Supply Chain Propagation of Textual Signals: LLM Embeddings and Cross-Sectional Return Predictability

Asef Y{\i}lk{\i}

arxiv: 2606.29290 · v1 · pith:UOVQ6UJRnew · submitted 2026-06-28 · 💱 q-fin.PR

Supply Chain Propagation of Textual Signals: LLM Embeddings and Cross-Sectional Return Predictability

Asef Y{\i}lk{\i} This is my paper

Pith reviewed 2026-06-30 02:05 UTC · model grok-4.3

classification 💱 q-fin.PR

keywords supply chainLLM embeddingstextual analysisreturn predictabilityasset pricingknowledge graphFinBERTcross-sectional returns

0 comments

The pith

Propagating LLM embeddings through supply chain networks yields a cross-sectional return predictor with 0.86 Sharpe ratio.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper establishes that signals extracted from annual report text become stronger return predictors when spread across supply chain connections. Researchers embed 10-K MD&A sections with FinBERT for S&P 500 firms, then propagate those embeddings along a knowledge graph of inter-firm links to form network-augmented factors. The resulting net_pc_5 factor shows significant negative predictability in Fama-MacBeth regressions after standard controls and supports a long-short portfolio with 0.86 annualized Sharpe and 7.27 percent Fama-French five-factor alpha. A reader would care because the result points to information traveling between firms in ways that markets have not yet fully priced.

Core claim

The paper claims that the network-augmented factor net_pc_5 carries significant return predictability with a Newey-West t-statistic of -2.64 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 percent per year with t-statistic 2.30. Predictive power survives out-of-sample tests, placebo experiments, sector neutralization, and subsample analysis, indicating that inter-firm network structure supplies pricing-relevant information beyond isolated firm-level textual disclosures.

What carries the argument

Supply chain knowledge graph propagation, which augments firm-level LLM embeddings by spreading them across documented inter-firm linkages.

If this is right

Network-augmented embeddings retain predictive power after momentum, volatility, and size controls.
Long-short portfolios on net_pc_5 earn positive risk-adjusted returns that survive multiple robustness checks.
Direct LLM embeddings alone leave economically relevant signals uncaptured.
Inter-firm linkages contain incremental pricing information relative to standalone textual disclosures.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If propagation works here, the same graph-augmentation step could be applied to earnings-call transcripts or news text to test whether other disclosure channels also benefit from network structure.
Portfolio construction rules that ignore supply-chain neighbors may leave exploitable mispricing on the table when textual signals are used.
Models that treat firms as isolated nodes may systematically understate the speed at which textual information diffuses into prices.

Load-bearing premise

The supply chain knowledge graph accurately captures the economically relevant inter-firm linkages and the chosen propagation rule does not create spurious cross-sectional correlations.

What would settle it

A new long-short portfolio sorted on net_pc_5 that produces no statistically significant alpha when constructed on a later sample period or when the supply chain graph is replaced by a random network with the same degree distribution.

Figures

Figures reproduced from arXiv: 2606.29290 by Asef Y{\i}lk{\i}.

**Figure 1.** Figure 1: Empirical Pipeline: Network-Augmented LLM Embeddings (NALE) The figure illustrates the four-step empirical framework. Step ① collects three data inputs: 2,365 annual 10-K MD&A filings from SEC EDGAR covering 255 S&P 500 firms over 2011–2025, a supply chain knowledge graph (KG) with 48 nodes and 51 directed edges, and monthly stock returns from Yahoo Finance. Step ② encodes each MD&A filing using FinBERT (A… view at source ↗

**Figure 2.** Figure 2: Long-Short Portfolio Performance: Network-Augmented LLM Embeddings (NALE), 2013–2025 Notes: Panel A plots cumulative returns of equal-weighted quintile long-short portfolios sorted monthly on pc_5 (Model 1, blue) and net_pc_5 (Model 2, green), alongside the S&P 500 (gray dashed). Panel B plots the rolling 24-month annualized Sharpe ratio. Model 2 peaks at 2.51 in 2020–2021. Sample: February 2013 – December… view at source ↗

**Figure 3.** Figure 3: Decile Portfolio Returns and Sector-Neutral Strategy: Network KG Factor (net_pc_5) Notes: Panel A reports annualized average returns for equal-weighted decile portfolios sorted monthly on net_pc_5. D1 (short leg) has the lowest net_pc_5 scores; D10 (long leg) has the highest. The D1–D10 spread is −8.6% per year. Panel B plots cumulative returns of the raw KG long-short portfolio (dashed) and a sector-neutr… view at source ↗

read the original abstract

This paper proposes a novel asset pricing framework that augments large language model (LLM) embeddings of annual report disclosures with supply chain knowledge graph (KG) propagation. Using FinBERT embeddings of 10-K MD&A sections for 255 S&P 500 firms over 2011-2025, two sets of return predictors are constructed: direct LLM embeddings and network-augmented embeddings, where firm-level signals propagate through inter-firm linkages. Fama-MacBeth cross-sectional regressions reveal that the network-augmented factor (net_pc_5) carries significant return predictability with a 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. The findings suggest that inter-firm network structure contains pricing-relevant information beyond firm-level textual disclosures.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Network-augmented FinBERT embeddings show cross-sectional return power after controls, but the abstract leaves the propagation mechanics and graph construction too opaque to judge if the signal is real or artifactual.

read the letter

The main result is that propagating FinBERT embeddings of 10-K MD&A sections through a supply chain knowledge graph produces a net_pc_5 factor that enters Fama-MacBeth regressions with a Newey-West t-stat of -2.64 after momentum, volatility, and size controls, and supports a long-short portfolio with 0.86 Sharpe and 7.27% annual FF5 alpha (t=2.30). The claim survives the listed robustness checks.

