Decision Trees and Decisions Under Uncertainty

Introduction

Organizations routinely make decisions before the future is known. A retailer must decide how much inventory to order before demand is observed; a bank must approve or reject a loan before the borrower’s repayment behavior is certain; and a firm may need to invest in a product before market acceptance is established. These are decisions under uncertainty because the consequences depend partly on events outside the decision maker’s control.

Decision trees provide a structured way to represent such problems. They show the sequence of decisions and uncertain events, attach probabilities to uncertain outcomes, assign monetary or utility consequences to complete paths, and support systematic comparison of policies. Their central computational technique is rollback analysis, also called backward induction: values are calculated from the terminal outcomes back toward the initial decision.

A decision tree is not a substitute for managerial judgment. It is a transparent model of judgment. Its quality depends on the quality of the alternatives, probabilities, payoffs, timing assumptions, and risk preferences incorporated into it.

Learning Objectives

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

1. Distinguish decision, chance, and terminal nodes.
2. Construct a decision tree that represents sequential decisions and uncertain events.
3. Assign probabilities consistently and calculate expected monetary value (EMV).
4. Apply rollback analysis to select an optimal policy.
5. Use dominance, conditional paths, sensitivity analysis, and risk profiles.
6. Explain the difference between an optimal action and an optimal contingent policy.
7. Communicate recommendations, assumptions, limitations, and governance implications.

1. The Logic of Decision Trees

A decision tree is a directed, acyclic representation of a decision problem. It is read from left to right, usually beginning with an initial decision and ending with terminal outcomes.

Node types

A branch is an alternative leaving a node. A path is a sequence of branches from the initial node to a terminal node. A policy specifies what the decision maker will do at every decision node that might be reached. This distinction is important: an action describes one immediate choice, whereas a policy includes contingent future actions.

For a chance node with mutually exclusive and collectively exhaustive outcomes, probabilities must satisfy:

Here, p_i is the probability of outcome i. Probabilities must refer to the information available when the chance event is modeled. If probabilities depend on earlier events, conditional probabilities should be used rather than copied from unrelated situations.

Additivity and timing

A tree should represent events in their chronological order. If a firm chooses whether to test a product before deciding whether to launch it, the test decision must appear before the launch decision. A chance event should not be placed before a decision if the outcome will only be known afterward.

Payoffs should be measured at a consistent point in time. If cash flows occur in different periods, they should be discounted when the decision is based on present value. For a cash flow C_t received at time t, with discount rate r, its present value is:

This formula assumes a discrete-period model, a constant discount rate, and cash flows known at their stated times. Units must be consistent: for example, annual cash flows require an annual rate unless the rate is converted appropriately.

2. Expected Monetary Value

When outcomes are expressed in monetary units and the decision maker is risk neutral, the principal evaluation criterion is expected monetary value:

where p_i is the probability of outcome i, V_i is its monetary value, and the sum covers all mutually exclusive outcomes. EMV has the same monetary unit as V_i. It is an average over repeated, comparable situations; it is not a forecast that must occur in any single case.

For a chance node with two outcomes, high and low, the calculation is:

The condition p_L = 1 − p_H applies only when there are exactly two exhaustive outcomes.

At a decision node, the value is the maximum branch value for a value-maximizing decision maker:

For a cost-minimizing problem, the minimum is selected instead. The maximization rule assumes that all relevant consequences have been represented in the payoff and that the decision maker accepts risk neutrality or is using EMV as an approximation.

Worked example: capacity choice

A manufacturer must choose between a small and a large plant. Demand may be high with probability 0.60 or low with probability 0.40. Payoffs, in thousands of dollars, are as follows:

For the small plant:

For the large plant:

The large plant has the higher EMV, $156,000 versus $148,000. The recommendation is therefore to choose the large plant under the stated probabilities, payoffs, and risk-neutral criterion. However, the result is not overwhelmingly superior: the difference is only $8,000, while the large plant has a much worse low-demand outcome.

3. Constructing a Decision Tree

A defensible construction process consists of the following steps.

Step 1: Define the decision problem

Specify the decision maker, objective, decision horizon, feasible alternatives, and relevant performance measure. “Choose the best strategy” is insufficient unless “best” is defined as profit maximization, cost minimization, service-level improvement, risk-adjusted return, or another criterion.

