Expected Value and Expected Utility: How to Compare Decisions Under Risk and Uncertainty

Last Updated June 5, 2026

Expected value and expected utility are foundational concepts in decision science because they provide formal ways to evaluate choices under uncertainty. Expected value calculates the probability-weighted average of outcomes, while expected utility transforms outcomes through a utility function so that risk attitude, preference, and the subjective value of consequences can be represented explicitly.

Expected Value and Expected Utility examines how these concepts structure reasoning under risk, why they became central to decision theory, how they differ from ordinary intuition, and why their practical use requires care. Expected value is mathematically clear, but it assumes that outcomes can be compared linearly. Expected utility is more flexible because it accounts for diminishing marginal value and risk preference, but it depends on assumptions about probabilities, utilities, preference stability, and the decision-maker’s objectives.

These concepts form part of the analytical core of decision theory and remain central to decision science more broadly. They provide a structured way to compare alternatives when outcomes are uncertain, enabling decision-makers to move beyond intuition toward explicit evaluation. At the same time, they are not automatic answer machines. Their real value lies in making uncertainty, consequence, preference, and trade-off assumptions visible enough to inspect.

Why Expected Value and Expected Utility Matter

Expected value and expected utility matter because decisions often require action before outcomes are known. A decision-maker may face multiple options, each with different possible results and different probabilities. Intuition may focus on the best case, the worst case, the most vivid case, or the most emotionally salient case. Expected value and expected utility discipline that process by forcing the decision-maker to represent outcomes, probabilities, and valuation explicitly.

Expected value answers a relatively simple question: what is the probability-weighted average outcome of this option? Expected utility asks a more subtle question: how valuable is the probability-weighted structure of outcomes to a particular decision-maker, given preferences and risk attitude? The difference matters because a dollar, life-year, cost saving, policy benefit, reputational gain, or operational loss may not have the same practical value in every context.

For a risk-neutral decision-maker with repeated exposure and reliable probabilities, expected value can be an appropriate benchmark. For a decision-maker facing large downside exposure, irreversible consequences, limited wealth, public responsibility, ethical constraints, or unequal effects across stakeholders, expected value alone may be inadequate. Expected utility helps explain why the same gamble can be attractive to one actor and unacceptable to another.

Both concepts are foundational, but neither should be used mechanically. Their value depends on whether the probabilities are credible, outcomes are well specified, preferences are meaningful, and the decision context supports the assumptions being made.

Back to top ↑

Expected Value: A Probabilistic Foundation

Expected value is the simplest formal method for evaluating uncertain outcomes. It represents the weighted average of possible outcomes, where each outcome is multiplied by its probability of occurrence. In mathematical terms, it converts a probability distribution over outcomes into a single summary quantity.

The appeal of expected value lies in its clarity. If probabilities are known and outcomes can be expressed in a common unit, the option with the highest expected value can be identified as the best choice for a risk-neutral decision-maker. This makes expected value especially useful in gambling, finance, insurance, engineering reliability, operations research, portfolio management, and other domains where repeated exposure or aggregation makes average outcomes meaningful.

For example, an option with a 50 percent chance of gaining 200 and a 50 percent chance of gaining 0 has an expected value of 100. A guaranteed gain of 90 has a lower expected value. A risk-neutral decision-maker who can repeat the gamble many times may prefer the risky option. But a decision-maker who cannot tolerate the downside, who faces irreversible consequences, or who values certainty may prefer the guaranteed outcome.

Expected value is powerful because it makes probability-weighted consequence explicit. It is limited because it treats outcomes only in terms of their numerical magnitude. It does not ask whether a gain of 100 matters the same way to every decision-maker, whether a large loss threatens survival, or whether downside risk is ethically or institutionally acceptable.

That limitation is one reason expected utility became central to decision theory and decision analysis.

Back to top ↑

Expected Utility: Incorporating Preference and Risk

Expected utility extends expected value by introducing the idea that outcomes are not evaluated purely as objective quantities. Instead, outcomes are transformed through a utility function that reflects the decision-maker’s valuation of those outcomes. This allows decision analysis to represent risk attitude, diminishing marginal value, preference intensity, and context-dependent valuation.

The conceptual move is simple but profound. Instead of multiplying probabilities by raw outcomes, expected utility multiplies probabilities by utility-transformed outcomes. If utility is linear, expected utility and expected value produce similar rankings. If utility is concave, large gains add less utility than proportional small gains, and losses or low outcomes may weigh heavily. If utility is convex over a relevant range, the decision-maker may prefer riskier prospects.

Expected utility is historically associated with Daniel Bernoulli’s response to the St. Petersburg paradox and later formal developments in utility theory. Bernoulli argued that the value of money should not be treated as purely linear. The usefulness of an additional amount of money depends on the wealth and condition of the person receiving it. This insight made it possible to explain why real decision-makers may reject gambles with high expected monetary value.

Expected utility therefore allows decision-makers to model different attitudes toward risk:

• Risk-averse: preferring more certain outcomes even when expected monetary value is lower.
• Risk-neutral: focusing on probability-weighted average outcome.
• Risk-seeking: preferring higher-variance outcomes with potentially larger gains.

