Construction of Credit Risk Assessment System Based on Deep Reinforcement Learning and Machine Learning--Taking P2p Lending Platform as An Example

Credit risk assessment plays a crucial role in FinTech, especially in peer-to-peer (P2P) lending, where traditional credit scoring methods have limitations in dealing with the dynamic behaviour and economic fluctuations of borrowers. This paper proposes an advanced credit risk assessment framework integrating machine learning and deep reinforcement learning techniques using the Bondora 2009-2024 dataset to investigate the ability of each model to assess credit risk. The paper compares the predictive performance of Random Forest, Graph Convolutional Network (GCN), Deep Q Network (DQN) and Proximal Policy Optimisation (PPO) models in delinquency rate estimation. The results show that Random Forest is suitable for operations that require high accuracy due to its excellent performance in various aspects such as accuracy and precision. PPO has the highest recall and is very effective for risk-averse lending strategies. GCN shows balanced predictive behaviour with good identification capability using graph-based borrower relationships, while DQN exhibits a conservative bias in default identification. The results suggest that a hybrid approach combining multiple models can increase the effectiveness of credit risk assessment and improve risk control and business scaling strategies for P2P online lending. Future research should explore the real-time deployment of these models and integration into business credit assessment.