Scholarlink

<p>In the context of financial decision-making, this research explores how diverse features affect credit scoring through the application of machine learning and deep learning techniques, particularly in scenarios with imbalanced data. We focus on six models—Logistic Regression, LightGBM, Random Forest, Multi-Layer Perceptron (MLP), Deep Wide Network, and an Attention Classifier—across varied feature sets from different domains. These models are assessed individually and within a multi-model ensemble framework, leveraging both soft voting and a novel Kendall’s tau-based heuristic to quantify model agreement and maximize predictive synergy. This study further investigates the influence of diverse feature sets and sampling techniques, including under sampling, over sampling, and stratified sampling, to mitigate data imbalance and enhance generalization. Our novel heuristic utilizes Kendall’s tau rank correlation to quantify pairwise model agreement, contributing both as a core evaluation metric and a mechanism within the aggregation strategy. The analysis focuses on balancing prediction accuracy and model interpretability through a comprehensive examination of model outcomes, while ensuring robustness via extensive bootstrapped resampling techniques. This careful balancing of factors is especially vital when working with large, real-world financial datasets that present complex risk dynamics. Using a dataset of 44,119 commercial vehicle loan applications, along with their corresponding default and non-default statuses, collected over a span of 10 years by a leading financial institution, the proposed methodology enhances credit scoring outcomes by integrating interpretable machine learning with complex deep learning architectures. Kendall’s Tau-based agreement analysis revealed a concordance range of 0.39 to 0.90 across models, two deep learning architectures exhibiting the highest agreement in stratified sampling. Incorporating autoencoder and RQA embeddings enhanced the Gini scores across models, with LightGBM demonstrating the highest performance in the fusion framework. These findings contribute to more robust and transparent decision-making in financial risk assessment, and highlight the potential of integrating advanced ensemble techniques with interpretability-focused heuristics to improve credit scoring.</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

Evaluating model concordance in a multi-model binary scoring framework: feature-specific and sampling-driven agreement analyses using Kendall’s tau rank correlation for credit risk assessment

  • 期刊论文

  • 发表日期：

    2026年3月26日

查看全文
  • Fuat Kaan Mirza,
  • Arif Selçuk Öğrenci,
  • Ayşe Hümeyra Bilge,
  • Aslı Togan,
  • Chifaa Kardous,
  • Btissam Hajjaoui,
  • Berkay Çiçek,
  • Burak Yüksel,
  • Özden Gebizlioğlu Özvural,
  • Talat Kasımoğlu

摘要

In the context of financial decision-making, this research explores how diverse features affect credit scoring through the application of machine learning and deep learning techniques, particularly in scenarios with imbalanced data. We focus on six models—Logistic Regression, LightGBM, Random Forest, Multi-Layer Perceptron (MLP), Deep Wide Network, and an Attention Classifier—across varied feature sets from different domains. These models are assessed individually and within a multi-model ensemble framework, leveraging both soft voting and a novel Kendall’s tau-based heuristic to quantify model agreement and maximize predictive synergy. This study further investigates the influence of diverse feature sets and sampling techniques, including under sampling, over sampling, and stratified sampling, to mitigate data imbalance and enhance generalization. Our novel heuristic utilizes Kendall’s tau rank correlation to quantify pairwise model agreement, contributing both as a core evaluation metric and a mechanism within the aggregation strategy. The analysis focuses on balancing prediction accuracy and model interpretability through a comprehensive examination of model outcomes, while ensuring robustness via extensive bootstrapped resampling techniques. This careful balancing of factors is especially vital when working with large, real-world financial datasets that present complex risk dynamics. Using a dataset of 44,119 commercial vehicle loan applications, along with their corresponding default and non-default statuses, collected over a span of 10 years by a leading financial institution, the proposed methodology enhances credit scoring outcomes by integrating interpretable machine learning with complex deep learning architectures. Kendall’s Tau-based agreement analysis revealed a concordance range of 0.39 to 0.90 across models, two deep learning architectures exhibiting the highest agreement in stratified sampling. Incorporating autoencoder and RQA embeddings enhanced the Gini scores across models, with LightGBM demonstrating the highest performance in the fusion framework. These findings contribute to more robust and transparent decision-making in financial risk assessment, and highlight the potential of integrating advanced ensemble techniques with interpretability-focused heuristics to improve credit scoring.