Chapter One — 1.1 Background to the Study

The migration of banking and commerce onto mobile and internet platforms has lowered barriers to financial participation and compressed settlement times, but the same open architecture that makes those gains possible has handed determined criminals an ever-widening surface to attack. Financial fraud has not merely persisted through this digital transition; it has accelerated with it, with occupational fraud losses estimated at roughly five percent of annual revenues across organisations worldwide, adding up to hundreds of billions of dollars each year.

For a long time, the financial sector's answer to these threats was the rule-based expert system — a transaction above a certain amount, from an unusual location, in a suspicious merchant category, gets flagged. That approach can only catch what somebody already thought to write a rule about, generates a heavy volume of false positives, and increasingly cannot keep pace with transaction volumes scaling into the billions per day. Machine learning offers a different kind of answer, learning what normal behaviour looks like directly from data rather than being told in advance, and adaptive systems — drawing on reinforcement learning, online learning, and deep learning to update themselves continuously — have shown they can sustain detection capability without needing human-initiated retraining cycles. Despite this evidence, most deployed fraud detection systems, including machine-learning-based ones, remain static once trained and deployed.

1.2 Statement of the Problem

Rule-based and heuristic fraud detection systems cannot learn without human intervention, becoming outdated as criminals find ways around their thresholds while imposing heavy operational burdens through high false positive rates. Machine learning addresses many of these limitations, but a widely overlooked problem remains: most deployed ML-based fraud detection systems are themselves static, trained on historical data and left to run without any provision for ongoing learning, degrading reliably over time in a domain where transaction patterns and fraud typologies shift continuously. A further, less-discussed problem sits at the architectural level — individual elements of adaptive detection such as online learning, feedback loops, drift detection, and ensemble methods have each been studied separately, but no complete, end-to-end architecture puts these pieces together into a single deployable system. This project addresses that gap directly, developing and evaluating such an architecture.

1.3 Objectives of the Study

1. To design and implement an adaptive fraud detection model that integrates supervised, unsupervised, and reinforcement learning within a continuous learning framework.
2. To train the developed model on a publicly available financial transaction dataset and optimise its parameters for fraud detection performance.
3. To evaluate the trained model against established performance benchmarks using standard metrics including AUC-ROC, AUC-PR, precision, recall, F1-score, and false positive rate.
4. To deploy the model within a simulated fraud detection pipeline and assess its operational performance under conditions of temporal drift and evolving fraud patterns.
5. To compare the performance of the adaptive model against static baseline models to quantify the practical benefit of continuous learning architectures in financial fraud detection.

1.4 Research Questions

1. How should an adaptive fraud detection model be designed and implemented to integrate supervised, unsupervised, and reinforcement learning within a single continuous learning framework?
2. How does the trained model perform against established benchmarks (AUC-ROC, AUC-PR, precision, recall, F1-score, false positive rate) on a real financial transaction dataset?
3. How does the model's detection performance hold up under simulated conditions of temporal drift and evolving fraud patterns?
4. What practical benefit, if any, does the adaptive model demonstrate over static baseline models in financial fraud detection?

1.5 Significance of the Study

Theoretically, the study contributes a unified adaptive architecture for financial fraud detection that brings together supervised, unsupervised, and reinforcement learning within a single continuous learning system, operationalising a conceptual framework for adaptive detection into a fully specified architecture with empirically grounded component choices. Practically, the work is relevant to any institution handling financial transactions at scale: a system that updates itself as fraud patterns change is qualitatively different from one that requires periodic manual retraining, and its embedded explainability mechanisms speak to the growing regulatory expectation that automated financial decisions come with justifications a human can examine and audit.

1.6 Scope of the Study

The project concerns fraud in digital transaction environments; physical fraud detection falls outside its scope, and the system design is institution-agnostic rather than built for any specific bank or payment processor's infrastructure. Regulatory frameworks such as the GDPR and CCPA are addressed from an engineering design standpoint rather than a legal one. Three limitations are acknowledged: public benchmark datasets, though standard in this field, may not fully replicate the complexity of live institutional transaction data; model evaluation takes place in a simulated environment rather than a live production system, which may behave differently; and fraud evolves quickly enough that findings which hold today may need revisiting as new attack methods emerge.