Abstract: This paper constructs a stock momentum spillover factor using the BGE semantic embedding model. Analyst reports are transformed into high-dimensional vectors, and stock-level semantic similarity is measured through mixed pooling to build a stock association matrix. The momentum spillover factor is then obtained by weighting these associations with past one-week returns. Empirical results show that the factor effectively captures momentum spillovers among related firms and significantly predicts target firms’ returns in the following week, with the corresponding long-short portfolio generating significant annualized excess returns. Further analysis shows that the BGE-based factor outperforms the co-coverage factor, and that spillover effects are stronger when stock associations are derived from reports issued by non-star analysts. The findings suggest that large language models have strong potential in financial text analysis and asset pricing.

Key words: BGE model, momentum spillover effect, analyst reports, textual data