Expected Value of Perfect Information (EVPI)

Table of Contents

  1. Definition
  2. How EVPI Relates to Marketing
  3. How to Calculate EVPI
  4. How to Utilize EVPI
  5. Comparison to Similar Approaches
  6. Best Practices
  7. Future Trends
  8. FAQs
  9. Related Terms
  10. Sources

Definition

Expected Value of Perfect Information (EVPI) is the maximum amount a rational decision-maker should be willing to pay to remove all uncertainty from a decision before making it. It compares two worlds: one where the choice is made under the current state of knowledge, and one where the choice is made knowing exactly how the uncertain variables will resolve. The gap between those two expected outcomes is the EVPI.

The concept comes out of decision analysis, formalized by Ronald A. Howard in his 1966 paper Information Value Theory and codified earlier by Raiffa and Schlaifer in their work on statistical decision theory. Szaniawski in 1967 defined EVPI as “the highest price the decision-maker would be prepared to pay for perfect information”. ScienceDirect

A few properties are baked into the definition. The value of information can never be less than zero since the decision-maker can always ignore the additional information and make a decision as if such information is not available. No other information gathering/sharing activities can be more valuable than that quantified by EVPI. So EVPI is always non-negative, and it always sits above the value of any real-world study, which can only ever deliver imperfect information. Wikipedia

Back to Top

How EVPI Relates to Marketing

Marketing runs on bets made under uncertainty. Will a creative concept land? Will a price increase erode volume more than it lifts margin? Will a new segment respond to a different value proposition? Research, testing, and data spend exist to shrink that uncertainty — and EVPI is the tool that tells marketers how much that shrinkage is actually worth.

The framing is simple but uncomfortable. If a $250,000 campaign decision has an EVPI of $12,000, then no piece of research about that decision is worth more than $12,000. Doesn’t matter how thorough the methodology is. Doesn’t matter how reputable the vendor. The ceiling is the ceiling.

EVPI also separates two things marketers tend to conflate: how uncertain a forecast is, and how much that uncertainty matters. A wildly uncertain input that doesn’t change the action under any plausible scenario has an EVPI of zero. A modestly uncertain input that would flip the go/no-go decision can be worth a lot. The math forces the question of which kind a given uncertainty actually is.

Common marketing settings where EVPI shows up — sometimes explicitly, more often implicitly — include media-mix planning, pricing tests, segmentation studies, brand tracker procurement, attribution platform evaluation, and the perennial question of when to stop researching and just launch.

Back to Top

How to Calculate EVPI

The formula:

EVPI = EVwPI − EMV*

Where:

  • EVwPI = Expected Value with Perfect Information (the average payoff when the best option is chosen for each possible state of nature)
  • EMV* = the highest Expected Monetary Value among the available choices under current uncertainty

The expected value of perfect information is related to the EMV method and the EOL method, since the expected value of perfect information is the same as the minimum expected opportunity loss. MathCracker

A worked marketing example. A team is deciding whether to launch a premium subscription tier. Three possible states of demand:

Demand stateProbabilityProfit if launchedProfit if not launched
High0.3$800,000$0
Medium0.5$200,000$0
Low0.2−$400,000$0

EMV of launching = (0.3 × $800K) + (0.5 × $200K) + (0.2 × −$400K) = $240K + $100K − $80K = $260K. EMV of not launching = $0.

So the best choice under uncertainty is launch, with EMV* = $260K.

With perfect information, the team would launch in high and medium demand, and skip the launch in low demand: EVwPI = (0.3 × $800K) + (0.5 × $200K) + (0.2 × $0) = $240K + $100K + $0 = $340K.

EVPI = $340K − $260K = $80K.

That’s the ceiling. Any market study, segmentation, or concept test about demand for this tier should cost less than $80K, or it doesn’t pay for itself even in the best case.

A second property worth knowing: in a decision problem with two choices, and two possible outcomes for the random variable, EVPI cannot exceed half the maximum difference in the reward under each outcome. A useful sanity check. Wikipedia

Back to Top

How to Utilize EVPI

Budgeting research. Before commissioning a study, model the underlying decision as a payoff table and compute EVPI. If the number comes back at $15K and a research firm quotes $80K, that’s a real conversation — either the scope is wrong, or the decision isn’t the one driving the spend.

Prioritizing experiments. With a backlog of A/B test ideas, EVPI ranks them by how much the answer would actually change the next decision. High-traffic surface, high-stakes choice, real disagreement about the prior — those are high-EVPI tests. Low-traffic surface, decision already largely made — low EVPI.

Evaluating data platforms. Attribution tools, customer data platforms, and brand trackers all carry real recurring costs. EVPI gives a way to ask: across the decisions this platform is meant to inform, how often would better information change the action, and by how much? Sometimes the answer is uncomfortable.

Knowing when to stop. Three rounds of concept testing have been run. Does a fourth round materially shrink decision uncertainty, or does it mostly confirm what’s already there? EVPI gives a non-political answer.

Justifying skipping research. Sometimes the most useful EVPI calculation is the one that says: just launch. If the decision is essentially settled regardless of what new information would show, the rational move is to act now.

