Neural Predicates Enhance Investor Views in Black-Litterman Model.

The Black-Litterman model, a cornerstone of portfolio construction, traditionally requires investors to subjectively define their views on asset returns and associated uncertainties. This process is often difficult to scale and lacks formal structure. This paper introduces a novel approach that leverages neural predicates to automate and formalize this view generation. The proposed method processes structured financial analysis data through a compositional hierarchy of neural predicates. These predicates output probability distributions representing various market stances, which are then directly mapped to the Black-Litterman model's key components: the pick matrix, view return vector, and view uncertainty matrix. Crucially, view confidence is derived directly from these predicate output distributions, offering a data-driven alternative to subjective uncertainty elicitation. The resulting framework is both interpretable, allowing portfolio weights to be traced back to underlying data, and fully differentiable, enabling end-to-end learning.