation between parents and offspring introduced in the mutation process. The proposed distance-proportional mutation (DPM) constrains the mutation size to be a) small enough that the searching proceeds gracefully, while b) large enough to avoid being trapped into local optima. The hypothesis is that the DPM outperforms a random mutation or a constrained mutation in which only the component that is the closest to the one being mutated can be selected, and provides a better decision-making support for the stock classification problem. Experiments were implemented to test the hypothesis. DPM is also compared with other classifiers, such as decision trees. The results support the hypothesis and shed light on future directions to further delineate the theory of how evolutionary computation can gradually build on the body of human knowledge.