Back to Top

Comparison to Similar Approaches

ApproachWhat it measuresWhen to use itLimitation
EVPIMaximum justifiable spend for information that resolves uncertainty completelySizing research budgets, decision auditsAssumes perfect information, which isn’t real
EVSI (Expected Value of Sample Information)Value of a specific, imperfect studyComparing actual research proposalsRequires modeling the study design
EVPPI (Expected Value of Partial Perfect Information)Value of resolving one variable while others stay uncertainPicking which uncertainty to resolve firstMore computationally involved
Sensitivity analysisHow much outputs shift when inputs shiftIdentifying which inputs matterSays what matters, not what it’s worth
ROI on past researchRealized return after the factVendor evaluation, post-mortemBackward-looking
Cost-benefit analysisTotal expected benefits vs. costsGo/no-go on initiativesDoesn’t isolate the value of specific uncertainties

EVPI is the ceiling. EVSI is what you’d actually compute for a real study. EVPPI is what you’d compute when you want to know which single variable is most worth resolving. Most marketing teams start and stop at EVPI because it’s the easiest of the three and answers the most common question.

Back to Top

Best Practices

Start with the decision, not the data. If no plausible research result would change the next action, EVPI is zero. Calculate it before writing the brief, not after.

Be honest about priors. EVPI math is only as good as the probabilities going in. Marketers anchor — often hard — on the outcome they expect or want. Get a second opinion on the probability estimates from someone without skin in the game.

Use EVPI as a ceiling, not a target. EVSI for a real study is always less than EVPI, sometimes by a lot. A $40K EVPI doesn’t justify a $40K study.

Watch the structure of the decision. EVPI is sensitive to how the problem is framed. A binary launch/no-launch decision and a three-way launch-now/launch-later/no-launch decision can produce very different numbers from the same underlying data.

Document the assumptions. A VoI estimate that can’t be reconstructed in six months is just a number on a slide. Keep the payoff table, probabilities, and decision tree with the research plan.

Don’t ignore correlated uncertainties. When two uncertain variables move together — say, demand and competitor response — calculating EVPI on each in isolation can mislead. Joint resolution is sometimes worth more, sometimes less, than the sum of the parts.

Back to Top

Probabilistic modeling has gotten cheap. What used to require a decision-science consultant can now be drafted in a notebook in an afternoon, and increasingly, in marketing-mix modeling platforms that expose EVPI-style outputs alongside their forecasts.

Privacy changes are pushing EVPI from theoretical to practical. As third-party cookies disappear and consent requirements tighten, every piece of customer data costs more to acquire. EVPI gives marketing teams a defensible answer to “is this dataset worth the cost of collecting it?”

Experimentation programs at scale — hundreds of tests per quarter at large platforms — have created a real prioritization bottleneck. EVPI-based scoring is starting to appear in test backlogs the same way ICE and RICE scoring did a decade ago, sometimes under different names.

Bayesian methods are quietly displacing frequentist A/B testing in some marketing analytics stacks, and Bayesian inference pairs naturally with EVPI calculations. As one goes mainstream, the other tends to follow.

Back to Top

FAQs

Can EVPI be negative? No. The value of information can never be less than zero since the decision-maker can always ignore the additional information and make a decision as if such information is not available. Information can have zero value when it wouldn’t change the decision under any outcome. Never negative. Wikipedia

What’s the difference between EVPI and EVSI? EVPI assumes information that resolves the uncertainty completely — a theoretical upper bound. EVSI estimates the value of a specific, imperfect study (a survey of 500 respondents, a two-week A/B test, etc.). EVSI is always less than or equal to EVPI for the same underlying uncertainty.

Does EVPI assume risk neutrality? The basic formula does. In risk-sensitive settings the value of information has been defined as the expected utility increase of the revelation, or the buying price for the revelation. For risk-averse decision-makers, the calculation uses expected utility rather than expected dollars, and the resulting number can be meaningfully different. INFORMS

Do I need a decision tree? Not for simple cases. A payoff table in a spreadsheet works. Decision trees become useful when there are sequential decisions, multiple chance nodes, or when the same uncertain event affects several branches.

What if I don’t know the probabilities? Estimate them. The discipline of putting numbers on “how likely is this campaign to work” is often more useful than the resulting EVPI itself. Use ranges and sensitivity analysis if a point estimate feels too false.

Can EVPI handle qualitative inputs? Yes, but they have to be translated into probability shifts. A finding from focus groups might inform whether the prior on segment receptivity moves from 0.4 to 0.6. The translation is loose, but the math still works.

Is EVPI the same as the value of a study after it’s done? No. EVPI is forward-looking — the maximum justifiable spend before doing research. The realized value of a study after the fact is a different (and often less informative) calculation.

How does EVPI relate to marketing-mix modeling? MMM platforms produce estimates with confidence intervals. EVPI on those estimates asks how much shrinking the confidence intervals would actually change budget allocation decisions. Often less than the analyst team expects.

Does EVPI work for ongoing measurement like brand trackers? It can, but the decision being informed has to be made explicit. A tracker that doesn’t inform any specific decision has an EVPI of zero — even if the data is interesting.

Where did EVPI originate? Foundational work came from Howard Raiffa and Robert Schlaifer in the 1950s and 1960s on statistical decision theory. Ronald A. Howard formalized the broader framework in Information Value Theory (1966) and The Foundations of Decision Analysis (1968).

Back to Top
Back to Top

Sources

Back to Top

Was this helpful?

Previous

Double Diamond Design Process

Next

FAST Goals