Step 2: Identify decisions and information arrival

List decisions in chronological order. For every decision, state what information is available. A later decision may depend on observed demand, a market signal, a diagnostic test, or a competitor’s action.

Step 3: Identify chance events

List events that are not controlled by the decision maker but affect the outcome. Outcomes at a chance node must be exhaustive and mutually exclusive. Avoid combining outcomes with different consequences merely to simplify the tree.

Step 4: Assign probabilities

Use historical data, statistical models, expert elicitation, market research, or Bayesian updating. Document the source, reference population, time period, and uncertainty in the estimate. Probabilities should be conditional when earlier information changes the likelihood of later events.

Step 5: Assign terminal payoffs

Calculate profit, net present value, cost, utility, or another suitable consequence for every complete path. Include relevant fixed and variable costs, opportunity costs, taxes, penalties, salvage values, and implementation costs. Do not mix revenue on one branch with profit on another.

Step 6: Check the tree

Verify that:

• every decision branch is feasible;
• chance probabilities sum to one at each chance node;
• no future information is used prematurely;
• every path ends in a payoff;
• units and timing are consistent; and
• duplicated or impossible paths have been removed.

4. Rollback Analysis

Rollback evaluates the tree from right to left.

1. At each terminal node, record the payoff.
2. At each chance node, calculate the probability-weighted average of downstream values.
3. At each decision node, select the best available alternative.
4. Continue until the initial decision is reached.

The selected branches collectively define the recommended policy. A branch that is not selected at a decision node is discarded for purposes of the optimal policy, although it remains useful for sensitivity and risk analysis.

Sequential example: a market test

A firm is considering a product launch. It may launch immediately or conduct a market test costing $20,000. Without testing, the probability of strong demand is 0.50. Launch payoffs before test cost are $240,000 under strong demand and −$100,000 under weak demand. If the firm does not launch, the payoff is $0.

The EMV of immediate launch is:

Thus, immediate launch dominates immediate non-launch under EMV.

Suppose the test produces either a favorable or unfavorable signal. The test has the following conditional probabilities:

• P(favorable signal | strong demand) = 0.80
• P(favorable signal | weak demand) = 0.20

The prior probabilities are P(strong) = 0.50 and P(weak) = 0.50. The probability of a favorable signal is:

Using Bayes’ rule:

Similarly:

After a favorable signal, the EMV of launching is:

After an unfavorable signal, the EMV of launching is:

Therefore, the optimal contingent policy after testing is: launch if the signal is favorable and do not launch if it is unfavorable. The expected value before paying for the test is:

After subtracting the test cost:

Because $66,000 is less than the $70,000 EMV of immediate launch, the test is not worthwhile under these assumptions, even though it is informative. The recommended policy is immediate launch.

This example illustrates an important principle: information has value only when it changes an action sufficiently to justify its cost. A signal that is accurate but too expensive, or that does not alter the decision, may have little practical value.

5. Dominance and Simplification

An alternative is dominated if another alternative is at least as good in every relevant state and strictly better in at least one state. A dominated alternative can be removed before detailed calculation.

For example, suppose three suppliers have costs under three demand conditions:

For cost minimization, Supplier B is dominated by Supplier A because A costs less in every state. Supplier B can be excluded. Supplier C is not dominated because it is best in low and high demand but worse in medium demand.

Dominance may be unconditional or conditional. A future action can be dominated only after a particular signal or state is observed. Removing dominated branches reduces tree size and improves communication, but the comparison must use the same states, units, constraints, and timing.

6. Conditional Paths and Policies

A decision tree often represents decisions that depend on information. The correct question is not simply “Which first action has the greatest average value?” It is “Which policy specifies the best action for each information state that may occur?”

Suppose a firm can choose a conservative or aggressive marketing budget after observing a demand forecast. The optimal policy might be:

• favorable forecast → aggressive budget;
• neutral forecast → conservative budget;
• unfavorable forecast → no campaign.

This policy may be superior to any single action applied regardless of the forecast. A common modeling error is to calculate the value of the best action separately for each forecast and then compare those values without weighting by the probability of receiving each forecast. Rollback prevents this error by combining conditional decisions with the probability of reaching them.

