EST. 2026

The Archive

Data Analysis · REF. TA-21578

A Hybrid Machine Learning Framework for Detecting Insurance Claims Fraud in Nigeria Using XGBoost and Anomaly Detection Techniques

Abstract

Insurance claims fraud in Nigeria — inflated repair estimates, staged accidents, duplicate claims filed across insurers, and fabricated documentation — is typically caught, if at all, by manual claims-adjuster review, a process that scales poorly against rising claim volumes and depends heavily on individual adjuster experience and vigilance. This project develops a hybrid machine learning framework that combines a supervised XGBoost classifier, trained on historical claims labelled as fraudulent or legitimate, with an unsupervised anomaly detection layer that flags claims resembling no prior pattern in the training data at all — a combination intended to catch both fraud that resembles known schemes and genuinely novel fraud typologies that a purely supervised model, trained only on what has been seen before, would miss. The framework is built and evaluated against a structured dataset of Nigerian motor and health insurance claims constructed to reflect realistic claim features — claimant history, claim amount relative to policy value, time-to-claim after policy inception, and documentation completeness — with fraud patterns seeded at rates consistent with industry estimates. Performance is evaluated on detection accuracy, false-positive rate, and the specific contribution of the anomaly detection layer: how many fraudulent claims it catches that the XGBoost classifier alone misses. The project's findings are directly relevant to Nigerian insurers under growing pressure to control claims-fraud losses without proportionally expanding manual review capacity, offering evidence on whether a hybrid supervised/unsupervised approach meaningfully outperforms a supervised-only model in a Nigerian claims context specifically.

Chapter One — 1.1 Background to the Study

The Nigerian insurance sector has expanded claim volumes considerably as motor and health insurance penetration has grown, placing increasing strain on the manual claims-adjuster review process most insurers still rely on as their primary fraud-detection mechanism. That process depends on an individual adjuster noticing inconsistencies in a claim — a repair estimate inflated relative to the damage described, a claim filed unusually soon after policy inception, documentation that doesn't quite match the incident narrative — and its effectiveness varies with adjuster experience, caseload, and the sheer volume of claims moving through the pipeline at any given time.

Machine learning approaches to claims-fraud detection have matured considerably in markets with larger historical claims datasets and established fraud-labelling practices, with supervised classifiers such as XGBoost achieving strong results on the kinds of fraud patterns well-represented in training data. That approach has a structural limitation, though: a supervised model trained only on previously identified fraud will not reliably catch a fraud pattern that has not yet appeared in its training data — precisely the kind of novel or evolving scheme that determined fraudulent claimants have an incentive to develop specifically because it hasn't been flagged before.

1.2 Statement of the Problem

Nigerian insurers face a claims-fraud detection gap on two fronts: the manual review process most currently rely on scales poorly against growing claim volumes, and even where machine learning has been introduced, a purely supervised model is structurally limited to catching fraud patterns resembling ones already seen. Neither approach alone addresses both the volume problem and the novel-pattern problem at once. This project addresses that gap directly, developing and evaluating a hybrid framework combining a supervised XGBoost classifier with an unsupervised anomaly detection layer specifically to catch both known and novel fraud patterns within a Nigerian claims context.

1.3 Objectives of the Study

  1. To examine current claims-fraud detection practices and their limitations within the Nigerian insurance sector.
  2. To construct a structured motor and health insurance claims dataset reflecting realistic Nigerian claim features and fraud patterns.
  3. To design and train a supervised XGBoost classifier for claims-fraud detection on that dataset.
  4. To design and implement an unsupervised anomaly detection layer to flag claims resembling no prior pattern in the training data.
  5. To evaluate the hybrid framework's detection accuracy and false-positive rate, and specifically the incremental fraud catches contributed by the anomaly detection layer beyond the supervised classifier alone.

1.4 Research Questions

  1. What are the limitations of current claims-fraud detection practices within the Nigerian insurance sector?
  2. How should a hybrid supervised/unsupervised machine learning framework be designed to detect both known and novel insurance-claims fraud patterns?
  3. How accurately does the resulting framework detect fraudulent claims, and at what false-positive rate?
  4. What proportion of detected fraud does the anomaly detection layer catch that the supervised XGBoost classifier alone would have missed?

1.5 Significance of the Study

This project is significant to Nigerian insurers seeking to control claims-fraud losses without proportionally expanding manual adjuster capacity, offering an evaluated hybrid framework rather than a general case for 'machine learning' as a fraud-detection solution. It also contributes to the applied machine learning literature on claims fraud, which has so far drawn mainly on markets with more mature claims-labelling infrastructure than Nigeria's, by evaluating a hybrid approach specifically against Nigerian claim characteristics and fraud patterns.

1.6 Scope of the Study

The project is limited to the design, implementation and evaluation of a hybrid XGBoost and anomaly-detection fraud framework against a structured motor and health insurance claims dataset built for this study, rather than deployment against a live insurer's production claims system. It covers model design, training and evaluation on that dataset, and does not extend to integration with an insurer's live claims-management system, real-time claim scoring at the point of submission, or claim types beyond motor and health insurance.

Chapters Two through Five, references and appendices are available for a one-time fee of ₦75,000.

Unlock Full Document