Expected Value: Weighing the Whole Bet

A slot machine knows exactly what it is doing, and so does the casino that bought it. The machine pays out most of the money it takes in — 92 cents on the dollar, say — which feels almost generous until you notice the casino keeps the other 8 cents on every dollar, forever, across millions of pulls. No single spin tells you this. A spin is a coin flip with confetti: sometimes you win big, usually you win nothing, and the noise drowns out the trend. The casino doesn’t watch the spin. It watches the average, and the average is a machine for printing 8% of everything that walks in the door.

That average is expected value, and learning to compute it is the difference between being the player and being the house. So far this course taught you to read a probability (lesson 01) and to anchor it to its base rate (lesson 02). Now we put a payoff on each outcome and combine the two into a single verdict: across many tries, is this bet worth it?

The one-line way to value a gamble

Here’s the analogy first. Imagine you could play a bet ten thousand times and then split the total winnings evenly across all ten thousand plays. The per-play share you’d walk away with — that’s the bet’s expected value. It’s not what happens on any single try (you might win, you might lose); it’s the long-run average each try is worth.

Precisely: the expected value (EV) of a decision is the sum of every possible outcome’s payoff, each weighted by the probability that it happens. Its home discipline is probability and decision theory, and the formula is one line:

$EV = \sum_i p_i \times x_i$

where $x_i$ is the payoff of outcome $i$ and $p_i$ is its probability (a number between 0 and 1, so that all the $p_i$ sum to 1). You take each payoff, multiply it by how likely it is, and add up the results. That’s the whole operation. The $\times$ is doing the real work: it shrinks payoffs you’re unlikely to get and protects payoffs you probably will.

A tiny first example. You pay nothing to roll a fair six-sided die, and you’re paid in dollars equal to the number that comes up. Each face has probability $1/6$. Step through it:

The EV is $3.50 — which is not a payoff you can ever actually roll. There’s no 3.5 on a die. That’s the first thing to internalize: expected value is an average, not a prediction. No single roll pays $3.50, yet $3.50 is exactly what each roll is worth in the long run. If someone charged you $3 to play this game, you should play all day; at $4, you should walk away.

Now turn the same machine on the lottery from the pretest. The numbers are designed to feel generous and to bleed you anyway.

Read the contribution column. The jackpot’s payoff is gigantic — $10,000,000 — but multiplied by its probability ($0.0000001$) it contributes just $1 to the sum. The “nothing” outcome contributes $0. So the EV of the ticket is $1, and it costs $2. Net expected value: −$1 per ticket. Buy a ticket a day for a lifetime and the long-run average is that you set fire to a dollar each morning. This is why a lottery is sometimes called a tax on people who can’t (or won’t) compute expected value — the state sells a vivid $10M dream for $2 and pockets $1 of it, on average, every single time.

EV is what you’d average over many repeats

Why should one number born from multiplication tell you how to act on a single, one-time bet? Because of the law of large numbers: as you repeat an independent random trial more and more times, the average of your actual results converges to the expected value. One play is pure noise — you win or you don’t, and EV looks like an abstraction. But stack up the plays and the noise cancels while the EV accumulates.

Picture the die game at $3 a play. Roll once and you might get a $1 (a $2 loss) or a $6 (a $3 win) — wild swings. But your average payoff per roll marches steadily toward $3.50:

The casino lives in the bottom row. It doesn’t care that you might hit a jackpot tonight — across its millions of plays, its average result is nailed to the EV, and it set the EV in its favor. The gambler lives in the top row, where the noise is big enough to feel like a “system.” Expected value is the promise that, given enough repetitions, the math always shows up to collect.

Negative-EV vs positive-EV bets

Sort every bet by the sign of its net expected value and you get a clean taxonomy. A positive-EV bet pays you to take it (over the long run); a negative-EV bet charges you for the privilege. The whole game of rational gambling, investing, and insuring is: take positive-EV bets, refuse negative-EV ones — with one important asterisk we’ll get to.

The lottery and the casino are obvious: their entire business model is selling you negative-EV bets dressed as fun. The roulette wheel pays $36 on a $1 bet but has 38 pockets, so its EV is about 94.7 cents per dollar — a 5.3% house edge that, by the law of large numbers, is as reliable as gravity.

Insurance is the interesting one. The premium is also negative-EV — that’s how the insurer makes money; on average you pay more than you’ll ever collect. And yet buying it is often the rational move. Why? Because EV isn’t the only thing you care about when an outcome can ruin you. A house fire isn’t a −$300 average annoyance you can shrug off over many repetitions — it’s a single, un-repeatable catastrophe that can wipe you out. You happily pay a small, certain, negative-EV premium to delete a rare but unsurvivable tail. That’s not a failure to compute EV; it’s recognizing that you don’t get to “average out” a loss you can’t come back from — a thread we’ll pull hard in the lessons on ranges and on decision-versus-outcome.

