Extracting Supply-Chain Relationships From Text
NLP models can mine filings, transcripts, and news for supplier-customer mentions, building a map of which companies depend on which others without needing anyone to publish that map directly.
Companies rarely publish a clean list of who their suppliers and customers are, but that information leaks out constantly — in 10-K risk-factor sections ("we rely on a single supplier for..."), in earnings-call transcripts ("our largest customer, [Company], accounts for..."), and in trade press. Information extraction models scan this text for supplier-customer language patterns and build a directed graph connecting companies to the other companies they depend on, at a scale no analyst could match by reading manually.
This matters for investing because economically linked companies tend to have related fortunes: a factory shutdown at a key supplier can crimp a customer's output well before the customer's own numbers show it, and news that moves one node in the graph often should move its close neighbours too, with a predictable lag.
Supply-chain graphs built by mining filings and transcripts for mentions of suppliers and customers reveal economic links that aren't otherwise public, and price moves tend to propagate along those links with a delay the market doesn't always immediately price in.
Worked example
An NLP pipeline scans a semiconductor customer's 10-K and flags the sentence "approximately 30% of our chip supply is sourced from a single foundry located in Taiwan," matching it to a specific supplier already in the company database. A week later, news breaks of a production disruption at that foundry. A strategy using the extracted supply-chain graph immediately flags the customer as exposed and can act before the customer's own quarterly filing confirms any impact months later.
Further reading
- Cohen and Frazzini, 'Economic Links and Predictable Returns' (Journal of Finance)