Scholarlink

<p>Decentralized Finance (DeFi) lending requires credit assessment without identity-linked borrower data, relying instead on transparent on-chain behavior. We present an interpretable DeFi-native predictive hurdle model that decomposes expected loss into Probability of Default (PD), Liquidation Severity (LS, used as an LGD proxy), and Exposure at Default (EAD). Unlike traditional hurdle formulations that model zero versus positive outcomes of the same target, our design separates rare-event liquidation incidence on the full wallet universe from conditional severity ranking on the loss tail, and then recombines them with EAD into a deployable credit score. The model uses a hybrid feature set combining wallet-level financial attributes and relational signals from a directed, weighted heterogeneous transaction graph. On Compound V2 data (340,737 transactions, 37,332 wallets; 2019–2025), the full feature set achieves strong PD performance (AUC-ROC 0.8658, AUC-PR 0.2395), while graph-only features are most effective for LS ranking (MAP 0.8302). The integrated score reaches AUC-ROC 0.8738 and concentrates realized risk effectively: the top 100 riskiest wallets (0.4% of scored wallets) are all liquidated in the label period, capturing 19.6% of PD-positive wallets (50.9-fold lift over the baseline liquidation rate). These findings show that graph-augmented ensemble models can provide practical and interpretable credit triage for DeFi lending protocols.</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

Transaction graph-based predictive hurdle model for credit scoring in DeFi lending protocols

  • 期刊论文

  • 发表日期：

    2026年4月3日

查看全文
  • Achutha M,
  • Bhoomika R. Hegde,
  • Bhaskarjyoti Das

摘要

Decentralized Finance (DeFi) lending requires credit assessment without identity-linked borrower data, relying instead on transparent on-chain behavior. We present an interpretable DeFi-native predictive hurdle model that decomposes expected loss into Probability of Default (PD), Liquidation Severity (LS, used as an LGD proxy), and Exposure at Default (EAD). Unlike traditional hurdle formulations that model zero versus positive outcomes of the same target, our design separates rare-event liquidation incidence on the full wallet universe from conditional severity ranking on the loss tail, and then recombines them with EAD into a deployable credit score. The model uses a hybrid feature set combining wallet-level financial attributes and relational signals from a directed, weighted heterogeneous transaction graph. On Compound V2 data (340,737 transactions, 37,332 wallets; 2019–2025), the full feature set achieves strong PD performance (AUC-ROC 0.8658, AUC-PR 0.2395), while graph-only features are most effective for LS ranking (MAP 0.8302). The integrated score reaches AUC-ROC 0.8738 and concentrates realized risk effectively: the top 100 riskiest wallets (0.4% of scored wallets) are all liquidated in the label period, capturing 19.6% of PD-positive wallets (50.9-fold lift over the baseline liquidation rate). These findings show that graph-augmented ensemble models can provide practical and interpretable credit triage for DeFi lending protocols.