Using credit cards is a fundamental way for financial transactions. However, their widespread use also exposes users to significant risks, particularly with the increasing prevalence of credit card fraud. To address this challenge, financial institutions and their customers must continuously adapt, as fraudsters develop increasingly sophisticated tactics. In line with this need, we have developed an effective fraud detection method emphasizing the selection of the most relevant features. Our approach leverages both bio-inspired techniques, such as genetic algorithms such as bee colony and ant colony, alongside classical methods, including sequential forward selection, Pearson’s correlation, and mutual information. This research revealed that bio-inspired methods outperform classical feature selection techniques in identifying the optimal number of features for fraud detection. These selected features were subsequently used to train various ensemble learning models, including AdaBoost, Bagging, Gradient Boosting, and Stacking. Among these, the Stacking model consistently demonstrated superior performance. Notably, when the number of features ranged between 10 and 20, models based on bio-inspired methods achieved higher accuracy, F1-score, and area under the curve (AUC). However, when the feature count dropped below 10, models utilizing features selected by classical methods also delivered competitive performance, showcasing their effectiveness in scenarios with limited feature sets.