An Explainable Machine Learning Framework with Large Language Model Features for Predicting Monetary Redress in Bank Fraud and Scam Complaints

Abstract Liability for scam losses is shifting towards banks, yet few interpretable tools can anticipate which fraud and scam complaints will end in redress. This study develops SCAM-XD, an explainable framework that predicts whether a complaint to the US Consumer Financial Protection Bureau (CFPB) flagged as possible fraud or a scam closes with monetary relief reported by the company. From the CFPB narratives archive, the study builds a benchmark of these complaints, about half of which human coders confirm as fraud or scams. A large language model (LLM) codes each narrative for scam type, payment rail, authorisation, manipulation cues and signals of the bank's response, agreeing closely with human coding on scam type. These features are combined with structured fields, company identity and text representations in a stacking ensemble, explained with SHapley Additive exPlanations (SHAP) and linked to a cost-sensitive triage layer. Tested on a later period, SCAM-XD improves significantly on strong baselines across seeds, including a reimplemented topic model approach. LLM features yield a compact, interpretable model that nearly matches a high-dimensional text baseline but do not significantly improve the full model. Company identity is the strongest driver of predicted redress, yet firm identity alone predicts far less well, and payment rail and bank response cues act in directions consistent with liability rules. Under assumed costs, thresholds selected on validation data substantially reduce modelled expected cost relative to standard thresholds, whereas model choice matters much less. Redress depends on both the firm and the facts that determine liability.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-16
DOI
https://doi.org/10.21203/rs.3.rs-11721933/v1
Primary Topic
Explainable Artificial Intelligence (XAI)
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

An Explainable Machine Learning Framework with Large Language Model Features for Predicting Monetary Redress in Bank Fraud and Scam Complaints

Urvashi Rahul Saxena, Raza Nowrozy, Tahsien Al‐Quraishi, Jemal Abawajy (13075551) et al.
1 links
Research Square
Explainable Artificial Intelligence (XAI)
preprint

An Explainable Machine Learning Framework with Large Language Model Features for Predicting Monetary Redress in Bank Fraud and Scam Complaints

Urvashi Rahul Saxena, Raza Nowrozy, Tahsien Al‐Quraishi, Jemal Abawajy (13075551), Naseer Ali Hussien, O. S. Albahri, Islam Mohammad Tahidul
preprint en

Abstract

Abstract Liability for scam losses is shifting towards banks, yet few interpretable tools can anticipate which fraud and scam complaints will end in redress. This study develops SCAM-XD, an explainable framework that predicts whether a complaint to the US Consumer Financial Protection Bureau (CFPB) flagged as possible fraud or a scam closes with monetary relief reported by the company. From the CFPB narratives archive, the study builds a benchmark of these complaints, about half of which human coders confirm as fraud or scams. A large language model (LLM) codes each narrative for scam type, payment rail, authorisation, manipulation cues and signals of the bank's response, agreeing closely with human coding on scam type. These features are combined with structured fields, company identity and text representations in a stacking ensemble, explained with SHapley Additive exPlanations (SHAP) and linked to a cost-sensitive triage layer. Tested on a later period, SCAM-XD improves significantly on strong baselines across seeds, including a reimplemented topic model approach. LLM features yield a compact, interpretable model that nearly matches a high-dimensional text baseline but do not significantly improve the full model. Company identity is the strongest driver of predicted redress, yet firm identity alone predicts far less well, and payment rail and bank response cues act in directions consistent with liability rules. Under assumed costs, thresholds selected on validation data substantially reduce modelled expected cost relative to standard thresholds, whereas model choice matters much less. Redress depends on both the firm and the facts that determine liability.

Research Square
Federation University (AU), Deakin University (AU), Al-Ayen Iraqi University (IQ), Melbourne Institute of Technology (AU), Victoria University (AU)
Peace, justice, and strong institutions
Explainable Artificial Intelligence (XAI)
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon