Scholarlink

<p>Financial institutions face increasing challenges in accurately assessing credit risk owing to the coexistence of structured financial records and unstructured textual information. This study proposes a hybrid predictive modelling framework that integrates Machine Learning (ML) and Deep Learning (DL) techniques for financial risk assessment using heterogeneous data sources. Structured attributes were derived from the Home Credit Default Risk dataset (307,511 applicant records, 122 features), whereas contextual signals were extracted from financial news and social media via sentiment analysis. A unified preprocessing pipeline incorporating missing value imputation, label encoding, feature scaling, and SMOTEENN resampling addressed data quality and class imbalance. The ML pipeline employs XGBoost, LightGBM, and Random Forest within voting and stacking ensembles, whereas the DL pipeline utilises a CNN–LSTM architecture to model high-order feature interactions. Experimental evaluation using Accuracy, Precision, Recall, F1-score, ROC-AUC, and confusion matrix analysis indicated that sentiment-derived features provided modest yet consistent predictive refinement. The Stacking Classifier achieved ROC-AUC = 0.832 and F1-score = 0.61, outperforming the CNN–LSTM model (ROC-AUC = 0.821 and F1-score = 0.57). The results suggest that ensemble learning offers superior discrimination capability under fused feature conditions. Sentiment variables function as complementary contextual indicators alongside the dominant structured financial predictors.</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

Enhanced predictive modeling for financial risk assessment using hybrid AI (ML & DL) on structured and unstructured data

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月16日

查看全文
  • Yogesh Kumar Jain,
  • Shashi Kant Gupta,
  • Deema Mohammed Alsekait,
  • Mohammed Yahya Albeshri,
  • Diaa Salama AbdElminaam

摘要

Financial institutions face increasing challenges in accurately assessing credit risk owing to the coexistence of structured financial records and unstructured textual information. This study proposes a hybrid predictive modelling framework that integrates Machine Learning (ML) and Deep Learning (DL) techniques for financial risk assessment using heterogeneous data sources. Structured attributes were derived from the Home Credit Default Risk dataset (307,511 applicant records, 122 features), whereas contextual signals were extracted from financial news and social media via sentiment analysis. A unified preprocessing pipeline incorporating missing value imputation, label encoding, feature scaling, and SMOTEENN resampling addressed data quality and class imbalance. The ML pipeline employs XGBoost, LightGBM, and Random Forest within voting and stacking ensembles, whereas the DL pipeline utilises a CNN–LSTM architecture to model high-order feature interactions. Experimental evaluation using Accuracy, Precision, Recall, F1-score, ROC-AUC, and confusion matrix analysis indicated that sentiment-derived features provided modest yet consistent predictive refinement. The Stacking Classifier achieved ROC-AUC = 0.832 and F1-score = 0.61, outperforming the CNN–LSTM model (ROC-AUC = 0.821 and F1-score = 0.57). The results suggest that ensemble learning offers superior discrimination capability under fused feature conditions. Sentiment variables function as complementary contextual indicators alongside the dominant structured financial predictors.