Probability represents belief under uncertainty
A distribution assigns relative plausibility to possible values or events rather than forcing one forecast.
Business · Strategy
Probabilistic Reasoning, Distributions and Bayesian Networks
A practical introduction to probability distributions, conditional probability, independence and Bayesian networks for representing uncertainty in business decision models.
Conditioning updates context
Conditional probability expresses how belief changes when evidence or a condition is known.
Networks exploit structure
A Bayesian network represents conditional dependencies so a complex joint distribution can be factored into smaller local relationships.
Represent uncertainty with random variables
A random variable maps uncertain outcomes to values that can be reasoned about mathematically.
Use discrete variables for alternatives such as pass/fail, supplier state or demand category, and continuous variables for quantities such as cost, lead time or sales volume. A probability distribution assigns probability mass or density across the possible values. The distribution is more informative than a single “best estimate” because it retains uncertainty and supports risk-sensitive decisions.
Choose variables that have a clear business meaning. “Market strength” may be too vague unless defined by observable indicators; “monthly orders” or “conversion probability” may be easier to estimate and validate. The model should represent the uncertainty relevant to the decision rather than every measurable feature in the environment.
Joint and conditional probability
Complex decisions involve several uncertain quantities whose relationships matter.
Conditional probabilityP(A | B) = P(A ∩ B) ÷ P(B), when P(B) > 0.
A joint distribution describes combinations of variables. Conditional probability asks how belief about one variable changes after another is known. For example, the probability of a late project may be different given that a critical supplier is delayed. Conditioning is the foundation of evidence-based updating.
Bayes ruleP(H | E) = P(E | H) P(H) ÷ P(E). Prior belief × evidence likelihood → posterior belief after normalisation.
Bayes rule is valuable because it reverses the direction of a known probability. If historical data tells how likely an inspection result is under good and bad process states, the rule can update the probability that the process is bad after observing the result. The quality of the update still depends on the model and data; mathematics does not repair a poorly chosen prior or an unreliable likelihood.
Independence and conditional independence
Independence assumptions can dramatically reduce model complexity, but unjustified independence creates hidden error.
Two variables are independent when knowledge of one does not change the probability of the other under the model. Conditional independence is more useful in real systems: two outcomes may become independent once a common cause is known. For example, two defect indicators may be correlated because both depend on a machine condition; once machine condition is included, their residual dependence may be small.
Document important independence assumptions. They are not just mathematical convenience; they are statements about the system. Test them with domain knowledge and data where possible. If an assumption is wrong, the model can become overconfident because it effectively counts related evidence multiple times.
Bayesian networks as a compact dependency model
A Bayesian network is a directed acyclic graph in which nodes represent variables and directed edges encode conditional dependence structure.
Each node has a local conditional distribution given its parent variables. Together, the graph and local distributions define a joint distribution. This factorisation lets analysts represent a large system without enumerating every combination explicitly. The graph also encourages causal thinking, although a statistical Bayesian network is not automatically a proven causal model.
Network factorisationP(X₁,…,Xₙ) = ∏ᵢ P(Xᵢ | Parents(Xᵢ)).
DriversExternal demand, environment, supplier state or policy.
Intermediate statesCapacity, quality condition, workload or system health.
ObservationsMeasurements, alarms, survey results or inspections.
OutcomeCost, delivery, customer result or failure event.
Keep the graph interpretable. If every variable points to every other variable, the network loses much of its structural advantage. Build from domain relationships, then validate whether the conditional model reproduces observed patterns.
Specify distributions responsibly
Probability tables and continuous distributions turn the graphical structure into a quantitative model.
Use data where sufficient, but incorporate expert judgement transparently where data are sparse. Define elicitation questions around observable frequencies or ranges rather than asking experts for abstract probabilities without context. When using parametric distributions, verify whether shape assumptions are reasonable; when they are not, non-parametric or empirical distributions may be preferable.
Represent parameter uncertainty where it matters. A model with an estimated failure probability of 0.02 should not necessarily behave as if 0.02 were known exactly. Bayesian parameter models can preserve uncertainty about the probability itself, which is especially useful when observations are limited.
Worked example: quality risk network
Consider a generic production line where final defect risk depends on material state, machine condition and operator setup, while inspections provide noisy evidence.
A simple network might place material state and machine condition as parents of process stability; setup and process stability as parents of defect state; and defect state as parent of inspection result. Before inspection, the model provides a prior probability of defect. After a failed inspection, evidence is propagated backward through the network, increasing belief in possible underlying causes according to their likelihood.
The network helps decide where additional inspection or maintenance information would be most valuable. It also avoids a common error: treating every failed indicator as independent evidence when several indicators are caused by the same underlying process condition.
Common modelling failures
Probability models fail when structure and meaning are neglected.
| Failure | Effect | Control |
|---|---|---|
| Point forecast only | Hides tail risk and decision sensitivity. | Use distributions or scenarios. |
| Unjustified independence | Produces overconfident probabilities. | Model shared causes or test dependence. |
| Ambiguous variables | Experts and data refer to different concepts. | Define states and measurement rules. |
| Probability without calibration | Numbers look precise but are unreliable. | Back-test and use calibration plots or scoring methods. |
| Causal interpretation without evidence | Intervention decisions may be wrong. | Distinguish association from causal claims. |
Do Bayesian networks require huge datasets?
Not necessarily. They can combine structure and prior knowledge with data, but sparse data increases parameter uncertainty and makes validation more important.
Is a Bayesian network always causal?
No. Directional edges can represent factorisation rather than proven intervention effects. Causal use requires additional assumptions and evidence.
Model review record
A technically correct method still needs an auditable operating translation.
When a probability model is used for a real decision, keep a compact review record showing the variable definitions, data period, prior assumptions, conditional relationships and calibration evidence. Record which probabilities are directly estimated and which depend mainly on judgement. This prevents later users from treating all numbers as equally certain. Revisit the model after material new evidence or a change in the operating process, because a once-calibrated dependency can become stale when customer mix, equipment, policy or measurement methods change. For important decisions, ask a second reviewer to reconstruct the reasoning from the documentation. If they cannot explain why a dependency exists or how a probability was obtained, the model is not yet sufficiently auditable for high-consequence use.
Application checklist
- Define random variables and states in operational terms.
- Use distributions instead of one-point forecasts where uncertainty matters.
- Apply conditional probability to incorporate evidence.
- Document independence and conditional-independence assumptions.
- Design Bayesian-network structure around defensible dependencies.
- Estimate local distributions from appropriate data or transparent elicitation.
- Preserve parameter uncertainty when data are sparse.
- Validate calibration and do not infer causality without justification.
Source basis. Decision-analysis source set: probabilistic reasoning, sequential decisions, learning, state uncertainty and multiagent methods. This page is an original handbook synthesis of the supplied materials. Named people, organisations and identifying case details from the sources have been removed. Numerical examples are labelled as illustrative where used.