How Do You Run a Supply-Chain Event Study Without Fooling Yourself?
September 14, 2026
Altsets
Research by Altsets Research
Use the pre-event relationship graph to define treated firms, construct controls before observing returns, choose event windows from the transmission mechanism, and test whether stronger economic exposure produces stronger abnormal responses.
Data used:Altsets Supply Chain Intelligence: 90k+ entities, 400k+ relationships, 20+ years of history.
Key findings
- The supplied network provides different evidence levels for Microsoft-linked companies, including a quantified 561M USD HPE-Microsoft relationship and a structural Nvidia-Microsoft relationship, so event-study treatment should preserve that distinction.
- A defensible event study freezes the historical treatment set before the event, uses matched or exposure-ranked controls, flags overlapping news, and chooses event windows for economic reasons rather than the window that produces the largest result.
Run a supply-chain event study by defining the event, connected firms, exposure thresholds, comparison group, return window, and exclusions before examining the outcome. The network defines the treated firms in advance; the event study then tests whether they behaved abnormally without confusing market moves, sector news, overlapping earnings, or post-hoc selection with genuine propagation.
Define treatment from the historical relationship state
The supplied Altsets data shows a 561M USD relationship between HPE and Microsoft, while Nvidia also maps to Microsoft as a customer relationship in the supplied network without a displayed economic metric on that edge. That creates two different evidence levels for a hypothetical Microsoft earnings event study. HPE can be treated as a quantified connected firm, while Nvidia can remain in a structural-only connected group. The researcher should not pretend those relationships are equally strong simply because both companies appear next to Microsoft in the graph.
The treatment definition should be frozen using the relationship snapshot available before the event. If the researcher waits until after Microsoft reports and then chooses the suppliers or customers that happened to move most, the study becomes circular. A stronger design specifies the treated set using point-in-time relationships, then compares realized returns, volume, volatility, revisions, or another outcome around the event. MacKinlay's classic event-study framework is useful because it emphasizes measuring abnormal security behavior relative to an estimated normal-return process rather than treating raw price moves as the event effect.
The control group determines what the event study really tests
An HPE move around Microsoft earnings could reflect Microsoft-specific information, enterprise technology news, a broad market move, or unrelated HPE information released at the same time. A useful control group tries to remove those alternative explanations. One design can match HPE with companies from the same industry and similar size that are not connected to Microsoft. Another can compare more-exposed and less-exposed Microsoft-linked firms while controlling for ordinary sector returns. The strongest design depends on the hypothesis, but the control group should be chosen before the result is known.
A network can improve control construction because it identifies economically similar but differently exposed companies. For example, a researcher can match firms on industry, size, and momentum, then use supply-chain relationships to separate those with a Microsoft connection from those without one. The event study is no longer asking whether technology stocks moved after Microsoft earnings. It is asking whether companies with a documented Microsoft dependency behaved differently from comparable companies without that path.
Event windows should match the mechanism
A one-day event window can be appropriate if the hypothesis is rapid information transmission. A wider window may be necessary if the event occurs after market close, if international securities trade in different time zones, or if the hypothesis concerns slow attention spillover. The researcher should choose the window because of the expected transmission mechanism rather than because one particular window produced the strongest result. Testing many windows and reporting only the best one creates another multiple-testing problem.
Supply-chain events can also overlap. A supplier may report its own earnings near the customer event, another large customer may issue guidance, or an industry conference can introduce simultaneous information. Those contaminated events should be flagged or excluded according to a rule defined before the final test. The more connected the firms are, the more likely some events cluster in time. That makes clean event selection an important part of network research rather than an administrative detail.
Exposure gradients can make the test more informative
A binary connected-versus-unconnected indicator throws away information when directional economic metrics are available. The researcher can ask whether firms with larger supplier revenue exposure to the reporting customer exhibit larger abnormal responses. The relationship percentage should not be treated as a predicted stock return, but it can define an exposure gradient for testing whether event sensitivity increases with economic dependence.
The same idea can be used without forcing structural edges into fake numbers. Quantified relationships can form ordered exposure buckets, while structural-only edges can remain a separate category. If the strongest abnormal response appears among the most economically exposed firms and weakens across lower-exposure groups, the result is more consistent with a relationship mechanism than a simple market-wide reaction. If no gradient appears, that is valuable evidence against the strongest version of the hypothesis.
The conclusion is that the graph defines who should care before the event happens
A supply-chain event study should begin with a pre-event network, a clearly defined treatment group, a defensible control group, and event windows chosen for economic reasons. The relationship data is useful because it identifies which securities have a plausible path from the event before the returns are observed. The event study then tests whether that path actually mattered in prices, volatility, revisions, or another outcome instead of assuming that every connected stock must react.
The earnings lead-lag guide covers the broader idea of information arriving at connected firms at different times. The market-knowledge timing guide explains why the relationship used to define treatment must itself have been observable before the event.
For relationship definitions and evidence limits, read the Altsets methodology.