What is new is the specific step of taking LLM embeddings and diffusing them across inter-firm linkages to create the predictor. Earlier papers have either used raw text signals or network centrality separately; the augmentation itself is the incremental move.

The work is straightforward on the empirical side: it reports portfolio sorts, regression coefficients, out-of-sample performance, and sector-neutral results on 255 S&P 500 firms from 2011-2025. That is useful for anyone tracking textual or supply-chain factors.

The soft spots sit in the missing construction details. No equations or description appear for the exact propagation operator, whether the adjacency matrix is lagged and time-varying, how the five principal components are selected, or the precise embedding aggregation. Without those, the reported predictability could arise from the graph structure mechanically correlating embeddings within industries or size groups rather than from genuine information flow. The concern about spurious cross-sectional correlations therefore lands on the current text.

This is for quant-finance readers who already work with text embeddings or supply-chain data and want to see one concrete combination tried. A serious referee would be appropriate because the empirical pattern is testable and the framing is clear enough to evaluate once the methods are filled in. I would send it to review rather than desk-reject, with the expectation that the first round focuses on the propagation rule and data-construction choices.

Referee Report

3 major / 2 minor

Summary. The paper proposes a novel asset pricing framework that augments FinBERT embeddings of 10-K MD&A sections for 255 S&P 500 firms (2011-2025) with propagation through a supply chain knowledge graph. It constructs direct LLM embeddings and network-augmented embeddings (net_pc_5), then shows via Fama-MacBeth regressions that net_pc_5 predicts returns with Newey-West t-statistic -2.64 after controls for momentum, volatility, and size; a long-short portfolio on net_pc_5 yields annualized Sharpe ratio 0.86 and FF5 alpha of 7.27% (t=2.30). The predictive power is claimed to survive out-of-sample tests, placebo experiments, sector-neutralization, and subsample analysis, implying that inter-firm network structure contains pricing-relevant information beyond firm-level textual disclosures.

Significance. If the propagation step is shown to reflect genuine lagged information flow without mechanical artifacts, the result would be significant for textual asset pricing by demonstrating that supply-chain linkages amplify cross-sectional return predictability from disclosures. The reported economic magnitudes (Sharpe 0.86, alpha t=2.30) and robustness checks would strengthen the case for network-augmented textual factors, provided the construction avoids circularity with the tested returns.

major comments (3)

[Methodology / Embedding Propagation] The manuscript provides no equations or algorithmic description for the KG propagation step that produces net_pc_5 from FinBERT embeddings (including adjacency-matrix construction, normalization, time-variation, lag structure, or principal-component selection). This is load-bearing for the central claim, as the reported NW t-statistic of -2.64 and FF5 alpha could arise from graph-induced correlations rather than information propagation.
[Results / Fama-MacBeth Regressions] It is unclear whether net_pc_5 is constructed in a manner that uses only lagged supplier data without look-ahead bias or whether the propagation operator mechanically correlates embeddings within industries or by size even after the listed controls; the abstract's claim of survival after sector-neutralization does not resolve this without the explicit rule.
[Abstract / Empirical Design] The abstract reports t-statistics and alphas after controls but supplies no details on embedding construction, exact propagation algorithm, principal-component selection, or data-exclusion rules, preventing assessment of whether the predictability reduces to quantities fitted on the same return data used for testing.

minor comments (2)

[Data] The sample period 2011-2025 extends into the future; clarify the exact end date of the data and any forward-looking elements.
[Variable Construction] Clarify the precise definition of 'net_pc_5' (e.g., which principal components are retained and why five).

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight areas where additional methodological transparency will strengthen the paper. We address each point below and will incorporate revisions to provide the requested details on the propagation algorithm, bias safeguards, and abstract description.

read point-by-point responses

Referee: [Methodology / Embedding Propagation] The manuscript provides no equations or algorithmic description for the KG propagation step that produces net_pc_5 from FinBERT embeddings (including adjacency-matrix construction, normalization, time-variation, lag structure, or principal-component selection). This is load-bearing for the central claim, as the reported NW t-statistic of -2.64 and FF5 alpha could arise from graph-induced correlations rather than information propagation.

