Expected Value and the One-Shot Problem
Expected value is the probability-weighted average of a choice’s possible outcomes — the full definition and a couple of worked examples are covered in the practical guide to expected value. This piece zooms in on the single limitation that matters most and gets skipped most often: an average is a statement about many repetitions of the same gamble, and a great many of the decisions people actually reach for expected value to justify are not repeated at all. They happen once. Understanding exactly what that means — not just that it matters, but what specifically breaks — is the difference between using expected value well and using it as a false comfort.
A worked example: repair or replace
Your washing machine has developed a fault. A repair costs $220 and, if it holds, saves you the cost of replacing the machine outright, which would run about $700 — a net benefit of $480 compared with buying new right now. But there is a real chance the repair does not hold: say a 35% chance it fails again within a year, at which point you will have spent $220 for nothing and still need to buy the $700 replacement, a net loss of $220 relative to just replacing it today. Feeding those two outcomes into the expected value calculator — a 65% chance of a $480 net benefit, a 35% chance of a $220 net loss — returns an expected value of exactly $235, with the tool separately reporting the best outcome ($480, if the repair holds) and the worst ($−220, if it doesn’t).
On the strength of that $235, repairing looks like the better move. And in one specific, narrow sense it is: if this exact decision, with these exact probabilities, played out a hundred times, the repair option would leave you further ahead than always replacing outright, on average, over those hundred trials. But you are not making this decision a hundred times. You have one washing machine, one repair decision, and you are about to find out which of exactly two things happens: you keep $480, or you lose $220. The number $235 is not a third possible outcome sitting between them — it is not, in fact, a value that any single trial of this decision can actually produce. It is a description of the population of trials you are not going to run.
What "on average" actually requires
The law of large numbers — the mathematical fact that an average outcome converges toward the expected value as the number of trials grows — requires trials. Specifically, it requires many independent instances of the same gamble, with the same probabilities and the same payoffs, played out enough times for the good and bad draws to balance each other out. An insurer selling thousands of similar policies gets this. A shop pricing thousands of similar items gets this. You, deciding once whether to repair one specific washing machine, do not get this, no matter how sound your probability estimate is or how carefully you computed the arithmetic.
It is tempting to reach for a broader kind of repetition as a workaround — “well, I make lots of decisions like this over my lifetime, so it evens out eventually.” That argument works for genuinely similar, genuinely repeated small decisions: which of two short commute routes to take most days, say, where you really are facing close to the same gamble over and over. It works far less well the further a decision drifts from being truly repeatable in substance, not just in category. Repairing this washing machine, moving to this city, taking this specific job offer — each belongs to a category of decision you may face several times across a life, but each individual instance has its own specific probabilities, its own specific stakes, and its own specific one-time outcome. Averaging across a category of superficially similar decisions is not the same as averaging across repeated trials of one identical gamble, and treating the two as interchangeable is where the one-shot problem usually sneaks in unnoticed.
Reading the calculator's own output as evidence against itself
There is a useful trick for keeping this limitation visible: look at the calculator’s best and worst fields, not just the headline expected value. In the repair example, those are $480 and $−$220. Notice that the expected value, $235, is not close to either of them in any meaningful sense — it sits between two outcomes that are $700 apart, and your one actual trial is going to land on one of the two ends, not on the number in the middle. For a decision you will only make once, the honest way to read the tool’s output is: “I am going to get $480 or I am going to lose $220, with roughly a two-in-three chance of the better one.” That sentence describes your actual situation. “I am going to get $235” does not describe anything that can actually happen to you on a single repair. The expected value calculator will always report this middle number front and center, because that number is genuinely the most useful single summary of the decision as a category — but for a one-shot instance of it, the best/worst pair next to it deserves at least equal weight in your reading of the result.
How to tell whether you are actually facing a one-shot decision
Two questions do most of the work. First: will you make this exact choice, with roughly these odds and these payoffs, again? Not a similar-sounding decision in some other context — this one, close to as-is. If the honest answer is no, you are looking at a one-shot instance, whatever broader category it belongs to. Second, and more important for how much the distinction should worry you: is the worst-case outcome one you can absorb without lasting damage? A one-shot decision with a survivable downside is not especially dangerous to reason about using expected value — you might land on the worse of two outcomes, but you will be fine either way, and the EV number remains a reasonable tiebreaker among otherwise similar-looking options. A one-shot decision where the worst case is close to ruinous is the genuinely dangerous combination, because a good average is doing nothing to protect you from the one draw you are actually going to get.
A spectrum, not a binary
It is worth heading off an objection here: strictly speaking, no two decisions are ever perfectly identical, so in the strictest sense every decision is a one-shot event. That objection is true but not useful, because it erases a real and important distinction. Deciding which of two nearly identical commute routes to take is, for all practical purposes, the same gamble played out day after day — the traffic pattern varies, but the structure of the choice barely changes, and a hundred repetitions really do behave like a hundred draws from one distribution. Deciding whether to relocate for a specific opportunity is nothing like that; there is no meaningful sense in which you get to run that decision a hundred times and see how it averages out. Treat repeatability as a spectrum rather than a strict yes-or-no test, and place a given decision on it by asking how much the probabilities and payoffs would actually vary if you somehow got to repeat it. The closer a decision sits to the truly-repeated end, the more confidently you can lean on its expected value. The closer it sits to the truly-singular end, the more the spread of outcomes around that average deserves your attention.
What to reach for instead, on a genuine one-shot call
None of this means expected value is useless for a one-shot decision — it is still a clarifying way to organize your thinking about probabilities and payoffs, and it remains one useful input among several. But on a genuine one-shot call, especially one with a severe downside, it should not be the only input. A few complements are worth reaching for specifically because they do not collapse the spread of outcomes into a single number the way an average does. If you cannot confidently attach probabilities to the possible futures at all, a regret-minimization approach sidesteps the need for probabilities entirely and instead asks which option leaves you with the smallest worst-case regret. If the decision is one you could restructure to make more reversible — a trial period, a staged commitment, a pilot — the piece on reversible versus irreversible decisions covers how to convert a one-shot bet into something closer to a repeatable one, which is often the single best way to defuse the one-shot problem rather than merely reasoning around it. And if the honest issue is that the dollar values in your EV calculation do not reflect how much each outcome would actually matter to you, revisit the expected-utility discussion in the main expected value guide, which addresses exactly that gap.
The short version: expected value tells you what a category of similar gambles would average out to. On a genuine one-shot decision, you are not buying the average — you are buying one specific outcome from a specific distribution, and the size of that distribution's spread deserves at least as much of your attention as the number sitting in its middle.