Tracking specific parts makes simulated automotive supply shocks four to five times larger

Knowing who supplies whom may not be enough to assess supply-chain risk. A new preprint suggests that knowing exactly what changes hands—and how readily it can be replaced—can substantially alter estimates of how a disruption spreads.

A supplier that makes a similar product is not necessarily a replacement for the particular component a customer has lost. That distinction drives the central finding of a preprint by Benjamin Rolf, Hiroyasu Inoue and Sebastian Lang: requiring product-specific matches makes simulated automotive supply disruptions substantially larger than a broader supplier-matching rule does.

Across the main comparisons involving the failure of a highly connected supplier, the product-specific rule produces cascades four to five times larger, with other model mechanics held fixed. These are results from an agent-based simulation, not measurements of actual automotive shutdowns.

What counts as a replacement?

The researchers assembled a network from the commercial MarkLines database containing 76,986 firms and 197,077 directed supplier–customer relationships. They conducted approximately 119,000 simulations across their experiments.

The broader substitution rule treats a firm as a potential replacement if its product portfolio overlaps with that of the failed supplier. On an industry-wide network, this can admit candidates whose shared product is unrelated to the customer's missing input.

The alternative starts with the specific product delivered along the failed relationship. It uses numerical representations of product names to identify semantically similar products, then searches for firms offering them. In the main comparison, the median lost product has roughly 31,000 candidate suppliers under portfolio overlap, versus about 830 under the product-specific rule. Candidate status does not guarantee spare capacity or technical suitability.

Each firm acts as an agent. Customers losing supplies search for replacements and can form new supplier relationships. In the extended model, they contact three candidates per weekly cycle. Suppliers have limited spare capacity, while inventory-based budgets limit search time. Customers can also enter the disrupted state immediately when no untried candidate remains, even if inventory would last longer.

Separating the rule from other changes

The authors did not simply compare an older model with a package of new assumptions. They introduced an intermediate configuration that retains portfolio overlap while using the extended model's capacity, inventory and search mechanics. Comparing this configuration with the product-specific version isolates candidate selection.

For failures drawn from the most-connected tenth of suppliers, one main setting produces an average of 3.37 additional firm removals under portfolio overlap and 16.11 under product-specific matching. Across the main hub scenarios, the increase is a factor of 4.2 to 4.8. Cascade size counts firms removed beyond the initial failure.

The original baseline configuration produces no additional removals in these comparisons. Thus, the calibrated mechanics first make cascades possible; product-specific matching then increases their extent. The experiment does not separately establish whether capacity limits or inventory constraints explain the first change.

Finding the suppliers that matter

The simulations also distinguish a supplier's reach from the difficulty of replacing its products. Having many customers exposes more firms to a failure, but scarce substitutes make exposed customers more likely to fail.

The researchers built a criticality index combining estimated propagation probabilities with paths extending three supply relationships downstream. Among the 50 most damaging suppliers under random disruptions, a topology-only downstream-reach measure identifies 20. Adding product substitutability identifies 32.

Applied without refitting to a regional shock scenario, the index identifies 30 of the 50 most damaging previously unseen seed firms, compared with 20 for downstream reach. This demonstrates transfer between simulated scenarios, not validation against observed supplier failures. The index and simulation share the same substitutability proxy, so their agreement is partly built into their construction.

An earthquake-inspired scenario, not a reconstruction

The regional experiment uses the geographical pattern of seismic intensities from the 2011 Tohoku earthquake, but applies it to supply relationships from the early 2020s. It is not a reconstruction of the historical disaster.

All five primary Japanese vehicle manufacturers have simulated disruption probabilities of at least 94% across the tested severities. They cannot be initial failures because their database records lack locations; their disruptions arise entirely through suppliers. These probabilities concern unresolved input losses under the model, not permanent business closures.

Missing products leave substantial uncertainty

Only 6.3% of supply relationships have observed product annotations. The rest are filled through statistical sampling. A hold-out check finds that the main uniform-sampling approach makes products appear easier to replace than the observed annotations suggest, although those annotated relationships are not representative of the whole network.

Across alternative imputations and similarity thresholds, the hub-cascade multiplier ranges from 1.2 to 17.7. Product-name similarity also cannot establish compatibility, qualification or tooling requirements. MarkLines itself has incomplete, biased coverage.

Finally, removed firms never recover in the model. Prolonged capacity losses therefore must not be read as forecasts of production recovery. The preprint's central implication is narrower: within these simulations, product-level information changes both estimated cascade sizes and which suppliers appear most critical.