Evolutionary Methods in Financial Modeling

This study introduces the Genetic Tree (GT) meta-classifer, tailored for the challenges of noisy settings like those often found in Financial Machine Learning (FML). In a rigorous evaluation protocol, the study pits (white-box) GT against (black-box) TPOT, a notable Automated Machine Learning (AutoML) tool, with C4.5, a straightforward yet classic machine learning classifer, serving as a baseline. The research conducts 88 tests on each classifer, covering diverse areas such as economics, life sciences, and competitive games. Imitating the real-world conditions of FML, the evaluation injects large quantities of Gaussian noise into the datasets. These problems originate from a curated version of the UCI Machine Learning repository amended by two large fnancial market datasets. The evaluation protocol uses the F-Score, a binary performance metric for classifers computed using a 3-fold cross-validation approach, as the key performance metric. Results reveal that TPOT signifcantly outperformed in 20 tests and GT in 13, while the remaining 55 tests yielded no defnitive winner between GT and TPOT. These outcomes highlight the robustness of GT, even when compared to a more complex algorithm. Notably, GT and TPOT demonstrated similar performance metrics when tested on the FML-specifc and most other datasets under consideration. The study ends by emphasizing the pivotal role of feature engineering in machine learning, especially in noisy conditions, and questions the idea that higher computational complexity automatically results in superior performance.