Who Are a Firm's Economic Peers? Mapping a Latent Information Field with Language Models

arXiv:2608.29669 · q-fin.ST · Submitted 2026-08-30 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.

Ruth: We're looking at the paper: "Who Are a Firm's Economic Peers? Mapping a Latent Information Field with Language Models".

Warren: Here's the gist of the paper. Spatial asset-pricing models typically require researchers to specify inter-firm interaction *ex ante*.

Ruth: This paper shifts this paradigm by inferring the structure of inter-firm interaction from firms’ information environments using language-model representations, thereby providing a measurement framework for constructing economically interpretable relations from unstructured information rather than specifying them beforehand.

Warren: The core methodology involves representing each firm as a distribution of news-article embeddings.

Ruth: The target-anchored Wasserstein barycentric reconstruction is employed to select, for every firm, the weighted combination of other firms whose information footprints jointly reconstruct its own.

Paper summary: Ruth: The paper, Who Are a Firm's Economic Peers? Mapping a Latent Information Field with Language Models, essentially gives us a new way to map out inter-firm interaction by treating the interaction field as an estimated object with an explicit construction and economic interpretation.

Warren: I think the main point is using language-model representations as a measurement instrument for inter-firm structure, not just as a predictor of returns or a source of factors in the first place.

Iris: And it shows that distributional similarity in news content captures peer relations that explicit co-mentions only partly reflect, which is something we need to keep in mind when thinking about news networks.

Ruth: This construction is applicable wherever you have firm-level text corpus, an embedding model, and a return panel data set. It moves the analysis upstream by making the interaction field an estimated object with a clear way to build it.

Warren: So we can now construct these economic relations from the distribution and composition of firms’ information environments instead of just specifying those relations beforehand.

Iris: That’s what it boils down to, giving us a framework for constructing economically interpretable relations from unstructured information through this kind of peer field construction.

Conclusion: Ruth: So, we've looked at how they built this whole picture of who firms interact with based on their news content.

Warren: It’s about inferring those interactions instead of just being told what they are beforehand. They use language models to figure out the structure of peer relationships.

Iris: The core idea is that each firm is represented by a distribution of news article embeddings, and they use something called target-anchored Wasserstein barycentric reconstruction to pick which other firms' information pieces fit together best to form its own picture.

Ruth: And what this does for us is it creates this directed peer field, which they then plug into a quadratic exposure-adjustment model. This model balances how much a firm's own information suggests it should be exposed versus what the market says about its peers.

Warren: The math involves solving for spatial equilibrium in these peer-adjusted exposures, and there’s this feedback coefficient that tells you how much the market adjusts based on those peer misalignments.

Iris: The results they found suggest this constructed field organizes cross-sectional return dependence in a way that goes beyond just standard factors or momentum. It has two layers of structure depending on how you weight the peers, and it uses news embeddings as a different channel than just looking at explicit news links.

Ruth: The authors are saying that using these language models is a new measurement tool for understanding inter-firm structure in finance, moving the analysis upstream to construct these relations rather than specifying them ex ante.

Warren: It's robust because they tested it with different embedding models, and the results held up, showing that this construction gives predictive performance similar to other methods.

Iris: The paper suggests we can now build economically interpretable relations just by looking at how firms’ information is distributed, which opens up a new way to look at market structure.

Ruth: So, they've moved from being told the interaction to figuring out what the interaction is based on the data itself. That leads us to thinking about how this framework applies when we try to model something different, like how firms change their minds over time.

University of Cape Town

q-fin.ST

Submitted: 2026-08-30

Updated: 2026-11-05

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 89/100

The gist: Spatial asset-pricing models typically require researchers to specify inter-firm interaction *ex ante*.

Key concepts

Distribution of News-Article Embeddings
Each firm is represented as a set of numerical vectors (embeddings) derived from all its news articles. These vectors capture the semantic meaning and content of the firm's information footprint, allowing researchers to quantify what kind of news a firm generates.
Target-Anchored Wasserstein Barycentric Reconstruction
This is the core method used to find peers. It aligns a target firm's articles with those of every other candidate peer using optimal transport, then selects non-negative weights that sum to one. This process identifies the weighted combination of other firms whose information best reconstructs the target firm's own information.
Peer Field and Exposure Adjustment Model
The reconstruction process yields a 'peer field,' which is then used in a model where each firm has an implied exposure based on its own news. An adjustment balances this stand-alone exposure against how well it aligns with the exposures of its information peers, resulting in spatial equilibrium.
Language-Model Representations
These are numerical representations derived from language models that capture the complex relationships and structure within text data. The paper uses these representations as a measurement tool to uncover hidden inter-firm structures that are not obvious just by looking at explicit news links.

