How Should You Bootstrap a Supply-Chain Trading Strategy?
September 14, 2026
Altsets
Research by Altsets Research
Resample time blocks and preserve relevant customer, supplier, or event clusters instead of independently shuffling stock rows that are economically connected.
Data used:Altsets Supply Chain Intelligence: 90k+ entities, 400k+ relationships, 20+ years of history.
Key findings
- The supplied Micron-SK Hynix-Nvidia cluster illustrates why independently resampling stock rows can separate supplier returns from the customer shocks the original hypothesis was designed to test.
- Block-bootstrap and multidimensional cluster-bootstrap methods provide more defensible starting points when a strategy contains both temporal and network dependence.
Resample contiguous time blocks and preserve customer, supplier, or event clusters when the hypothesis depends on them. Bootstrapping individual stock-days independently can destroy the same temporal and network dependence a supply-chain strategy is supposed to measure, producing confidence intervals for a synthetic problem rather than the strategy actually being tested.
A naive bootstrap can break the economic event
The supplied Altsets network maps both Micron and SK Hynix to Nvidia as a customer. If a bootstrap independently resamples Micron days, SK Hynix days, and Nvidia days, the synthetic sample can pair a Nvidia event from one historical period with supplier returns from unrelated periods. The resample no longer represents the economic sequence the original hypothesis was testing.
The same problem appears cross-sectionally. One customer event can affect several connected suppliers simultaneously. Independently resampling company rows can separate that cluster and make the synthetic portfolio appear more diversified than it was during the original shock. Traditional block-bootstrap methods were developed because iid resampling is inappropriate for dependent time series. A supply-chain model adds an observed network layer that can create dependence across firms at the same time.
Preserve time first, then decide which network dependence matters
A practical starting point is to resample contiguous time blocks while keeping the whole cross-section aligned inside each block. That preserves customer announcements, supplier responses, market regimes, and contemporaneous cross-stock behavior better than shuffling individual dates. The block length remains a modeling choice, so the researcher should test reasonable values rather than search for the one length that produces the narrowest interval or strongest result.
Some hypotheses also require cluster-aware resampling across firms. A customer-event study can treat the customer event as the cluster. A shared-supplier strategy may need supplier-centered groups. A network-factor test may need communities or economically coherent graph clusters. Research on bootstrap methods with multi-dimensional cluster dependence is relevant because supply-chain observations can share time, customer, supplier, and industry dependence simultaneously.
Bootstrap the statistic the research claim actually uses
A useful bootstrap does not have to operate only on the final portfolio return series. The researcher can re-estimate an event-study abnormal return, an exposure-gradient coefficient, a difference between connected and unconnected firms, or the performance gap between a baseline model and a supply-chain-enhanced model. Re-estimating the feature and model inside each dependence-preserving resample captures more of the uncertainty than bootstrapping a finished Sharpe ratio after every research choice has been frozen.
That distinction matters when network features are estimated quantities. A centrality score, peer group, or customer-concentration rank can change when the sampled graph changes. If the uncertainty being studied includes feature construction, the feature should be recomputed inside the bootstrap where feasible. If the feature is intentionally treated as fixed, the research report should say that the interval is conditional on the historical graph and feature pipeline.
Model selection creates a second layer of uncertainty
A researcher who tried dozens of exposure thresholds, event windows, or graph features before selecting the final strategy has already used information from the historical sample. A bootstrap that freezes the winning specification measures uncertainty around that selected strategy, not uncertainty around the entire discovery process. Repeating model selection inside every resample is computationally expensive, but it is closer to the question of how stable the research process itself is.
At minimum, the quant should separate the two claims. One bootstrap can ask how uncertain performance is if the chosen specification were fixed in advance. Other robustness tests can ask whether nearby thresholds, feature definitions, and cluster assumptions lead to similar conclusions. A network signal that survives only one exact resampling and specification choice is much less convincing than one that remains plausible across several defensible dependence structures.
The conclusion is to resample the dependencies you say matter
A supply-chain strategy begins from the claim that companies are economically connected. The bootstrap should preserve the time blocks and network clusters relevant to that claim rather than randomly shuffling the graph apart and then reporting artificially precise confidence intervals. If the result survives dependence-preserving resamples, the uncertainty estimate is far more aligned with the actual strategy.
The network-dependence guide explains why connected stock observations should not automatically be treated as independent. The event-study guide explains why several supplier reactions to one customer event still belong to one event cluster.
For relationship definitions and evidence limits, read the Altsets methodology.