Applying An Adaptive Multi-Population Optimization Algorithm to Enhance Machine Learning Models for Computational Finance

Developing effective machine learning models to maximize the generalization capability of the learned data patterns while minimizing the risk of overfitting and also the ultimate computational costs of the trained models are very challenging yet significant tasks in many real-world applications. Due to the high complexity and huge amount of data involved in financial applications, the critical task of constructing appropriate machine learning models will surely become more challenging. Typically, data scientists or researchers may employ various optimizers to enhance the feature selection and/or tuning of the involved hyper-parameters in order to improve the overall performance of the obtained learning models. In this paper, a very flexible optimizer, namely the Adaptive Multi-Population optimization Algorithm (AMPOA), utilizing multiple populations with different search strategies is considered in which two of its variants are carefully evaluated with the other state-of-theart optimizers to enhance the feature selection process of the commonly used K-nearest-neighbor classifier approach on a set of 12 well-known benchmark problems. The first variant named the BAMPOA-R approach employs a rounding technique to determine on the binary decision of choosing an input feature or not for the underlying classifier while the other BAMPOA- T approach utilizes a specific transform function to decide on the feature selection. To demonstrate the feasibility of the proposed optimizer for financial applications, the BAMPOA- R approach with the best performance in the previous test is integrated with the Support Vector Regression (SVR) as the BAMPOA-R-SVR algorithm to compare against other regressors on 10 major financial market indexes. The evaluation results clearly demonstrate the remarkable performance of the proposed AMPOA approach to enhance the prediction performances of the underlying machine learning models through more vigorous feature selection for numerous real-world applications including those challenging problems in computational finance.