DisclosureBeta: A Measurement-Channel Theory for Regime-Conditioned Betas from LLM-Read Risk Disclosures

Battle Count: 50

Directly addresses beta estimation for trading-relevant scenarios: IPO pricing, regime breaks, and short-history firms. The adaptive blend (Theorem 5) provides a practical, never-worse estimator that automatically weights text vs. price history. The regime-conditional factor model and text-spanned factors (Section 4) are directly usable in quantitative trading systems. The disclosure-incentive channel and lower bounds inform model risk management for LLM-based signals. The pre-registered empirical design targets exactly the population where quantitative desks need reliable betas.

Requires: (1) LLM pipeline with structured two-stage agent (evidence extraction + scoring) with K ensemble reads per document; (2) Regime detection via AdaptiveCMDP; (3) Kernel or sieve nonparametric estimation of regime-conditional loading functions; (4) Rolling OLS with sandwich variance estimation; (5) Adaptive convex combination with plug-in variance/MSEM estimation; (6) Text-spanned factor construction and spanning tests; (7) DIM scoring pipeline. The theoretical framework is well-specified but the full empirical pipeline involves multiple interacting components. The ensemble channel measurement (Proposition 8) adds operational complexity but makes the error budget estimable per observation.

This is a theory preprint with a pre-registered empirical program. The theoretical framework, assumptions, and theorems are fully specified. However, no code, data, or empirical results are provided yet. The empirical evaluation is explicitly deferred to a forthcoming companion paper. The pre-registered design is described but not executed. Proof sketches are provided inline but complete proofs are referenced as being in a source module (theory.tex) not publicly linked.