Scholarlink

This study evaluates whether machine learning can improve the early prediction of SME failure using routinely available indicators. Using a balanced sample of 400 U.K. service-sector SMEs (200 failed, 200 active) over 2020–2025 and a feature set of 27 financial and non-financial ratios, we compare four classifiers: deep learning with an LSTM component (DPL), artificial neural networks (ANN), decision trees (DT), and logistic regression (LR). Model training follows a hold-out scheme for DPL/ANN and K-fold cross-validation for LR/DT, and evaluation uses accuracy, sensitivity, specificity, Type I/II errors, and AUC. We additionally conduct non-parametric significance testing (Friedman with Bonferroni–Dunn post-hoc). Results show DPL achieves the strongest overall performance—highest AUC and accuracy on both training and testing—while ANN and DT are competitive and LR provides a transparent baseline. Practical implications include integrating such models into credit screening and supervisory early-warning workflows. Scope is limited to U.K. service SMEs and ratio-based features; future extensions could incorporate additional sectors, countries, and alternative data.

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

Predicting SME’s Failure Using Machine Learning

  • 章节

  • 发表日期：

    2026年

查看全文
  • Aram Khalaf Nawaiseh,
  • Maha Shehadeh,
  • Noor Majed Esaifan

摘要

This study evaluates whether machine learning can improve the early prediction of SME failure using routinely available indicators. Using a balanced sample of 400 U.K. service-sector SMEs (200 failed, 200 active) over 2020–2025 and a feature set of 27 financial and non-financial ratios, we compare four classifiers: deep learning with an LSTM component (DPL), artificial neural networks (ANN), decision trees (DT), and logistic regression (LR). Model training follows a hold-out scheme for DPL/ANN and K-fold cross-validation for LR/DT, and evaluation uses accuracy, sensitivity, specificity, Type I/II errors, and AUC. We additionally conduct non-parametric significance testing (Friedman with Bonferroni–Dunn post-hoc). Results show DPL achieves the strongest overall performance—highest AUC and accuracy on both training and testing—while ANN and DT are competitive and LR provides a transparent baseline. Practical implications include integrating such models into credit screening and supervisory early-warning workflows. Scope is limited to U.K. service SMEs and ratio-based features; future extensions could incorporate additional sectors, countries, and alternative data.