Probability, Bayesian Updating and Value of Information

Introduction and context

Prescriptive analytics recommends actions under uncertainty. Unlike descriptive analytics, which summarizes what has happened, and predictive analytics, which estimates what may happen, prescriptive analytics asks: What should the organization do, given uncertain outcomes, competing objectives, and limited information?

Probability supplies a formal language for uncertainty. Bayesian analysis provides a disciplined method for revising beliefs when evidence arrives. Value-of-information analysis determines whether obtaining better information—through experiments, market research, pilots, inspections, or additional data—is economically justified.

These methods are central to decisions such as launching a product, approving a loan, expanding capacity, selecting a supplier, or investing in a diagnostic system. Their purpose is not to eliminate uncertainty. Rather, they help decision makers quantify uncertainty, distinguish information from action, and avoid paying more for information than it is worth.

Learning objectives

By the end of this chapter, you should be able to:

1. Define events, conditional probability, independence, and joint probability.
2. Apply the law of total probability and Bayes’ theorem.
3. Use discrete and continuous probability distributions in decision analysis.
4. Construct and evaluate decision trees using expected monetary value.
5. Calculate expected value of perfect information (EVPI) and expected value of sample information (EVSI).
6. Assess whether research, experimentation, or additional data justify their cost.
7. Identify practical, ethical, and governance risks in probabilistic decision systems.

1. Probability as a decision language

An outcome is a possible state of the world, such as high demand or low demand. The sample space, denoted Ω, is the set of all relevant outcomes. An event is a subset of Ω. A probability assigns each event A a number P(A) between 0 and 1.

The fundamental rules are:

Here Aᶜ is the complement of A. If A and B cannot occur together, they are mutually exclusive and:

For general events, the addition rule is:

The subtraction prevents the intersection from being counted twice.

Conditional probability and dependence

Conditional probability measures the probability of A given that B has occurred:

Rearranging gives the multiplication rule:

Events A and B are independent when learning that B occurred does not change the probability of A:

Equivalently, for events with positive probabilities:

Independence is a substantive assumption, not a default. Two customer purchases, machine failures, or survey responses may be correlated because they share seasonality, income, geography, or operating conditions.

Bayes’ theorem

Bayes’ theorem reverses a conditional relationship. If B₁, B₂, …, Bₖ are mutually exclusive and exhaustive hypotheses, then:

P(Bᵢ) is the prior probability, P(A|Bᵢ) is the likelihood, P(A) is the evidence probability, and P(Bᵢ|A) is the posterior probability. The denominator is found using the law of total probability:

For two events, a commonly used form is:

Bayesian updating is especially important when a signal is imperfect. A positive test, favorable survey, or abnormal sensor reading is evidence—not proof.

Worked example: an imperfect signal

Suppose 20% of prospective customers are genuinely high-value. A market-research signal is positive for 80% of high-value customers and for 30% of other customers. Let H denote high value and + denote a positive signal.

The posterior probability is:

Thus, a positive result raises the probability from 20% to 40%, but does not establish that the customer is high value. The result reflects both the signal’s accuracy and the base rate.

2. Probability distributions for business decisions

A probability distribution specifies how probability is allocated across possible values of a random variable. A discrete random variable takes countable values; a continuous random variable can take values over an interval.

For a discrete variable X with possible values xᵢ and probabilities pᵢ:

The expected value is a probability-weighted average, not necessarily an outcome that will occur. Variance measures dispersion:

The standard deviation is σ = √Var(X), measured in the same units as X. Expected monetary value (EMV) applies the same principle to payoffs:

where a is an action, s is a state of nature, and V(a,s) is the payoff from action a in state s. EMV is appropriate when probabilities are credible, payoffs are comparable, and the decision maker is willing to use an expected-value criterion.

Common distributions include:

Distribution choice should follow the data-generating process, not convenience. For example, a normal model can assign negative probability to demand, making it unsuitable where negative values are impossible unless a justified approximation is being used.

