Supply Chain Propagation of Textual Signals: LLM Embeddings and Cross-Sectional Return Predictability

Battle Count: 65

Directly addresses cross-sectional return predictability with a novel factor (net_pc_5) achieving 0.86 gross Sharpe ratio and 7.27% annual FF5 alpha. The long-short portfolio construction, transaction cost analysis (20bps), and out-of-sample validation are all directly relevant to systematic equity strategies. The investor inattention mechanism provides a clear economic rationale for the signal. However, the small KG (48 nodes) and S&P 500-only sample limit immediate deployability. The signal is strongest in mid-size firms and post-2020 period, suggesting regime-dependent applicability.

Moderate complexity. FinBERT embedding extraction is straightforward via Hugging Face. PCA is standard. The main complexity lies in: (1) constructing the supply chain knowledge graph from regulatory filings and industry reports (manual, labor-intensive), (2) implementing the weighted propagation scheme with alpha parameter, (3) Fama-MacBeth two-step regression with Newey-West HAC standard errors, and (4) portfolio construction with proper temporal alignment (12-month forward return windows). No GPU-intensive training required since FinBERT is used as a fixed embedding generator. The pipeline is transparent and interpretable by design.

FinBERT is open-source (Hugging Face), SEC EDGAR data is publicly accessible, and Yahoo Finance provides return data. However, the supply chain knowledge graph (48 nodes, 51 edges) is manually curated from regulatory filings and industry reports, making exact replication difficult. The PCA transformation and Fama-MacBeth procedure are standard. No code repository is provided. The propagation weight alpha=0.4 is specified but the KG construction methodology lacks full transparency.