Do news distances shape stock returns?

A recent study found that how financial news portrays a company can affect how its stock returns move together with other firms. Instead of summarizing each firm with a single number, the researchers treated firm information as collections of news articles, disclosures, and analyst reports. They used language models to turn these texts into numerical vectors, called embeddings, and measured the distance between two firms' profiles using the quadratic Wasserstein distance. This metric captures how different the distributions of firm characteristics are when compared as whole sets, not just as single numbers. Mathematically, under certain assumptions, this distance imposes a ceiling on the amount of systematic covariance that the firms' returns can share. In plain terms, the more diverse the narratives about companies, the less room there is for their prices to move together in a coordinated way. To test this, the authors applied the method to hundreds of Nasdaq firms from 2018 to 2022. They calculated the pairwise article‑embedding W2 distance and found that larger distances were linked to weaker return co‑movement: the coefficient was 1.92, with a 95 % confidence interval of [1.41, 2.47]. The effect persisted when returns were measured in a later window, though the construction is not point‑in‑time and the evidence remains reduced‑form. The work shows how language data can become a powerful lens for exploring hidden links between corporate stories and market dynamics.

The paper leverages quadratic Wasserstein (W2) geometry to measure the separation between latent distributions of firm characteristics. By defining a randomized bi‑Lipschitz characteristic‑to‑exposure map, the authors introduce a bounded risk‑coordinate slack and a return‑covariance bridge. Under these assumptions, the observed characteristic‑side distance yields a conditional interval for the systematic covariance ceiling: larger transport distance leaves less room for aligned systematic exposures. A non‑negative transport excess term records the gap between realized and maximum‑covariance arrangements. This envelope is a scenario‑analysis tool rather than an expected‑return, no‑arbitrage, or identified structural model. Empirically, encoder‑only language‑model embeddings are extracted from financial news articles, disclosures, and analyst reports for Nasdaq firms over 2018‑2022. In a dyadic regression with symmetric firm effects, greater pairwise article‑embedding W2 distance is associated with weaker return co‑movement: the coefficient is 1.920, with a 95 % confidence interval of [1.412, 2.468]. The estimate retains the same sign when returns are measured in a later window, although the feature construction is not point‑in‑time and the evidence remains reduced‑form. The findings suggest that linguistic narratives capture pricing‑relevant information that shapes systematic covariance, opening a new avenue for quantitative finance research and feature engineering.