Asymmetric bets

Most of the time payoffs and probabilities roughly trade off — big prizes are rare, small prizes are common. An asymmetric bet breaks that symmetry on purpose: the downside is small and capped, while the upside is enormous (or the mirror image — a tiny chance of a catastrophic loss). When the structure is lopsided, a single term in the EV sum can dominate everything else, and the bet can be wildly positive-EV even though you lose most of the time.

This is the logic behind a cheap experiment, a startup bet, an out-of-the-money option, a cold email to a dream employer. You’ll probably get nothing. But “probably nothing” multiplied by a small, fixed cost is cheap, while “rarely, an enormous win” multiplied even by a tiny probability can be large. Consider a $500 experiment with a 4% chance of returning $30,000 and otherwise returning $0 — concretely, a small ad test for a product idea.

Run the numbers: $30,000 × 0.04 = $1,200 of expected value, minus the $500 cost = net +$700 per experiment. You will fail 24 times out of 25 — and you should still take this bet every chance you get, because the 1-in-25 win is large enough to pay for all the failures and then some. The trick is that the downside is capped at $500 (you can’t lose more than the test costs) while the upside is 60× that.

Two connections make this rigorous rather than romantic. First, base rates (lesson 02): that 4% is doing enormous work, and you don’t get to wish it higher. If your real base rate for “ad test for a random idea hits” is 0.5%, the EV becomes $30,000 × 0.005 = $150 against a $500 cost — net −$350, a bad bet. The probability you plug in must come from honest base rates, not optimism, or the asymmetry flips against you. Second, opportunity cost: the $500 and the time aren’t free — the right comparison is this experiment’s net EV against the net EV of the best thing you’d do with the same money. Asymmetric bets are powerful precisely because they’re cheap; that’s what keeps their opportunity cost low.

Where EV lies to you

Expected value is the right default, and it still misleads in three specific ways. Knowing them is what separates a thoughtful EV user from a reckless one.

1. One-shot games where you can’t survive to “average out.” EV is a promise about the long run, redeemable only if you’re still in the game to collect. If a bet has a positive expected value but a real chance of wiping you out, you may never reach the long run. The classic trap: a coin flip where heads doubles your entire net worth and tails loses all of it. The EV is positive (you gain 50% of your wealth on average per flip), yet repeat it and you go broke with near-certainty, because a single tails ends the sequence forever. This is the ruin problem, and it’s why a positive EV is necessary but not sufficient: never risk what you can’t afford to lose, even at attractive odds. Variance — how wildly outcomes swing around the EV — matters enormously when the downside is fatal, even though the EV formula ignores it completely.

2. Garbage-in probabilities. EV is only as good as the $p_i$ and $x_i$ you feed it. The arithmetic is flawless and the inputs are often guesses. We saw it above: nudge the experiment’s hit rate from 4% to 0.5% and a +$700 bet becomes a −$350 one. A precise-looking EV computed from invented probabilities is more dangerous than no number at all, because it launders a guess into a decision. Anchor every $p_i$ to a base rate (lesson 02) and treat a single point-estimate EV with suspicion — better to compute it across a range of plausible probabilities, which is exactly the next lesson.

3. Un-aggregatable, unique stakes. EV assumes outcomes are commensurable — that you can put every payoff on one shared scale and add them up. Some stakes refuse. The “payoff” of a medical decision, a marriage, a once-in-a-lifetime opportunity isn’t always a number you can multiply, and there’s no “many repetitions” to average over. When the stake is unique and the payoffs aren’t truly on the same scale, the EV number is a metaphor at best, and leaning on it as if it were a bank balance is its own failure mode.

When to reach for it

Reach for expected value whenever a decision is repeated or poolable and its payoffs are quantifiable on a shared scale: pricing, betting, insuring, A/B testing, building a portfolio of bets, any choice you’ll face many times. In those settings EV is the correct comparison — rank your options by net expected value, not by their best case (which flatters the lottery) or their worst case (which would forbid every insurable risk). Then layer on the survival check: among the positive-EV options, prefer the ones whose downside you can absorb. EV chooses which bets are worth making; the ruin question decides how much to stake on each.

The handle to carry forward

You now have the one number that combines probability and payoff: expected value, the sum of each outcome’s payoff weighted by its chance, equal to what each play is worth averaged over the long run. You’ve seen it expose the lottery as a −$1 tax, justify a negative-EV insurance premium, and bless a bet you lose 96% of the time. And you’ve seen its three blind spots — ruin, garbage inputs, and unique stakes — which is why “maximize EV” is a default, not a commandment.

The biggest of those blind spots is that a single EV number hides the spread it was averaged from. Two bets can share an EV of $700 while one is a near-sure thing and the other is a coin-flip between glory and ruin. Next, in Thinking in Ranges, we stop collapsing the spread to one number and start reasoning about the whole distribution — best case, worst case, and how much the outcome can swing — which is exactly the missing piece that tells you how much to bet, not just whether to.