Grounding Investor Views: Neural Predicates in the Black-Litterman Model

This paper proposes using neural predicates to generate structured, probabilistic investor views for the Black-Litterman portfolio model. Financial analysis data passes through a compositional hierarchy of predicates whose output distributions over market stances are mapped to the model’s pick matrix P, view return vector q, and uncertainty matrix Ω. Confidence is derived from predicate distributions instead of being elicited subjectively. The authors describe the approach as interpretable because portfolio weights can be traced through predicate logic to underlying data, and differentiable for end-to-end learning. The provided abstract does not report empirical results or benchmark comparisons.