How Do You Know a Customer-Supplier Spillover Is Not Just Industry Momentum?
September 14, 2026
Altsets
Research by Altsets Research
A credible spillover test needs industry controls, unrelated-event placebos, fake-network benchmarks, pre-trend checks, reverse-direction tests, and exposure gradients before a customer-supplier link is treated as the source of predictability.
Data used:Altsets Supply Chain Intelligence: 90k+ entities, 400k+ relationships, 20+ years of history.
Key findings
- The supplied Nvidia relationships with Micron and SK Hynix create a plausible customer-information channel, but semiconductor comovement remains an obvious confound that must be controlled rather than assumed away.
- The strongest network explanation should survive industry-matched controls, random-edge or unrelated-event placebos, pre-event trend tests, reverse-direction checks, and a relationship between economic exposure strength and measured spillover.
Customer-supplier return spillovers are one of the most interesting quantitative ideas in supply-chain research, and also one of the easiest to overstate. Cohen and Frazzini documented predictable returns across economically linked firms, and newer work continues to study momentum propagation through customer-supplier networks. A backtest that finds connected firms moving after one another is therefore plausible, but plausibility is not causality. The researcher still needs to show that the relationship adds explanatory power beyond industry momentum, shared factors, simultaneous news, and data-mining choices.
Start with a mechanism that predicts more than comovement
The supplied Altsets network maps both Micron and SK Hynix to Nvidia. If Nvidia receives positive demand information, a limited-attention hypothesis predicts that economically exposed suppliers may incorporate some of that information with a delay. That mechanism produces several testable implications. More-exposed suppliers should generally display stronger spillover than weakly exposed or unrelated firms, the effect should be strongest after information that plausibly changes supplier demand, and the result should weaken after the customer information is fully incorporated.
Those predictions are more informative than finding positive return correlation between Nvidia and memory stocks. The companies share an industry theme, macro sensitivities, and AI expectations, so comovement is expected even without a direct customer channel. A causal story needs to explain why the documented relationship changes the behavior relative to otherwise similar companies. The relationship graph should therefore be used to define differences in exposure rather than simply to label every connected stock as treated.
Industry and common-factor placebos are essential
The supplied Shin-Etsu Chemical data makes the confounding problem obvious. Samsung, TSMC, and Intel are separate customers, but all participate in semiconductor manufacturing. A supplier connected to those firms can move with them because of semiconductor-cycle news even when no customer-specific information is propagating. A spillover test should therefore compare connected firms with industry-matched firms that share the broad cycle but do not share the specific relationship being tested.
Another placebo can use events from economically unrelated firms. If Micron appears to react after every large technology company's earnings regardless of whether that company is a customer, the Nvidia relationship is less convincing as the mechanism. Randomly rewiring edges while preserving broad network degree can provide another benchmark: if the actual network does not outperform fake networks with similar topology, the apparent spillover may come from generic industry structure rather than the true customer-supplier links.
Timing and pre-trends can reveal reverse stories
A clean spillover hypothesis usually implies an event ordering. Customer information arrives first, then the supplier incorporates it. If the supplier begins moving before the customer event, the result may reflect common information that both markets were already processing rather than delayed transmission from the customer. Examining pre-event abnormal returns and volatility can therefore falsify some versions of the spillover story.
The researcher should also consider reverse causality. A supplier can contain information about the customer's future production, and both firms can respond to third-party news from a common product market. Directional relationship data helps organize the possibilities, but the arrow in the commercial graph does not automatically identify the direction of stock-price information flow. Customer-to-supplier and supplier-to-customer lead-lag tests should be compared rather than assumed from the business relationship alone.
Exposure gradients and event specificity strengthen the case
A stronger design can test whether spillover magnitude increases with economic dependence. The relationship percentage should not be converted directly into expected return, but it can define ordered exposure groups. If customer news produces larger subsequent effects among suppliers for which that customer represents a larger share of revenue, the result is more consistent with an economic transmission mechanism than a simple binary edge test.
Event specificity matters too. Demand guidance, product ramps, capital spending, and purchasing changes should plausibly matter more to suppliers than unrelated corporate news such as a legal settlement with no connection to purchasing. If all customer announcements generate the same measured spillover, the strategy may be capturing attention or market beta rather than supply-chain economics. A causal interpretation becomes stronger when the response varies in the direction predicted by the business mechanism.
The conclusion is to make the network hypothesis capable of failing
Supply-chain links are a plausible channel for information propagation, and decades of research give quants good reasons to investigate them. That is exactly why the tests should be demanding. A convincing spillover result should survive industry controls, unrelated-event placebos, fake-network benchmarks, pre-trend checks, reverse-direction tests, and exposure gradients. If the effect disappears under those tests, the researcher has learned that the relationship is economically real but not the source of the trading signal. If it survives, the network explanation becomes much harder to dismiss as ordinary sector momentum.
The event-study guide explains how to define treated and control firms before observing event returns. The hypothesis-testing guide explains why a rich relationship dataset needs explicit falsification and untouched out-of-sample evidence.
For relationship definitions and evidence limits, read the Altsets methodology.