Do Supply-Chain Signals Only Work in Certain Market Regimes?
September 14, 2026
Altsets
Research by Altsets Research
Many dependencies are dormant until a customer shock, supplier disruption, volatility regime, or known catalyst activates them, making conditional research potentially more appropriate than forcing every relationship into a continuous return signal.
Data used:Altsets Supply Chain Intelligence: 90k+ entities, 400k+ relationships, 20+ years of history.
Key findings
- Shared dependencies can remain economically real while producing little unconditional return signal, because the transmission path may become important only when the shared customer, supplier, or macro environment generates new information.
- Regime definitions should be economically motivated and fixed before final evaluation, with enough independent events to distinguish a repeatable conditional mechanism from one historically convenient episode.
Often, yes. Many supply-chain signals become informative only when an earnings surprise, shortage, demand shock, regulatory change, liquidity event, or broader risk regime activates the underlying dependency. A feature can therefore be economically real while showing little unconditional return predictability.
Structural features can be dormant until the relevant shock arrives
Two companies can share a major customer for years without their stocks displaying a reliable daily relationship. During ordinary periods, company-specific earnings, valuation changes, factor exposures, and unrelated news can dominate returns. When the shared customer becomes the source of a major demand surprise, however, the commercial link has a direct reason to matter. The supply-chain feature may therefore behave more like a conditional exposure than a constant alpha signal.
The Micron and SK Hynix relationships with Nvidia illustrate the idea. Both memory companies have current public links to Nvidia's AI roadmap, but that does not imply their returns should move together every day. A model may find more useful structure when conditioning on Nvidia earnings, product-roadmap changes, memory-demand shocks, or other events that plausibly activate the shared-customer channel. The economic relationship defines the transmission path, while the regime or event determines when the path deserves more weight.
Regimes can be market-wide or relationship-specific
A regime does not have to mean bull market versus bear market. Macro variables such as volatility, credit spreads, rates, or liquidity can change how investors price risk across many securities, and recent machine-learning research explicitly studies return prediction under shifting market regimes. Supply-chain signals may behave differently when market stress is high because investors care more about bottlenecks, balance-sheet resilience, and concentrated counterparties. A customer-concentration feature that is mostly ignored during calm markets can become important when financing or demand conditions deteriorate.
Relationship-specific regimes can be even more targeted. An upcoming customer earnings report, an export-control announcement, a product launch, or a capacity transition can create a temporary state in which one network path becomes unusually relevant. These event regimes are attractive because the mechanism is easier to explain than a broad hidden-state label. They also reduce the temptation to search blindly for statistical regimes until one makes the backtest look good.
Define the conditioning variable before inspecting the final returns
Regime research is extremely easy to overfit. A researcher can try several volatility thresholds, macro indicators, event windows, and clustering methods until a weak signal suddenly looks powerful. If the conditioning rule was chosen after studying the strategy's failures, the apparent regime dependence may simply encode the backtest's history. The economic mechanism should therefore narrow the regime definition before the final test period is examined.
A practical design can start with a small number of defensible states. For example, compare normal periods with pre-specified high-volatility periods, or compare ordinary days with windows surrounding the earnings reports of important customers. Keep the supply-chain feature definition fixed, keep the trading rule fixed, and ask whether the feature's information coefficient or portfolio spread changes consistently across states. The objective is to test a hypothesis about when dependencies matter, not to rescue an otherwise failed signal.
Conditional usefulness can matter more than unconditional Sharpe
Suppose a supplier-concentration feature is flat most of the year but reliably identifies stocks with worse downside around major customer disappointments. That feature may be useful for event-risk scaling even if a long-short portfolio built from it has mediocre unconditional returns. A quant who insists that every dataset become a continuous directional signal can miss this use case. Supply-chain data may be better suited to deciding when to reduce exposure, which securities to include in an event basket, or how aggressively to trust another model.
This suggests evaluating more than one objective. Measure unconditional returns, but also test conditional drawdowns, event-window dispersion, forecast errors, and changes in cross-sectional rank performance. A feature that improves a model only in stressed states can still create real portfolio value if those states are expensive enough. The economic role of the feature should determine the metric used to judge it.
Regime-conditioned models need enough independent events
The danger with event conditioning is sample size. A strategy can look spectacular around six historical disruptions simply because the events happened to line up with the researcher's intuition. The researcher should count independent events, examine whether one episode dominates the result, and test the mechanism across different sectors or time periods where possible. A signal that requires one very specific historical crisis to work is not the same as a robust conditional relationship.
This is another place where a large relationship graph can help. Instead of testing one famous supplier event, the researcher can define a repeatable rule for identifying customer shocks, supplier shocks, or shared-node events across many companies. The larger sample makes the hypothesis more falsifiable. It also forces the researcher to confront cases where the expected propagation did not occur.
The conclusion is that dependency can be a state variable rather than a permanent forecast
Many supply-chain relationships describe where risk or information could travel, not a guarantee that it is traveling every day. Quant research can respect that structure by asking when the path becomes active and whether the feature improves decisions specifically in those states. If a supply-chain signal looks weak unconditionally, the next question should not automatically be how to tune it harder. It may be whether the economic mechanism is conditional by nature.
The same-shock-different-outcomes guide explains why a common event can produce different company outcomes. The supply-chain catalyst-calendar guide shows how known external events can be attached to relationship-aware research.
For relationship definitions and evidence limits, read the Altsets methodology.