AI-Driven Credit Scoring in Community Banks: A Systematic Review of SME Lending Decisions and Default Risk Prediction

Patrick Botchwey, Victor Agbeve, Jerome Christopher Atisu and Rosemary Dosu
Additional contact information
Patrick Botchwey: KPMG, Ghana
Victor Agbeve: United Bank for Africa, Ghana
Jerome Christopher Atisu: Kwame Nkrumah University of Science and Technology, School of Business, Ghana
Rosemary Dosu: Systems Accountant, Finance and Accounts Department, Ghana National Gas Company Limited, Ghana

Post-Print from HAL

Abstract: The shorter financial statements, limited track records, and limited collateral that small and medium-sized enterprises (SMEs) often have can make it difficult for community banks to conduct traditional risk assessments of these ‘thin-file' borrowers, further reducing credit availability for SMEs. This systematic review examines the literature on the impact of SME lending decisions using artificial intelligence-based credit-scoring models (ACSMs), the prediction of default risk, types of algorithms and data sources, model validation quality, model explainability, fairness and privacy concerns, and applicability from a community banking perspective. In addition, a literature review was conducted on studies published in the last ten years (2015-2023) in Scopus, Web of Science, IEEE Xplore, ScienceDirect, Emerald Insight, ProQuest, and other studies retrieved from Google Scholar. A search of 479 records, full texts of 162 articles, and duplicates of 96 articles yielded a narrative synthesis including 98 studies, as no meta-analysis could be performed due to methodological differences among these studies. Predictivity of support vector machines, random forests, gradient-boosting models, neural networks, and hybrid systems typically yields higher performance than traditional scorecarding or statistical models. The achieved accuracy on the test data using a Support-Vector-Machine is 93.65%, compared to 55.56% for a backpropagation-based neural network. There is also an improvement in the AUROC of XGBoost (from 0.81 to 0.93 relative to logistic regression) and in the AUC with missing legal data (from 0.703 to 0.737 after incorporating legal data for XGBoost). Narratives from loan officers, informational sources, relationship networks, supply-chain information, and websites were used to address information asymmetry, but not uniformly. There was limited evidence due to geographical concentration, small and unbalanced samples, inconsistent metric reporting, lack of external validation, limited reporting of calibration, fairness, and privacy, and limited reporting of implementation outcomes. There were no studies that directly tied to prospective default reduction in community banks. Challenges 1 to 5 involve ensuring that AI is implemented responsibly, with measures in place to ensure it is used appropriately, is explained, can be monitored and overseen, and provides responsible data collection and management.

Keywords: Systematic review; Default risk prediction; SME lending; Community banks; Credit scoring; Artificial intelligence (search for similar items in EconPapers)
Date: 2024-12-29
References: Add references at CitEc
Citations:

Published in European Journal of Management, Economics and Business, 2024, 1 (3), pp.296-318. ⟨10.59324/ejmeb.2024.1(3).27⟩

There are no downloads for this item, see the EconPapers FAQ for hints about obtaining it.

Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.

Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text

Persistent link: https://EconPapers.repec.org/RePEc:hal:journl:hal-05727269

DOI: 10.59324/ejmeb.2024.1(3).27

Access Statistics for this paper

More papers in Post-Print from HAL
Bibliographic data for series maintained by CCSD ().