Scholarlink

This study compares S-ICAS with advanced machine learning (ML) techniques, including random forests, extreme gradient boosting, deep learning, and a stacked ensemble that integrates their predictions. Using a dataset of over 2.5 million firm-year observations from 2014 to 2023, we assess whether these methods improve discriminatory power relative to S-ICAS. Results show that tree-based ensembles outperform the financial and behavioural components of S-ICAS, while deep learning contributes additional information when combined within the stacked model. The meta-model consistently delivers higher predictive accuracy, particularly during periods of economic disruption, such as the Covid-19 crisis, indicating greater robustness than S-ICAS. However, interpretability challenges limit the stand-alone use of ML in regulatory contexts. To address this, we employ eXplainable Artificial Intelligence (XAI) tools, specifically Shapley values, to identify the drivers of discrepancies between ML and S-ICAS outputs. This enhances analysts’ ability to refine second-stage assessments and ensures transparency. Overall, our findings suggest that ML and XAI can complement, but not replace, traditional models, improving both the effectiveness and resilience of the ICAS process.

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

Credit Risk Assessment with Stacked Machine Learning

  • 章节

  • 发表日期：

    2026年

查看全文
  • Francesco Columba,
  • Manuel Cugliari,
  • Stefano Di Virgilio

摘要

This study compares S-ICAS with advanced machine learning (ML) techniques, including random forests, extreme gradient boosting, deep learning, and a stacked ensemble that integrates their predictions. Using a dataset of over 2.5 million firm-year observations from 2014 to 2023, we assess whether these methods improve discriminatory power relative to S-ICAS. Results show that tree-based ensembles outperform the financial and behavioural components of S-ICAS, while deep learning contributes additional information when combined within the stacked model. The meta-model consistently delivers higher predictive accuracy, particularly during periods of economic disruption, such as the Covid-19 crisis, indicating greater robustness than S-ICAS. However, interpretability challenges limit the stand-alone use of ML in regulatory contexts. To address this, we employ eXplainable Artificial Intelligence (XAI) tools, specifically Shapley values, to identify the drivers of discrepancies between ML and S-ICAS outputs. This enhances analysts’ ability to refine second-stage assessments and ensures transparency. Overall, our findings suggest that ML and XAI can complement, but not replace, traditional models, improving both the effectiveness and resilience of the ICAS process.