Corporate News Shows a Small Link Between Firm Repositioning and Stock Returns

A company expanding into a promising activity changes more than its own prospects: it can also reduce the margins available to competitors. Financial markets may respond to the anticipated consequences before factories, products or operating capacity change. Could those slower adjustments leave a measurable trail in corporate news?

In the preprint Information Games: Strategic Crowding and Firm Repositioning in Language-Model Space, Marcus Gawronsky and Chun-Sung Huang, of the University of Cape Town, develop a model and measurement method for investigating that possibility. Their historical application finds a positive but small association. Crucially, it does not establish that news movements reveal actual changes in operating capacity.

How competition could change a firm's profile

The model treats firms as allocating capacity across activities that serve overlapping customer demands. Expanding an activity can capture more of an available market opportunity, but maintaining and changing capacity carries costs. Competitors' choices also affect how profitable expansion becomes.

Under the model's assumptions, the authors establish a unique equilibrium. Introducing costs of changing inherited capacity makes adjustment gradual. Firms choose current operating payoffs net of adjustment costs, rather than explicitly optimizing all future decisions. Investors, meanwhile, value the resulting cash-flow path using a constant discount factor; financial prices do not feed back into investment.

The central directional prediction applies to a restricted economy with two opportunities. When one improves relative to the other, firms initially more exposed to it have larger return responses. Other firms increasing their exposure move toward those firms' initial reported profiles—provided coverage reflects business mix equally across activities.

This is a sufficient local mechanism under specific timing and error assumptions, not a general prediction for every industry or market shock.

Measuring movement through news

The researchers represent each article as a numerical vector using Qwen3-Embedding-4B, a language model designed for information retrieval. A firm's articles form a distribution of vectors, preserving differences across stories rather than collapsing everything into one average.

They compare distributions using squared Wasserstein distance. In accessible terms, this measures the minimum cost of rearranging one cloud of article vectors to match another. It can detect changes in composition that an average vector misses.

For each target firm, the method measures annual movement toward or away from other firms' news distributions held fixed at their 2018 positions. It compares those movements with the reference firms' daily excess returns, weighted by the fraction of the target's articles published afterward that year. These returns are added, not compounded. The comparison asks whether a target approaches some reference profiles more than others when those references have relatively higher weighted returns.

A consistent but small historical association

The application covers 100 Nasdaq-listed firms, five article years from 2018 through 2022, and four annual transitions. The expanded sample contains 43,720 articles; a matched sample contains 12,800. The authors compare full 2,560-dimensional vectors with compressed 64-dimensional representations. Because the embedding model postdates the articles, this is a retrospective analysis, not a demonstration of information available to investors at the time.

All four specifications produce positive associations. In the primary, full-dimensional expanded specification, normalized alignment is about 0.0363. Its square corresponds to approximately 0.13% of the within-sample variation in centered distance changes explained by this single-regressor relationship. The pattern therefore accounts for only a small component of observed news movement.

The primary estimate remains positive across individual years and checks removing firms or annual transitions. However, the financial variation supporting it is uneven: the 2020 transition contributes 41.4% of the pooled centered financial second moment.

Why news movement is not operating movement

Simulations expose substantial measurement difficulties. With equal reporting intensity, the underlying coefficient is positive in all 16 modeled shock histories. Yet at observed article coverage, only 53.1% of sampled panels yield a numerically resolved positive estimate. Even multiplying coverage sixteenfold raises that share only to 59.8%.

A separate counterfactual produces positive alignment by changing reporting intensity while holding capacities fixed. It retains financial responses from the modeled operating path rather than solving a separate fixed-capacity economy. The result demonstrates an observational ambiguity: shifting attention can resemble shifting activity.

Historical article-redraw ranges remain positive, but they measure article-selection sensitivity conditional on the retained archive, firms, returns and years. They are not population confidence intervals or causal evidence. A separate held-out exercise also finds that the restricted movement pattern improves reconstruction of news changes by only 1.47% in the primary specification; two alternatives perform worse than predicting zero centered change.

The preprint supports a limited directional correspondence, not a complete account of corporate repositioning. Distinguishing investment from attention would require independent, dated measures of operating activity alongside news and returns.