Modeling Dynamic Financial Risk Transmission with Adaptive Multi-Agent Reinforcement Learning

Authors
  • Shuyu Hu

    Columbia University, New York, USA

    Author

Abstract

This study addresses the challenges of multi-agent interaction complexity, high market dynamics, and data incompleteness in dynamic financial risk assessment by proposing a modeling approach based on adaptive multi-agent reinforcement learning. The financial system is abstracted as a Markov game process, where the multi-agent structure simulates the cooperation, competition, and strategic interactions of different financial entities. By constructing policy functions and a joint value function, the model captures the transmission mechanisms of risk signals at a global level and effectively reflects the accumulation of systemic risks. An adaptive mechanism dynamically adjusts the learning rate, value function, and policy distribution to enhance robustness under distribution shifts and market fluctuations. Sensitivity experiments on noise ratios, exploration rate decay, and missing value proportions systematically validate the stability of the method. The results show that the proposed approach achieves superior performance in RMSE, MAE, MSE, and R2, maintaining predictive accuracy and explanatory power under diverse uncertainty conditions. Overall, the method enables efficient characterization and dynamic modeling of financial risk features, offering a new technical pathway for risk assessment and understanding risk transmission mechanisms in financial systems.

Cover Image
Downloads
Published
2025-02-01
Section
Articles