Preferences complete the model
Probability tells what may occur; utility determines how the organisation ranks the consequences.
Business · Strategy
Utility, Expected Value and Value of Information
A handbook for converting uncertain outcomes into rational choices using preference models, expected utility, decision networks and value-of-information analysis.
Expected value is not always enough
Non-linear preferences, constraints and severe downside can make equal expected-dollar alternatives strategically different.
Information has economic value
A test, pilot or research activity is valuable only to the extent that it can improve a later decision enough to justify its cost and delay.
From prediction to decision
A forecast does not choose an action until outcomes are connected to preferences.
Suppose two projects have different probability distributions of cost and benefit. A probability model can describe those distributions, but the organisation still needs a rule for comparing them. If dollars are the only consequence and risk neutrality is appropriate, expected monetary value may be enough. If liquidity, safety, reputation or nonlinear strategic effects matter, a utility function or multi-attribute value model is more appropriate.
The source’s rational-preference framework formalises consistency conditions under which preferences can be represented by expected utility. In practice, the value is less about philosophical perfection and more about making trade-offs explicit and auditable.
Expected monetary value and expected utility
Expected value averages outcomes using their probabilities.
Expected valueE[X] = Σ P(xᵢ) xᵢ for discrete outcomes, or the corresponding integral for continuous outcomes.
Expected monetary value is appropriate when repeated exposure, financial scale and risk tolerance make dollars approximately linear in preference. For a one-off decision that could threaten solvency, linear monetary value may be inappropriate. A loss that forces closure has a consequence greater than its nominal dollar amount in a spreadsheet.
Expected utilityEU(a) = Σ P(x | a) U(x). Choose the action with the greatest expected utility under the specified model.
A concave utility over wealth represents aversion to variability in classical models; convex regions can represent risk-seeking behaviour. Rather than assuming a shape, elicit meaningful trade-offs and test sensitivity.
Multi-attribute decisions
Business outcomes often contain more than one consequence.
A sourcing decision can affect cost, quality, lead time, resilience and intellectual-property exposure. Combining all consequences into one arbitrary weighted score can conceal interactions and double counting. Define attributes carefully, ensure scales have clear meaning and examine whether trade-offs are independent enough for an additive model.
Where strategic or regulatory constraints are non-negotiable, treat them as constraints rather than compensable attributes. A severe safety requirement should not be traded away simply because another score is high. Preference modelling should reflect the actual governance rules of the organisation.
Decision networks
A decision network extends a probabilistic network with decision and utility nodes.
Chance nodes represent uncertain states, decision nodes represent choices and utility nodes represent value. Information arcs specify what will be known when a decision is made. This timing is crucial. A test result has no value if it arrives after an irreversible commitment, and a forecast should not be used in a decision node if it would not actually be available at that time.
Prior informationWhat is known before acting?
DecisionChoice available at that time.
Chance outcomesUncertain consequences after the action.
UtilityValue assigned to the resulting state.
Later observationNew evidence may enable a subsequent decision.
Value of perfect information
Perfect information is a useful upper bound on what any real information source can be worth.
EVPI conceptValue of perfect information = Expected value with the uncertain state known before deciding − Expected value under the best current decision.
If perfect information about an uncertainty would not change the chosen action, then no test about that uncertainty has decision value. This is a powerful screening rule: it can stop teams collecting data simply because it is available. If perfect information would materially change the choice, then real tests may be worth investigating.
Perfect information is hypothetical and rarely obtainable. Its value is an upper bound before considering the cost, delay and imperfection of an actual measurement.
Value of imperfect information
Real tests produce noisy signals rather than certainty.
Model the possible test results and how each result changes posterior beliefs. For each result, calculate the best subsequent action and its expected value. Average across possible results, then subtract the value of the best decision without the test. Finally subtract the cost of obtaining the information and include the strategic cost of delay where relevant.
Define the uncertain variable
Identify what knowledge could change the decision.
Define the test
List possible signals and their reliability.
Update beliefs
Calculate posterior probabilities for each signal.
Choose conditionally
Identify the best action after each possible signal.
Average future value
Weight those conditional values by the probability of each signal.
Compare with no-test decision
The difference is gross value of sample information.
Subtract cost and delay
Proceed only when the net value and practical timing justify the test.
Worked example: trial before rollout
A generic organisation considers a full rollout with uncertain customer adoption.
Without further data, the expected value of full rollout is slightly higher than not proceeding, but the downside under low adoption is material. A small trial can produce a high or low adoption signal. If a high signal would justify rollout and a low signal would justify stopping, the trial changes behaviour and therefore may have positive value.
If management would roll out regardless of the trial result, the trial has essentially no decision value even if it produces interesting information. Conversely, if the trial is so slow that competitors erase the opportunity, its delay cost may exceed its informational value. The calculation forces those trade-offs into the same decision frame.
Behavioural deviations and governance
The source also recognises that real decision-makers do not always behave like ideal expected-utility agents.
Framing, loss aversion, overconfidence, anchoring and inconsistent probability perception can affect choices. The remedy is not to assume people can become perfectly rational; it is to design process controls. Use pre-defined criteria, independent review, explicit probability ranges and decision records. Compare forecasts with outcomes to improve calibration.
Is expected value the same as forecast value?
No. A forecast may give the most likely outcome, while expected value averages across all modelled outcomes. The most likely outcome does not necessarily have the highest probability mass in a continuous distribution and does not capture tail consequences.
Can value of information be negative?
Gross information value cannot be negative when it can be ignored, but a real information-gathering activity can have negative net value after cost, delay or operational disruption.
Application checklist
- Separate probability beliefs from outcome preferences.
- Use expected monetary value only when linear monetary preference is appropriate.
- Represent hard constraints separately from compensable attributes.
- Ensure decision models respect what information will actually be available at each decision time.
- Calculate perfect-information value as an upper bound before commissioning research.
- Model real test accuracy and posterior updating for imperfect information.
- Include cost and delay when deciding whether to gather information.
- Use process controls to reduce behavioural bias and overconfidence.
Source basis. Decision-analysis source set: probabilistic reasoning, sequential decisions, learning, state uncertainty and multiagent methods. This page is an original handbook synthesis of the supplied materials. Named people, organisations and identifying case details from the sources have been removed. Numerical examples are labelled as illustrative where used.