Design and evaluation of machine learning - based framework For fraud detection in Kenya’s digital payments ecosystem

Abstract

The rapid growth of Kenya’s digital payments ecosystem has revolutionized financial transactions, enabling greater financial inclusion through fintech innovations such as Point-of-Sale (POS) systems, mobile wallets, and e-commerce platforms. However, this expansion has also increased exposure to fraud, as cybercriminals exploit digital vulnerabilities and high transaction volumes to execute increasingly sophisticated schemes. Conventional rule-based fraud detection systems, which rely on static thresholds and predefined patterns, have proven inadequate in addressing evolving fraud tactics. They often result in high false-positive rates and delayed responses, ultimately compromising customer trust and financial integrity. This study presents the design and evaluation of a machine learning based framework for fraud detection tailored to Kenya’s digital payments ecosystem. Using anonymized transaction data from Pesapal Ltd, a leading regional fintech provider, the research applies a range of supervised learning algorithms including Random Forest, Gradient Boosting, Logistic Regression, Naïve Bayes, Decision Trees, KNearest Neighbors, and Neural Networks to identify the most effective approach for real-time fraud detection. The study follows the CRISP-DM (Cross-Industry Standard Process for Data Mining) methodology and incorporates feature engineering, Synthetic Minority Oversampling Technique (SMOTE), cost-sensitive learning, and explainability techniques to enhance model robustness and interpretability. Model performance was evaluated using precision, recall, F1-score, AUC, and costbased metrics to capture both statistical accuracy and business implications. After applying SMOTE, Random Forest demonstrated a superior balance between detection sensitivity and false-positive control, achieving Precision = 52.54%, Recall = 52.77%, F1-score = 52.66%, and PR-AUC = 53.55%, outperforming other models in identifying fraudulent transactions. The SHAP-based explainability analysis further highlighted the dominant role of transaction geography, processing bank, and card type in predicting fraud. The results highlight the effectiveness of ensemble learning techniques for fraud detection in imbalanced financial datasets and demonstrate the value of explainable AI (XAI) in enhancing transparency and regulatory compliance. The study contributes to the ongoing discourse on financial cybersecurity in African fintech ecosystems by offering a scalable, interpretable, and contextually relevant fraud detection framework suited for real-time digital payment environments.