This makes expected utility more behaviorally and normatively flexible than expected value alone. It also helps explain why a decision that looks attractive on average may be unacceptable in practice.

Back to top ↑

Expected Value vs. Expected Utility

Expected value and expected utility are closely related, but they answer different questions. Expected value evaluates uncertain prospects by average outcome. Expected utility evaluates uncertain prospects by average utility. The distinction becomes important whenever outcomes do not have constant marginal value.

Suppose two options have the same expected value. One is a guaranteed outcome. The other is a high-risk gamble with a chance of a large gain and a chance of a severe loss. Expected value may treat them as equivalent. Expected utility may not. For a risk-averse decision-maker, the guaranteed option may have higher utility because the loss state is too damaging. For a risk-seeking actor, the gamble may be preferred because the large gain carries special value.

This distinction is not merely psychological. Institutions also have utility-like preferences. A hospital may value avoiding catastrophic patient harm more than maximizing average operational efficiency. A city may value service continuity and legitimacy more than narrow cost savings. A financial institution may avoid trades with attractive expected returns if they create unacceptable tail risk. A public agency may reject a policy with high average benefit if it imposes severe burdens on vulnerable groups.

The practical lesson is not that expected utility replaces expected value in every decision. Rather, expected value should often be treated as a baseline calculation, while expected utility asks whether that baseline is adequate for the decision context.

Back to top ↑

Utility Functions and Decision Behavior

A utility function translates outcomes into subjective or institutional value. Its shape determines how outcomes are evaluated under uncertainty. A linear utility function implies risk neutrality. A concave utility function implies diminishing marginal utility and risk aversion. A convex utility function implies risk-seeking behavior over the relevant range.

Utility functions can represent money, health, safety, time, reliability, reputation, service continuity, environmental quality, or any other consequence that can be ordered and valued. In formal decision theory, utility functions are often derived from preference axioms. In applied decision analysis, they may be elicited through expert judgment, stakeholder preference, institutional risk tolerance, or structured decision processes.

In practice, utility functions are powerful but difficult. Preferences may vary across domains, time horizons, wealth levels, moral contexts, and stakeholder groups. A person may be risk-averse about health, risk-neutral about small routine financial decisions, and risk-seeking in entrepreneurship. An institution may tolerate operational variance but be extremely risk-averse about legal liability, safety, or public trust.

Utility functions should therefore be treated as explicit modeling assumptions. A utility function is not a fact about the world in the same way a measured frequency may be. It is a representation of valuation. Its use should be documented, justified, and tested through sensitivity analysis.

Back to top ↑

Risk Aversion, Risk Neutrality, and Risk Seeking

Risk attitude describes how a decision-maker evaluates uncertain outcomes relative to certain ones. A risk-neutral decision-maker cares only about expected value. A risk-averse decision-maker prefers certainty or downside protection. A risk-seeking decision-maker may prefer a gamble even when a safer option has similar or higher expected value.

Risk attitude is not necessarily irrational. Risk aversion can be rational when losses threaten survival, public legitimacy, basic services, safety, or institutional continuity. Risk seeking can also be understandable in contexts where a decision-maker faces limited downside, needs breakthrough gains, or operates under competitive pressure. The key question is whether the risk attitude is appropriate to the decision context and explicitly represented.

In expected utility theory, risk attitude is represented by the curvature of the utility function. A concave utility function means each additional unit of outcome adds less utility than the previous unit. This makes risky prospects less attractive when they include low-outcome states. A convex utility function means marginal utility increases over the relevant range, making high-variance prospects more attractive.

Decision science does not assume one risk attitude is always correct. It asks whether risk attitude is coherent, defensible, transparent, and aligned with consequences.

Back to top ↑

The St. Petersburg Paradox and Bernoulli’s Intervention

The St. Petersburg paradox occupies a central place in the history of expected utility because it exposed the inadequacy of expected monetary value as a universal rule of rational choice. The classical gamble has an unbounded expected monetary value, yet ordinary decision-makers are not willing to pay an arbitrarily large amount to enter it.

The paradox arises because the monetary payoff can become extremely large with very small probability. Mathematically, the expected monetary value diverges. Behaviorally and practically, people do not treat the gamble as infinitely valuable. The gap between formal expectation and actual valuation revealed that expected monetary value was missing something important.

Bernoulli’s 1738 intervention proposed that the value of money should be understood through utility rather than nominal magnitude. The additional value of money declines as wealth increases. This idea made it possible to explain why a person might reject a gamble with enormous expected monetary value: the rare high payoff does not compensate for the structure of risk when evaluated through diminishing marginal utility.

The enduring lesson is broader than the paradox itself. Rational evaluation under uncertainty cannot always be reduced to arithmetic expectation. It must also consider how outcomes are valued, how risk is borne, and what downside or extreme outcomes mean in context.

Back to top ↑

Probability Quality and Model Assumptions

Expected value and expected utility both depend on probabilities. When probabilities are reliable, these methods can provide strong structure. When probabilities are weak, contested, unstable, or invented, the calculations can create false precision. Decision science must therefore distinguish between probability use and probability quality.

