Skip to content
Mental Models

The Value of Information

The Perfect-Info Ceiling (EVPI)

What a flawless answer — a literal crystal ball — would be worth, worked out on a real payoff table. EVPI is the most any test could ever be worth, and it equals the expected cost of the mistake you'd otherwise make.

14 min Updated Jul 11, 2026

In the last lesson you learned to ask would it change my action? — and to bin any test that fails it. Now we put a number on the tests that pass. We start at the top, with the best possible test: a clairvoyant who answers with perfect certainty. Whatever a real test could tell you, it can’t beat knowing for sure. So the value of perfect information — EVPI — is the ceiling: the absolute most any test, survey, or expert could ever be worth on this decision. Nail the ceiling first and every real test slots in beneath it.

Before you read — take a guess

Guess: what's the most a market survey could possibly be worth to a launch decision?

The set-up: should you launch?

Let’s make it concrete with a decision we’ll carry through the next three lessons. You can launch a new product or walk away. The world is either good (the market wants it) or bad (it doesn’t). The payoffs, in thousands of dollars:

Good (market wants it)Bad (market rejects it)
Launch+400−300
Walk away00

Your honest prior is that the market is good with probability 40% (so bad with 60%). That’s the whole problem: a payoff table and a prior. Everything else is arithmetic.

Info:

Reading a payoff table

Rows are the actions you control; columns are the states of the world you don’t. Each cell is what you get if you take that action and that state turns out to be true. “Walk away” pays 0 either way — walking away is the safe, boring floor every decision is measured against.

Step 1 — what’s your best move with no extra info?

Weigh each action by the prior. This is just expected value — each outcome times its probability, summed:

EV(Launch)=0.40×400+0.60×(300)=160180=20EV(\text{Launch}) = 0.40 \times 400 + 0.60 \times (-300) = 160 - 180 = -20

EV(Walk away)=0.40×0+0.60×0=0EV(\text{Walk away}) = 0.40 \times 0 + 0.60 \times 0 = 0

Walking away (0) beats launching (−20), if only barely. So with nothing but your prior, your best move is to walk away, and the value of that best prior action — call it V0V_0 — is 0. Hold onto that: right now you launch nothing and pocket nothing.

Warning:

Notice how CLOSE this is

Launching scores −20 against walking away’s 0. A hair’s breadth. That closeness is exactly the condition under which information gets valuable — we’ll make that precise in lesson four. A blowout decision (launch scores +500) wouldn’t leave much for a test to change.

Step 2 — what would a crystal ball be worth?

Now imagine a clairvoyant who tells you the true state before you decide, with perfect certainty. You’d use it perfectly: in every good world you’d launch and bank 400; in every bad world you’d walk away and take 0 instead of eating −300.

You don’t know in advance which message you’ll get, so weigh the clairvoyant-guided outcomes by the same prior:

EV(with crystal ball)=0.40×400launch, good+0.60×0walk, bad=160EV(\text{with crystal ball}) = 0.40 \times \underbrace{400}_{\text{launch, good}} + 0.60 \times \underbrace{0}_{\text{walk, bad}} = 160

With perfect foresight you’d expect to make 160. Without it you make 0. The value of the perfect information is the difference:

EVPI=1600=160\text{EVPI} = 160 - 0 = \mathbf{160}

Tip:

EVPI, defined

EVPI (Expected Value of Perfect Information) = the expected payoff of your best decision with a perfect answer, minus the expected payoff of your best decision without one. Here: 160 − 0 = 160. That’s the ceiling — no real test on this decision can be worth more than 160.

Where does the 160 of value actually come from? You already knew the payoffs and the prior.

EVPI is the expected cost of the mistake you’d otherwise make

There’s a second way to see the same number, and it’s the one to remember. Acting on your prior, you walk away from every world — including the 40% good ones where launching would have made 400. That’s the mistake: walking away from a winner. It costs you 400 each time it happens, and it happens 40% of the time:

expected cost of the mistake=0.40×400=160\text{expected cost of the mistake} = 0.40 \times 400 = 160

Same 160. EVPI is exactly the expected cost of the error your prior-based choice walks you into — because perfect information is precisely what lets you stop making that error. This is why value can never exceed the stakes: you can’t save more than the mistake was going to cost.

State what EVPI measures.

Pick the right option for each blank, then check.

EVPI is the value of a answer: the payoff of your best decision it minus the payoff of your best decision without it. It equals the expected cost of the your prior-based choice would otherwise make, so it is the on what any real test could be worth.

When the ceiling is the floor: EVPI = 0

Run the same machine on a lopsided prior and watch the value vanish. Suppose you were 95% sure the market is bad:

EV(Launch)=0.05×400+0.95×(300)=20285=265EV(\text{Launch}) = 0.05 \times 400 + 0.95 \times (-300) = 20 - 285 = -265

You walk away — emphatically. Now the crystal ball: it’s worth 0.05×400=200.05 \times 400 = 20 with perfect play, and your prior action already banks 0, so EVPI =200=20= 20 - 0 = 20. Push to 99% bad and it drops to 4. The more lopsided you are, the less a perfect answer can do, because you’re already taking the action you’d take in almost every world. At the extreme — 100% sure — EVPI is exactly 0: nothing left to learn, because nothing could change your mind.

Warning:

The pattern to file away

EVPI is largest when you’re genuinely torn and the payoffs are big; it shrinks toward zero as your prior gets lopsided or the stakes get small. A confident prior is a cheap prior to hold — no test can improve on a decision you were always going to make the same way. We’ll turn this into a full “when is it worth it?” rule in lesson four.

A friend is 99.5% sure their idea will flop and plans to shelve it. They ask whether to pay for a big market study first. Using EVPI, what do you tell them?

Why start at the ceiling?

Because it’s the cheapest sanity check in all of decision analysis. Before you haggle over which survey to buy or how accurate it is, compute EVPI. If the perfect answer is worth less than the price of the survey, you can stop right there — no real test can be worth buying, because every real test is worth less than perfect information. EVPI turns a fuzzy “should we research this?” into a hard upper bound you can compare against a price tag in thirty seconds.

Big picture

The perfect-info ceiling

  • EVPI
    • What it is
      • Value of a perfect, certain answer
      • Best decision WITH it − best decision WITHOUT it
    • Why it matters
      • The ceiling: no real test beats knowing for sure
      • Equals the expected cost of the mistake you’d make
    • When it’s zero
      • Lopsided prior — you’d act the same anyway
      • Tiny stakes — little mistake to avert

You now have the ceiling: on the launch decision, perfect information is worth 160, and that caps everything. But nobody sells perfect information. Next we drop from the crystal ball to a real, noisy test — a market survey that’s right most of the time but not always — and work out what that is worth. Spoiler baked into the maths: it’ll come in under 160, every time.

Mark lesson as complete