Research on an Automated Intraday Liquidity Scheduling Strategy for Finance Companies Based on Deep Reinforcement Learning

Abstract

This study rigorously formulates the complex fund-scheduling problem as a Markov decision process (MDP). It constructs a state space that integrates real-time and forecast information, an atomic action space that conforms to business logic, and a reward function that balances long-term returns against immediate risk. To address the curse of dimensionality and the credit-assignment problem in coordinated scheduling among multiple fund units, a multi-agent deep deterministic policy gradient (MADDPG) algorithm is adopted. Under a centralized-training and decentralized-execution framework, the algorithm reconciles global optimization with decentralized decision-making. In addition, a difference-reward mechanism and Kalman filtering are used to accurately measure each agent’s individual contribution and reduce the impact of environmental noise on reward signals. The results show that, compared with a static rule engine and a conventional linear programming method, the proposed deep reinforcement learning strategy reduces average daily funding costs by 50.4%, lowers the payment failure rate to 0.002%, and maintains a high liquidity buffer adequacy ratio. The strategy also demonstrates clear advantages in decision timeliness, collaborative handling of complex instructions, and self-adaptation potential, thereby providing an innovative pathway for finance-company fund scheduling to progress from intelligentization to automation.