Aug 2026· Journal of Intelligent Decision Making and Information Science· 0 citations· 64 references
TL;DR
It is argued that predictive accuracy and regulatory transparency are not competing objectives but complementary necessities for institutional survival in Nepal’s cooperative sector.
Abstract
The cooperative banking sector in Nepal constitutes a foundational pillar of financial inclusion, serving approximately 7.4 million members across 31,450 primary cooperatives and disbursing loans exceeding NPR 453 billion. Yet this sector is hemorrhaging credibility. The National Cooperative Bank Limited reported a non-performing loan ratio of 33.01 percent as of mid-July 2025, with its capital adequacy ratio collapsing to 0.82 percent — a figure that would trigger immediate regulatory intervention in any conventional banking jurisdiction. Against this backdrop, the question is no longer whether cooperative banks in Nepal need better risk assessment tools, but whether they can afford to continue without them.
This paper develops a comprehensive theoretical framework for integrating explainable machine learning into credit risk management systems of Nepalese cooperative banks. We derive the complete mathematical architecture of ensemble gradient boosting — specifically XGBoost and LightGBM — alongside post-hoc interpretability mechanisms grounded in cooperative game theory, namely SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations). The framework incorporates information-theoretic feature selection, SMOTE-based class imbalance correction, fairness constraints calibrated to Nepal’s socio-economic heterogeneity, and a rigorous convergence analysis of the boosting iteration. We prove consistency of the mutual-information feature selector, derive the bias-variance decomposition for gradient-boosted ensembles, establish generalization bounds via Rademacher complexity, and verify the four Shapley axioms for the TreeSHAP algorithm. Drawing upon Nepal Rastra Bank’s Financial Stability Report FY 2024/25, the National Cooperative Federation of Nepal’s sectoral statistics, and the Government of Nepal’s Economic Survey 2025/26, we situate the technical apparatus within Nepal’s regulatory architecture — the Cooperative Act 2017, the Financial Sector Development Strategy 2022–2026, and the National Financial Inclusion Roadmap. The paper argues that predictive accuracy and regulatory transparency are not competing objectives but complementary necessities for institutional survival in Nepal’s cooperative sector.
An Explainable Machine Learning (XML) framework for credit risk assessment that combines an ensemble classifier, integrating XGBoost, Random Forest, and LightGBM, with an integrated SHAP-and-LIME explainability layer is proposed and evaluated using a large-scale retail and priority-sector loan dataset drawn from public sector, private sector, regional rural, and small finance bank segments operating in India.
A. Agrawal, Vaibhav C. Gandhi· International journal of com...· 0 citations
Credit default prediction has become an important application of machine learning in the banking and financial sector, as it helps financial institutions identify potential loan defaulters and support informed lending decisions. Although machine learning models often provide high predictive performance, many of them function as black-box system, making it difficult for financial analysts and decision-makers to understand the reasoning behind their predictions. This lack of transparency can reduce user trust, particularly in high-stakes financial applications where explainable decisions are essential. To address this challenge, this study explores the use of Explainable Artificial intelligence (XAI) techniques to improve the interpretability of credit default prediction. A Random Forest classifier was developed using a publicly available credit default dataset containing financial attributes such as employment status, bank balance, annual salary, and loan default status. The dataset was preprocessed and partitioned into training and testing sets before model development. To explain the prediction process, SHAP (Shapley Additive Explanations) and LIME (Local Interpretable Model-Agnostic Explanations) were integrated with the trained Random Forest model. SHAP was used to provide both global and local explanations by identifying the overall importance and contribution of individual features, while LIME generated instance-level explanations to illustrate how specific features influenced individual predictions. The explanation results were presented through visualizations, including feature importance plots, waterfall plots, and local explanation graphs, allowing a clearer understanding of the model's decision-making process. The findings demonstrate that the combined use of SHAP and LIME enhances the transparency and interpretability of the Random Forest model by providing complementary perspectives on feature contributions. This study highlights the practical value of explainable machine learning in developing more understandable, trustworthy, and accountable credit risk assessment systems for real-world financial decision-making.
Muskan, B. Sidhu· International Journal of Com...· 0 citations
By empirically proving that high-performance algorithms can be mathematically blind to demographic biases, this framework directly advances SDG 10 (Reduced Inequalities) and provides the accountable, feature-level justifications required for secure and sustainable financial inclusion (SDG 8).