Some probabilities come from stable frequencies, controlled experiments, actuarial data, engineering reliability records, or well-calibrated forecasts. Others come from expert judgment, limited historical analogies, sparse data, or contested models. The same formula may be used in both cases, but the quality of the decision support differs greatly.

Probability quality matters especially in rare events, strategic decisions, climate futures, public health, geopolitical risk, AI system performance, infrastructure planning, and systemic financial risk. In these domains, probabilities may be uncertain not only because data are limited, but because the future system may differ from the past.

A high-quality expected value or expected utility analysis should document where probabilities came from, how uncertain they are, what assumptions support them, and how the decision changes if they are wrong.

Back to top ↑

Wealth, Baselines, and Reference Context

Expected utility depends on context. The same monetary gain or loss can have different utility depending on wealth, institutional capacity, baseline vulnerability, and reference point. This is one of the central insights behind diminishing marginal utility: outcomes do not have the same practical meaning for every decision-maker.

For an individual with limited resources, a loss may threaten basic stability. For a large institution, the same loss may be tolerable. For a public agency, a small average benefit may not justify severe distributional harm. For a hospital, a low-probability severe safety event may dominate a small expected efficiency gain. For a climate adaptation decision, the relevant baseline may include future risk, not only current cost.

Reference context also matters. A decision-maker may evaluate outcomes relative to current wealth, expected baseline, aspiration level, loss threshold, minimum service obligation, legal requirement, or ethical constraint. This affects how utility is represented.

This is why expected utility analysis should not treat the utility function as a decorative addition to expected value. The utility function encodes the decision-maker’s relationship to outcomes. That relationship may include wealth, risk capacity, mission, ethics, stakeholder obligations, and institutional resilience.

Back to top ↑

Limits of Expected Value and Expected Utility

Expected value and expected utility are powerful, but both have important limitations. First, they require outcome specification. If the relevant consequences are poorly defined, omitted, or measured in incompatible ways, the calculation may be misleading. A decision cannot be saved by a precise formula if the outcome frame is wrong.

Second, both methods require probabilities or probability-like beliefs. In many real-world situations, probabilities are uncertain, disputed, unstable, or impossible to estimate with confidence. When the probability model itself is fragile, expected value and expected utility become harder to apply cleanly.

Third, expected utility assumes that preferences can be represented coherently enough to support formal analysis. Yet preferences can be context-dependent, unstable, socially shaped, ethically constrained, and influenced by framing. In organizations, there may be no single unified preference function. Multiple stakeholders may value outcomes differently.

Fourth, these models often evaluate outcomes in a relatively contained way. In complex systems, outcomes may depend on feedback loops, adaptation, spillovers, delays, and path dependence. The decision problem may not be adequately represented by a fixed menu of states and payoffs alone.

The conclusion is not that expected value and expected utility should be abandoned. It is that they should be embedded in a broader decision architecture that includes uncertainty quality, sensitivity, values, systems awareness, and accountability.

Back to top ↑

From Expected Utility to Modern Decision Science

Modern decision science builds on expected value and expected utility rather than abandoning them. These concepts still provide the baseline grammar of formal choice under uncertainty. They clarify outcomes, probabilities, preferences, and risk. But contemporary decision science often extends them through sensitivity analysis, Bayesian updating, multi-criteria decision analysis, scenario comparison, robust decision-making, behavioral insight, and decision governance.

Instead of relying solely on a single expected value or expected utility calculation, decision-makers increasingly ask how the recommendation changes across assumptions, models, futures, and stakeholder values. This shift is especially important in environments characterized by deep uncertainty, contested probabilities, fragile models, and system complexity.

Expected utility remains foundational, but it no longer stands alone. It is one element in a wider architecture of judgment. That architecture should include decision framing, evidence quality, uncertainty representation, trade-off transparency, behavioral safeguards, systems awareness, decision records, and review triggers.

Expected value and expected utility therefore remain essential, but their best use is disciplined rather than mechanical. They clarify how uncertainty and valuation enter the decision. They do not remove the need for judgment.

Back to top ↑

Applications in Decision-Making

Expected value and expected utility remain central in many applied domains. They are especially useful where alternatives have uncertain consequences, probabilities can be estimated, and outcomes can be compared in a structured way.

In finance, expected value supports portfolio analysis, pricing, risk-return evaluation, and insurance logic. Expected utility helps explain why investors may prefer lower-variance portfolios, why downside protection matters, and why risk tolerance varies across wealth levels and institutional mandates. In healthcare, expected utility can incorporate patient values, quality-adjusted outcomes, treatment risks, and diagnostic uncertainty. In engineering, expected value supports reliability and cost-risk analysis, while expected utility can represent safety-critical risk aversion.

In public policy, expected value and expected utility can inform cost-benefit analysis, resource allocation, intervention design, and risk management. But public decisions often involve distributional consequences, legitimacy, rights, and intergenerational effects that cannot be reduced to a single expected monetary value without ethical loss. Expected utility may help, but multi-criteria and deliberative frameworks are often needed.

