Lineage-Aware Search Control Improves Diversity and Predictiveness in LLM-Based Evolutionary Factor Mining

01 ABSTRACT

Fact: The authors propose LineageEvo, which adds lineage-level search control to LLM-based evolutionary factor mining via lineage-conditioned LLM variation and lineage-stratified search budget allocation, evaluated on CSI 500 and S&P 500 for factor predictiveness, portfolio performance, search efficiency, and lineage/expression diversity. Author view: lineage can serve as a useful intermediate unit for controlling LLM-based evolutionary factor search, making semantic variation more directed and avoiding excessive branch concentration. Results report that LineageEvo achieves the strongest predictive performance on both markets, while LSPS yields a more balanced lineage distribution and a structurally broader factor pool.

02 KEY FINDINGS

1. Lineage-conditioned LLM variation distills Lineage-Scoped Search Experience into operator-specific priors and fuses them with global priors, providing persistent lineage-relevant guidance for generation.
2. Lineage-stratified search allocation distributes the search budget across lineages during evolution, balancing factor quality against lineage concentration and preventing a single branch from dominating.
3. On CSI 500 and S&P 500, LineageEvo reports the strongest predictive performance, and LSPS produces a more balanced lineage distribution and a structurally broader factor pool.