Expected Value & Opportunity Loss — Compute expected value and expected opportunity loss

The notion of systematically weighing uncertain outcomes stretches back to the correspondence between Blaise Pascal and Pierre de Fermat in 1654, where the two mathematicians wrestled with fair divisions of stakes in interrupted games of chance. Their exchange gave rise to the concept of expected value — the probability-weighted average of all possible payoffs — which became the cornerstone of rational decision-making under uncertainty. Over the next three centuries, economists, statisticians, and operations researchers refined this idea, eventually formalizing it within decision analysis, a discipline that pairs expected value with complementary measures such as expected opportunity loss (EOL) to guide choices when the state of nature is unknown.

The central question decision analysis addresses is deceptively simple: given several possible actions and several possible states of nature — each with an estimated probability — which action should a rational decision-maker choose? Expected value provides one lens: pick the action with the highest (or lowest, for costs) probability-weighted payoff. Expected opportunity loss provides another, equally powerful lens: pick the action whose average regret, measured against what you could have done if you knew the future, is smallest. As we will demonstrate, these two criteria always agree on the optimal action — a remarkable and elegant equivalence.

Before diving into calculations, it is essential to establish the structural vocabulary of decision analysis. Every problem begins with a payoff table (also called a decision matrix), which enumerates every combination of decision alternatives and states of nature. Each cell of the table records the payoff (profit, revenue, or cost) that results if a particular action is taken and a particular state occurs. An opportunity loss table is then derived from the payoff table by computing, column by column, the difference between the best possible payoff in that column and the payoff of each action — this difference is the regret incurred by not having chosen optimally for that particular state.

The diagram below illustrates the complete decision-analysis pipeline: we start with a payoff table on the left, derive the opportunity loss table in the center, and compute expected values and expected opportunity losses on the right. Follow the arrows to trace how each number is generated. Note that the column maxima in the payoff table become the reference points from which all opportunity losses are measured.

In the diagram above, notice how the opportunity loss for each cell is computed by subtracting the cell's payoff from the column maximum. For alternative a₁ under state s₂, the opportunity loss is 60 − 30 = 30 because, had the decision-maker known s₂ would occur, she would have chosen a₂ (earning 60 instead of 30). The expected opportunity loss for each alternative is then the probability-weighted sum of its row in the opportunity loss table. The alternative with the smallest EOL is optimal — and it is always the same alternative that has the largest expected value.

Let there be m decision alternatives a₁, a₂, …, aₘ and n mutually exclusive, collectively exhaustive states of nature s₁, s₂, …, sₙ with prior probabilities P(s₁), P(s₂), …, P(sₙ) that sum to 1. Let Vᵢⱼ denote the payoff when alternative aᵢ is chosen and state sⱼ occurs. The following formulas constitute the mathematical backbone of expected-value decision analysis.

To solidify the procedure, consider a three-alternative, three-state-of-nature scenario. A manufacturer must decide whether to build a small plant (a₁), a medium plant (a₂), or a large plant (a₃). The states of nature are low demand (s₁), moderate demand (s₂), and high demand (s₃) with prior probabilities 0.2, 0.5, and 0.3 respectively. Payoffs are net profits in millions of dollars.

The opportunity loss table derived from the payoff table above is shown next. Each cell is obtained by subtracting the alternative's payoff from the column maximum for that state.

Verification: EV|CI = 0.2(10) + 0.5(15) + 0.3(25) = 2 + 7.5 + 7.5 = 17.0. The best EV is EV(a₂) = 0.2(−2) + 0.5(15) + 0.3(15) = −0.4 + 7.5 + 4.5 = 11.6. Wait — let us recalculate carefully: EV(a₁) = 0.2(10) + 0.5(10) + 0.3(10) = 10.0; EV(a₂) = 0.2(−2) + 0.5(15) + 0.3(15) = −0.4 + 7.5 + 4.5 = 11.6; EV(a₃) = 0.2(−10) + 0.5(8) + 0.3(25) = −2 + 4 + 7.5 = 9.5. So EV* = 11.6, and EVPI = 17.0 − 11.6 = 5.4 = EOL(a₂). The identity EVPI = EOL* holds perfectly.

An investor must allocate funds among three portfolios: Conservative (a₁), Balanced (a₂), and Aggressive (a₃). An economist forecasts three possible economic conditions: recession (s₁, probability 0.3), stability (s₂, probability 0.5), and boom (s₃, probability 0.2). Annual returns (in thousands of dollars) are given below.

The EV and EOL criteria are powerful, but they are not the only frameworks for decision-making under uncertainty. Other classical criteria — such as maximax (optimistic), maximin (conservative), and minimax regret — do not require probability estimates at all. Understanding when each criterion is appropriate depends on the information available and the decision-maker's attitude toward risk.

The expected value and expected opportunity loss framework introduced in this lesson serves as the entry point to several deeper theoretical constructs. In a Bayesian decision-analysis course, you will learn how to update prior probabilities with new data (e.g., market research, diagnostic tests) using Bayes' theorem, which leads to revised expected values and a quantity called the expected value of sample information (EVSI). In expected utility theory, the linear payoff function is replaced by a concave utility function to model risk aversion, and the decision criterion becomes maximization of expected utility rather than expected monetary value. In multi-stage settings, decision trees and influence diagrams generalize the single-stage payoff table into a sequential framework.

Despite its simplicity, the EV/EOL framework provides the conceptual scaffolding for all of these extensions. The identity EVPI = EOL*, for example, generalizes naturally: in the Bayesian framework, EVSI can be computed as the reduction in expected opportunity loss achieved by observing sample data before making the final decision. Mastering the mechanics of payoff tables, opportunity loss tables, and expected-value computations is therefore essential preparation for the more sophisticated tools you will encounter in operations research and management science.