Can Supply-Chain Data Improve Statistical Arbitrage Pair Selection?
September 14, 2026
Altsets
Research by Altsets Research
Use shared customers, shared suppliers, and network similarity as an economic filter for stat-arb candidate generation, then require the same out-of-sample spread, liquidity, and cost tests as any other pairs strategy.
Data used:Altsets Supply Chain Intelligence: 90k+ entities, 400k+ relationships, 20+ years of history.
Key findings
- The supplied network maps both Micron and SK Hynix to Nvidia, providing an economically interpretable shared-demand reason to test the pair without claiming that the relationship itself produces a profitable spread.
- A defensible experiment should compare network-selected pairs with ordinary price-selected or industry-matched pairs using the same point-in-time, out-of-sample, transaction-cost, and stability tests.
Potentially. Supply-chain data can improve statistical-arbitrage pair selection by narrowing the search to companies with a shared customer, supplier, or dependency structure, making candidate pairs more economically interpretable. It does not prove that any pair will mean-revert.
With thousands of stocks, the number of possible pairs becomes enormous. A researcher can search for historically correlated or cointegrated pairs, but many of those relationships can be accidental, regime-specific, or difficult to explain economically.
Supply-chain data offers another way to build the candidate universe: start with companies connected by a plausible economic relationship, then test whether the prices contain a tradable relative-value structure.
Economic linkage can be a prior for pair selection
Micron and SK Hynix are a useful example.
The supplied Altsets network maps both companies to Nvidia as a customer, and current public announcements connect both memory suppliers to Nvidia's AI roadmap.
That does not mean Micron and SK Hynix form a profitable pair.
It does mean the companies share an economically interpretable demand node, giving a researcher a reason to test whether their relative behavior contains structure beyond ordinary semiconductor co-movement.
The graph can reduce the combinatorial search space
A naive pair search across 2,000 equities contains nearly two million unique pairs.
Most of those combinations have no obvious economic relationship.
A network-conditioned search can restrict candidates to companies that share important customers, share important suppliers, sit on adjacent network paths, or have similar dependency profiles.
That can reduce the number of hypotheses before the price data is examined.
Reducing the search space can also reduce the opportunity for accidental in-sample discoveries.
Supply-chain similarity should not replace the statistical test
The relationship graph is a candidate-generation layer.
The researcher still needs to test spread behavior, stability, factor exposures, liquidity, transaction costs, borrow constraints where relevant, and out-of-sample performance.
Two companies can share Nvidia and still have completely different product economics, capital structures, geographies, and valuation regimes.
The economic connection explains why the pair is worth investigating. It does not prove mean reversion.
Compare network-selected pairs with ordinary pair-selection methods
A useful experiment can form two candidate sets.
One set selects pairs through standard historical price similarity. Another selects or filters pairs using supply-chain relationships before applying the same statistical tests.
The researcher can then compare out-of-sample stability, turnover, decay, drawdowns, and sensitivity to market regimes.
This is more informative than reporting the best supply-chain-selected pair in isolation.
The question is whether the network improves candidate quality across many trials.
Relationship direction can create asymmetric hypotheses
A shared customer pair is not the only design.
A supplier-customer pair can create a lead-lag hypothesis. Two suppliers to the same customer can create a relative-demand hypothesis. Two customers sharing a critical supplier can create a common-bottleneck hypothesis.
Those structures suggest different spread models and event conditioning.
The network topology can therefore influence the form of the stat-arb test rather than merely selecting two tickers.
The pair should be point-in-time valid
A pair selected because two companies share a customer today should not automatically be tested ten years backward as though the shared customer existed throughout the period.
The relationship state and the pair-selection rule need to be reconstructed at each formation date.
Otherwise the researcher has introduced the same future-information problem that affects any historical relationship strategy.
This is especially important when companies enter or leave product ecosystems quickly.
The conclusion is to use the network as an economic filter
Classic pairs trading begins with historical price similarity.
Supply-chain data offers an additional route: identify companies with a reason to share an economic driver, then ask whether the market prices create a robust relative-value opportunity.
The graph can make the pair search more intentional. The backtest still has to prove that the spread is tradable.
The shared-customer relative-value guide covers fundamental comparison between companies tied to the same customer. The point-in-time backtesting guide explains why relationship formation dates must be historically valid.
For relationship definitions and evidence limits, read the Altsets methodology.