Work

Machine Learning for Credit Risk and Default Prediction in Retail and SME Lending : A Systematic Review

Year: 2020
Type: review
Study design: Systematic Review
Abstract: Credit risk and default prediction sit at the core of retail and small- and medium-enterprise (SME) lending, where the marginal cost of a wrong decision is asymmetric and material. Over the past two decades, machine learning (ML) has progressively displaced, augmented, or contested the long-standing dominance of logistic regression and scorecard methods. This systematic review consolidates the empirical evidence on ML for credit risk and default prediction in retail and SME lending, synthesising the literature published up to and including 2020. Following a PRISMA-informed protocol, we searched Scopus, Web of Science, IEEE Xplore, ScienceDirect, and the ACM Digital Library, screened records against pre-registered inclusion and exclusion criteria, and retained a corpus of peer-reviewed studies and seminal foundational works. We organise the synthesis around six themes: feature sets and data sources; the comparative performance of classical, ensemble, and deep-learning methods; the treatment of class imbalance; evaluation metrics spanning discrimination (AUC), calibration, and profit-based measures; interpretability and regulatory constraints; and issues specific to SME lending. The evidence indicates that heterogeneous ensemble methods, particularly gradient-boosted trees and random forests, consistently rank among the strongest performers, while advances in absolute accuracy over well-tuned logistic regression are frequently modest and dataset-dependent. Class imbalance, evaluation on discrimination alone, opacity, and thin or fragmented SME data emerge as recurrent limitations. We identify gaps in profit-centred evaluation, reproducibility, and interpretable modelling under regulatory scrutiny, and we set out a research agenda. The review offers lenders, regulators, and researchers a structured account of where ML adds value and where caution remains warranted.
Language: English
Subfield: Accounting
Open Access status: diamond
OpenAlex ID: W7207570072
FWCI: 0.664
Cites: 141
Cited by: 11
Related to: 10