Preprocessing is a key step in boosting the performance of machine learning models. Among the most important techniques are outlier removal and feature selection, which help eliminate irrelevant, redundant, and noisy data—making the model more efficient and accurate. Feature selection starts by identifying the most relevant features and then evaluating their impact on the model. Counterfeit banknotes remain a serious threat to economies, and with the rise of AI, detecting fake currency has become even more challenging. In this paper, we explore two feature selection methods—correlation-based selection and best-first search—applied to the banknote authentication dataset. These are followed by an ensemble evaluation using bootstrap aggregating. We tested five machine learning models, which are Random Forest, Logistic Regression, Multilayer Perceptron, Support Vector Machine, and K-Star. Among them, K-Star delivered the best performance, achieving 99.7% accuracy. The results clearly show that the proposed preprocessing steps significantly improve classification performance.