Htet Nge Nge Ko, Aung Htoo Khine, Shadab Kalhoro et al.· Journal of Risk and Financia...· 0 citations
Credit risk assessment forms a cornerstone of banking risk management and the stability of the wider financial system. Over the past decade, the rapid development of machine learning (ML) techniques has substantially enhanced traditional credit risk assessment methodologies. ML has now emerged as a core technological pillar for the banking sector, strengthening risk identification capabilities, optimising credit decision-making, and advancing financial inclusion. Conventional credit scoring models, dominated by logistic regression (LR) and scorecard approaches, offer inherent strengths in interpretability and regulatory compliance. However, constrained by their linear assumptions, these methods struggle to capture complex non-linear relationships within credit data and deliver insufficient predictive accuracy for the “credit-invisible” population lacking formal credit histories. This paper presents a systematic literature review (SLR) of ML applications in credit risk assessment (CRA), covering publications from January 2016 to May 2026. A total of 894 papers were retrieved from five digital libraries, and following a rigorous multi-stage screening process, 129 studies were selected for final inclusion. Our analysis reveals that tree-based ensemble models and deep learning (DL) architectures predominate in contemporary research in this field. Meanwhile, post hoc explanation methods and machine learning operations (MLOps) are gaining significant traction as solutions to address fairness, transparency, and system maintenance challenges in real-world production environments. We synthesise prevailing methodologies into a unified end-to-end credit risk modelling framework spanning data preprocessing, feature engineering, model training, evaluation, and operational deployment. Through a critical assessment of the advantages, limitations, and inherent trade-offs of existing approaches, this SLR not only identifies current research gaps and future directions for the academic community, but also provides practical guidance for the banking sector to build compliant, fair, and efficient intelligent risk assessment systems.
Bolun Zhang, Jun Luo, Ruobing Wu et al.· Journal of Risk and Financia...· 0 citations
Non-Performing Loans (NPL) are a fundamental indicator of a financial institution's asset health, reflecting loans that fail to meet interest or principal payment obligations as agreed. A high NPL ratio negatively impacts a bank's financial performance, such as decreased profitability as measured by Return on Assets (ROA) and decreased liquidity. Bank Indonesia sets an NPL tolerance limit of 5% of total credit provided by banking financial institutions. Therefore, a predictive model is needed that can detect the possibility of customers experiencing NPLs early. This study aims to identify relevant factors in predicting NPLs and create an NPL prediction model based on these factors. The contribution of this study lies in combining the results of three feature selection techniques: Chi-Square, Mutual Information, and Random Forest feature importance, using the average score eliminated by the Recursive Feature Elimination technique. Several ensemble algorithms, namely Random Forest, XGBoost, Gradient Boosting, and LightGBM, were explored to produce the best-performing model. Then, hyperparameter tuning was performed on the best model. The Random Forest model produced the best performance, with 92.17% accuracy, 78.1% precision, 98.1% recall, and 95.5% AUC. Hyperparameter tuning was shown to improve recall, thus improving the model's ability to measure how much positive data (Current class) was successfully predicted by the model. The results of this study can assist management in making credit decisions. Thus, it is hoped that it can help reduce the number of NPL cases.
David Jefri Aruan, Rusdah Rusdah, Ahmad Pudoli· IDEALIS : InDonEsiA journaL...· 0 citations
Banking is increasingly shaped by expanding data volumes, more complex borrower behaviour, and stricter credit risk management requirements. Under such conditions, scoring models are becoming especially relevant as instruments for the formalised assessment of creditworthiness, combining analytical accuracy, speed of decision-making, and the possibility of integration into the bank’s risk management system. The study compares traditional and modern scoring models in bank credit risk management and proposes an approach to their practical use in Ukrainian banking. Its focus is on scoring models as instruments for credit risk assessment. The study combines comparative analysis, matrix modelling, simulation, statistical modelling, and machine learning methods. Given limited access to primary banking information and confidentiality requirements, the empirical analysis was conducted on a synthesised demonstration dataset designed to reflect the structure of a real retail credit portfolio. For the analysis, a sample of 1,000 observations with a default share of 22.0% was constructed, and logistic regression, discriminant analysis, Random Forest, XGBoost, and a hybrid logit + ML re-ranking model were used for comparison. The results showed that XGBoost provided the highest predictive accuracy, with an AUC-ROC of 0.861, Gini of 0.722, Recall of 0.781, and Brier score of 0.141, whereas logistic regression demonstrated an AUC-ROC of 0.781 and retained advantages in terms of interpretability and suitability for validation. The hybrid model achieved an AUC-ROC of 0.848, Gini of 0.696, Recall of 0.773, and Brier score of 0.144, thus ensuring the best balance between accuracy, explainability, calibration, and practical applicability. Practically, the study offers an adaptive approach to selecting scoring models and a matrix for evaluating them under Ukrainian banking conditions, taking into account the requirements of the regulatory environment, data quality, and the instability of the operating conditions of Ukrainian banks.