Scholarlink

<p>Rural small and micro enterprises (RSMEs) struggle to borrow, and the reason is rarely a shortage of willing lenders. It is that the information a lender would normally rely on is missing, informal, or simply not written down. Conventional scorecards, built for firms with clean books, tend to falter here, where the data are heterogeneous, riddled with gaps, and shaped by nonlinear risk. We build a credit-risk model that pairs a convolutional network, a bidirectional LSTM, and an attention layer, and we ask it to read three things at once: structured financial indicators, temporal behavioural sequences, and softer signals that stand in for social capital. On 8,647 real loan applications the model reaches an AUC-ROC of 0.943, an AUC-PR of 0.867, an F1-score of 0.849, and 87.4% recall. We report raw accuracy (91.8%) only for completeness, since the 7.3% base default rate makes accuracy a weak guide to minority-class performance. Because oversampling distorts predicted probabilities, we recalibrate the outputs and check them against the observed default rate, and we probe generalization through temporal, province-level, and institution-level holdouts alongside an ablation study and a subgroup fairness probe. The attention weights give case-level explanations that, to our reading, fit a decision-support role rather than fully automated lending. The model code, the full preprocessing and evaluation pipeline, and an anonymized synthetic dataset are provided in the Related files. We frame the contribution as a rural-finance application of established methods, and we read the gains cautiously in light of the limitations set out below.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Intelligent credit risk assessment for rural small and micro enterprises based on hybrid deep learning architecture

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月7日

查看全文
  • Tao Xu,
  • Chen Liu

摘要

Rural small and micro enterprises (RSMEs) struggle to borrow, and the reason is rarely a shortage of willing lenders. It is that the information a lender would normally rely on is missing, informal, or simply not written down. Conventional scorecards, built for firms with clean books, tend to falter here, where the data are heterogeneous, riddled with gaps, and shaped by nonlinear risk. We build a credit-risk model that pairs a convolutional network, a bidirectional LSTM, and an attention layer, and we ask it to read three things at once: structured financial indicators, temporal behavioural sequences, and softer signals that stand in for social capital. On 8,647 real loan applications the model reaches an AUC-ROC of 0.943, an AUC-PR of 0.867, an F1-score of 0.849, and 87.4% recall. We report raw accuracy (91.8%) only for completeness, since the 7.3% base default rate makes accuracy a weak guide to minority-class performance. Because oversampling distorts predicted probabilities, we recalibrate the outputs and check them against the observed default rate, and we probe generalization through temporal, province-level, and institution-level holdouts alongside an ablation study and a subgroup fairness probe. The attention weights give case-level explanations that, to our reading, fit a decision-support role rather than fully automated lending. The model code, the full preprocessing and evaluation pipeline, and an anonymized synthetic dataset are provided in the Related files. We frame the contribution as a rural-finance application of established methods, and we read the gains cautiously in light of the limitations set out below.