In each domain, expected value and expected utility help structure uncertainty into comparable terms. Their usefulness depends on whether they are embedded in broader frameworks that account for evidence quality, model assumptions, stakeholder values, constraints, and review.

Back to top ↑

Behavioral Critiques and Descriptive Departures

One of the major developments in modern decision science is the recognition that expected utility does not always describe how people actually choose. Behavioral research has shown systematic deviations from expected-utility predictions, especially under framing, loss aversion, reference dependence, probability distortion, availability, anchoring, and overconfidence.

This does not make expected utility obsolete. It clarifies its role. Expected utility remains a powerful normative and analytical tool. It tells us how choices can be represented under coherent preference assumptions. Behavioral decision theory helps explain how actual human judgment often departs from those assumptions.

The contrast is productive. Expected utility provides a formal benchmark. Behavioral research reveals where real judgment is vulnerable. Together, they help decision science move from abstract rationality toward practical decision improvement.

Expected utility is therefore best understood as one layer of decision science. It clarifies the normative structure of choice under uncertainty. Behavioral research clarifies how real people and institutions often struggle to meet that structure.

Back to top ↑

Implications for Decision Science

Expected value and expected utility have several enduring implications for decision science. First, they show that uncertainty can be formalized. Outcomes under risk can be compared systematically rather than impressionistically. Second, they show that preferences matter. Value is not exhausted by raw monetary or numerical magnitude. Third, they show that risk attitude matters. The same gamble may be rationally treated differently under different utility structures.

They also show that formal clarity has limits. A clean formula can conceal weak probabilities, incomplete outcomes, unstable preferences, hidden values, or unmodeled system effects. Decision science therefore uses expected value and expected utility as analytical foundations, but not as the entire decision process.

The deeper implication is that structured judgment requires both calculation and interpretation. A decision-maker must ask: what outcomes are included, whose utility matters, how probabilities were estimated, what assumptions are fragile, what risks are unacceptable, and what trade-offs are being made?

• Uncertainty can be formalized: probability-weighted reasoning improves clarity.
• Preferences matter: raw outcomes do not exhaust value.
• Risk attitude matters: the same prospect can be evaluated differently by different decision-makers.
• Probability quality matters: formulas are only as credible as their assumptions.
• Formal clarity has limits: EV and EU must be connected to evidence, values, systems, and accountability.

These implications explain why expected value and expected utility remain indispensable. They are not the whole of decision science, but they are among its deepest foundations.

Back to top ↑

Summary Table: Expected Value, Expected Utility, and Decision Quality

The table below summarizes how expected value and expected utility fit into a broader decision-science framework.

The summary emphasizes a practical lesson: expected value and expected utility are not just formulas. They are structured ways of making assumptions about uncertainty and value visible.

Back to top ↑

Back to top ↑

Mathematical Lens: Expected Value, Expected Utility, Risk Attitude, and Certainty Equivalents

The mathematical lens clarifies the formal relationship between expected value, expected utility, risk attitude, and certainty equivalents.

The expected value of an uncertain prospect with outcomes \(x_1, x_2, \dots, x_n\) and probabilities \(p_1, p_2, \dots, p_n\) is:

Expected utility replaces raw outcomes with a utility transformation:

A common concave utility function is logarithmic utility:

A more general constant-relative-risk-aversion utility function is:

The certainty equivalent is the guaranteed amount that gives the same utility as a risky prospect:

The risk premium is the amount a decision-maker would give up to avoid the risk:

Decision comparison can then be extended beyond expected value alone:

The mathematical lens shows why expected utility is not merely expected value with extra notation. It changes the object of evaluation from outcome magnitude to valued consequence under uncertainty.

Back to top ↑

R Workflow: Expected Value, Utility Curves, Risk Premiums, and Sensitivity Analysis

The R workflow below compares stylized uncertain prospects under expected value, log utility, CRRA utility, certainty equivalents, and risk premiums. It also performs sensitivity analysis across risk-aversion values and exports reproducible decision summaries.

# expected_value_expected_utility_workflow.R
# Base R workflow for comparing uncertain prospects using:
# expected value, expected utility, utility curvature,
# certainty equivalents, risk premiums, and risk-aversion sensitivity.

args <- commandArgs(trailingOnly = FALSE)
file_arg <- grep("^--file=", args, value = TRUE)

if (length(file_arg) > 0) {
  script_path <- normalizePath(sub("^--file=", "", file_arg[1]), mustWork = TRUE)
  article_root <- normalizePath(file.path(dirname(script_path), ".."), mustWork = TRUE)
} else {
  article_root <- getwd()
}

setwd(article_root)

tables_dir <- file.path(article_root, "outputs", "tables")
figures_dir <- file.path(article_root, "outputs", "figures")

dir.create(tables_dir, recursive = TRUE, showWarnings = FALSE)
dir.create(figures_dir, recursive = TRUE, showWarnings = FALSE)

