Skip to article

Bayesian Inference for Supply Chain Risk: Uses, Examples and Limitations

How priors and new evidence combine, a worked example, and where the method stops helping

Many supply chain risk dashboards react after the event: the storm has made landfall, the port is congested, the supplier has gone quiet. Bayesian methods are one way to turn early signals into updated probability estimates before the full impact arrives. They are also easy to oversell. This article explains what Bayesian inference does for supply chain risk, works through a small example with the arithmetic shown, and sets out what the method cannot do.

What Bayesian Inference Is

Bayesian inference starts from a prior: a probability estimate you hold before seeing the latest evidence. It then combines that prior with a likelihood, which describes how probable the observed evidence would be under each possible state of the world, to produce a posterior. When new evidence arrives, the posterior becomes the next prior.

The prior can come from historical data, from expert judgement, or from both. That is a strength when history is thin, because you can state your assumptions openly and test how much they matter. It is also a weakness: a poorly chosen prior gives poor early estimates, and two analysts with different priors can reach different answers from the same data.

What This Article Does Not Claim: Frequentist Models Are Not Independence Models

A common argument says that frequentist or "traditional" statistics treat every event as independent and therefore cannot handle cascading failures. That is not accurate. Independence is a modelling assumption, not a property of a school of statistics. Both approaches can assume independence, and both can represent dependence:

  • Frequentist models routinely include dependence: regression with several predictors, time-series models such as ARIMA, survival models for time to failure, and multivariate models of correlated demand or lead times.
  • Bayesian models can also be simplistic. A naive Bayes classifier assumes the signals are conditionally independent given the outcome, and the worked example below makes that same assumption on purpose so that you can see its effect.
  • The main difference is in how uncertainty is represented and how prior information is used: a Bayesian model returns a distribution over unknowns and has a built-in place for prior knowledge, while a frequentist model reports estimates and confidence intervals that are interpreted over repeated sampling.

A frequentist model can also be re-estimated as new data arrives, and a Bayesian model that is never refreshed or checked will be as stale as any other. Updating continuously is something you build, not something either method gives you for free.

A Worked Example With Explicit Assumptions

The numbers below are hypothetical. They are chosen to show the mechanics, not to describe any real supplier, port or dataset.

Question. Is a supplier's inbound lane about to suffer a serious delay (more than two weeks) this quarter? Call that event D.

Assumptions.

  • Prior: P(D) = 0.10, a planner's estimate before any new signal.
  • Signal 1, a regional shipping-delay index rising above a threshold: P(S1 | D) = 0.80 and P(S1 | not D) = 0.20.
  • Signal 2, the supplier reporting a delay: P(S2 | D) = 0.70 and P(S2 | not D) = 0.30.
  • For the second update we assume S1 and S2 are conditionally independent given D. This is an assumption, and it is checked below.

Update 1: the index rises.

P(D | S1) = (0.80 × 0.10) / (0.80 × 0.10 + 0.20 × 0.90) = 0.08 / (0.08 + 0.18) = 0.08 / 0.26 ≈ 0.308.

One signal moves the estimate from 10% to about 31%, so it is informative but far from conclusive.

Update 2: the supplier also reports a delay. Use the posterior as the new prior. The prior odds are 0.308 / 0.692 = 4/9 ≈ 0.444. The likelihood ratio for S2 is 0.70 / 0.30 = 7/3 ≈ 2.33. The posterior odds are 4/9 × 7/3 = 28/27 ≈ 1.037, which is a probability of 28 / (28 + 27) = 28/55 ≈ 0.509.

Two signals take the estimate to about 51%.

StepProbability of a serious delay
Prior (hypothetical)10.0%
After signal 130.8%
After signals 1 and 2 (assuming conditional independence)50.9%

Check the assumptions. If both signals mostly reflect the same congestion event, they are not independent, and multiplying their likelihood ratios counts the same evidence twice. The real posterior would then be lower than 50.9%. A model that fits the joint behaviour of the two signals from data, in either a Bayesian or a frequentist framework, would handle this better than the simple product.

Check the prior. With the same likelihoods, a prior of 5% gives a posterior after signal 1 of 0.04 / (0.04 + 0.19) ≈ 17.4%, and a prior of 20% gives 0.16 / (0.16 + 0.16) = 50.0%. A prior that doubles or halves can move the answer a lot, so state it and test it.

