Can a Supply-Chain Graph Tell You Which Feature Interactions Are Worth Testing?
September 14, 2026
Altsets
Research by Altsets Research
Yes. Customer and supplier relationships can restrict interaction search to economically plausible cross-company combinations, reducing a huge feature space before the return data gets a chance to overfit meaningless pairs.
Data used:Altsets Supply Chain Intelligence: 90k+ entities, 400k+ relationships, 20+ years of history.
Key findings
- The supplied Shin-Etsu customer relationships create targeted interaction candidates such as TSMC information multiplied by TSMC exposure, instead of brute-force interactions with thousands of economically unrelated companies.
- Sparse and hierarchical interaction methods exist specifically to control the combinatorial feature problem, while a supply-chain graph can add an ex ante economic structure to that statistical hierarchy.
Yes. A supply-chain graph can reduce the feature-interaction search space by telling a quant which variables have an economic reason to interact before the backtest is run. Instead of testing every customer return against every stock, the researcher can test customer information only against actual suppliers and scale the interaction with relationship evidence where appropriate.
Shin-Etsu shows how the graph can generate targeted interactions
The supplied Altsets data maps Shin-Etsu Chemical to Samsung Electronics, TSMC, and Intel with displayed supplier revenue shares of 2.43%, 4.02%, and 1.79% respectively. A generic machine-learning model can be given returns, earnings revisions, implied volatility, and fundamentals for thousands of companies and asked to discover useful nonlinear interactions. That creates an enormous search space.
The graph provides a prior. TSMC earnings surprise interacted with Shin-Etsu's TSMC exposure is economically plausible. Intel capex revisions interacted with the Intel-Shin-Etsu relationship are plausible. A random airline's earnings surprise interacted with Shin-Etsu customer concentration is much harder to justify. The network can exclude most of those meaningless combinations before the return data gets a chance to overfit them.
Interaction selection is one of the main dimensionality problems in factor modeling
Financial factor modeling already struggles with feature selection and nonlinear interaction search because hundreds of predictors can create tens of thousands of pairwise combinations. Sparse and hierarchical interaction methods were developed to restrict that problem by requiring interactions to satisfy structural rules instead of allowing every combination.
Supply-chain data provides an economic hierarchy rather than only a statistical one. A customer-return interaction can be allowed only when the company is an observed customer. A supplier-volatility interaction can be allowed only when the company is an observed supplier. A customer-event interaction can be weighted or bucketed by supplier revenue share where that metric is available. The graph defines the candidate structure before the model estimates whether the interaction matters.
Directed relationships create directed feature interactions
The same pair of companies can support different features depending on direction. If Nvidia is a customer of Micron, Nvidia demand information can interact with Micron's supplier revenue exposure. Micron-specific production news does not represent the same economic mechanism in reverse.
This matters because unconstrained models can discover symmetric statistical interactions that are difficult to explain. The supply-chain graph allows the researcher to encode direction into feature construction and then test whether those directed interactions improve out-of-sample performance beyond ordinary sector and market controls.
The graph prior can reduce multiple testing without guaranteeing alpha
Restricting interactions to observed economic relationships reduces the number of hypotheses, but it does not make every remaining interaction valid. A TSMC-Shin-Etsu feature can still fail because the product relevance is weak, the event was already priced, or the relationship measure is noisy.
The benefit is research discipline. The model searches among interactions with an ex ante economic explanation rather than among every mathematically possible pair. That can reduce compute, improve interpretability, and lower the chance that the strongest interaction is a purely accidental historical pattern.
The conclusion is to let economic structure decide what deserves to interact
Yes, the supply-chain graph can guide feature engineering. Use observed customer and supplier relationships to define which cross-company variables are allowed to interact, preserve relationship direction and missingness, and then let out-of-sample testing decide whether those economically plausible interactions actually add value. The network acts as a hypothesis filter before the statistical model becomes involved.
The hypothesis-testing guide explains why a large feature universe increases multiple-testing risk. The incremental-alpha guide explains how network features and interactions should be tested after conventional predictors are already present.
For relationship definitions and evidence limits, read the Altsets methodology.