Explainable Artificial Intelligence for Real-Time

Financial Fraud Detection: A Systematic Literature Review

Ulugbek Inoyatov1,* Eugene Q. Castro1

1 Department of Computer Science, Central Asian University, Tashkent, Uzbekistan

Received: February 02, 2025 R e vis ed: April 04, 2025 A c cep ted: July 05, 2025 ⋆ C o rre sponding author

ABSTRACT

Financial fraud detection systems increasingly rely on machine learning to identify suspicious transactions at

scale. However, the opacity of many high-performing models raises significant concerns regarding trust, regulatory

compliance, and practical deployment in real-time financial environments. Explainable Artificial Intelligence

(XAI) has emerged as a promising solution to enhance transparency and accountability, yet its feasibility under

real-time constraints remains unclear. This systematic literature review examines empirical studies on explainable

AI approaches for financial fraud detection, with explicit focus on real-time applicability. Following PRISMA

guidelines, nineteen peer-reviewed empirical studies were selected and analyzed based on fraud domain, model

type, explainability technique, evaluation metrics, and evidence of real-time performance. Results show that posthoc

explanation methods, particularly SHAP and LIME, dominate the literature, while intrinsic explainability

and deployment-level latency reporting remain limited. Despite frequent claims of real-time applicability, only

one study provides quantitative runtime evidence. The findings highlight critical gaps: absence of explanation

latency evaluation, lack of deployment-oriented validation, and insufficient regulatory compliance integration. This

review reveals a systematic disconnect between real-time claims and empirical evidence, establishing the need for

standardized latency benchmarking in explainable fraud detection research.

Keywords: Explainable AI Financial Fraud Detection Systematic Literature Review PRISMA Real-Time Systems

Model Interpretability

1. INTRODUCTION

Financial fraud poses a persistent and evolving threat to banking

institutions, payment systems, insurance providers, and

digital marketplaces. As transaction volumes grow and fraud

strategies become increasingly sophisticated, machine learning

(ML) techniques have become central to automated fraud

detection systems. These models are capable of identifying

complex patterns in large-scale transactional data and have

demonstrated strong predictive performance across various

fraud domains, including credit card fraud, insurance fraud,

and online payment fraud.

Despite these advances, the deployment of machine learning

models in financial decision-making contexts introduces critical

challenges. Many high-performing models, particularly

ensemble methods and deep learning architectures, operate

as black boxes, offering limited transparency into how predictions

are generated. In regulated financial environments, such

opacity is problematic. Financial institutions are required to

justify automated decisions to regulators, auditors, and customers,

while fraud analysts must understand model outputs

to validate alerts and minimize false positives. As a result,

accuracy alone is insufficient; explainability has become a

fundamental requirement for trustworthy fraud detection sys-