Hierarchical Transformer Framework for Causal Discovery across Multi-Sector Financial Time Series

Causal discovery in financial time series presents significant challenges due to the complex temporal dependencies and multi-sector interactions that characterize modern financial markets. Traditional statistical methods and conventional deep learning approaches often fail to capture the hierarchical nature of causal relationships across different market sectors and temporal scales. This paper proposes a novel Hierarchical Transformer Framework for Causal Discovery (HT-CD) that leverages multi-level attention mechanisms to identify causal relationships across multi-sector financial time series. Our framework introduces a sector-aware hierarchical encoder that processes financial data at multiple granularities, from individual asset-level patterns to sector-wide dynamics, while maintaining temporal consistency through specialized positional encoding schemes. The architecture incorporates a novel multi-head causal attention mechanism that explicitly models both contemporaneous and lagged causal effects across different sectors. Experimental evaluation on comprehensive financial datasets spanning multiple sectors demonstrates that HT-CD significantly outperforms existing causal discovery methods, achieving 23% improvement in causal structure accuracy while maintaining computational efficiency. The framework successfully identifies cross-sector causal relationships that align with established economic theories and provides interpretable insights into market contagion effects during financial stress periods.