Paper Demonstrates Networked LLM Embeddings Predict Returns
A new arXiv paper reports that propagating FinBERT text embeddings of company 10-K filings across a supply-chain knowledge graph produces a statistically significant stock-return predictor, with a long-short portfolio achieving a 0.86 annualized Sharpe ratio and a 7.27% per-year Fama-French five-factor alpha (t = 2.30). The study, by researcher Asef Yilki, analyzed 255 S&P 500 firms' 10-K MD&A sections from 2011-2025, comparing plain LLM embeddings against network-augmented versions where firm-level textual signals spread through measured supplier and customer linkages; only the network-augmented factor (net_pc_5) showed significant predictive power (Newey-West t = -2.64) after controlling for momentum, volatility, and firm size. The results held up in out-of-sample tests, placebo experiments, sector-neutral portfolios, and subsample splits, though the paper is a single-author preprint awaiting peer review.
The interesting result for ML and quant teams is not that text embeddings predict returns, that is a known technique, it is that the predictive power only shows up once those embeddings are propagated through the supply-chain graph. Direct firm-level LLM embeddings alone were not the significant factor; the network-augmented version was. That is a concrete, reproducible example of relational structure adding information that a per-firm text model misses on its own.
What happened
Researcher Asef Yilki proposes an asset-pricing framework that augments LLM embeddings of annual report disclosures with supply-chain knowledge-graph propagation (arXiv:2606.29290). Using FinBERT embeddings of 10-K MD&A sections for 255 S&P 500 firms over 2011-2025, the paper builds two predictor sets: direct LLM embeddings and network-augmented embeddings where firm-level signals propagate through inter-firm supplier and customer linkages. Fama-MacBeth cross-sectional regressions find that the network-augmented factor, net_pc_5, carries significant return predictability (Newey-West t-statistic of -2.64) even after controlling for momentum, volatility, and firm size. A long-short portfolio sorted on net_pc_5 achieves an annualized Sharpe ratio of 0.86 and a Fama-French five-factor alpha of 7.27% per year (t = 2.30). The predictive power survives out-of-sample tests, placebo experiments, sector-neutralization, and subsample analysis.
Technical context
Prior financial-NLP work has generally treated firms as independent text-generating units; this paper's contribution is explicitly modeling inter-firm exposure via a measured supply-chain graph and testing whether propagated signals carry pricing-relevant information beyond each firm's own disclosures. The two implementable pieces are (1) extracting dense FinBERT embeddings from regulatory filings, and (2) applying a graph-propagation or diffusion operator to spread those embeddings along measured supplier/customer links before using them as a factor.
For practitioners
Teams building financial-NLP or factor-research pipelines get a concrete template here: the paper quantifies economic magnitudes (0.86 Sharpe, 7.27% alpha) large enough to justify follow-up research, and it isolates the network-propagation step, not just the embedding step, as the source of the effect. That is a useful, falsifiable claim to test against an independent dataset or embedding model before relying on it in production.
What to watch
This is a single-author arXiv preprint (submitted June 28, 2026) that has not yet been peer-reviewed, so independent replication on other datasets, embedding families (beyond FinBERT), and propagation kernels is the key next step. Watch for follow-up work testing whether the effect persists across different market regimes and whether it survives transaction-cost-adjusted backtesting, since a 0.86 Sharpe on a long-short factor can look very different after realistic trading frictions.
Key Points
- 1A new arXiv paper finds network-propagated FinBERT embeddings, not raw firm-level embeddings, predict stock returns significantly.
- 2A long-short portfolio built on the network-augmented factor achieved a 0.86 Sharpe ratio and 7.27 percent annual alpha.
- 3Results held across out-of-sample, placebo, and sector-neutral tests, but the single-author preprint still awaits peer review.
Scoring Rationale
A methodologically concrete, well-tested (out-of-sample, placebo, sector-neutral) paper showing that network-propagated LLM text embeddings, not raw embeddings, produce a statistically and economically significant return factor. Notable for financial-ML practitioners, but it remains a single-author, non-peer-reviewed preprint with backtested (not live) results.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