Joint and predictive distributions

When uncertainty concerns multiple variables, their joint distribution matters. Marginal probabilities describe variables separately; conditional distributions describe one variable given another. In Bayesian modeling, a predictive distribution averages over parameter uncertainty:

Here D is observed data, θ is an unknown parameter, P(θ|D) is its posterior distribution, and P(y|θ) is the model for a future observation. This distinction is important: uncertainty about the parameter and randomness in the future outcome are not identical.

3. Decision trees and sequential decisions

A decision tree represents decisions, uncertain events, and consequences in chronological order. A square is conventionally used for a decision node, a circle for a chance node, and terminal branches for final payoffs. Each chance branch must have probabilities that sum to 1, and each terminal payoff must be defined on a consistent basis—for example, net present value after research cost.

Figure: A decision tree separates controllable decisions from uncertain states of nature.

The standard solution method is rollback or foldback analysis. At each chance node, calculate the probability-weighted value of downstream branches. At each decision node, select the action with the greatest value, subject to constraints. For multistage trees, probabilities are conditional on the information available at that point.

A tree can clarify timing, but it can also become unwieldy. Simulation, influence diagrams, dynamic programming, or stochastic optimization may be preferable when there are many periods or correlated uncertainties.

4. Information and its economic value

Information has value only because it can change an action or improve the expected consequence of an action. The value of information must therefore be measured relative to the best decision without that information.

Expected value without additional information

For actions a ∈ A and states s ∈ S:

V₀ is the value of the best action using current information.

Expected value with perfect information

Perfect information reveals the true state before the decision is made. The decision maker then selects the best action separately in each state:

The expected value of perfect information is:

EVPI is an upper bound on what any additional information can be worth. If information costs more than EVPI, it cannot be economically justified under the specified model.

Expected value with sample information

Sample information is imperfect evidence Y, such as a survey result. Its predictive probability is:

After observing y, the posterior state probability is:

The expected value with sample information is:

The expected value of sample information is:

EVSI cannot exceed EVPI when both are calculated using the same actions, states, probabilities, and payoffs. The expected net benefit of information is:

where C is the total cost of obtaining and using the information. Proceed when ENBI is positive, while also considering timing, risk, strategic value, and ethical constraints.

Worked example: whether to commission research

A firm must choose between launching a product or not launching. The probability of high demand is 0.40 and low demand is 0.60. Payoffs, in thousands of dollars, are:

Without research:

Therefore V₀ = 60 and the firm launches.

With perfect information, the firm launches under high demand and does not launch under low demand:

Thus, $60,000 is the maximum justifiable price of perfect information.

Now suppose a research study produces a favorable result F or unfavorable result U. Its accuracy is:

P(F|high) = 0.75 and P(F|low) = 0.25.

The probability of a favorable result is:

The posterior probability of high demand after F is:

The launch EMV after F is:

So the firm launches after F. For U:

The launch EMV after U is approximately −27.3, so the firm does not launch. Hence:

The research is worth at most $15,000 before considering its cost. If it costs $10,000, expected net benefit is $5,000; if it costs $20,000, it should not be purchased on financial value alone.

This example illustrates a critical principle: the study is valuable because it changes the decision for some evidence outcomes. If research would lead to the same action regardless of its result, its decision value is zero, even if it is scientifically interesting.

5. Designing experiments and market research

Value-of-information analysis should precede data collection when possible. A practical workflow is:

1. Define the decision, decision deadline, feasible actions, and objective.
2. Specify uncertain states and current probabilities.
3. Estimate payoffs, costs, constraints, and risk consequences.
4. Identify what information could change the action.
5. Model the signal’s accuracy or sampling distribution.
6. Calculate EVPI and EVSI before committing expenditure.
7. Include acquisition, analysis, delay, implementation, and opportunity costs.
8. Perform sensitivity analysis on probabilities, payoffs, accuracy, and costs.
9. Establish a stopping rule and a plan for acting on each possible result.

A pilot experiment may have value beyond the immediate decision because it improves future decisions, reveals operational constraints, or reduces implementation risk. Those benefits should be modeled explicitly rather than added informally. Conversely, information may lose value when the decision window closes, competitors act, or the environment changes.

