QACO: An adaptive multi-phase ant colony optimization approach for robust feature selection

High-dimensional financial data contains numerous redundant or irrelevant indicators, which significantly reduce prediction accuracy and increase computational costs. Although ant colony optimization (ACO) has been widely adopted for wrapper feature selection, its traditional form often suffers from premature convergence when the search space is large and complex. Inspired by the ecological dynamics of real ant colonies, we introduce QACO, a multi-phase ACO framework that explicitly models three biological mechanisms: (i) dynamic colony resizing that reflects the changing queen/worker ratio, (ii) phase-based pheromone updating that follows the colony's life cycle, and (iii) adaptive offspring generation that gradually freezes high-impact features while preserving diversity. Extensive experiments on ten real-world Chinese corporate bankruptcy datasets and four tree-based learners, using a rigorous 30-run ten-fold cross-validation protocol, show that QACO achieves the highest average accuracy and the narrowest confidence intervals, outperforming eight state-of-the-art meta-heuristics in 88% of 360 statistical comparisons. Ablation analyses indicate that dynamic colony sizing speeds up early convergence by a factor of 3.4, while phase-based pheromone updating enhances recall by up to 4.8 percentage points without losing precision. The bio-inspired mechanisms proposed can be easily integrated into any swarm-based optimizer, offering a reliable and reusable solution for high-dimensional feature selection in financial risk prediction and beyond.