BUSINESS STATISTICS • PROBABILITY MODELS

Decision Trees & Expected Value — Decision Trees and Expected Value (Intro)

Structuring uncertainty so that every strategic choice has a quantifiable payoff.

Historical Context & Motivation

Managers have always confronted uncertainty—whether to launch a new product line, enter a foreign market, or invest in R&D—but for centuries those choices relied almost entirely on intuition and experience. The formal tools we now use to map, quantify, and compare uncertain outcomes evolved across several disciplines over hundreds of years. Decision trees and expected value sit at the intersection of probability theory, economics, and operations research, offering a structured visual and mathematical framework that converts ambiguity into actionable numbers.

1654
The Pascal–Fermat Correspondence
Blaise Pascal and Pierre de Fermat exchanged letters on the problem of points, laying the foundation for probability theory and the concept of mathematical expectation—what we now call expected value.
1738
Bernoulli's Expected Utility
Daniel Bernoulli introduced expected utility theory in his St. Petersburg Paradox paper, arguing that decision-makers weight outcomes by subjective value (utility) rather than raw monetary amounts—a refinement still central to modern finance.
1944
Von Neumann–Morgenstern Axioms
John von Neumann and Oskar Morgenstern published Theory of Games and Economic Behavior, axiomatizing rational choice under uncertainty and demonstrating that consistent preferences can always be represented by an expected utility function.
1960s
Decision Analysis at Harvard & Stanford
Howard Raiffa and Ronald Howard formalized decision analysis as an academic discipline, popularizing tree-structured diagrams as a practical tool for corporate strategy, medical decisions, and public policy.
2000s–Present
Computational Decision Trees
With spreadsheet software, Monte Carlo simulation, and machine-learning classification trees, decision-tree methodology has expanded from boardroom strategy into data science, algorithmic trading, and supply-chain optimization.

The central question these pioneers addressed remains the same one you face in any business course case study: given multiple alternatives, each leading to uncertain outcomes with different probabilities, which option yields the highest long-run value? Decision trees provide the visual scaffolding, and expected value provides the arithmetic engine to answer that question rigorously.

Core Principles & Definitions

Before constructing a decision tree, it helps to internalize the handful of foundational ideas that give the technique its power. Each principle addresses a different facet of how we decompose, represent, and evaluate strategic choices under uncertainty.

1

Decision Node (□)

A square node where the decision-maker chooses among two or more alternatives. Each branch leaving the square represents a distinct course of action—for example, 'Launch Product A' versus 'Launch Product B.'
2

Chance Node (○)

A circular node representing an uncertain event outside the decision-maker's control. Each branch leaving the circle carries a probability, and these probabilities must sum to 1.0.
3

Terminal Payoff (△ or value)

The endpoint of a path through the tree, showing the monetary outcome (profit, cost, or net present value) if that exact sequence of decisions and chance events occurs.
4

Expected Value (EV)

The probability-weighted average of all possible payoffs at a chance node. EV allows you to collapse uncertainty into a single number, making otherwise complex comparisons straightforward.
5

Rollback (Folding Back)

The algorithm used to solve a decision tree: start at the terminal payoffs and work backward—computing EVs at chance nodes and selecting the best option at decision nodes—until you reach the root.
KEY TAKEAWAY
Think of a decision tree like a GPS navigation app that shows every possible route to your destination. Each fork in the road is either your choice (decision node) or traffic conditions beyond your control (chance node). Expected value is like the estimated arrival time the app calculates by weighting each possible delay by its likelihood—it gives you a single number to compare routes, even though no single trip will match the estimate exactly.

Anatomy of a Decision Tree

The diagram below illustrates a classic business scenario: a firm must decide whether to launch a new product or stay with its current portfolio. If the firm launches, market demand is uncertain—it can be high, moderate, or low. Each path through the tree ends at a terminal payoff denominated in thousands of dollars of profit. Study the structure carefully: notice how squares represent choices the manager controls, circles represent chance events, and the rightmost values represent financial outcomes.

