Expected Value Is a Lie Your Brain Can't Process

Expected value is supposed to be simple. Multiply each outcome by its probability, sum the results, done. A coin flip where heads pays $150 and tails costs $100 has an expected value of $25. Take that bet every time. Casino textbooks love this example.

Photo by Brett Jordan on Pexels.

Here's the problem: almost nobody actually behaves this way, and in many real situations, they're not wrong to deviate.

Expected value is a tool built for a world where you can repeat a bet infinitely many times. In that world, it works perfectly, the law of large numbers guarantees your average outcome converges to the expected value. But you don't live in that world. You live in a world with one career, one retirement account, one liver, and a finite number of decisions before you die.

When repetition isn't available, expected value can actively mislead you.

The St. Petersburg Trap

In 1738, Daniel Bernoulli posed a puzzle that still doesn't have a clean resolution. A casino flips a coin until it lands heads. If heads appears on the first flip, you win $2. Second flip: $4. Third: $8. The payout doubles each round. Your expected winnings from this game are infinite, the series diverges.

So how much would you pay to play?

Most people say somewhere between $20 and $50. Very few say more than $100. Almost nobody says infinity, which is what pure expected value theory demands.

Bernoulli's fix was to introduce utility, the idea that an extra dollar means less to you as you get richer. That's a reasonable patch, but it just shifts the problem: now you need a utility function, and choosing the wrong one produces equally absurd results.

The deeper issue isn't psychological. It's mathematical. Expected value averages over an ensemble of parallel universes. You experience exactly one.

Ergodicity: The Word That Changes Everything

Physicist Ole Peters has spent years making a simple but devastating point about how economics applies probability. A process is ergodic when your time average (what happens to you over time) matches the ensemble average (what happens across many simultaneous versions of the scenario). For ergodic processes, expected value is the right guide. For non-ergodic processes, which describes most of actual life, it's a category error.

Consider a gamble where you gain 50% on heads and lose 40% on tails, coin flip each round. Expected value per round: +5%. Sounds good. But watch what happens over time to a single player:

After one heads and one tails (in either order), your multiplier is 1.5 × 0.6 = 0.9. You've lost 10% of your stake despite a positive expected value. Repeat this long enough and you go broke with probability 1, even though the ensemble average keeps rising, driven by a small fraction of players who hit long winning streaks and inflate the mean.

The mean, again, lying to you.

graph TD
    A[Single Bet: +EV] --> B{Repeated Over Time?}
    B --> |Yes, ergodic| C(Time avg = Ensemble avg)
    B --> |No, non-ergodic| D{Ruin possible?}
    D --> |Yes| E[/Positive EV can still destroy you/]
    D --> |No| F(EV is a safe guide)

What This Actually Means for Decisions

None of this means expected value is useless. It means you need to know when it applies.

Use expected value when: outcomes are truly repeatable, stakes are small relative to your resources, and no single loss can take you out of the game entirely. Insurance companies and casinos operate at scale, expected value governs their world.

Be suspicious of expected value when: you're making a one-shot decision, potential losses are catastrophic and irreversible, or the variance is so high that even favorable bets carry serious ruin risk. Your decision to take on debt, change careers, or launch a company does not become correct just because someone multiplied some probabilities together and got a positive number.

Kelly's Criterion gets at this more honestly than raw EV: size your bets as a fraction of your bankroll proportional to your edge. The math behind Kelly is explicitly about maximizing long-run growth rate, not expected value, and it tells you to bet far less than expected value alone would suggest. That conservatism isn't irrationality. It's the correct response to living one life instead of a thousand.

The Real Takeaway

Expected value gives you the right answer to a question you're rarely actually asking. The question it answers is: what would happen to the average outcome if this situation were repeated indefinitely across a population? The question you're usually asking is: what should I do, given that I have to live with the result?

Those are different questions. Treating them as identical is how smart people take bets they shouldn't, skip insurance they need, and build models that look rigorous while quietly ignoring ruin.

Run the math. Then ask whether the math is answering your actual question. Often, it isn't.