The probability of a complete path is the product of conditional probabilities along that path. For events A and B:

This multiplication rule assumes that the second probability is conditional on the first event. Independence is a special case in which P(B | A) = P(B). It should not be assumed merely because two events occur at different times.

7. Risk Profiles and Decision Quality

EMV alone does not reveal the distribution of outcomes. A risk profile lists each possible payoff and its probability under a selected policy. Two policies may have the same EMV but very different downside exposure.

Useful risk measures include:

Expected value

Variance and standard deviation

Variance is measured in squared monetary units, while standard deviation has the same unit as the payoff. These measures summarize dispersion but do not fully describe asymmetry or tail risk.

Downside probability

For a threshold T, the probability of loss below the threshold is:

where I(·) equals 1 when the condition is true and 0 otherwise. A threshold may be zero for loss probability or a minimum acceptable return.

Returning to the plant example, the small plant has a 0% probability of a negative payoff, while the large plant has a 40% probability of a −$60,000 payoff. Although the large plant has a higher EMV, a risk-averse manager may prefer the small plant.

Risk aversion and expected utility

If monetary value does not adequately represent preferences, use utility. Let U(V) be the decision maker’s utility of payoff V. The expected utility is:

The preferred alternative maximizes expected utility rather than EMV. A risk-averse utility function is concave, meaning that an additional dollar generally provides less utility as wealth increases. Utility must be elicited carefully; a mathematically convenient function is not automatically behaviorally valid.

8. Sensitivity, Thresholds, and Value of Information

Decision-tree recommendations should be stress-tested. A one-way sensitivity analysis changes one input while holding others constant. A scenario analysis changes several coherent inputs together. A probabilistic sensitivity analysis assigns distributions to uncertain inputs and repeatedly evaluates the model, often through simulation.

For the plant example, let p be the probability of high demand. The EMVs are:

The large plant is preferred when:

Thus, the large plant is preferred only when the probability of high demand exceeds approximately 57.14%. This threshold is more informative than simply reporting the result for one estimated probability.

The expected value of perfect information (EVPI) measures the maximum amount worth paying for certainty about the state of nature. For a maximization problem:

The first term is the expected value if the state is known before the decision; the second is the best EMV without such information. EVPI is an upper bound: no imperfect information can be worth more than perfect information.

The expected value of sample information (EVSI) is the improvement from a specific imperfect information system before its cost. It is calculated using rollback with updated probabilities and contingent actions, less the value without information. Information should be purchased only if its expected benefit exceeds its acquisition and implementation cost, subject to strategic and regulatory considerations.

9. Case Application: Regional Distribution Center

The following is an illustrative case.

Background

Northstar Home Goods is considering a regional distribution center. It can build a small facility, build a large facility, or lease capacity for two years while demand becomes clearer. Management estimates that regional demand will be high with probability 0.45 and low with probability 0.55.

Problem

The company must choose a policy that balances expected profit, downside risk, and the value of learning. Monetary figures are present values in millions of dollars.

Alternatives and assumptions

The leasing option includes a later decision after demand information becomes available. Its payoffs already include lease expenses and expansion costs, so they should not be added again.

Analysis

Small facility:

Large facility:

Lease policy:

The small facility has the highest EMV. It also avoids the large facility’s negative low-demand outcome. The lease provides flexibility but has the lowest EMV because the value of flexibility does not compensate for the cost of waiting under the stated assumptions.

Outcome and recommendation

Northstar should choose the small facility if it is using expected monetary value and the assumptions are credible. The recommendation should be communicated conditionally: if the probability of high demand rises sufficiently, if the large facility’s low-demand loss falls, or if capacity constraints make unmet demand especially costly, the decision may change.

Learning points

1. Flexibility has value only when future information can change future action.
2. A high-capacity option may have attractive upside but unacceptable downside.
3. Payoff definitions must prevent double counting.
4. The recommendation should include assumptions and trigger conditions for review.

