The increasing availability of retail data offers substantial possibilities to ascertain customer behavior and facilitate data-driven decisions through predictive analysis and modelling. This research uses a derived retail dataset consisting of 200 customers, described with ten principal attributes, including demographic, financial, and behavioral signals, for the intent of customer profiling/demand forecasting. Prior to the modelling stage, various preprocessing procedures such as missing value management of the observed data, normalizing numerical features, and encoding categorical variables, were established to assist learning effectiveness. Various ensemble learning methods for modelling- Bagging Tree, Boosting Tree, Random Forest, Extra Tree, & Voting Classifier, were employed to examine model validity & predictability. Experiment processes showed that the Voting Classifier produced the best results overall accuracy of 1.00 for all overall metrics, outperformance all previous models. Bagging Tree methods were held valid with 0.97 accuracy and a 0.96 F1 composite, compared to the Boosting method which demonstrated inferior capability to generalize. As illustrated by the results, ensemble fusion certainly stabilizes classification in easily interpreted metrics even when moderate sized data is employed in modelling procedures. This type of analysis gives a comprehensive base for data driven strategies in retail for effective customer segmentation, targeted customized communications, or reliable demand forecasting to make informed decisions in competitive settings.