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

Area of Science:

• Financial network analysis
• Systemic risk forecasting
• Machine learning in finance

Background:

• Financial markets are complex networks where institutional holdings influence information flow and risk propagation.
• Forecasting stock network evolution is crucial for stability but challenging due to data complexity and non-stationarity.

Purpose of the Study:

• To develop an interpretable LLM-based forecasting framework for financial networks.
• To predict market-wide concentration and firm-level anomalies for early risk detection.

Main Methods:

• Constructed time-indexed bipartite fund-stock graphs from 2016-2024.
• Computed network indicators: degree centralization (cen_d) and residual density (den).
• Developed an LLM framework integrating time series, textual disclosures, and retrieval-augmented context for multi-step forecasting.

Main Results:

• The LLM framework significantly outperformed ARIMA, Prophet, and Temporal Fusion Transformer benchmarks in accuracy.
• Attention analysis revealed model importance for quarters with fund co-movement or policy shocks.
• Demonstrated reduced errors and improved directional accuracy in forecasting network indicators.

Conclusions:

• LLM-driven forecasting offers early warnings for systemic risk and interpretable insights.
• Language-informed graph forecasting presents a new paradigm for financial market surveillance and policy.