A Multi-Objective Evolutionary Wrapper for Feature Selection across Heterogeneous Financial Risk Domains

Feature selection in financial risk modeling is a complex discrete combinatorial optimization problem. We propose a multi-objective evolutionary wrapper using NSGA-II to explore the global Pareto Front, simultaneously optimizing predictive power and structural parsimony. Evaluated across ten heterogeneous financial datasets, the methodology demonstrated robust generalizability. In the US Stocks benchmark, navigating a 2178 search space, it identified a Pareto-optimal
core of 36 variables (a 79.8% dimensionality reduction). Applying a Multi-Criteria Decision-Making (MCDM) schema, the evolved subset outperformed local search heuristics. Constrained to 36 features, Recursive Feature Elimination (RFE) achieved a 0.310 Gini coefficient, whereas NSGA-II reached 0.345, proving the discovery of non-linear synergies. Furthermore, probability calibration maximized the F1-score at a 0.3 decision boundary (AP=0.7157). By outperforming
greedy methods, this framework provides a mathematically grounded decision support system for deploying low-complexity AI in high-stakes finance.