What the Research Shows, and What It Does Not

Bayesian networks have been used to model supply chain risk, resilience and ripple effects. A literature review by Hosseini and Ivanov (2020) surveys that work, covering research articles published from 2007 to 2019. These are models of dependence between risk events, which is one reason they are attractive.

A 2025 paper in the International Journal of Production Research builds a multilayer Bayesian network for resilient and sustainable supply chains. We have not been able to read its full text, so we do not repeat its specific findings here.

A hierarchical Bayesian model for payment delay prediction uses partial pooling, so that suppliers with little history borrow strength from similar suppliers. In the aerospace supply chain case study reported in the paper, it predicted more accurately than non-hierarchical models. That is one case study with one dataset; it does not show that the approach wins in general.

A 2025 conference paper by Merendino and Monacelli (Boston Fed) is a macroeconomic model. It assumes firms treat energy prices as noisy signals of delivery delays for a specialised input such as microchips and update their beliefs by Bayesian learning. In that model the learning raises perceived marginal costs and increases how strongly energy shocks pass through to prices. It illustrates the mechanism; it is not evidence that firms using Bayesian methods respond better to disruptions.

Limitations

  • Data quality. Noisy, biased or sparse signals give unreliable posteriors, whichever method you use.
  • Model specification. The probabilities in a Bayesian network, and the structure connecting the nodes, are assumptions that need justification and review.
  • Calibration and validation. A useful model gives probabilities that match observed frequencies. Test this on data the model did not see, and compare it with a simpler baseline before relying on it.
  • Unprecedented events. A model can only update on the kinds of evidence it was built to read. A genuinely new kind of disruption is not foreseen by a method that has no variable for it.
  • Cost and skills. Building, maintaining and explaining probabilistic models takes data engineering and statistical skill, and the output still has to be tied to a decision.

When It Is Worth Considering

Bayesian approaches suit problems where there is some prior knowledge, evidence arrives in pieces, data on rare events is limited, and a decision-maker needs probabilities and not just point estimates. A simpler model may be a better first step where data is plentiful and the relationships are stable. A sensible way to start is a small pilot on one decision, such as expediting a critical component, with the model's probabilities compared against a simple baseline over time.

You can try decision-making under uncertainty in a simulation: play FMCG Chapter 1 free. The game does not teach Bayesian methods.

Frequently Asked Questions

Do frequentist methods assume events are independent?

No. Independence is a modelling assumption that either a frequentist or a Bayesian model can make or relax. Frequentist regression, time-series and survival models can represent dependence, and a Bayesian model can assume independence between signals.

How does Bayesian inference differ from other statistical approaches?

Bayesian inference combines a prior with observed evidence to give a posterior distribution, and it treats unknowns as having probability distributions. Frequentist methods report estimates and confidence intervals interpreted over repeated sampling. Both can be updated as new data arrives, and both depend on data quality, model specification, calibration and validation.

Does Bayesian inference predict unprecedented disruptions?

Not automatically. A Bayesian model updates on the evidence it was designed to read. A new kind of event that the model has no variable or prior for will not be foreseen, although signals that correlate with it may still raise the estimated risk.

When should organizations consider Bayesian approaches to supply chain risk?

When prior knowledge or expert judgement matters, data on rare events is sparse, evidence arrives in pieces, and decisions need probabilities. Start with a small pilot and compare it against a simple baseline.

Sources

  1. Hosseini, S. and Ivanov, D. (2020). Bayesian networks for supply chain risk, resilience and ripple effect analysis: a literature review. Repository record.
  2. Lu, J., Wu, D. and Dolgui, A. (2025). Construction of resilient and sustainable supply chain based on multilayer Bayesian network. International Journal of Production Research 63(23). Publisher page.
  3. A hierarchical Bayesian model for payment delay prediction in supply chain financing. Publisher page.
  4. Merendino, A. and Monacelli, T. (2025). Supply chain uncertainty, energy prices, and inflation. Conference paper, Federal Reserve Bank of Boston, 21–22 November 2025. PDF.

Share this article

LinkedInXEmail

Spotted an error or an outdated fact? Suggest a correction — corrections are reviewed by the author.