BUSINESS ANALYTICS • PRESCRIPTIVE ANALYTICS

Decision Trees & Expected Value — Decision trees and expected value decision-making

A structured framework for choosing optimal strategies under uncertainty by quantifying outcomes and their probabilities.

Historical Context & Motivation

Every business leader faces decisions under uncertainty — whether to launch a new product, enter a foreign market, or invest in emerging technology. Before formal analytical methods existed, these choices relied on intuition, political negotiation, and experience, often producing inconsistent and suboptimal outcomes. The desire to bring rigor and repeatability to strategic choice gave rise to decision analysis, a discipline that merges probability theory, economics, and graphical reasoning to help decision-makers evaluate alternatives systematically. At the heart of decision analysis lie two powerful tools: the decision tree and the expected value criterion.

1654
Foundations of Probability
Blaise Pascal and Pierre de Fermat exchange letters on the problem of points, establishing the mathematical foundations of probability theory and the concept of expected value as a basis for rational wagering.
1738
The St. Petersburg Paradox
Daniel Bernoulli publishes his resolution of the St. Petersburg paradox, introducing diminishing marginal utility and showing that expected monetary value alone cannot always explain human preferences — a distinction that remains central to modern decision theory.
1944
Game Theory & Formal Decision Models
John von Neumann and Oskar Morgenstern publish Theory of Games and Economic Behavior, formalizing expected utility theory and providing the axiomatic framework upon which modern decision trees are built.
1968
Decision Analysis Emerges
Howard Raiffa publishes Decision Analysis: Introductory Lectures on Choices Under Uncertainty, popularizing decision trees as a practical tool for business managers and policy analysts. This work codifies the fold-back (rollback) method still taught today.
2000s
Software & Big Data Integration
Decision-tree software integrates with Monte Carlo simulation and Bayesian updating, allowing analysts to incorporate real-time data into prescriptive models. Tools like TreeAge, PrecisionTree, and open-source Python libraries bring decision analysis into everyday business analytics workflows.

The central question that decision trees and expected value address is deceptively simple: Given multiple alternatives, each leading to uncertain outcomes, which choice maximizes long-run payoff? This question arises in product launches, capital budgeting, supply chain design, marketing campaigns, and virtually every strategic domain. Understanding the historical arc — from Pascal's gambling puzzles to Raiffa's managerial frameworks — reveals why expected value reasoning remains the default quantitative benchmark against which more sophisticated methods are compared.

Core Principles & Definitions

Before constructing a decision tree, it is essential to master the vocabulary and conceptual building blocks that give the method its analytical power. A decision tree is a graphical representation of a sequential decision problem, mapping every possible action, every random event, and every resulting payoff into a single, unified structure. The tree leverages two distinct types of nodes and terminates in payoff leaves, making it both comprehensive and computationally tractable.

1

Decision Node (□)

Represented by a square, a decision node is a point where the decision-maker chooses among two or more alternatives. Each branch emanating from the square represents a distinct strategic option — for example, 'Launch Product A' versus 'Launch Product B.' The decision-maker has full control at these nodes.
2

Chance Node (○)

Represented by a circle, a chance node models randomness or uncertainty. Each branch carries a probability (all branches from one chance node must sum to 1.0). Examples include 'High demand (0.6)' or 'Low demand (0.4).' The decision-maker has no control; nature determines the path.
3

Terminal / Payoff Node (△)

Each terminal node reports the payoff — typically expressed in monetary units (profit, cost, net present value) — associated with the specific sequence of decisions and events leading to that endpoint. These values are the raw inputs for expected value calculations.
4

Expected Value (EV)

The expected value of an alternative is the probability-weighted average of all possible payoffs reachable from that alternative. Formally, EV = Σ(pᵢ × Vᵢ), where pᵢ is the probability and Vᵢ is the payoff of outcome i. The EV criterion selects the alternative with the highest (or lowest, for costs) expected value.
5

Rollback (Fold-Back) Method

The rollback method solves a decision tree from right to left. At each chance node, compute the expected value; at each decision node, choose the branch with the best EV. This backward induction propagates optimal values to the root, revealing the recommended strategy.
KEY TAKEAWAY
Think of a decision tree like a GPS navigation system for strategic choices. Just as a GPS evaluates every possible route by considering distance, traffic (uncertainty), and your destination (payoff), a decision tree maps every combination of your actions and uncertain events, then recommends the path with the best expected outcome. The rollback method is the algorithm running behind the screen, calculating from the destination backward to your current position.