prospects <- data.frame(
  prospect = c(
    "Safe Option",
    "Balanced Gamble",
    "High-Risk Gamble",
    "Catastrophic Downside",
    "Resilient Moderate Upside"
  ),
  outcome_1 = c(100, 180, 400, 260, 150),
  probability_1 = c(1.00, 0.60, 0.25, 0.45, 0.70),
  outcome_2 = c(100, 40, 0, -120, 70),
  probability_2 = c(0.00, 0.40, 0.75, 0.55, 0.30),
  stringsAsFactors = FALSE
)

validate_probabilities <- function(p1, p2) {
  if (any(abs((p1 + p2) - 1) > 1e-8)) {
    stop("Each prospect's probabilities must sum to 1.")
  }
}

validate_probabilities(prospects$probability_1, prospects$probability_2)

shift_outcomes <- function(x, floor_value = -150, offset = 151) {
  x + offset
}

log_utility <- function(x) {
  log(shift_outcomes(x))
}

crra_utility <- function(x, rho, offset = 151) {
  z <- x + offset
  if (any(z <= 0)) stop("Shifted outcomes must be positive.")
  if (abs(rho - 1) < 1e-8) {
    return(log(z))
  }
  (z^(1 - rho) - 1) / (1 - rho)
}

inverse_crra <- function(u, rho, offset = 151) {
  if (abs(rho - 1) < 1e-8) {
    return(exp(u) - offset)
  }
  ((u * (1 - rho) + 1)^(1 / (1 - rho))) - offset
}

prospects$expected_value <- (
  prospects$outcome_1 * prospects$probability_1 +
  prospects$outcome_2 * prospects$probability_2
)

prospects$log_expected_utility <- (
  prospects$probability_1 * log_utility(prospects$outcome_1) +
  prospects$probability_2 * log_utility(prospects$outcome_2)
)

prospects$log_certainty_equivalent <- exp(prospects$log_expected_utility) - 151
prospects$log_risk_premium <- prospects$expected_value - prospects$log_certainty_equivalent

risk_levels <- c(0.25, 0.75, 1.00, 1.50, 2.25)

sensitivity_rows <- data.frame()

for (rho in risk_levels) {
  expected_utility <- (
    prospects$probability_1 * crra_utility(prospects$outcome_1, rho) +
    prospects$probability_2 * crra_utility(prospects$outcome_2, rho)
  )

  certainty_equivalent <- inverse_crra(expected_utility, rho)
  risk_premium <- prospects$expected_value - certainty_equivalent

  temp <- data.frame(
    risk_aversion = rho,
    prospect = prospects$prospect,
    expected_value = prospects$expected_value,
    expected_utility = expected_utility,
    certainty_equivalent = certainty_equivalent,
    risk_premium = risk_premium,
    stringsAsFactors = FALSE
  )

  temp$ev_rank <- rank(-temp$expected_value, ties.method = "min")
  temp$eu_rank <- rank(-temp$expected_utility, ties.method = "min")
  temp$ce_rank <- rank(-temp$certainty_equivalent, ties.method = "min")

  sensitivity_rows <- rbind(sensitivity_rows, temp)
}

ranking_instability <- aggregate(
  eu_rank ~ prospect,
  data = sensitivity_rows,
  FUN = function(x) max(x) - min(x)
)

names(ranking_instability) <- c("prospect", "expected_utility_rank_range")

summary_table <- merge(prospects, ranking_instability, by = "prospect")
summary_table <- summary_table[order(-summary_table$expected_value), ]

write.csv(prospects, file.path(tables_dir, "expected_value_expected_utility_profiles.csv"), row.names = FALSE)
write.csv(sensitivity_rows, file.path(tables_dir, "risk_aversion_sensitivity.csv"), row.names = FALSE)
write.csv(summary_table, file.path(tables_dir, "expected_utility_summary_table.csv"), row.names = FALSE)

png(file.path(figures_dir, "expected_value_by_prospect.png"), width = 1200, height = 800)
barplot(
  prospects$expected_value,
  names.arg = prospects$prospect,
  las = 2,
  main = "Expected Value by Prospect",
  ylab = "Expected value"
)
grid()
dev.off()

png(file.path(figures_dir, "risk_premium_log_utility.png"), width = 1200, height = 800)
barplot(
  prospects$log_risk_premium,
  names.arg = prospects$prospect,
  las = 2,
  main = "Risk Premium Under Log Utility",
  ylab = "Risk premium"
)
grid()
dev.off()

png(file.path(figures_dir, "certainty_equivalent_by_risk_aversion.png"), width = 1200, height = 800)
plot(
  sensitivity_rows$risk_aversion,
  sensitivity_rows$certainty_equivalent,
  type = "n",
  xlab = "Risk aversion",
  ylab = "Certainty equivalent",
  main = "Certainty Equivalent Across Risk Aversion"
)

for (prospect_name in unique(sensitivity_rows$prospect)) {
  subset_rows <- sensitivity_rows[sensitivity_rows$prospect == prospect_name, ]
  lines(subset_rows$risk_aversion, subset_rows$certainty_equivalent, type = "b")
}

legend(
  "topright",
  legend = unique(sensitivity_rows$prospect),
  bty = "n",
  cex = 0.8
)
grid()
dev.off()

print(summary_table)
print(sensitivity_rows)

This R workflow shows how a ranking can change when risk attitude changes. The option with the highest expected value is not necessarily the option with the highest expected utility, certainty equivalent, or acceptable risk profile.

