A good decision can have a bad result, and a bad decision can have a good one, and if you can't tell the two apart, you'll spend a career punishing sound thinking and rewarding lucky guesses. Expected value is the one idea that keeps them straight.

The quick version

  • Expected value (EV) is the probability-weighted payoff: for each possible outcome, multiply what you'd gain or lose by how likely it is, then add it all up. Pick the option with the highest total.
  • Judge the bet, not the result. A sound call can still lose on a bad draw. Grade the decision on what you knew at the time, separately from how it turned out.
  • Size it so you can't be ruined. A good average is no comfort if one bad outcome ends the game. Never take a bet you can't survive, however good the odds look.
  • Keep a decision journal. Write down the call, the odds you gave it, and your reasons, before you know the outcome. It's the only honest defence against hindsight rewriting the story.

The idea in depth

Strip a hard decision down and the skeleton is the same every time: a set of options, each leading to several possible futures, each future with a payoff and a probability. Expected value is just the probability-weighted average of those payoffs, multiply each outcome by its chance of happening, add them up. The option with the highest expected value is, on average and over many repetitions, the better bet.

The maths is old. In 1738 Daniel Bernoulli, wrestling with a gambling puzzle now called the St. Petersburg paradox, showed that rational people don't maximise raw money, they maximise utility, and the utility of money diminishes as you get more of it (a thousand pounds means more to you than to a billionaire). Two centuries later John von Neumann and Oskar Morgenstern, in Theory of Games and Economic Behavior (1944), turned that intuition into an axiom system: if your preferences are consistent, you behave as if you are maximising expected utility. That work is the formal backbone of modern decision theory, the reason "weight the payoff by its probability" is a theorem, not a productivity hack.

The move: before a consequential call, force the implicit numbers into the open. You don't need three decimal places. You need to say out loud, "If this works, maybe a 60% chance, we gain roughly this; if it fails, we lose roughly that." Estimating the probability and the payoff separately is most of the value. It converts a gut feeling into a structure you can argue with, write down, and revisit.

flowchart LR
    A(["Launch the pilot?"]) --> B(["Two outcomes"])
    B -->|"60% works"| C("+£500k")
    B -->|"40% fails"| D("−£150k")
    C --> E(["EV = .6 × 500k − .4 × 150k = +£240k"])
    D --> E
A one-branch decision tree: weight each payoff by its probability, sum, and you have the expected value. A positive EV says "take it" even though four times in ten it loses money. Leaders Loop

Why outcomes lie, and process tells the truth

The reason EV matters for leaders isn't the arithmetic; it's what it protects you from. Human judgement is systematically distorted under uncertainty, and the distortions are well evidenced. Daniel Kahneman and Amos Tversky's prospect theory (Econometrica, 1979, among the most-cited papers the journal has ever published) showed that we don't feel gains and losses symmetrically: a loss hurts roughly twice as much as an equivalent gain feels good. That loss aversion quietly tilts leaders toward defending sunk costs and away from positive-EV bets that carry a visible downside.

Annie Duke, the former professional poker player turned decision scientist, gives the resulting trap a name: "resulting", grading a decision by its outcome rather than by the process that produced it. A manager makes a sound, positive-EV call; it loses on the unlucky 40%; the organisation treats the loss as proof of bad judgement; everyone learns to avoid bets that could ever look bad in hindsight. Duke's fix, borrowed straight from the poker table where the best players lose hands constantly, is to separate the two axes entirely: the quality of the decision, and luck.

"We are uncomfortable with the idea that luck plays a significant role in our lives.", Annie Duke, Thinking in Bets (2018)

So the move is: in your reviews, grade the decision and the outcome on separate lines. Ask "Was this a good bet given what we knew at the time?" before you ask "Did it work?" The two-by-two below is about the cheapest leadership tool there is, it costs one whiteboard and the willingness to admit that a win can come from a bad bet.

quadrantChart
    title Decision quality vs. outcome ("resulting")
    x-axis "Bad outcome" --> "Good outcome"
    y-axis "Bad process" --> "Good process"
    quadrant-1 "Earned it"
    quadrant-2 "Dodged a bullet, fix the process"
    quadrant-3 "Deserved it, fix the process"
    quadrant-4 "Unlucky, keep the process"
    "Sound bet, flood hit": [0.25, 0.8]
    "Reckless bet, lucky win": [0.78, 0.2]
Good process can still lose; bad process can still win. Reward the top half, not the right half. The flooded-supplier call below sits in "Unlucky, keep the process." Leaders Loop

Getting the probabilities right, and the limit you can't ignore

The obvious objection: EV is only as good as the probabilities you feed it, and most of us are terrible at probabilities. True, but it's a trainable skill, not a fixed trait. Philip Tetlock's Good Judgment Project, the forecasting tournament run with the US intelligence community's IARPA from 2011, found that a subset of ordinary volunteers, his "superforecasters", out-predicted professional analysts with access to classified material. What set them apart wasn't IQ or insider knowledge; it was calibration (when they said 70%, it happened about 70% of the time), thinking in fine-grained probabilities, and keeping score so they could learn from misses.

