Large language model-driven time-series forecasting of financial network indicators

Introduction: Financial markets operate as dynamic networks in which institutional cross-holdings shape the diffusion of information and the propagation of risk. Forecasting the evolution of stock information networks is critical for anticipating herding behavior and safeguarding systemic stability, yet remains challenging due to high-dimensional heterogeneity, structural non-stationarity, and the need for economically interpretable predictions. Methods: Using a quarterly fund-stock holding panel from 2016 to 2024, we construct time-indexed bipartite fund-stock graphs and project them onto the stock layer. From these graphs, we compute two key network indicators: degree centralization (cen_d), capturing market-wide concentration, and residual density (den), reflecting firm-level anomalies. We then develop a large language model (LLM)-enhanced forecasting framework that transforms numeric time series and textual fund disclosures into promptable sequences, incorporates retrieval-augmented historical context, and performs multi-step forecasting of both cen_d and abnormal den spikes. Results: Extensive experiments show that the proposed LLM-based framework significantly reduces mean absolute error and root mean square error, and improves directional accuracy, compared with ARIMA, Prophet, and Temporal Fusion Transformer benchmarks. Attention-weight analysis further indicates that the model assigns higher importance to historical quarters characterized by sharp fund co-movement or policy shocks. Discussion: These findings demonstrate that LLM-driven time-series forecasting can provide early warnings of systemic risk and generate economically interpretable insights for investors and regulators. The results highlight the broader potential of language-informed graph forecasting as a new paradigm for financial market surveillance and policy design.