Back to top ↑

Python Workflow: Simulating Expected Utility, Risk Aversion, and Choice Reversal

The Python workflow below simulates uncertain prospects across different risk-aversion levels. It calculates expected value, expected utility, certainty equivalent, risk premium, rank changes, and simulated choice reversal. It uses only the Python standard library.

# expected_value_expected_utility_simulation.py
# Standard-library workflow for expected value, expected utility,
# certainty equivalents, risk premiums, and risk-aversion sensitivity.

from __future__ import annotations

from dataclasses import dataclass
from pathlib import Path
import csv
import json
import math
from statistics import mean

ARTICLE_ROOT = Path(__file__).resolve().parents[1]
TABLES = ARTICLE_ROOT / "outputs" / "tables"
RECORDS = ARTICLE_ROOT / "outputs" / "decision_records"


@dataclass(frozen=True)
class Prospect:
    name: str
    outcomes: tuple[float, ...]
    probabilities: tuple[float, ...]


def validate_probabilities(prospect: Prospect) -> None:
    total = sum(prospect.probabilities)
    if not math.isclose(total, 1.0, abs_tol=1e-9):
        raise ValueError(f"Probabilities for {prospect.name} sum to {total}, not 1.")


def expected_value(prospect: Prospect) -> float:
    validate_probabilities(prospect)
    return sum(outcome * probability for outcome, probability in zip(prospect.outcomes, prospect.probabilities))


def shifted_outcome(x: float, offset: float = 151.0) -> float:
    z = x + offset
    if z <= 0:
        raise ValueError("Shifted outcome must be positive for utility calculation.")
    return z


def crra_utility(x: float, rho: float, offset: float = 151.0) -> float:
    z = shifted_outcome(x, offset)
    if math.isclose(rho, 1.0, abs_tol=1e-9):
        return math.log(z)
    return (z ** (1.0 - rho) - 1.0) / (1.0 - rho)


def inverse_crra(u: float, rho: float, offset: float = 151.0) -> float:
    if math.isclose(rho, 1.0, abs_tol=1e-9):
        return math.exp(u) - offset
    return ((u * (1.0 - rho) + 1.0) ** (1.0 / (1.0 - rho))) - offset


def expected_utility(prospect: Prospect, rho: float) -> float:
    validate_probabilities(prospect)
    return sum(
        probability * crra_utility(outcome, rho)
        for outcome, probability in zip(prospect.outcomes, prospect.probabilities)
    )


def certainty_equivalent(prospect: Prospect, rho: float) -> float:
    return inverse_crra(expected_utility(prospect, rho), rho)


def risk_premium(prospect: Prospect, rho: float) -> float:
    return expected_value(prospect) - certainty_equivalent(prospect, rho)


def rank_descending(rows: list[dict[str, object]], field: str, rank_field: str) -> None:
    sorted_rows = sorted(rows, key=lambda row: float(row[field]), reverse=True)
    for rank, row in enumerate(sorted_rows, start=1):
        row[rank_field] = rank


def write_csv(path: Path, rows: list[dict[str, object]]) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    if not rows:
        raise ValueError(f"No rows to write: {path}")
    with path.open("w", encoding="utf-8", newline="") as handle:
        writer = csv.DictWriter(handle, fieldnames=list(rows[0].keys()))
        writer.writeheader()
        writer.writerows(rows)


def write_json(path: Path, payload: dict[str, object]) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    path.write_text(json.dumps(payload, indent=2), encoding="utf-8")


def main() -> None:
    prospects = [
        Prospect("Safe Option", (100.0,), (1.0,)),
        Prospect("Balanced Gamble", (180.0, 40.0), (0.60, 0.40)),
        Prospect("High-Risk Gamble", (400.0, 0.0), (0.25, 0.75)),
        Prospect("Catastrophic Downside", (260.0, -120.0), (0.45, 0.55)),
        Prospect("Resilient Moderate Upside", (150.0, 70.0), (0.70, 0.30)),
    ]

    risk_levels = [0.25, 0.75, 1.0, 1.50, 2.25]
    rows: list[dict[str, object]] = []

    for rho in risk_levels:
        batch: list[dict[str, object]] = []

        for prospect in prospects:
            ev = expected_value(prospect)
            eu = expected_utility(prospect, rho)
            ce = certainty_equivalent(prospect, rho)
            rp = risk_premium(prospect, rho)

            batch.append({
                "risk_aversion": rho,
                "prospect": prospect.name,
                "expected_value": round(ev, 4),
                "expected_utility": round(eu, 6),
                "certainty_equivalent": round(ce, 4),
                "risk_premium": round(rp, 4),
            })

        rank_descending(batch, "expected_value", "expected_value_rank")
        rank_descending(batch, "expected_utility", "expected_utility_rank")
        rank_descending(batch, "certainty_equivalent", "certainty_equivalent_rank")
        rows.extend(batch)

    summary_rows: list[dict[str, object]] = []