The value of information ratio is sometimes reported as:

A ratio above 1 indicates positive expected financial value under the model, but it is not a substitute for ENBI. It can obscure the scale of benefits and should not be interpreted as a universal investment rule.

6. Critical assumptions and limitations

Probability quality

Probabilities may be estimated from historical frequencies, statistical models, expert judgment, or hierarchical Bayesian models. Historical frequencies can be misleading under structural change. Expert assessments can suffer from anchoring, overconfidence, availability bias, and motivated reasoning. Elicitation should document the source, time period, population, and uncertainty around each estimate.

Model risk

A precise calculation can still be wrong if the state space omits relevant outcomes, payoffs ignore implementation costs, or dependencies are misrepresented. Correlated risks must not be treated as independent merely because independence simplifies computation. Sensitivity analysis, scenario analysis, stress testing, and model comparison are essential.

Risk preferences

EMV assumes that expected monetary value is an appropriate objective. A risk-averse organization may instead maximize expected utility:

where u(·) is a utility function. For a risk-averse decision maker, u is typically concave, so a loss may matter more than an equal-sized gain. EVPI and EVSI should then be calculated using utility rather than raw money if that reflects the decision objective.

Information quality and bias

Research may be affected by selection bias, nonresponse, measurement error, confirmation bias, p-hacking, or conflicts of interest. A predictive signal can also be unstable when the population, market, or policy changes. Bayesian updating does not automatically correct biased evidence; it formalizes updating conditional on the specified likelihood.

Ethical and governance issues

Organizations should not use probabilistic scores as unquestionable facts. Governance requires clear ownership, documentation of assumptions, audit trails, validation, monitoring, and a process for challenging outputs. Decisions involving credit, employment, healthcare, insurance, or public services may create disparate impacts even when a model has strong average accuracy. Privacy, consent, data minimization, fairness, explainability, and applicable law must be considered before collecting or deploying information.

A high EVSI does not justify unethical experimentation. Participants must not face disproportionate harm, and research should have a legitimate purpose, appropriate oversight, and secure data handling. Decision makers should distinguish uncertainty from ignorance: a narrow probability estimate based on poor data can create false confidence.

7. Contemporary analytical practice

Modern analytics can estimate posterior distributions using Bayesian hierarchical models, Markov chain Monte Carlo, sequential Monte Carlo, and approximate inference. These methods can pool information across regions, products, or customer segments while preserving group-level differences. Bayesian optimization and contextual bandits extend the value-of-information perspective to repeated experimentation, where each action trades off immediate reward against learning.

Machine-learning systems often generate predictive probabilities, but calibration matters. A classifier that assigns 0.7 to many cases should be correct about 70% of those cases, subject to the calibration definition and population. Decision analysis then combines those probabilities with action-specific payoffs, constraints, and costs. A highly accurate prediction can have low decision value if it does not alter the recommended action; a modestly accurate prediction can be valuable near a decision threshold.

Automation should therefore support, not replace, accountable judgment. Human review is particularly important when consequences are irreversible, data are sparse, or the model operates outside its validated population.

Illustrative case: a retailer’s regional launch

Background. A fictional retailer is considering a regional launch of a subscription service. The contribution payoff is $1.2 million if adoption is high and a loss of $0.5 million if adoption is low. Current assessment assigns a 35% probability to high adoption.

Problem. Management can launch immediately, defer, or commission a six-week customer study. Deferral reduces the available selling season and is valued separately; the central question is whether the study changes the launch decision enough to justify its $120,000 cost.

Analysis. The analytics team first defines high and low adoption, checks that payoffs are net of fulfillment and marketing costs, and estimates the study’s likelihood from comparable campaigns. It then builds a tree: research decision, study result, posterior adoption probability, launch decision, and terminal payoff. The team calculates EMV without research, EVPI, and EVSI. It also runs sensitivity analysis on the 35% prior, study accuracy, the launch loss, and the cost of delay.