10. Limitations, Governance, and Ethics

Decision trees are powerful but can create false precision. The principal limitations are:

• Probability uncertainty: estimated probabilities may be unstable, biased, or based on small samples.
• Model risk: omitted states, dependencies, constraints, or costs can distort results.
• Payoff uncertainty: forecasts of revenue, cost, and behavior may be more uncertain than the tree suggests.
• Risk-preference ambiguity: EMV may be inappropriate for individuals or organizations with strong downside constraints.
• Tree explosion: many sequential events can create an unwieldy number of branches.
• Dependence and feedback: competitors, regulators, and customers may react to the decision, violating fixed-state assumptions.
• Behavioral effects: managers may overweight salient outcomes or probabilities and reject a mathematically sound recommendation.

Governance requires documenting the model owner, data sources, assumptions, version history, approval authority, and review date. Material decisions should have independent challenge, reproducible calculations, and an audit trail. If the tree uses customer, employee, health, or credit data, privacy, fairness, and applicable data-protection rules must be considered.

An apparently optimal policy can be unethical if its payoffs exclude harms imposed on vulnerable groups or the environment. Organizations should include relevant externalities, legal constraints, safety thresholds, and distributional consequences rather than treating profit as the only legitimate payoff.

11. Technology and Contemporary Practice

Modern analytics platforms can generate and evaluate large trees, connect probability estimates to databases, and simulate thousands of scenarios. Machine-learning models may estimate demand probabilities or conditional transition probabilities, while optimization systems may select actions under constraints.

Automation does not eliminate the need for human oversight. Predictive accuracy is not the same as decision quality: a well-calibrated probability model can still support a poor policy if costs are wrong or fairness constraints are ignored. Analysts should monitor calibration, drift, subgroup performance, and realized outcomes. When model outputs affect people, explanations should identify the relevant variables, uncertainty, and recourse options rather than presenting an opaque recommendation.

Best Practices for Communicating a Recommendation

A professional decision-tree report should state:

1. the decision and objective;
2. the recommended initial action and contingent policy;
3. the EMV or expected utility and units;
4. the principal assumptions and probability sources;
5. key downside outcomes and risk measures;
6. sensitivity thresholds and monitoring triggers;
7. information, implementation, and governance costs; and
8. conditions under which the recommendation should be revisited.

Avoid reporting a single number without the policy that generated it. A manager needs to know not only what to do now, but also what to do when new information arrives.

Chapter Summary

Decision trees represent sequential choices under uncertainty through decision nodes, chance nodes, branches, and terminal payoffs. A valid tree preserves chronology, uses exhaustive and mutually exclusive chance outcomes, assigns probabilities that sum to one, and measures payoffs consistently.

Expected monetary value is the probability-weighted average of monetary outcomes and is appropriate under risk neutrality or as a transparent baseline. Rollback analysis evaluates chance nodes by weighted averages and decision nodes by selecting the best feasible branch. In sequential problems, rollback produces a contingent policy rather than merely a first action.

Dominance can simplify a tree by removing alternatives that are never better. Conditional probabilities and Bayes’ rule support decisions that respond to signals or tests. Risk profiles, variance, standard deviation, downside probability, and expected utility reveal considerations that EMV alone may hide. Sensitivity analysis identifies thresholds at which a recommendation changes, while EVPI and EVSI quantify the potential value of information.

Finally, decision trees are models, not objective truths. Their recommendations depend on data quality, payoff definitions, risk preferences, dependencies, ethical constraints, and governance. Effective analysts communicate both the recommendation and the assumptions, risks, monitoring triggers, and circumstances requiring revision.

Key Terms

Decision tree: A directed representation of sequential decisions, uncertain events, and outcomes.

Decision node: A point at which the decision maker chooses among feasible alternatives.

Chance node: A point at which one of several uncertain outcomes occurs.

Terminal node: The endpoint of a path with an assigned payoff or utility.

Branch: An alternative action or outcome leaving a node.

Path: A sequence of branches from the initial node to a terminal outcome.

Policy: A complete specification of actions, including contingent future actions.

Expected monetary value: The probability-weighted average of monetary outcomes.

Rollback analysis: Backward evaluation of a decision tree from terminal nodes to the initial decision.

Dominance: A relationship in which one alternative is never worse and sometimes better than another.

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

Bayes’ rule: A method for updating probabilities using new evidence.