    for prospect in prospects:
        prospect_rows = [row for row in rows if row["prospect"] == prospect.name]
        eu_ranks = [int(row["expected_utility_rank"]) for row in prospect_rows]
        risk_premiums = [float(row["risk_premium"]) for row in prospect_rows]

        summary_rows.append({
            "prospect": prospect.name,
            "expected_value": round(expected_value(prospect), 4),
            "best_expected_utility_rank": min(eu_ranks),
            "worst_expected_utility_rank": max(eu_ranks),
            "rank_range": max(eu_ranks) - min(eu_ranks),
            "average_risk_premium": round(mean(risk_premiums), 4),
        })

    summary_rows = sorted(summary_rows, key=lambda row: float(row["expected_value"]), reverse=True)

    write_csv(TABLES / "expected_utility_sensitivity.csv", rows)
    write_csv(TABLES / "expected_utility_summary.csv", summary_rows)

    write_json(
        RECORDS / "expected_value_expected_utility_decision_record.json",
        {
            "article": "Expected Value and Expected Utility",
            "decision_context": "Comparison of uncertain prospects under expected value and expected utility.",
            "modeling_principles": [
                "Expected value is a risk-neutral benchmark.",
                "Expected utility incorporates risk attitude through utility curvature.",
                "Certainty equivalents translate expected utility into interpretable outcome terms.",
                "Risk premiums measure willingness to give up expected value for certainty.",
                "Choice rankings can change when risk aversion changes.",
                "Probability and utility assumptions should be documented and sensitivity-tested.",
            ],
            "summary": summary_rows,
        },
    )

    print("Expected value and expected utility workflow complete.")
    print(TABLES / "expected_utility_sensitivity.csv")
    print(TABLES / "expected_utility_summary.csv")
    print(RECORDS / "expected_value_expected_utility_decision_record.json")


if __name__ == "__main__":
    main()

This workflow illustrates how a prospect can look attractive under expected value but become less attractive as risk aversion increases. It also shows why certainty equivalents and risk premiums can make expected utility easier to interpret in practical decision settings.

Back to top ↑

GitHub Repository

The companion repository for this article supports reproducible exploration of expected value, expected utility, risk aversion, utility functions, certainty equivalents, risk premiums, probability assumptions, sensitivity analysis, and decision-record documentation.

articles/expected-value-and-expected-utility/
├── python/
│   ├── expected_value_expected_utility_simulation.py
│   ├── expected_value_calculator.py
│   ├── utility_function_profiles.py
│   ├── certainty_equivalent_calculator.py
│   ├── risk_premium_analysis.py
│   ├── risk_aversion_sensitivity.py
│   ├── probability_quality_audit.py
│   ├── decision_record_exporter.py
│   └── run_all_expected_utility_workflows.py
├── r/
│   ├── expected_value_expected_utility_workflow.R
│   ├── expected_value_profiles.R
│   ├── utility_curve_sensitivity.R
│   ├── certainty_equivalent_report.R
│   ├── risk_premium_profiles.R
│   ├── probability_assumption_audit.R
│   └── run_all_expected_utility_workflows.R
├── julia/
│   ├── high_performance_expected_utility_scan.jl
│   ├── crra_sensitivity_surface.jl
│   └── certainty_equivalent_frontier.jl
├── sql/
│   ├── schema_expected_utility.sql
│   ├── prospects.sql
│   ├── outcomes.sql
│   ├── probabilities.sql
│   ├── utility_models.sql
│   ├── model_runs.sql
│   └── decision_records.sql
├── rust/
│   └── expected_utility_diagnostics_cli.rs
├── go/
│   └── expected_utility_score_runner.go
├── cpp/
│   ├── expected_value_core.cpp
│   └── crra_expected_utility.cpp
├── fortran/
│   └── numerical_expected_utility_model.f90
├── c/
│   └── expected_value_core.c
├── docs/
│   ├── article_notes.md
│   ├── modeling_principles.md
│   ├── expected_value.md
│   ├── expected_utility.md
│   ├── utility_functions.md
│   ├── certainty_equivalents.md
│   ├── risk_premiums.md
│   ├── probability_quality.md
│   ├── responsible_use.md
│   └── assumptions_and_limitations.md
├── data/
│   ├── synthetic_prospects.csv
│   ├── synthetic_outcomes.csv
│   ├── synthetic_probability_sets.csv
│   ├── synthetic_utility_models.csv
│   ├── synthetic_risk_aversion_levels.csv
│   └── synthetic_decision_records.csv
├── outputs/
│   ├── README.md
│   ├── figures/
│   ├── tables/
│   └── decision_records/
└── notebooks/
    ├── python_expected_utility_walkthrough.ipynb
    └── r_expected_utility_placeholder.ipynb

This repository structure reflects the article’s central argument: expected value and expected utility are not merely formulas. They are reproducible tools for clarifying uncertain outcomes, probability assumptions, utility assumptions, risk preferences, and accountable decision reasoning.