The square node (D) is the decision node where the manager chooses. The circle node (C) is the chance node where uncertain market demand resolves. Terminal payoffs appear in colored rectangles on the right. The rollback calculation boxes show that launching yields an EV of $230K, exceeding the certain payoff of $150K from staying.

Several conventions appear consistently across business textbooks and industry practice. Decision nodes are always drawn as squares because a square implies a deliberate, controlled action. Chance nodes are drawn as circles to suggest the wheel-of-fortune nature of stochastic events. Trees are read from left to right (or top to bottom), tracing the chronological sequence of decisions and uncertainties. When solving the tree, however, we work in the opposite direction—right to left—using the rollback algorithm, because we need to know the value of downstream outcomes before we can evaluate upstream choices.

Mathematical Framework

The quantitative backbone of decision-tree analysis rests on the concept of expected value (EV), which is the probability-weighted sum of all possible payoffs associated with a chance event. The expected value does not predict what will happen on any single occasion; rather, it represents the long-run average outcome if the same decision were repeated many times under identical conditions. In business settings where decisions are one-shot, EV still serves as a rational benchmark for comparing alternatives.

EXPECTED VALUE AT A CHANCE NODE
EV = Σᵢ pᵢ × Vᵢ = p₁V₁ + p₂V₂ + … + pₙVₙ
where pᵢ = probability of outcome i, Vᵢ = payoff (or value) if outcome i occurs, n = number of possible outcomes, and Σpᵢ = 1.
DECISION NODE RULE
V*(decision node) = max { EV(alternative₁), EV(alternative₂), … , EV(alternativeₖ) }
At a decision node the rational manager selects the branch with the highest expected value (for profits) or the lowest expected cost (for cost-minimization problems).
EXPECTED VALUE OF PERFECT INFORMATION (EVPI)
EVPI = EV(with perfect information) − EV(without perfect information)
EVPI quantifies the maximum amount the decision-maker should be willing to pay for a perfect forecast. It represents the gap between the payoff achievable with clairvoyance and the payoff from the best decision under uncertainty.

The rollback procedure applies these two rules systematically. Starting at every terminal payoff on the far right of the tree, you replace each chance node with its EV and each decision node with the value of its optimal branch. Propagating these values leftward eventually yields a single number at the root—the expected value of the entire decision problem—and the path that produced it is the optimal strategy.

The Rollback Algorithm — Step by Step

The rollback algorithm (also called backward induction or folding back) is the standard method for solving any decision tree, regardless of its size or complexity. The diagram below generalizes the process into a flowchart that you can apply to trees with many stages of decisions and chance events.

The flowchart captures the iterative nature of the rollback algorithm. At each unsolved node the procedure branches: compute EV at chance nodes or select the best branch at decision nodes. The loop continues until the root node carries the optimal expected value for the entire problem.
  1. Identify all terminal payoffs — these are the rightmost endpoints, representing the financial result of each complete path through the tree.
  2. Move one step left to the nearest unsolved chance or decision node.
  3. Chance node → compute EV by multiplying each branch's probability by its downstream value, then summing.
  4. Decision node → pick the best branch (highest EV for profit problems; lowest EV for cost problems).
  5. Repeat until the root node carries a single value—the expected value of the optimal strategy.

Worked Example — Market Expansion Decision

A regional coffee chain is evaluating whether to expand into a neighboring city. The expansion requires a $120K upfront investment. If the local economy is strong (probability 0.4), the new store generates $300K in net revenue. If the economy is moderate (probability 0.35), revenue is $180K. If the economy is weak (probability 0.25), revenue is only $60K. Alternatively, the chain can invest the same $120K in upgrading existing locations, which generates a certain net profit of $80K after all costs. Should the firm expand?

