It is crucial to predict the credit risk of small and medium-sized enterprises (SMEs) accurately for the success of supply chain finance (SCF). However, most of the existing research ignore the fact that the data distribution is usually imbalanced, that is, the proportion of default SMEs is much smaller than that of non-default SMEs. To fill this research gap, we propose a novel approach called DRL-Risk to deal with the imbalanced credit risk prediction (ICRP) of SMEs in SCF with deep reinforcement learning (DRL). Specifically, we formulate the ICRP problem as a Markov decision process and suggest an instance-based reward function to incorporate financial loss into the reward function with consideration of the actual loss caused by misclassification in the ICRP of SMEs. Then, we recommend a deep dueling neural network for decision policy to predict the credit risk of SMEs. With deep reinforcement learning, the DRL-Risk approach can prioritize the learning on the SMEs that would lead to great financial losses. Experimental results demonstrate that the DRL-Risk approach can significantly improve the performance of credit risk prediction of SMEs in SCF compared with the baseline methods in recall, G-mean, and financial loss. We have also identified management implications for the decision-makers participating in SCF.

准确预测中小企业（SME）的信用风险对于供应链金融（SCF）的成功至关重要。然而，现有研究大多忽视了数据分布通常不平衡的事实，即违约中小企业的比例远小于非违约中小企业的比例。为了填补这一研究空白，我们提出了一种名为 DRL-Risk 的新方法，通过深度强化学习 (DRL) 来处理 SCF 中中小企业的不平衡信用风险预测 (ICRP)。具体来说，我们将 ICRP 问题表述为马尔可夫决策过程，并提出基于实例的奖励函数，将财务损失纳入奖励函数，同时考虑到中小企业 ICRP 错误分类造成的实际损失。然后，我们推荐用于决策政策的深度决斗神经网络来预测中小企业的信用风险。通过深度强化学习，DRL-Risk 方法可以优先考虑会导致巨大财务损失的中小企业的学习。实验结果表明，与基线方法相比，DRL-Risk方法在召回率、G均值和财务损失方面能够显着提高SCF中中小企业信用风险预测的性能。我们还确定了对参与 SCF 的决策者的管理影响。