Back to top ↑

Back to top ↑

Common Pitfalls

Expected value and expected utility are often misused when decision-makers treat formulas as substitutes for judgment. The most common pitfalls involve weak probabilities, incomplete outcomes, hidden utility assumptions, and overconfident rankings.

The most dangerous misuse is treating expected value or expected utility as objective because the formula is formal. The calculation is only as good as the outcomes, probabilities, utility assumptions, and decision frame behind it.

Back to top ↑

Why Expected Value and Expected Utility Still Matter

Expected value and expected utility remain foundational because they provide disciplined ways to compare uncertain prospects. Expected value offers a clear probabilistic benchmark. Expected utility adds preference, risk attitude, and subjective valuation. Together, they help decision-makers make uncertainty and value assumptions explicit.

Their deepest value is not that they eliminate judgment. It is that they structure judgment. They force decision-makers to specify outcomes, probabilities, preferences, risk attitudes, and trade-offs. They reveal when a high-average option is fragile, when a safer option has greater utility, and when a recommendation depends on contested assumptions.

In modern decision science, expected value and expected utility are best used as part of a broader architecture of judgment. They should be connected to sensitivity analysis, scenario comparison, behavioral insight, stakeholder values, robustness, and decision records. Used this way, they remain among the most important tools for reasoning under uncertainty.

Back to top ↑

Related Articles

• Decision Science
• Decision Science vs. Decision Theory
• Why Uncertainty Changes Decision-Making
• Core Principles of Decision Science
• Decision Records and Accountable Judgment
• Decision Trees and Structured Choice
• Risk Analysis and Probabilistic Reasoning
• Bayesian Decision-Making
• Judgment Under Uncertainty
• Behavioral Decision Theory
• Multi-Criteria Decision Analysis
• Robust Decision-Making

Back to top ↑

Further Reading

• Bernoulli, D. (1954) “Exposition of a New Theory on the Measurement of Risk.” Econometrica, 22(1), pp. 23–36. Translation of the 1738 essay. Stable bibliographic discussion available at: https://plato.stanford.edu/archives/spr2022/entries/paradox-stpetersburg/
• Howard, R.A. and Abbas, A.E. (2023) Foundations of Decision Analysis. Harlow: Pearson. Available at: https://www.pearson.com/en-us/subject-catalog/p/foundations-of-decision-analysis/P200000003532/9780137981878
• Kahneman, D. (2013) Thinking, Fast and Slow. New York: Farrar, Straus and Giroux. Available at: https://us.macmillan.com/books/9780374533557/thinkingfastandslow/
• Keeney, R.L. (1992) Value-Focused Thinking: A Path to Creative Decisionmaking. Cambridge, MA: Harvard University Press.
• Raiffa, H. (1968) Decision Analysis: Introductory Lectures on Choices Under Uncertainty. Reading, MA: Addison-Wesley. Bibliographic record available at: https://books.google.com/books/about/Decision_Analysis.html?id=zvm3AAAAIAAJ
• Von Neumann, J. and Morgenstern, O. (1944) Theory of Games and Economic Behavior. Princeton, NJ: Princeton University Press.
• Savage, L.J. (1954) The Foundations of Statistics. New York: Wiley.
• Tversky, A. and Kahneman, D. (1974) “Judgment under Uncertainty: Heuristics and Biases.” Science, 185(4157), pp. 1124–1131. Available at: https://www.science.org/doi/10.1126/science.185.4157.1124

Back to top ↑

References

• Bernoulli, D. (1954) “Exposition of a New Theory on the Measurement of Risk.” Econometrica, 22(1), pp. 23–36. Translation of the 1738 essay. Stable bibliographic discussion available at: https://plato.stanford.edu/archives/spr2022/entries/paradox-stpetersburg/
• Howard, R.A. (1966) “Decision Analysis: Applied Decision Theory.” Proceedings of the Fourth International Conference on Operational Research.
• Howard, R.A. and Abbas, A.E. (2023) Foundations of Decision Analysis. Harlow: Pearson. Available at: https://www.pearson.com/en-us/subject-catalog/p/foundations-of-decision-analysis/P200000003532/9780137981878
• Kahneman, D. (2013) Thinking, Fast and Slow. New York: Farrar, Straus and Giroux. Available at: https://us.macmillan.com/books/9780374533557/thinkingfastandslow/
• Keeney, R.L. (1992) Value-Focused Thinking: A Path to Creative Decisionmaking. Cambridge, MA: Harvard University Press.
• Raiffa, H. (1968) Decision Analysis: Introductory Lectures on Choices Under Uncertainty. Reading, MA: Addison-Wesley. Bibliographic record available at: https://books.google.com/books/about/Decision_Analysis.html?id=zvm3AAAAIAAJ
• Savage, L.J. (1954) The Foundations of Statistics. New York: Wiley.
• Tversky, A. and Kahneman, D. (1974) “Judgment under Uncertainty: Heuristics and Biases.” Science, 185(4157), pp. 1124–1131. Available at: https://www.science.org/doi/10.1126/science.185.4157.1124
• Von Neumann, J. and Morgenstern, O. (1944) Theory of Games and Economic Behavior. Princeton, NJ: Princeton University Press.

Back to top ↑