Anatomy of a Decision Tree

The diagram below illustrates a classic business scenario: a firm must decide whether to invest in a large-scale expansion or a small-scale expansion. Each path leads to a chance node where market demand can be high or low, with associated probabilities and payoffs. Study the structure carefully — the square on the left is the decision node, the circles are chance nodes, and the rectangles on the right are terminal payoff nodes.

The decision tree reads left to right. The purple square (D) is the decision node. The cyan circles (C) are chance nodes. Terminal payoffs appear on the right, and the rollback expected values are summarized in the shaded boxes. Because $220,000 > $170,000, the optimal choice is the large expansion.

Notice how the tree imposes a clear chronological structure: the firm first makes its investment decision, and then the uncertain market conditions resolve. This temporal sequencing is critical because it reflects the real information structure of the problem — you must commit resources before you learn how the market will respond. Each complete path from the decision node to a terminal node represents one possible scenario, and the probabilities on the chance branches allow us to weight these scenarios when computing expected values.

Mathematical Framework

The mathematical engine behind decision trees is surprisingly compact. Two formulas suffice for solving any decision tree: one for chance nodes and one for decision nodes. These are applied iteratively through the rollback procedure, working from the terminal payoffs backward to the root. Mastering these equations allows you to solve trees of arbitrary complexity.

EXPECTED VALUE AT A CHANCE NODE
EV(Chance Node) = Σᵢ pᵢ × Vᵢ = p₁V₁ + p₂V₂ + … + pₙVₙ
Where pᵢ is the probability of outcome i (with Σpᵢ = 1), and Vᵢ is the payoff (or downstream EV) associated with outcome i. This formula computes the probability-weighted average of all branches emanating from the chance node.
OPTIMAL VALUE AT A DECISION NODE
V*(Decision Node) = max { EV(Alternative₁), EV(Alternative₂), …, EV(Alternativeₘ) }
At a decision node, the rational decision-maker selects the alternative with the highest expected value (when maximizing profit) or the lowest expected value (when minimizing cost). The non-optimal branches are pruned (marked with ‖).
EXPECTED VALUE OF PERFECT INFORMATION (EVPI)
EVPI = EV(with perfect information) − EV(without perfect information)
EVPI measures the maximum amount a decision-maker should be willing to pay for a crystal ball that perfectly predicts the uncertain event. It is calculated by computing the expected payoff if you could always choose the best alternative after observing the state of nature, then subtracting the EV of the best alternative under uncertainty.

To derive EVPI more concretely, suppose there are two states of nature — high demand (p = 0.6) and low demand (p = 0.4). With perfect information, the firm would choose the large expansion when demand is high ($500,000) and the small expansion when demand is low ($50,000). Thus, EV(with PI) = 0.6 × $500,000 + 0.4 × $50,000 = $320,000. The best EV without perfect information is $220,000 (large expansion), so EVPI = $320,000 − $220,000 = $100,000. The firm should pay no more than $100,000 for a perfect market research study that eliminates all demand uncertainty.

⚠️ Why Probabilities Must Sum to 1
A common error in constructing decision trees is assigning probabilities that do not sum to 1.0 at a chance node. Since the branches at a chance node represent a collectively exhaustive and mutually exclusive set of events, their probabilities must satisfy the axiom of total probability. If your probabilities sum to less than 1.0, you have an unmodeled scenario; if they sum to more, your model is logically inconsistent.

Types of Decision Trees & Extensions

Decision trees come in several forms, ranging from single-stage models to elaborate multi-stage structures that model sequential decisions unfolding over time. The basic framework extends naturally to accommodate additional complexity such as the value of imperfect information, risk preferences, and multi-objective payoffs. Understanding the taxonomy of decision trees helps analysts select the right level of modeling sophistication for each business problem.

Decision trees range from simple single-stage models to multi-stage sequential models, and can incorporate Bayesian updating for imperfect information. Key extensions include EVPI, EVSI, sensitivity analysis, and risk profiles.

The sensitivity analysis extension deserves special attention for business practitioners. In most real-world applications, probabilities are estimated — not known with certainty. Sensitivity analysis asks: "At what probability does my optimal decision change?" By solving for the crossover probability p*, we identify the threshold at which two alternatives yield equal expected values. If our probability estimate is far from p*, we can be confident in our decision even if our estimates are somewhat imprecise. If it is close to p*, additional market research may be warranted — and EVSI quantifies exactly how much that research is worth.