Terminology

Summary

Spatial asset-pricing models typically require researchers to specify inter-firm interaction ex ante. This paper shifts this paradigm by inferring the structure of inter-firm interaction from firms’ information environments using language-model representations, thereby providing a measurement framework for constructing economically interpretable relations from unstructured information rather than specifying them beforehand. p. 0

The core methodology involves representing each firm as a distribution of news-article embeddings. p. 0 The target-anchored Wasserstein barycentric reconstruction is employed to select, for every firm, the weighted combination of other firms whose information footprints jointly reconstruct its own. p. 0 This process is defined by first aligning the target’s articles with those of every candidate peer via optimal transport, and then choosing non-negative weights that sum to one to approximate the target’s footprint using these aligned peer footprints. p. 1 This yields a directed peer field, which is subsequently incorporated into a quadratic exposure-adjustment model. p. 0 In this model, each firm has a stand-alone exposure (xi i) implied by its own information, and an adjustment balances the departure from xi i against misalignment with the exposures of its information peers. p. 1 Solving this yields a spatial equilibrium in peer-adjusted exposures (B i) with a feedback coefficient rho = lambda / 1 + lambda, where lambda is the penalty on peer misalignment relative to the penalty on departing from stand-alone exposure. p. 1

Empirical findings demonstrate that this constructed field organizes cross-sectional return dependence beyond standard factors and momentum, raising the held-out mean Gaussian quasi-log score relative to a matched factor-only model, with paired intervals above zero. p. 0 The content of the field is attributed to joint reconstruction, exhibiting two layers: under the same factor-conditioned specification, a field that converts pairwise Wasserstein distances into proximity weights recovers less than a third of the barycentric field’s held-out gain, and equal weighting of peers recovers less than half. p. 1 Furthermore, language-model representations serve as a distinct channel from explicit news links under the primary factor-conditioned specification. p. 1 The result is robust to measurement design; linear and quadratic transport generate nearly identical peer-return signals and equivalent held-out predictive performance. p. 0 The barycentric–proximity ordering persists across alternative embedding models, and a pre-period encoder preserves the held-out advantage under the primary specification. p. 0

The paper contributes to three literatures: for spatial finance, it moves the analysis upstream by making the interaction field an estimated object with an explicit construction and economic interpretation. p. 2 For text-based finance, it uses language-model representations as a measurement instrument for inter-firm structure rather than as a predictor of returns or a source of factors. p. 2 For news networks, it shows that distributional similarity in news content captures peer relations that explicit co-mentions only partly reflect. p. 2 The construction is applicable wherever firm-level text corpus, an embedding model, and a return panel data exist. p. 2 The final conclusion posits that language-model representations provide a measurement framework for constructing economically interpretable relations from unstructured information, expanding the measurement frontier of empirical finance by allowing researchers to construct economic relations from the distribution and composition of firms’ information environments instead of specifying those relations ex ante.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Measurement of inter-firm information structure: The system can infer that structure from firms’ information environments using language-model representations, which are a measurement instrument for latent inter-firm information structure in capital markets.

  2. Peer relevance definition: The system can define peer relevance not as pairwise similarity but as target specific and compositional, allowing it to identify which combinations of firms jointly represent one another rather than just nearest neighbors.

  3. Spatial model application: The system can apply the resulting field to a quadratic exposure-adjustment model, where the spatial coefficient indexes alignment with information peers relative to standalone exposure.

  4. Model comparison and selection: The system can compare different peer definitions, such as pairwise distance weighting and equal weighting of the same peers, against the constructed barycentric field to determine which captures more return content, noting that target-specific joint reconstruction is superior.

  5. Channel allocation: The system can identify how total model-implied adjustment divides across distinct information channels by using a two-field adjustment problem, resulting in a result where the barycentric definition receives the larger share of adjustment.

  6. Robustness to measurement design: The system's peer field construction is robust to the specific embedding model used, as the barycentric field attains a higher quasi-log-likelihood than its matching RBF field across various encoders and training vintages.

  7. Factor model integration: The system can integrate the spatial field into existing factor models such that reduced-form factor betas and factor fitted means equal to firm-by-firm least squares, meaning the field organizes the joint distribution of residual returns.

  8. Risk measure derivation: The system can derive a spatial attenuation certificate to bound how much characteristic-implied exposure heterogeneity survives peer adjustment, providing a structural interpretation of risk dispersion.

Sources

Related papers