Risk profile: The set of possible outcomes and their probabilities under a policy.

Expected utility: The probability-weighted utility of possible outcomes.

Sensitivity analysis: Examination of how results change when model inputs change.

EVPI: Expected value of perfect information.

EVSI: Expected value of sample or imperfect information.

Model risk: The possibility that model structure, assumptions, or inputs produce misleading conclusions.

Review Questions

Short Answer Questions

1. What are the three principal types of nodes in a decision tree?
2. Why must probabilities at a chance node sum to one?
3. What is the difference between an action and a policy?
4. When can a dominated alternative be removed from a tree?
5. Why might a manager prefer a lower-EMV alternative?

Long Answer Questions

1. Explain the process of constructing and validating a decision tree from a business problem.
2. Describe rollback analysis and explain how it produces a contingent policy.
3. Compare expected monetary value with expected utility as decision criteria.
4. Explain how sensitivity analysis and threshold probabilities improve decision communication.
5. Discuss the principal ethical and governance risks associated with automated decision-tree models.

Analytical / Application Questions

1. A firm can choose Project A with payoffs of $120,000 under success and $20,000 under failure, or Project B with payoffs of $220,000 under success and −$80,000 under failure. If success has probability 0.60, calculate both EMVs and recommend an option under risk neutrality.
2. A diagnostic test is positive with probability 0.70 when a condition is present and 0.10 when it is absent. If the prior probability of the condition is 0.20, calculate the probability that the condition is present given a positive result.
3. Design a decision tree for a retailer deciding whether to order regular or expanded inventory after receiving a favorable or unfavorable demand forecast. State the nodes, probabilities, payoffs, policy, and at least two sensitivity questions.

Multiple Choice Questions with Answers

Question 1

Which node represents a point at which the decision maker chooses among alternatives?

A. Terminal node

B. Chance node

C. Decision node

D. Payoff node

Correct Answer: C. Decision node

Question 2

A chance node has outcomes with probabilities 0.30, 0.25, and 0.45. What is the expected monetary value if the payoffs are $100, $200, and −$40, respectively?

A. $66

B. $82

C. $94

D. $120

Correct Answer: B. $82

Question 3

What is the correct order for rollback analysis?

A. Evaluate the initial decision first, then terminal nodes

B. Evaluate terminal nodes, chance nodes, and decision nodes moving backward

C. Average all terminal payoffs without using probabilities

D. Select the most likely path only

Correct Answer: B. Evaluate terminal nodes, chance nodes, and decision nodes moving backward

Question 4

An alternative is dominated when another alternative is:

A. Better in every state and strictly better in at least one state

B. Better only in the most likely state

C. More expensive but less risky

D. Preferred by one manager

Correct Answer: A. Better in every state and strictly better in at least one state

Question 5

If P(A) = 0.4 and P(B | A) = 0.5, what is P(A and B)?

A. 0.10

B. 0.20

C. 0.40

D. 0.90

Correct Answer: B. 0.20

Question 6

In the plant example, the large plant is preferred to the small plant when the probability of high demand is approximately:

A. Below 40%

B. Exactly 50%

C. Above 57.14%

D. Above 90%

Correct Answer: C. Above 57.14%

Question 7

Which measure directly describes the probability that payoff V is below a specified threshold T?

A. Expected value

B. Variance

C. Downside probability

D. EVPI

Correct Answer: C. Downside probability

Question 8

Expected utility is especially useful when:

A. All outcomes are certain

B. The decision maker’s risk preferences matter

C. Probabilities cannot be assigned at all

D. All payoffs have identical values

Correct Answer: B. The decision maker’s risk preferences matter

Question 9

The expected value of perfect information is best interpreted as:

A. The guaranteed profit from gathering information

B. The maximum expected amount worth paying for certainty about the state

C. The cost of every market test

D. The variance of the best policy

Correct Answer: B. The maximum expected amount worth paying for certainty about the state

Question 10

In the market-test example, why was immediate launch preferred to testing?

A. The test had no predictive value

B. Testing guaranteed a loss

C. The test’s expected value after cost was lower than immediate launch

D. The firm could not make a contingent decision

Correct Answer: C. The test’s expected value after cost was lower than immediate launch