ABSTRACT The proliferation of mobile banking has been accompanied by a surge in sophisticated financial fraud, necessitating detection systems that go beyond traditional methods. This paper designs and validates a multi- faceted, hybrid machine learning framework that synergizes behavioral biometrics (e.g., typing speed, swipe patterns) with transactional data (e.g., amount, geolocation) for high-accuracy fraud detection. We evaluate the progression of models, demonstrating that while unsupervised autoencoders are effective at profiling normal behavior, they fail to detect over 65% of fraudulent activities. Supervised Long Short-Term Memory (LSTM) networks, capturing temporal sequences, significantly improve performance, achieving fraud recall rates as high as 97%. However, gradient-boosting models (LightGBM and XGBoost) yield the most balanced standalone performance, with 98% recall and 94% precision. Feature importance analysis from these models confirms that the framework's predictive power is derived from a hybrid of both behavioral and transactional features. The framework culminates in a stacked ensemble model that optimizes the precision-recall trade-off, achieving 97% accuracy, 97% fraud recall, and 95% fraud precision. This final model registers the lowest false positive rate, presenting a robust, reliable, and deployable solution that maximizes fraud capture while minimizing unnecessary friction for legitimate users. KEYWORDS: Fraud Detection, Mobile Banking, behavioral Biometrics, Transactional Pattern.