Evolutionary Algorithms Based on Effective Search Space Reduction for Financial Optimization Problems

Abstract

This thesis presents evolutionary algorithms incorporated with effective search space reduction for financial optimization problems. Typical evolutionary algorithms try to find optimal solutions in the original, or unrestricted search space. However, they can be unsuccessful if the optimal solutions are too complex to be discovered from scratch. This can be relieved by restricting the forms of meaningful solutions or providing the initial population with some promising solutions. To this end, we propose three evolution approaches including modular, grammatical, and seeded evolutions for financial optimization problems. We also adopt local optimizations for fine-tuning the solutions, resulting in hybrid evolutionary algorithms. First, the thesis proposes a modular evolution. In the modular evolution, the possible forms of solutions are statically restricted to certain combinations of module solutions, which reflect more domain knowledge. To preserve the module solutions, we devise modular genetic operators which work on modular search space. The modular genetic operators and statically defined modules help genetic programming focus on highly promising search space. Second, the thesis introduces a grammatical evolution. We restrict the possible forms of solutions in genetic programming by a context-free grammar. In the grammatical evolution, genetic programming works on more extended search space than modular one. Grammatically typed genetic operators are introduced for the grammatical evolution. Compared with the modular evolution, grammatical evolution requires less domain knowledge. Finally, the thesis presents a seeded evolution. Our seeded evolution provides the initial population with partially optimized solutions. The set of genes for the partial optimization is selected in terms of encoding complexity. The partially optimized solutions help genetic algorithm find more promising solutions efficiently. Since they are not too excessively optimized, genetic algorithm is still able to search better solutions. Extensive empirical results are provided using three real-world financial optimization problems: attractive technical pattern discovery, extended attractive technical pattern discovery, and large-scale stock selection. They show that our search space reductions are fairly effective for the problems. By combining the search space reductions with systematic evolutionary algorithm frameworks, we show that evolutionary algorithms can be exploited for realistic profitable trading.