What Language Models Add to Healthcare-Equity Rankings, and What They Could Have Known

Abstract

In healthcare equities, much of what moves a single name arrives as text: filings, trial and regulatory records, news and analyst commentary. We evaluate how screened non-market features and language-model representations improve market-based equity rankings for US healthcare firms over 2020-2025. Adding balance-sheet and attention features improves the resulting hybrid's price- and liquidity-neutral rank information coefficient (IC) at one, three and twelve months. At twelve months, neutral IC rises from 0.063 to 0.093, a 48% increase, while raw IC rises from 0.289 to 0.298. Language models contribute most clearly at three months: among company-months with scientific-text coverage, a current embedding model raises the hybrid's neutral IC from 0.046 to 0.056. This increment is nominally significant (t = 2.8; paired 95% bootstrap interval [+0.006, +0.015]) and remains exploratory. A larger post hoc gain occurs in months containing trial or regulatory records, consistent with text complementing prices and financial statements at that horizon. Across all text, pre-2020 and current representations yield small positive three-month increments in both neutral and raw IC but no established twelve-month gain. Much of their ranking performance survives within-company time shuffling, consistent with a persistent company component. Rater-cutoff comparisons find negligible pre-cutoff advantages for current models over the pre-2020 rater at one and three months that do not shrink after the cutoffs; twelve-month hindsight attribution remains inconclusive. At twelve months, balance-sheet signals lead after rallies and in volatile markets, while attention leads after declines. Retrospective state-dependent blending improves on a static blend without an established advantage over the full hybrid. Results apply to a survivor-only sector sample and remain unadjusted for the full exploratory search.