Artificial Intelligence and Fraud Detection in Digital Financial Services: A Hybrid Analytical Framework

Abstract

Fraud in digital financial services is rare, adaptive and costly, and no single detection approach handles all three properties: supervised models learn known fraud patterns but miss new ones, anomaly detection flags unusual transactions regardless of whether they are fraudulent, and expert rules encode domain knowledge but are brittle. This paper proposes a hybrid analytical framework that combines the three through rank-based fusion and examines it in a simulation of 40,000 transactions with a 1.6% fraud rate, of which 19% follow a pattern absent from the training data. Under a 1% alert budget, the supervised model attains precision of 14% and recall of 9% but catches none of the novel fraud; the anomaly detector attains 19% precision and 13% recall and catches 15% of novel fraud; rules attain 11% and 7%; and the hybrid attains the highest precision (22%) and recall (14%) while retaining 9% recall of novel fraud. The paper explains why fusion improves on its components, how the component weights should be set from the organisation’s fraud mix and alert capacity, and how the framework supports the feedback loop in which analyst decisions retrain the supervised component and anomaly findings become new rules. Implications for alert management, model governance and the evolution of fraud detection as attackers adapt are discussed.