Keeping score is the practical heart of it. Tetlock's forecasters ran unflinching post-mortems; Duke recommends the same instrument by a homelier name, the decision journal. Before the outcome is known, you write down the call, the probabilities you assigned, and why. Later you reread it. This is the only honest defence against hindsight bias, which quietly rewrites your memory of what you "obviously" knew at the time.

EV is not a universal law, though, and selling it as one would be its own bad bet. The deeper limit is ruin. Expected value averages over many repetitions, but some bets you only get to make once, and some losses you don't come back from. Nassim Taleb's image is the sharpest: never cross a river because it is on average four feet deep. A positive average is no comfort if one stretch is over your head and there's no second attempt. The practical rule sits on top of EV, not against it: maximise expected value only among options that cannot ruin you. Size the bet so the worst case is survivable; then, and only then, chase the best odds.

So here's the move: make "what's the worst realistic case, and can we survive it?" the standing first question, before the EV maths, not after. If the answer is "no," the expected value is irrelevant. This is the same instinct behind the "decide well, revisit honestly" philosophy in our profile of product leader Rob Alford, and it's the bridge to EV's sibling discipline of moves and counter-moves in game theory and strategic interaction.

A worked example: the manager and the flood

A regional manager signs off on a new supplier. The price is 12% lower, the references check out, the contract has an exit clause. Six weeks in, the supplier's warehouse floods and a shipment is lost. The deal looks like a disaster, and in the post-mortem she's quietly written off as the person who picked the wrong vendor. Here's the uncomfortable part: she made a good decision.

Before signing, she'd made the bet explicit. The new supplier saves £500k a year if it holds (she put that at 80%); the realistic downside, switching back to the incumbent plus disruption, costs about £150k (she put a serious failure at 20%). Expected value: 0.8 × £500k − 0.2 × £150k ≈ +£370k (figures illustrative). Crucially, she also asked the ruin question: could any single failure sink the division? No, the exit clause and a backup supplier capped the loss. A clearly positive-EV bet with no path to catastrophe. She should take it every time it's offered.

Then the warehouse floods, the bad draw inside the 20%. Under "resulting," she's the villain. Under expected-value thinking, the review says: the process was sound, the odds were read fairly, the downside was capped, and the org got unlucky. Keep the process; don't fire the thinker. The decision journal she wrote six weeks earlier is what makes that conversation possible, without it, everyone in the room "always knew" the supplier was dodgy.

flowchart LR
    A(["Size the bet, EV + ruin check"]) --> B(["Pre-mortem, how could this fail?"])
    B --> C(["Decide"])
    C --> D(["Journal, odds + reasons"])
    D --> E(["Review on process, not outcome"])
    E -. "learn" .-> A
The leader's decision loop. The arrow back from review to sizing is the whole point: calibration only improves if you close the loop. Leaders Loop

Frequently asked questions

Isn't this just gambling?

It's the opposite. Gambling is taking negative-EV bets for the thrill; this is the discipline of only taking bets where the odds are in your favour and the downside can't ruin you. The casino isn't the gambler, it's the house, and the house thinks in expected value. So should you.

How do I estimate probabilities I don't actually know?

You estimate anyway, out loud, and then keep score. A rough number you can revise beats a confident hunch you never test. Tetlock's research shows calibration improves with feedback: assign the odds, write them down, check later how often your "70%" calls actually land. Over a few quarters your guesses get measurably better.

What's the difference between a bad decision and a bad outcome?

A bad decision is a poorly-reasoned bet, wrong odds, ignored downside, no plan B. A bad outcome is simply the unlucky result of any bet, including a good one. The flood was a bad outcome from a good decision. Judge people on the first; they don't control the second.

When does expected value break down?

Two cases. First, when you genuinely can't estimate the probabilities, a market that doesn't exist yet, a true one-off, the numbers can be guesses dressed as data. Second, and more dangerously, when a loss can ruin you: averages assume you get to play again, and ruin ends the game. In both cases EV is a thinking aid, not an oracle.

Do I have to do maths for every decision?

No. Reserve the explicit arithmetic for consequential, repeatable calls. For everything else, the habit is what matters: name the odds roughly, name the downside, check it can't sink you. The structure is the value, not the decimal places.

Where to go next

If you remember nothing else, remember this: a good decision is a good bet, not a good result. The flood, the failed launch, the hire that didn't work out, none of them tell you, on their own, whether the call was right. Reconstruct the bet, the odds, the payoff, the survivable downside, and you build the rarest thing in any organisation: a culture that backs its sound thinkers through a run of bad luck, and questions its lucky gamblers before the luck runs out.

Related in the Toolkit