Market Expansion — Rollback Solution
1
Step 1 — Define Payoffs (Net of Investment)Because both alternatives require a $120K investment, we compute net profit (revenue minus investment cost) for each outcome of the expansion. Strong economy: $300K − $120K = $180K. Moderate economy: $180K − $120K = $60K. Weak economy: $60K − $120K = −$60K. The upgrade alternative yields a certain net profit of $80K (already stated net of its $120K cost). All terminal values are therefore expressed on a consistent net-profit basis.
Expansion net payoffs: $180K (strong), $60K (moderate), −$60K (weak). Upgrade net payoff: $80K.
2
Step 2 — Compute EV at the Chance Node (Expansion)EV(Expand) = 0.4 × $180K + 0.35 × $60K + 0.25 × (−$60K) = $72K + $21K + (−$15K) = $78K.
EV(Expand) = $78K
3
Step 3 — Evaluate the Decision NodeCompare the EV of expansion ($78K) with the certain net profit of the upgrade ($80K). Since $80K > $78K, the upgrade alternative is marginally preferred under an EV criterion.
Optimal decision: Upgrade existing locations (EV = $80K)
4
Step 4 — Compute EVPITo compute EV with perfect information, we determine the best action for each possible state of the economy before it is revealed. If strong (p = 0.4): choose expand ($180K) over upgrade ($80K), so payoff = $180K. If moderate (p = 0.35): choose upgrade ($80K) over expand ($60K), so payoff = $80K. If weak (p = 0.25): choose upgrade ($80K) over expand (−$60K), so payoff = $80K. EV(with perfect information) = 0.4 × $180K + 0.35 × $80K + 0.25 × $80K = $72K + $28K + $20K = $120K. EVPI = EV(with PI) − EV(best decision without PI) = $120K − $80K = $40K.
EVPI = $40K — the maximum the firm should pay for a perfect economic forecast
💡 Practical Insight
Notice how a $2K difference in expected value ($80K vs. $78K) could easily swing the other way if probabilities shift even slightly. In practice, managers complement EV analysis with sensitivity analysis—varying the probability estimates to see at what threshold the optimal decision flips. If a ±5% change in the strong-economy probability reverses the choice, the firm knows to invest in better market research before committing.

Strengths & Limitations of Decision Trees

Like every analytical tool, decision trees offer powerful advantages in certain contexts while carrying inherent limitations that the practitioner must acknowledge. The table below distills the most commonly cited points so that you can gauge when the technique is appropriate and when you should supplement or replace it with alternative methods.

Decision Tree Strengths vs. Limitations
DimensionStrengthsLimitations
TransparencyThe visual layout makes assumptions explicit—every probability, payoff, and decision is visible to stakeholders.Large trees become cluttered; more than 4–5 stages may require software to remain readable.
Sequential LogicNaturally represents multi-stage decisions where later choices depend on earlier outcomes (e.g., phased investment).Assumes a fixed sequence; poorly suited for simultaneous or continuous decisions.
Quantitative RigorForces explicit assignment of probabilities and payoffs, preventing 'gut-feel' vagueness.Garbage in, garbage out—if probability estimates are unreliable, the EV calculation is misleading.
Risk NeutralitySimple and intuitive when decision-makers are approximately risk-neutral or when stakes are small relative to wealth.Standard EV ignores risk aversion; a −$60K loss may hurt more than a +$60K gain helps. Must extend to expected utility for risk-sensitive decisions.
CommunicationExcellent for boardroom presentations—stakeholders can trace the logic of each branch and verify the assumptions behind every probability and payoff.May oversimplify continuous distributions into discrete scenarios, losing nuance.
KEY TAKEAWAY
Decision trees are the Swiss Army knife of business decision analysis: compact, portable, and highly useful for a wide range of problems. But just as a Swiss Army knife is no substitute for a full toolbox, decision trees work best when combined with sensitivity analysis, scenario planning, and Monte Carlo simulation for complex, high-stakes decisions.

Connection to Advanced Decision Theory

The introductory decision-tree framework presented in this lesson assumes risk neutrality and discrete probability distributions. As you progress through your business statistics and finance courses, you will encounter several extensions that address these simplifications. The table below maps the basic concepts you have learned to their more sophisticated counterparts.