Worked Example — Product Launch Decision

Consider a technology startup deciding between three strategies for a new software product: Full Launch (high investment), Pilot Launch (moderate investment), or Abandon (no investment, $0 payoff). Market response can be Strong (probability 0.4), Moderate (probability 0.35), or Weak (probability 0.25). The payoff table is as follows: Full Launch yields $800K / $200K / −$400K; Pilot Launch yields $350K / $150K / $20K.

Solving the Product Launch Decision Tree
1
Step 1 — Structure the TreeDraw the decision node (square) at the root with three branches: Full Launch, Pilot Launch, and Abandon. Each of the first two branches leads to a chance node (circle) with three branches for Strong, Moderate, and Weak demand. The Abandon branch leads directly to a terminal node with a payoff of $0. Label every branch with its probability and every terminal node with its payoff.
2
Step 2 — Compute EV at Full Launch Chance NodeApply the expected value formula: EV(Full Launch) = 0.40 × $800,000 + 0.35 × $200,000 + 0.25 × (−$400,000). Calculating each term: $320,000 + $70,000 + (−$100,000).
EV(Full Launch) = $290,000
3
Step 3 — Compute EV at Pilot Launch Chance NodeEV(Pilot Launch) = 0.40 × $350,000 + 0.35 × $150,000 + 0.25 × $20,000. Calculating: $140,000 + $52,500 + $5,000.
EV(Pilot Launch) = $197,500
4
Step 4 — Roll Back to the Decision NodeCompare the EVs of all three alternatives at the decision node: Full Launch ($290,000), Pilot Launch ($197,500), and Abandon ($0). The decision-maker selects the alternative with the maximum EV.
Optimal Decision: Full Launch with EV = $290,000
5
Step 5 — Compute EVPIWith perfect information, the firm would choose Full Launch when demand is Strong ($800,000), Full Launch when Moderate ($200,000), and Pilot Launch when Weak ($20,000 > −$400,000 > $0). Thus EV(PI) = 0.40 × $800,000 + 0.35 × $200,000 + 0.25 × $20,000 = $320,000 + $70,000 + $5,000 = $395,000. EVPI = $395,000 − $290,000.
EVPI = $105,000 — the maximum the firm should pay for perfect market intelligence.

Strengths, Limitations & Comparisons

Decision trees with expected value analysis are among the most widely used prescriptive analytics tools, yet they are not without limitations. Understanding both the power and the boundaries of the method allows analysts to deploy it appropriately and to recognize when more sophisticated techniques — such as utility theory, Monte Carlo simulation, or real options analysis — might be warranted.

Strengths and limitations of decision tree / expected value analysis
DimensionStrengthsLimitations
TransparencyVisual structure makes assumptions explicit; stakeholders can trace every path and verify logic without specialized training.Large trees with many stages and states become visually cluttered and difficult to interpret — the 'bushy tree' problem.
Sequential LogicNaturally captures decisions that unfold over time, including the ability to adapt strategy based on intermediate outcomes.Assumes a fixed sequence of decisions and events; cannot easily model continuous processes or real-time adaptation.
Quantitative RigorExpected value provides a single, defensible metric for ranking alternatives, grounded in probability axioms.EV is risk-neutral: it ignores the decision-maker's risk aversion. A −$10M loss and a +$10M gain are treated symmetrically.
Data RequirementsRequires only payoff estimates and probability assessments, which can come from data, expert judgment, or both.Probabilities are often subjective; garbage-in, garbage-out applies. Sensitivity analysis is essential but not always performed.
ScalabilitySoftware tools handle thousands of nodes computationally; rollback is O(n) in the number of nodes.Number of paths grows exponentially with the number of stages and states, making exhaustive enumeration impractical for very complex problems.
CONTEXT IN PRESCRIPTIVE ANALYTICS
Decision trees occupy a sweet spot in the prescriptive analytics toolkit: they are more structured than informal pro/con lists, yet more interpretable than mathematical programming or simulation models. Think of the EV criterion as a first approximation — much like how accountants use straight-line depreciation as a useful default before exploring more nuanced methods. When risk aversion matters significantly, extend the framework to expected utility; when probability distributions are complex, layer in Monte Carlo simulation.

Connection to Advanced Decision Theory

The expected value framework you have studied in this lesson is the foundation upon which more advanced prescriptive methods are built. As you progress in business analytics, you will encounter techniques that relax the assumptions of the basic model — particularly the assumption of risk neutrality and the requirement for discrete, well-defined outcomes. The table below maps each limitation of the basic EV approach to the advanced method that addresses it.

