Introduction
Small and Medium-sized Enterprises (SMEs) are vital to global economic health but frequently struggle to secure financing due to a lack of traditional collateral. To overcome this, many SMEs form "mutual guarantee" networks, where companies act as guarantors for each other's loans. While this practice facilitates credit access, it introduces a systemic vulnerability: a single default can propagate through the network, leading to a chain reaction of financial failures.
The Networked-Loan Challenge
Traditional credit scoring models typically evaluate borrowers in isolation, failing to account for the structural dependencies inherent in mutual guarantee agreements. These "networked-loan" structures create two primary risks that standard assessments miss: cascading defaults and sensitivity to external shocks.
When one node in the network fails, the repayment obligation shifts to its guarantors. If those guarantors are already financially stretched, they too may default, triggering a domino effect across the entire ecosystem. Furthermore, external factors like social media sentiment and market panics can accelerate these collapses faster than traditional financial indicators can signal.
The RisQNet Framework
To address these complexities, we present RisQNet, a deep-learning framework designed for robust networked-loan risk assessment. RisQNet integrates three specialized modules to capture both internal network dynamics and external market influences:
- Multisource Risk Learning Module: This module extracts risk features from public news using Semantic Role Labeling and employs Large Language Models (LLMs), specifically GPT-4, to analyze professional financial reports. By role-playing as financial experts, the LLM assigns risk ratings and calculates a "Cascading Risk" score based on a Likelihood-Impact matrix.
- Graph Attention Learning Module: Utilizing a Graph Attention Network (GAT), this module maps the structural relationships between SMEs. It identifies central nodes within the guarantee network that pose the highest systemic risk.
- Default Prediction Module: RisQNet employs an iTransformer (Inverted Transformer) architecture. By treating independent time series as variate tokens, the iTransformer is better equipped to capture multivariate correlations in noisy financial data over time.
Experimental Results
The framework was evaluated using a massive real-world dataset from a major East Asian commercial bank, spanning 2010–2016. The data included 480,000 SMEs and 760,000 guarantee contracts totaling USD 3 trillion, supplemented by over 230,000 news items and research reports.
RisQNet demonstrated state-of-the-art performance, significantly outperforming traditional methods like XGBoost and standard Graph Neural Networks:
- AUC: 0.871 (vs. 0.677 for traditional scorecards).
- F-Score: 0.768.
- KS Score: 0.621.
Ablation studies highlighted that the inclusion of media sentiment and cascading risk modules provided a 3-5% improvement in AUC, proving that external shocks and network topology are critical for accurate prediction.
Conclusion
RisQNet represents a significant advancement in financial risk management by bridging the gap between individual credit assessment and systemic network analysis. By leveraging LLMs for qualitative data and inverted transformers for temporal patterns, it provides a powerful tool for banks and regulators to identify and mitigate the risks of catastrophic financial contagion. Future work will explore expanding this framework to supply chain finance and broader regulatory monitoring.
Summary updated