Can Supply-Chain Data Build Better Stock Peer Groups Than Industry Codes?
September 14, 2026
Altsets
Research by Altsets Research
Shared customers, shared suppliers, and similar dependency profiles can define dynamic economic peer groups for normalization, relative value, stat arb, and risk modeling that compete directly with fixed sector classifications.
Data used:Altsets Supply Chain Intelligence: 90k+ entities, 400k+ relationships, 20+ years of history.
Key findings
- The supplied network creates economically meaningful peer relationships that standard labels can miss, including HPE and Nvidia through Microsoft and Micron and SK Hynix through Nvidia.
- A 2026 Journal of Financial Economics study finds predictive information in economically motivated peer networks beyond firm-level machine-learning characteristics, providing a strong benchmark for testing network-defined peer groups against fixed industry peers.
Sometimes. Supply-chain data can build better peer groups when shared customers, suppliers, or dependency profiles explain the comparison better than industry labels, but the improvement must be tested for the specific valuation, risk, or trading task.
Economic peers can cross standard industry boundaries
The supplied Altsets network shows HPE connected to Microsoft with a 561M USD relationship, while Nvidia also maps to Microsoft as a customer relationship in the supplied graph. HPE and Nvidia are not obvious conventional peers. They have different products, different margins, and different sector narratives. From the perspective of one Microsoft-driven catalyst, however, they can belong to the same economic peer set because both have a documented relationship path to the same outside company.
The reverse pattern also matters. Micron and SK Hynix look like obvious semiconductor peers, but their real economic similarity can change depending on which customers, products, and suppliers are active. A network-defined peer score can use shared Nvidia exposure, other overlapping customers, upstream dependencies, and relationship magnitude to decide how similar the companies are at a specific point in time. The result can be more granular than assigning every memory company the same peer weight because they share an industry code.
Peer groups can be used without predicting returns directly
A supply-chain peer model can improve several quantitative tasks even if it never becomes an alpha factor. Cross-sectional features can be normalized against economically similar firms rather than an entire sector. Valuation multiples can be compared with companies that share important customers instead of companies that merely sell a similar product. A residual-return model can remove a peer component built from connected firms before testing company-specific signals. A portfolio can constrain exposure to one dependency cluster even when the constituent stocks span several industries.
Recent research provides a useful benchmark for this idea. A 2026 Journal of Financial Economics paper constructs a Peer Index from economically motivated peer networks and reports predictive power for stock returns and earnings surprises that is not subsumed by firm-level machine-learning characteristics. That result does not validate any particular Altsets peer construction, but it supports the broader research question: economically defined peer networks can contain information that standard firm characteristics do not fully capture.
Relationship direction and strength can improve peer construction
A binary shared-customer indicator is only the simplest version. The supplied Altsets data can distinguish supplier revenue share from customer cost share, which allows the peer definition to ask whether two companies depend on the same outside firm in comparable ways. Two suppliers that both sell to Nvidia may be closer economic peers if Nvidia is financially important to both than if one relationship is minor and the other is central. Structural-only edges can still contribute to topology without being assigned fake numeric weights.
The peer score can also be asymmetric. Company A can view Company B as a useful peer because they share several important customers, while B's best peer set can differ because its wider customer network is much broader. That is not a flaw. Economic similarity does not have to satisfy the same symmetry as an industry classification. A directed graph can support peer definitions tailored to the feature or event being modeled.
Dynamic peer groups can adapt when business relationships change
Industry codes are relatively sticky. Supply-chain peer groups can change when a company wins a new customer, loses an old supplier, enters a different product ecosystem, or becomes more economically dependent on one external node. That creates a dynamic peer model that may be better suited to fast-changing technology sectors. It also raises the historical-validation burden because today's peers cannot be attached to old observations as though the relationships always existed.
A point-in-time implementation can reconstruct the peer set at each formation date, calculate the peer benchmark from only contemporaneously observable relationships, and then test whether peer-relative features improve out-of-sample performance. If the dynamic network peer set changes before standard classifications do, that can be the source of incremental information. If the result disappears after industry and ordinary return-similarity controls, the network may not be adding enough to justify the additional complexity.
The conclusion is to let economic relationships compete with labels
A GICS or NAICS peer group is useful because it is simple, stable, and interpretable. Supply-chain data should not replace it by assumption. The quant question is whether companies that share economically important customers, suppliers, and dependency structures provide a better benchmark for the specific modeling task than companies grouped only by industry. That can be tested directly through residual variance, forecast accuracy, stat-arb stability, valuation dispersion, or out-of-sample return prediction.
The stat-arb pair-selection guide explains how economic relationships can narrow a relative-value search space. The incremental-alpha guide explains how to test whether network information adds value after industry and conventional predictors are already present.
For relationship definitions and evidence limits, read the Altsets methodology.