From Introductory to Advanced Decision Models
Introductory ConceptAdvanced ExtensionKey Difference
Expected Value (EV)Expected Utility (EU)EU applies a concave utility function to payoffs, capturing risk aversion. A risk-averse manager may reject a positive-EV gamble if the downside is catastrophic.
Discrete chance nodesContinuous probability distributionsOutcomes modeled via normal, triangular, or beta distributions; solved analytically or via Monte Carlo simulation.
Static tree (one-time decision)Real options analysisTreats flexibility (delay, expand, abandon) as an option with quantifiable value, using option-pricing models from finance.
Known probabilitiesBayesian updatingProbabilities are revised as new data arrives, allowing the tree to reflect learning over time.
Single decision-makerGame trees (strategic interaction)Extends the tree to include decisions by competitors, modeled via game theory and Nash equilibrium.

Understanding the introductory framework thoroughly is essential because every advanced extension builds directly on it. Expected utility replaces the payoff Vᵢ with U(Vᵢ) but still uses the same weighted-sum formula. Bayesian updating modifies the probabilities pᵢ but leaves the tree structure intact. Mastering the basics ensures that each new layer feels like a natural generalization rather than a wholly new concept.

Practice Problems

PROBLEM 1CONCEPTUAL
Explain the difference between a decision node and a chance node in a decision tree. Why is this distinction critical for the rollback algorithm?
PROBLEM 2BASIC CALCULATION
A startup considers two strategies. Strategy A has three possible outcomes: profit of $400K (p = 0.25), $100K (p = 0.50), or a loss of $200K (p = 0.25). Strategy B yields a certain profit of $90K. Compute the expected value of Strategy A and identify the optimal strategy under an EV criterion.
PROBLEM 3INTERMEDIATE
A pharmaceutical company must decide whether to (1) develop Drug X at a cost of $5M or (2) license an existing drug for $2M. If Drug X is developed, clinical trials succeed with probability 0.6 (payoff $15M in revenue) or fail with probability 0.4 (payoff $0). If the company licenses, market reception is strong (p = 0.5, revenue $8M) or weak (p = 0.5, revenue $3M). Construct the decision tree mentally, compute net payoffs (revenue minus cost), find the EV of each alternative, and identify the optimal decision.
PROBLEM 4APPLIED
A logistics firm is deciding whether to build a large warehouse ($10M) or a small warehouse ($4M). If demand is high (p = 0.5), the large warehouse generates $18M in present-value revenue and the small warehouse generates $9M. If demand is low (p = 0.5), the large warehouse generates $7M and the small warehouse generates $6M. Additionally, if the firm builds small and demand turns out high, it can later expand for an additional $5M, which would bring revenue to $15M. Draw the tree structure, compute all EVs, and recommend the optimal strategy.
PROBLEM 5CRITICAL THINKING
A manager tells you: 'The expected value of our risky project is $2M, versus $1.8M for the safe project, so we should always pick the risky one.' Critically evaluate this statement. Under what conditions might the safe project be rationally preferred despite a lower EV? Discuss at least three factors that the expected value criterion alone does not capture.

Lesson Summary

A decision tree is a graphical framework that maps every possible sequence of choices and uncertain outcomes in a business problem. Decision nodes (squares) represent deliberate managerial choices, while chance nodes (circles) represent stochastic events with assigned probabilities. Each complete path through the tree terminates at a terminal payoff expressed in monetary terms.

The expected value (EV) formula, EV = Σpᵢ × Vᵢ, collapses uncertainty at each chance node into a single probability-weighted average. The rollback algorithm solves the tree from right to left—computing EVs at chance nodes and selecting the best alternative at decision nodes—until the root carries the optimal strategy's value. The expected value of perfect information (EVPI) measures the maximum worth of eliminating uncertainty entirely. While the EV criterion is a powerful rational benchmark, real-world decisions often require extensions such as expected utility theory, sensitivity analysis, and Bayesian updating to account for risk preferences and evolving information.

Varsity Tutors • Business Statistics • Decision Trees & Expected Value