Does Supply-Chain Network Centrality Predict Risk or Just Company Size?
September 14, 2026
Altsets
Research by Altsets Research
Degree, weighted degree, betweenness, and other graph measures can encode economically interesting network position, but they need point-in-time construction and controls for company size, industry, liquidity, analyst attention, and data coverage.
Data used:Altsets Supply Chain Intelligence: 90k+ entities, 400k+ relationships, 20+ years of history.
Key findings
- The supplied Tesla and Shin-Etsu views illustrate why simple degree and economically weighted degree answer different questions, but local screenshots are not enough to claim a global centrality ranking.
- Academic studies have linked supply-chain centrality with crash risk and stock-price efficiency, making centrality a legitimate feature family while still requiring replication against simple size, degree, and coverage baselines.
It can predict either, so the distinction must be tested rather than assumed. Supply-chain centrality is a distinct risk or return signal only if it adds out-of-sample information after controlling for company size, industry structure, liquidity, and uneven relationship-data coverage.
Local importance and global importance are different
The supplied Tesla network provides an intuitive example of local connectivity. The graph shows Tesla connected with several upstream companies, including LG Energy Solution, Samsung Electronics, Lens Technology, Huayu Automotive Systems, Toyota Industries, and other nodes. That does not establish that Tesla ranks highly on a global centrality statistic because the supplied view is only a local slice of the network. It does show why simple degree and economically weighted degree can answer different questions even before a global graph measure is calculated.
The supplied Shin-Etsu Chemical view offers another pattern. Shin-Etsu connects to Samsung, TSMC, and Intel with different supplier revenue and customer cost shares. A pure degree measure would count three displayed customer links equally, while a weighted measure could preserve the fact that TSMC is the largest of those three displayed relationships by both directional metrics. Neither representation is universally correct. Degree measures breadth; weighted degree adds economic magnitude; global centrality tries to capture position in the wider network.
Centrality can be correlated with size and disclosure coverage
Large public companies naturally have more opportunities to appear in observed customer and supplier datasets. They may also disclose more relationships, receive more analyst attention, and sit in industries with better public documentation. A high centrality score can therefore reflect economic importance, company size, dataset coverage, or all three at once. A quant who finds that central firms have different returns or risk needs to determine whether the effect survives controls for market capitalization, industry, liquidity, analyst coverage, and relationship-data completeness.
Existing academic work shows that supply-chain centrality has been associated with financial outcomes. Research on Chinese listed manufacturing firms has linked network centrality with stock-price crash risk, while newer work studies how degree and structural-hole centrality relate to stock-price efficiency. Those findings make centrality a legitimate candidate feature, but they do not establish that any particular centrality measure will create tradable alpha in another market or dataset. The replication needs to begin from the local data-generating process.
Different centrality measures imply different economic mechanisms
A high-degree company may simply interact with many counterparties. A high-betweenness company can sit on paths connecting otherwise separated parts of the network, which suggests a different shock-transmission role. Eigenvector-style centrality can emphasize relationships with other central firms, potentially identifying companies embedded in economically important clusters. These measures should not be thrown into one model as interchangeable graph statistics because each one corresponds to a different story about information or risk transmission.
The hypothesis should determine the metric. If the researcher expects widely connected companies to absorb diversified information, degree may be relevant. If the hypothesis concerns bottlenecks and shock propagation, betweenness or directed path measures may be more natural. If the question is whether a firm's customers are themselves economically central, an eigenvector-style measure may be appropriate. The centrality feature becomes more defensible when the mathematical definition maps to an economic mechanism.
Temporal centrality creates another leakage risk
A company's network position changes as customers, suppliers, and corporate entities change. Calculating centrality on today's full graph and attaching the score to historical observations introduces future edges into the past. The leakage can be more severe than with a single relationship feature because a centrality score depends on the structure of many other nodes at the same time. One future relationship can alter several companies' graph statistics through indirect paths.
Point-in-time graph reconstruction is therefore essential. At each historical formation date, the quant should calculate centrality using only the nodes and edges that belong to that date's information set. The feature should then be compared with simpler contemporaneous measures such as firm size, industry degree, or quantified relationship counts. If the sophisticated centrality measure cannot outperform those simpler representations out of sample, its extra graph complexity may not be justified.
The conclusion is that centrality is a hypothesis family, not one magic factor
Supply-chain networks naturally support centrality research, and the literature gives good reasons to study how network position relates to information efficiency, risk, and return comovement. The challenge is separating true network position from size, industry, and coverage artifacts. A useful centrality factor needs an economic mechanism, point-in-time graph construction, simple size and degree baselines, and an out-of-sample benefit that survives ordinary controls. Otherwise the researcher may have calculated an impressive graph statistic without learning anything new about the stock.
The GNN guide explains when more complex graph representations may be justified. The missing-data quant guide explains why observed graph density can reflect coverage as well as economics.
For relationship definitions and evidence limits, read the Altsets methodology.