Does Supply-Chain Data Add Anything After Momentum, Value, and Industry?
September 14, 2026
Altsets
Research by Altsets Research
Test whether network features provide incremental out-of-sample information after conventional predictors are already present instead of rewarding supply-chain variables for rediscovering industry, size, momentum, or ordinary fundamentals.
Data used:Altsets Supply Chain Intelligence: 90k+ entities, 400k+ relationships, 20+ years of history.
Key findings
- A credible incremental test keeps the universe, horizon, costs, and portfolio construction fixed while comparing a reasonable baseline model with and without the supply-chain feature family.
- Supply-chain variables should be tested against industry and conventional-factor explanations because a network feature can appear predictive simply by rediscovering a sector or common economic cycle.
It can, but only an incremental test can show whether it does. A supply-chain feature adds value after momentum, value, and industry only if it improves out-of-sample predictions or portfolio results beyond a credible baseline containing those controls. A feature that predicts returns only in isolation may simply repackage information the model already has.
Begin with a baseline that would exist without the network
The cleanest test starts with a model the researcher would have been willing to run before seeing the supply-chain data. That baseline might include price momentum, volatility, liquidity, size, valuation measures, industry controls, and macro variables appropriate to the horizon. Gu, Kelly, and Xiu show why this matters: in large cross-sectional machine-learning tests, familiar signals such as momentum, liquidity, and volatility remain important even when flexible models are allowed to capture nonlinear interactions. A supply-chain feature should therefore be asked to improve an already credible benchmark rather than compete against an artificially weak model.
The next step is to add a small, economically motivated set of relationship features without changing everything else at the same time. A researcher might add customer concentration, supplier concentration, shared-customer counts, relationship-change features, or measures of dependency overlap. If the network model uses a different universe, different return horizon, new hyperparameters, and a new cost model at the same time, it becomes impossible to know which change created the improvement. Incremental research works best when the supply-chain layer is the main experimental variable.
Measure improvement in the same out-of-sample windows
The comparison should be made on identical historical periods using the same securities, rebalance dates, costs, and portfolio construction rules. If the baseline is evaluated on one sample and the network model is evaluated after several additional years of data became available, the comparison is not meaningful. The quant should record whether the supply-chain model improves rank correlation, forecast error, realized spread returns, drawdowns, turnover, or whatever metric actually matters to the strategy. An uplift in one in-sample statistic is much weaker evidence than repeated improvement across several untouched out-of-sample windows.
This is also where simple portfolio sorts can be useful even when the final model is machine learned. If a customer-concentration feature appears predictive only inside a large nonlinear ensemble, the researcher should still ask whether simple high-versus-low concentration portfolios display any economically coherent pattern. The simple test does not need to reproduce the full model result, but it can reveal whether the feature has a visible standalone mechanism or whether the machine-learning system may be exploiting fragile interactions. Complex models are valuable when they capture real nonlinearities, not when they make the source of the result impossible to inspect.
Residualization can reveal whether the feature is just another industry signal
Supply-chain structure is naturally correlated with industry. Semiconductor companies often share customers, suppliers, manufacturing technologies, and capital-spending cycles, so a naive network feature can become an expensive way to rediscover that two companies are in semiconductors. One test is to neutralize or residualize the feature against industry, size, or other obvious exposures before asking whether it retains predictive value. Another is to compare stocks within industries or form double sorts that hold a conventional factor approximately constant while varying the network feature.
The supplied Shin-Etsu Chemical network illustrates the problem. Samsung Electronics, TSMC, and Intel are distinct customers with different directional relationship weights, but all three are connected to semiconductor manufacturing. A feature based on Shin-Etsu's customer network may contain real relationship information while also carrying a strong semiconductor-cycle component. A quant model should determine whether the network feature adds something after that common industry exposure is accounted for instead of crediting the entire effect to supply-chain data.
Feature importance is not the same as incremental alpha
Tree models and neural networks can assign high importance to a feature that improves fit without producing a robust tradable improvement after costs. Importance measures can also become unstable when predictors are correlated, which is common when network features overlap with industry, size, customer concentration, and fundamentals. The useful test is not whether a model says the graph feature mattered. The useful test is whether excluding the feature materially degrades out-of-sample performance under the same research design.
Ablation studies are particularly useful here. Train the same model with and without the supply-chain family, then repeat the comparison across time periods, sectors, and market regimes. If the network variables help only during one narrow window, that may still be useful, but the result should be described as conditional rather than universal. If they improve several different model classes and remain helpful after obvious exposures are controlled, the case for genuinely incremental information becomes stronger.
Incremental value can appear in risk even when return prediction barely changes
A network feature does not need to increase raw return forecasts to justify itself. It may reduce concentration, improve drawdowns, identify hidden event clusters, or prevent a portfolio optimizer from stacking several positions behind the same customer. Those benefits can improve realized portfolio efficiency even when the standalone feature has a weak information coefficient. A researcher who evaluates supply-chain data only through next-period return prediction can therefore miss some of its most plausible quantitative uses.
That suggests testing the network twice: once as a predictive feature and once as a portfolio constraint or conditioning variable. A customer-overlap feature might add little to expected-return forecasts but materially change the portfolio chosen from those forecasts. A supplier-risk feature might be more useful for scaling exposure around certain events than for predicting unconditional returns. The correct benchmark depends on the economic role the data is supposed to play.
The real test is whether the network changes the model after everything obvious is already there
Alternative data earns its place when it contributes information the existing stack did not already capture. Supply-chain data should be held to that same standard. Build a credible baseline, add the relationship features without changing the rest of the experiment, compare identical out-of-sample windows, neutralize obvious exposures, and examine both prediction and portfolio construction. If the network still changes the decision after momentum, industry, fundamentals, and ordinary risk controls are present, then the researcher has evidence of incremental value rather than a renamed conventional factor.
The quantitative-factor guide covers ways to convert relationships into systematic features. The alternative-data comparison explains what economic structure can contribute beside sentiment, transaction, physical-activity, and market-pricing datasets.
For relationship definitions and evidence limits, read the Altsets methodology.