Evolution from basic EV to advanced decision methods
Basic EV ApproachLimitation AddressedAdvanced Extension
EV = Σ pᵢVᵢ (risk-neutral)Ignores risk aversion; treats $1M gain same as $1M lossExpected Utility Theory — replace Vᵢ with U(Vᵢ) using a concave utility function
Discrete outcomes onlyPayoffs and probabilities are continuous in realityMonte Carlo Simulation — sample thousands of scenarios from continuous distributions and compute the mean and risk metrics
Single probability estimateUncertainty about the probabilities themselvesBayesian Decision Analysis — update prior probabilities with data using Bayes' theorem; computes EVSI
Commit-or-abandon choicesIgnores the value of waiting for new informationReal Options Analysis — values managerial flexibility (defer, expand, contract, abandon) using option pricing theory
Single-criterion payoff (profit)Many decisions involve multiple objectives (profit, sustainability, brand equity)Multi-Attribute Utility Theory (MAUT) — defines utility functions over multiple attributes and computes a weighted EV across dimensions

Despite these more sophisticated alternatives, the basic decision tree with expected value remains the starting point for nearly every real-world decision analysis engagement. Major consulting firms, pharmaceutical companies evaluating drug pipelines, and energy companies planning exploration portfolios all begin with a simple decision tree and only layer in complexity where the problem demands it. Mastering expected value reasoning therefore gives you a versatile tool that scales from classroom exercises to boardroom strategy.

Practice Problems

PROBLEM 1CONCEPTUAL
Explain the fundamental difference between a decision node and a chance node in a decision tree. Why is this distinction important for correctly solving the tree using the rollback method?
PROBLEM 2BASIC CALCULATION
A retailer must decide whether to order 500 units or 1,000 units of a seasonal product. If demand is High (p = 0.55), profits are $30,000 for the 500-unit order and $70,000 for the 1,000-unit order. If demand is Low (p = 0.45), profits are $20,000 for 500 units and −$10,000 for 1,000 units. Compute the expected value for each alternative and identify the optimal order quantity.
PROBLEM 3INTERMEDIATE
Using the retailer problem from Problem 2, compute the Expected Value of Perfect Information (EVPI). Interpret the result in a sentence that a store manager could understand.
PROBLEM 4APPLIED
A pharmaceutical company is evaluating a drug development project with two sequential decision points. At Stage 1, the company can invest $5M in a clinical trial or abandon (payoff $0). If the trial succeeds (p = 0.3), the company reaches Stage 2 where it can commercialize (costing an additional $15M) or license the drug (receiving a flat $10M). If commercialized and the market is favorable (p = 0.5), revenue is $80M; if unfavorable (p = 0.5), revenue is $20M. If the Stage 1 trial fails (p = 0.7), the entire $5M is lost. Construct the decision tree, solve it via rollback, and state the optimal strategy.
PROBLEM 5CRITICAL THINKING
A startup faces an investment decision with two alternatives. Alternative A has a 90% chance of yielding $100,000 and a 10% chance of yielding −$800,000 (EV = $10,000). Alternative B yields a certain $0. The expected value criterion recommends Alternative A, yet many rational managers would choose Alternative B. Explain this apparent contradiction, discuss at least two behavioral or theoretical reasons why EV may fail here, and suggest an analytical modification that would capture the manager's preference.

Lesson Summary

Decision trees provide a visual, structured framework for analyzing choices under uncertainty by mapping decision nodes (where the manager chooses), chance nodes (where nature acts), and terminal payoffs into a single diagram. The expected value (EV) of an alternative is computed as the probability-weighted sum of all reachable payoffs (EV = Σ pᵢ × Vᵢ), and the rollback method solves the tree from right to left — computing EVs at chance nodes and selecting the optimal branch at decision nodes — to identify the strategy that maximizes long-run payoff.

Key extensions include the Expected Value of Perfect Information (EVPI), which caps what a decision-maker should pay for eliminating uncertainty, and sensitivity analysis, which identifies the crossover probability at which the optimal decision changes. While the EV criterion assumes risk neutrality and requires probabilities that sum to 1.0, it serves as the analytical foundation for advanced methods including expected utility theory, Monte Carlo simulation, and real options analysis. Mastering decision trees equips business analysts with a versatile, transparent prescriptive tool that scales from simple go/no-go choices to complex multi-stage investment strategies.

Varsity Tutors • Business Analytics • Decision Trees & Expected Value