Supply Chain Knowledge Graph

Alpha Hypothesis

Language models are effective at extracting relationships from filings and transcripts. That lets you maintain a live map of who sells to whom, who depends on which inputs, and which firms are exposed to the same end-market demand.

• Company A buys from Company B
• Company C depends on commodity D
• Company E competes with Company F in segment G

When a shock hits one node, you propagate expected impact through first- and second-order links before the market fully prices the chain. This is where the graph becomes useful: not for static visualization, but for dynamic shock translation.

Data Inputs

• 10-K business descriptions and risk factors
• Earnings call transcripts and investor-day transcripts
• Major customer and supplier commentary
• Trade, tariff, and export-control announcements
• Commodity, freight, and logistics updates
• M&A releases and divestiture announcements

The institutional edge comes from maintaining these links continuously. A stale graph is worse than no graph because it gives false confidence. Timestamps and decay matter.

Graph Schema

{
  "source_company": "NVDA",
  "target_company": "SMCI",
  "relationship": "demand_beneficiary",
  "strength": 0.82,
  "direction": "positive",
  "economic_channel": "AI server demand",
  "timestamp": "2026-01-29",
  "evidence": "quote or filing span"
}

Useful Relationship Types

• supplier_of
• customer_of
• competes_with
• shares_end_market_with
• input_cost_exposed_to
• demand_beneficiary_of
• regulatory_peer_of

How to Build the Graph

1. Extract company names, products, regions, and demand channels from text.
2. Infer relationship type and direction from surrounding context.
3. Assign confidence and economic relevance score.
4. Store each edge with evidence spans and timestamps.
5. Refresh or decay edges as new documents arrive.

Why GPT Helps Here

Most raw documents do not literally say "A is a second-order beneficiary of B." They describe products, customers, channels, and dependencies in fragmented language. GPT-style models are good at normalizing those fragments into a usable relationship graph if you force structured extraction and preserve evidence.

Shock Propagation

Once the graph exists, you need a propagation model. A simple version works surprisingly well:

1. Classify the shock type: demand, pricing, regulation, litigation, supply disruption, input-cost move.
2. Assign direct impact to the primary node.
3. Propagate impact through linked nodes with decaying weight by distance.
4. Reduce scores for stale, low-confidence, or crowded edges.
5. Rank beneficiaries and victims against actual market movement.

Shock: airline calls out weakening premium travel demand
Graph impact:
    negative -> airline peers
    negative -> travel platforms with high premium exposure
    negative -> jet fuel refiners with demand beta
    positive -> defensives if macro read-through broadens

Trade:
    short weakest exposed basket
    hedge with sector ETF or market beta neutral overlay

Trade Construction

The cleanest expressions tend to be basket-based and relative-value oriented.

• Beneficiary basket: names with positive propagated score and weak initial price reaction
• Victim basket: names with negative propagated score but delayed market repricing
• Relative-value pair: long beneficiary, short exposed peer or sector ETF
• Time horizon: 2 days to 6 weeks depending on whether the shock is tactical or structural

The best hunting grounds are areas where the relationship web is real but underappreciated: industrial chains, semiconductors, specialty retail, transports, energy services, and commodity-linked sectors.

Risks and Controls

• Decay edges: a supplier relationship from three years ago may be wrong now
• Tag economic materiality: not every relationship moves earnings
• Avoid story traps: require price, volume, or revision confirmation
• Use baskets: single-node linkages are noisy
• Watch crowding: in mega-cap themes, the graph edge often gets priced immediately