Earnings-Call NLP Tells You What Was Said. Supply-Chain Data Tells You Whose Call Matters.
September 14, 2026
Altsets
Research by Altsets Research
Transcript models extract tone, semantic change, and management language. Supply-chain data defines the economically relevant outside companies whose calls should enter another stock's research process.
Data used:Altsets Supply Chain Intelligence: 90k+ entities, 400k+ relationships, 20+ years of history.
Key findings
- The supplied 561M USD HPE-Microsoft relationship and Nvidia's 17.62% share of Micron revenue provide economic reasons to include Microsoft and Nvidia calls in HPE and Micron research respectively.
- Academic transcript research finds useful information in tone, question-and-answer language, and semantic features, while the graph solves the separate problem of which outside transcripts deserve attention.
Earnings-call NLP tells you what management said and how it said it. Supply-chain data tells you which other companies' calls deserve to be read for the stock you are researching. The datasets solve different problems: text analysis extracts information from a document, while the relationship graph chooses the economically relevant document universe.
A transcript model still needs a reason to care about another company's call
Research on earnings calls finds that textual tone, question-and-answer language, and semantic features can contain information about future returns, analyst revisions, and earnings outcomes. Modern NLP and LLM methods can process thousands of words that would be slow for a human analyst to compare manually.
The problem is selection. There are thousands of public-company calls every quarter. A model can summarize all of them, but an investor researching one stock still needs a principled reason to decide which outside transcripts matter.
That is where the graph adds a different layer.
HPE gives the transcript model an external reading list
The supplied Altsets data shows a 561M USD relationship between Hewlett Packard Enterprise and Microsoft, compared with 203M USD for Swisscom, 109M USD for Home Depot, and 64.5M USD for Volkswagen across the four displayed HPE customer relationships.
If the investor is researching HPE, Microsoft is therefore not just a popular technology company whose transcript might contain vaguely relevant enterprise commentary. It is the largest relationship in that displayed customer set by relationship size.
A transcript system can now read Microsoft's call with a specific question in mind: did management say anything about Azure, hybrid infrastructure, enterprise demand, AI infrastructure, licensing, or customer spending that could alter the HPE thesis? The relationship data creates the reading priority before the NLP model interprets the language.
Micron and Nvidia show the same idea with a directional percentage
The supplied data shows Nvidia representing 17.62% of Micron revenue. That makes Nvidia's call economically relevant to Micron in a way that an unrelated semiconductor or software transcript is not.
The NLP system can still extract tone, semantic change, uncertainty, management responsiveness, or specific product language from Nvidia's transcript. The graph tells the system why that text belongs in the Micron research process.
This separation is important because a positive transcript score is not useful merely because two companies share an industry. The economic relationship provides a mechanism through which the information could matter.
The strongest workflow uses text to interpret the edge rather than invent the edge
An LLM reading a transcript can hallucinate or overstate commercial relationships if it tries to infer the entire supply chain from management language alone. Company executives can speak broadly about customers, markets, partners, and ecosystems without naming the exact economic counterparties or quantifying their importance.
A normalized relationship graph can provide the company pair, direction, metric, evidence state, and historical context first. The transcript model can then answer narrower questions about what changed around that known relationship.
This is especially valuable for repeated monitoring. The graph changes slowly. The text changes every quarter. One layer defines the economic structure while the other layer updates the current narrative.
The conclusion is that the graph chooses whose words matter
Earnings-call NLP is powerful at answering what a company said, how tone changed, and what themes appeared in the discussion. Supply-chain data answers which outside companies' words should enter another stock's research process in the first place. The supplied HPE-Microsoft and Micron-Nvidia relationships turn transcript analysis from a broad text-mining exercise into relationship-specific investment research.
The connected-company filings guide explains why investors should read disclosures from companies they do not own. The analyst-network comparison explains how economic relationships and information-attention relationships can be combined without treating them as the same network.
For relationship definitions and evidence limits, read the Altsets methodology.