Authors: We agree that the current draft lacks sufficient detail on the propagation operator. In the revised version we will insert a new subsection (3.2) containing the full specification: the adjacency matrix is constructed from lagged supplier-customer links extracted from 10-K filings and FactSet data up to t-1; the matrix is row-normalized with a decay factor for multi-hop paths; propagation is applied as a matrix multiplication on the prior-period embedding vectors; net_pc_5 is the fifth principal component of the resulting network-augmented matrix, chosen by eigenvalue threshold on the training window only. These equations will make clear that the operator uses strictly lagged information and is not a mechanical within-industry smoother. revision: yes
Referee: [Results / Fama-MacBeth Regressions] It is unclear whether net_pc_5 is constructed in a manner that uses only lagged supplier data without look-ahead bias or whether the propagation operator mechanically correlates embeddings within industries or by size even after the listed controls; the abstract's claim of survival after sector-neutralization does not resolve this without the explicit rule.

Authors: The construction uses only lagged supplier data: the knowledge graph at each rebalancing date t is built exclusively from disclosures filed by t-1. We will add an explicit statement of this timing rule together with two further robustness tables: (i) size-neutral long-short portfolios and (ii) industry-size double-sorted portfolios. These will be reported alongside the existing sector-neutral results to demonstrate that residual predictability is not an artifact of mechanical correlation induced by the graph. revision: yes
Referee: [Abstract / Empirical Design] The abstract reports t-statistics and alphas after controls but supplies no details on embedding construction, exact propagation algorithm, principal-component selection, or data-exclusion rules, preventing assessment of whether the predictability reduces to quantities fitted on the same return data used for testing.

Authors: We will revise the abstract to include one concise sentence describing the lagged KG propagation and the selection of net_pc_5 via out-of-sample principal components. The full algorithmic description and data-exclusion criteria (S&P 500 membership, filing-date alignment, and minimum supplier coverage) will remain in the main text, but the abstract will now signal that the network step is deterministic given the lagged graph and is not estimated on the test-period returns. revision: yes

Circularity Check

0 steps flagged

No circularity: predictors constructed from independent textual and network inputs, tested on returns

full rationale

The derivation constructs firm-level signals from FinBERT embeddings of 10-K MD&A text and propagates them via an external supply-chain KG to form net_pc_5; these are then used as regressors in Fama-MacBeth and portfolio tests. No equation or step equates the predictor to a fitted function of the same returns being predicted, no self-citation supplies a load-bearing uniqueness result, and the construction steps rely on disclosure text plus graph structure rather than on the target return series. The reported t-statistics and alphas are therefore not forced by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard asset pricing regression assumptions and the validity of the supply chain graph; no new entities are postulated and the propagation step is the main addition whose details are not visible in the abstract.

free parameters (1)

net_pc_5
The subscript 5 likely indexes a principal component or lag choice whose selection affects the reported predictor.

axioms (1)

domain assumption Fama-MacBeth cross-sectional regressions produce unbiased estimates of risk premia under the maintained controls
Invoked when claiming the t-statistic survives controls for momentum, volatility, and size.

pith-pipeline@v0.9.1-grok · 5732 in / 1308 out tokens · 34237 ms · 2026-06-30T02:05:16.110601+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages · 1 internal anchor

[1]

FinBERT: Financial Sentiment Analysis with Pre-trained Language Models

Ahern, K.R., and J. Harford (2014). The importance of industry links in merger waves. Journal of Finance, 69(2), 527–576. Araci, D. (2019). FinBERT: Financial sentiment analysis with pre-trained language models. arXiv preprint arXiv:1908.10063. Barrot, J. -N., and J. Sauvagnat (2016). Input specificity and the propagation of idiosyncratic shocks in produc...

work page internal anchor Pith review Pith/arXiv arXiv 2014
[2]

Chronologically consistent large language models, 2025

Fama, E.F., and K.R. French (1993). Common risk factors in the returns on stocks and bonds. Journal of Financial Economics, 33(1), 3–56. Fama, E.F., and K.R. French (2015). A five -factor asset pricing model. Journal of Financial Economics, 116(1), 1–22. Fama, E.F., and J.D. MacBeth (1973). Risk, return, and equilibrium: Empirical tests. Journal of Politi...

work page arXiv 1993

[1] [1]

FinBERT: Financial Sentiment Analysis with Pre-trained Language Models

Ahern, K.R., and J. Harford (2014). The importance of industry links in merger waves. Journal of Finance, 69(2), 527–576. Araci, D. (2019). FinBERT: Financial sentiment analysis with pre-trained language models. arXiv preprint arXiv:1908.10063. Barrot, J. -N., and J. Sauvagnat (2016). Input specificity and the propagation of idiosyncratic shocks in produc...

work page internal anchor Pith review Pith/arXiv arXiv 2014

[2] [2]

Chronologically consistent large language models, 2025

Fama, E.F., and K.R. French (1993). Common risk factors in the returns on stocks and bonds. Journal of Financial Economics, 33(1), 3–56. Fama, E.F., and K.R. French (2015). A five -factor asset pricing model. Journal of Financial Economics, 116(1), 1–22. Fama, E.F., and J.D. MacBeth (1973). Risk, return, and equilibrium: Empirical tests. Journal of Politi...

work page arXiv 1993