Multi-Resolution and Multi-Modal Temporal Graphs for Financial Forecasting

Accurate financial forecasting requires the ability to capture the complex temporal and relational dependencies inherent in market behavior. In this paper, we propose a novel framework based on dynamic graph neural networks (DGNN) for financial forecasting. Our approach introduces a new structured representation, the Temporal Finance Dynamic Graph, which integrates heterogeneous data sources, including market data, company relationships, and financial news, to model the evolving structure and dynamics of financial markets. This is paired with a novel DGNN architecture capable of modeling the different time-scales of the information available in the TFDG, allowing to simultaneously capture short-term shocks and long-term dependencies. This enables a comprehensive representation of the financial market, combining structural, temporal, and external information for robust predictive performance. We evaluate our model on real-world financial forecasting tasks, demonstrating significant improvements in both link prediction and stock price forecasting, compared to state-of-the-art baselines.