Approach and solution. The model indicates that research has positive expected net benefit only if its accuracy is at least a specified threshold and the study can be completed before the seasonal deadline. Management therefore approves a limited pilot with a pre-specified decision rule: launch if the posterior probability of high adoption exceeds the threshold implied by the payoffs; otherwise do not launch. The pilot’s sample is stratified by region to reduce the risk that early respondents misrepresent the target market.

Outcome. The pilot provides useful evidence, but its value is lower than the initial estimate because the launch window narrows during data collection. The final decision uses the updated EVSI and includes the cost of delay. The retailer avoids treating a favorable survey as a guarantee and documents the assumptions for later model review.

Learning points. Information value depends on timing, action thresholds, signal quality, and consequences. A research proposal should specify how each possible result will affect action before data are collected. Sensitivity analysis is not an optional presentation feature; it reveals whether the recommendation is robust or depends on fragile assumptions.

Best-practice checklist

• Separate decisions, uncertain states, observations, and payoffs.
• Use conditional probabilities rather than reversing probabilities informally.
• Check that probabilities are coherent and exhaustive.
• State whether outcomes are independent and justify the assumption.
• Include all relevant costs, including delay and implementation.
• Calculate the best current action before valuing information.
• Treat EVPI as an upper bound, not a likely purchase price.
• Model sample-information accuracy and posterior updating explicitly.
• Use utility or risk constraints when expected money is inadequate.
• Test recommendations with sensitivity analysis and scenario analysis.
• Document data provenance, model limitations, and governance controls.
• Establish how evidence will be acted upon before collecting it.

Chapter Summary

Probability provides rules for representing uncertainty, while conditional probability describes how beliefs change when information is known. Bayes’ theorem combines prior beliefs with the likelihood of observed evidence to produce posterior beliefs. Probability distributions support expected-value, variance, simulation, and predictive analysis.

Decision trees organize sequential choices, chance events, and payoffs. The best current decision is the action with the greatest expected value, subject to the decision maker’s objective and constraints. Perfect information reveals the true state and establishes an upper bound through EVPI. Imperfect sample information is evaluated through posterior probabilities and EVSI. Information should be purchased only when its expected benefit exceeds its total cost and when ethical, timing, strategic, and governance considerations are acceptable.

The central managerial lesson is that information is valuable only when it can improve a decision. Sound prescriptive analytics therefore combines mathematical calculation with careful problem definition, credible probabilities, transparent assumptions, sensitivity analysis, and responsible governance.

Key Terms

Bayes’ theorem: A rule for updating the probability of a hypothesis using observed evidence.

Conditional probability: The probability of an event given that another event has occurred.

Decision tree: A graphical representation of decisions, chance events, and consequences over time.

Expected monetary value: The probability-weighted average payoff of an action.

Expected utility: The probability-weighted value after transforming monetary outcomes through a utility function.

EVPI: Expected value of perfect information; the gain from knowing the true state before acting.

EVSI: Expected value of sample information; the gain from observing an imperfect signal before acting.

Evidence: Observed information used to revise beliefs.

Independence: A condition in which knowledge of one event does not change the probability of another.

Likelihood: The probability of observed evidence conditional on a hypothesis or state.

Marginal probability: The probability of one variable or event after averaging over other variables.

Posterior probability: The updated probability after incorporating evidence.

Prior probability: A probability assigned before incorporating the current evidence.

Predictive distribution: The probability distribution of a future observation, accounting for parameter uncertainty.

Random variable: A numerical representation of an uncertain outcome.

Sample information: Imperfect information obtained from a study, test, experiment, or data source.

Sensitivity analysis: Examination of how results change when assumptions or inputs vary.

State of nature: An uncertain external condition that affects outcomes but is not controlled by the decision maker.

Value of information: The expected improvement in decision value produced by additional information.

Review Questions

Short Answer Questions

1. Distinguish between a prior probability, a likelihood, and a posterior probability.
2. State the multiplication rule for two events and identify its relationship to conditional probability.
3. Why is EVPI an upper bound on the value of imperfect information?
4. What does it mean for two events to be independent?
5. Identify two reasons why a high-accuracy prediction may have low decision value.

