FraudBench-NG: A NIBSS-Calibrated Synthetic Dataset and Generator for Nigerian Inter-Bank Fraud Detection
Research on inter-bank fraud detection in Nigeria is limited by the lack of shareable transaction data, and synthetic benchmarks can mislead if they leak information or hide their assumptions. FraudBench-NG is a reproducible framework for studying fraud detection under explicitly documented modelling assumptions. It is a parametric simulation calibrated to 2023 NIBSS aggregates, with a released dataset of 2.1~million transactions from 25{,}000 customers and 0.1\% fraud. Fraud is generated as incidents, matching NIBSS counts and loss per case; every parameter is marked as NIBSS or assumed; and every behavioural feature uses only prior transactions, verified by row-by-row recomputation and a label-permutation test. Logistic regression, random forest and XGBoost reach a PR-AUC of 0.016--0.036 (no-skill 0.0007), with wide customer-level bootstrap intervals, so the models cannot be ranked. At a 1\% false-positive rate they raise about 70 false alerts per fraud row caught and flag only about 19\% of incidents at their first transaction. Results depend strongly on the assumptions: a model trained with takeover bursts transfers only partly to data with a different burst probability, and performance differs between burst settings mainly because of how an incident's loss is split across its transactions. Fraud prevalence is a design choice that cannot be compared with the NIBSS case-to-transaction ratio, and no real Nigerian transactions were available for validation.
Authors
- Otu Jacob
- Colin Kuka
- Adam Knowles
- Cyrus Orji
Institutions
- Wrexham University (GB)
- Imo State Polytechnic Omuma (NG)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-18
- DOI
- https://doi.org/10.21203/rs.3.rs-11431606/v1
- Primary Topic
- Imbalanced Data Classification Techniques
- Type
- preprint