Multi-Agent Adversarial Time Series Forecasting

Ye Qiao, Haoxiang Zhang, Linfeng Wang, Cheng Chen, Tong Niu, Xueru Bai, Yang Xiang

Keywords: Multi-agent prediction; Time Series; Joint Distribution Modeling; Distribution shift

Abstract: Time series forecasting is critical across finance, energy, and healthcare, yet remains challenged by the complexity and non-stationarity of real-world data. Although deep learning has advanced performance, single-model architectures often struggle with temporal volatility and limited generalization. Multi-agent collaborative training offers a promising path forward by leveraging diverse model strengths; however, existing methods mostly rely on simple ensembles, lacking deeper structural interaction and probabilistic alignment. In this paper, we propose Multi-Agent Adversarial Time Series Forecasting (MAA-TSF), a framework that orchestrates heterogeneous generators and discriminators into a dynamic, competitive–cooperative system, akin to a multi-force formation adapting to evolving terrains. It integrates intra-group dynamic knowledge alignment and cross-group adversarial training to enhance joint distribution modeling and resilience to distribution shifts, while solving adversarial baseline instability. By evaluating nineteen real-world financial assets in six distinct market categories and six well-known datasets, we find that it consistently outperforms both the ERM and GAN under different time-specific backbones , achieving MAE reductions of 10% – 70%, while delivering 5% – 25% gains in the accuracy of directional prediction across most datasets and models, verifying adversarial multi-agent coordination as a robust paradigm for complex time series.

Primary Area: learning on time series and dynamical systems

Submission Number: 11157