Explainable Fraud Risk Scoring

A research paper, currently in draft, written with Dorsa Salimi and Khadijeh Khodabakhshi Parijan, where I am the corresponding author. The question behind it: auditors already know which internal controls failed on a journal entry (segregation of duties, approval status, authorisation limits), so what happens when those controls become features of the risk model, next to the usual accounting red flags?

The pipeline scores a synthetic ledger of about 20,000 entries over 24 months on a Jalali fiscal calendar, with six injected fraud scenarios. It compares four models: rules traced to ISA 240, unsupervised Isolation Forest and LOF, and logistic regression with and without the control features. Features are point-in-time and the split is time-ordered, so no model ever sees the future.

On top sits an explanation layer. Claude receives only the triggered indicators and their weights, never ledger data, and writes a short explanation for each alert, which is then checked locally so it can only cite indicators that actually fired. I wrote the pipeline in Node.js with the maths (IRLS, Isolation Forest, LOF, PR-AUC, agreement statistics) implemented by hand and covered by tests, plus an Angular dashboard for reviewing alerts.