Long Answer Questions

1. Explain Bayes’ theorem using a business example involving an imperfect market signal. Interpret the base-rate effect.
2. Describe how a decision tree is constructed and solved using rollback analysis. Discuss common modeling errors.
3. Compare EVPI and EVSI, including their formulas, interpretation, and managerial use.
4. Discuss why expected monetary value may be inappropriate for a highly risk-averse organization.
5. Explain how ethical, privacy, fairness, and governance concerns affect the use of probabilistic decision models.

Analytical / Application Questions

1. An investment has a 30% probability of producing $500,000 and a 70% probability of losing $100,000. Calculate its EMV and compare it with doing nothing. What additional information would be worth investigating?
2. A diagnostic signal is positive 90% of the time when a condition is present and 10% of the time when it is absent. If the condition’s prior probability is 5%, calculate the probability of the condition given a positive signal. Interpret the result.
3. Design an EVSI analysis for a proposed customer survey. Specify the actions, states, prior probabilities, survey outcomes, likelihoods, payoffs, research cost, and decision rule. Explain what sensitivity analyses should be performed.

Multiple Choice Questions with Answers

Question 1

A firm assigns P(H) = 0.20, P(+|H) = 0.80, and P(+|Hᶜ) = 0.30. What is P(H|+)?

A. 0.20

B. 0.30

C. 0.40

D. 0.80

Correct Answer: C. 0.40

Question 2

Which expression defines expected monetary value for action a across states s?

A. maxₐ P(s)V(a,s)

B. ∑ₛ P(s)V(a,s)

C. ∑ₐ P(s)V(a,s)

D. V(a,s) ÷ P(s)

Correct Answer: B. ∑ₛ P(s)V(a,s)

Question 3

A project pays 300 in a high-demand state and −100 in a low-demand state. The probabilities are 0.40 and 0.60. What is its EMV?

A. 20

B. 40

C. 60

D. 120

Correct Answer: C. 60

Question 4

Which statement about EVPI is correct?

A. It is always equal to the research cost.

B. It is the value of the best current action without information.

C. It is the maximum expected improvement from knowing the true state before acting.

D. It measures only the accuracy of a survey.

Correct Answer: C. It is the maximum expected improvement from knowing the true state before acting.

Question 5

If a research study cannot change the recommended action for any possible result, its EVSI is:

A. Equal to EVPI

B. Zero, before considering broader non-decision benefits

C. Equal to the study cost

D. Always negative

Correct Answer: B. Zero, before considering broader non-decision benefits

Question 6

For mutually exclusive and exhaustive hypotheses Bᵢ, Bayes’ theorem requires which denominator?

A. P(Bᵢ) only

B. P(A|Bᵢ) only

C. ∑ⱼ P(A|Bⱼ)P(Bⱼ)

D. ∑ⱼ P(Bⱼ) ÷ P(A|Bⱼ)

Correct Answer: C. ∑ⱼ P(A|Bⱼ)P(Bⱼ)

Question 7

A positive signal has P(+|H) = 0.75, P(+|L) = 0.25, P(H) = 0.40, and P(L) = 0.60. What is P(+)?

A. 0.25

B. 0.40

C. 0.45

D. 0.75

Correct Answer: C. 0.45

Question 8

Under the expected-utility framework, the function u(·) is used to:

A. Remove all uncertainty from the model

B. Transform outcomes to reflect risk preferences

C. Calculate the likelihood of evidence

D. Guarantee that EVSI exceeds EVPI

Correct Answer: B. Transform outcomes to reflect risk preferences

Question 9

Which condition is required for two events A and B to be independent, assuming positive probabilities?

A. P(A|B) = P(B|A)

B. P(A ∩ B) = P(A)P(B)

C. P(A ∪ B) = 1

D. P(A) = P(B) = 0.5

Correct Answer: B. P(A ∩ B) = P(A)P(B)

Question 10

A study has EVSI of $15,000 and costs $10,000, excluding other considerations. What is its expected net benefit of information?

A. −$25,000

B. −$5,000

C. $5,000

D. $25,000

Correct Answer: C. $5,000