This study addresses the challenge of optimal feature selection for digital modulation classification in complex wireless environments, where high-dimensional signal representations often lead to increased computational costs and diminished classification performance. To overcome these limitations, we investigate the application of the Equilibrium Optimizer algorithm for selecting informative and non-redundant features from engineered signal descriptors. A balanced synthetic dataset comprising 600 digital modulation signals, each represented by an 18-dimensional normalized feature vector, is generated using four common modulation schemes and various signal-to-noise ratios. We employ a progressive dimensionality reduction strategy, where the Equilibrium Optimizer iteratively identifies Pareto-optimal feature subsets of increasing compactness, followed by supervised learning using multiclass support vector machines. Experimental evaluation demonstrates that the proposed approach maintains or enhances classification performance with fewer features: accuracy remains stable at approximately 66–69% and the macro F1-score peaks at 66.54% with only 12 selected features, compared to 66.67% and 61.48% using all features. The analysis also reveals a marked reduction in feature redundancy and consistently high Fisher scores for subsets of 8–12 features, indicating the preservation of discriminative information and class separability. The computational overhead incurred by the optimization is moderate and does not outweigh the benefits of reduced dimensionality. These results highlight the effectiveness of the Equilibrium Optimizer algorithm in producing efficient, interpretable, and high-performing models for digital modulation classification in wireless communication systems.