Expected value: how professionals make peace with uncertainty
The number you should optimize for isn't the most likely outcome — it's the probability-weighted average of all outcomes, and ignoring that difference is the hidden tax on every bad business decision.
A venture capital fund manager once told me that her job was to be wrong most of the time and still beat the market. Her fund’s last vintage had a 65 percent loss rate across portfolio companies — and returned 3.1x to investors. That sounds contradictory until you think in expected value.
The rest of the business world mostly doesn’t. It anchors to the most likely single outcome, which is a cognitive bias dressed up as prudence. It’s why marketing budgets get killed the first time a campaign doesn’t convert, why companies under-invest in redundancy until a data center burns down, and why smart people routinely turn down bets that would make them richer in any honest accounting.
What expected value actually is
Expected value (EV) is the probability-weighted average of all possible outcomes. Formally, if an event has outcomes O_1, O_2, ..., O_n with probabilities p_1, p_2, ..., p_n, then EV = p_1 * O_1 + p_2 * O_2 + ... + p_n * O_n. The probabilities must sum to 1.
That formula sounds like a statistics textbook. The intuition is simpler: if you ran the same decision a thousand times, what would the average result be? EV is that average.
Here’s the worked example that makes this concrete. Your growth team proposes a $50,000 marketing campaign. Internal analysis suggests a 40 percent chance it generates $200,000 of extra gross profit (the revenue left after the cost of goods sold), and a 60 percent chance it generates nothing. Should you run it?
The EV of the gross profit outcome is 0.40 * $200,000 + 0.60 * $0 = $80,000. Subtract the $50,000 campaign cost: net EV is +$30,000. The campaign costs half as much as its expected return. You should run it.
Notice what just happened: the most likely single scenario — 60 percent odds — is that the campaign returns nothing and you burn fifty grand. A manager anchored to that scenario would kill the campaign. An EV thinker runs it, because across the space of possibilities, it creates value.
Why this is genuinely hard to do
The obstacle isn’t mathematical — the arithmetic is fifth-grade level. The obstacle is psychological. Humans are loss-averse (a $50,000 loss feels roughly twice as bad as a $50,000 gain feels good, a finding from Kahneman and Tversky’s prospect theory work). We’re also bad at intuiting probabilities below about 10 percent, which means we systematically over-insure against tiny risks and under-prepare for moderate ones.
Organizational dynamics make it worse. A manager who runs the campaign and it returns nothing gets blamed. A manager who kills the campaign never faces the counterfactual. The incentive structure actively punishes EV thinking. This is the reason so many corporate decisions feel risk-averse in aggregate even when individual managers are nominally rational: the accountability structure is indexed to outcomes, not to the quality of the decision given the information available at the time.
There’s also the problem of probability estimation itself. Where did that 40 percent number come from? In practice, it comes from a spreadsheet someone built with assumptions baked in. EV analysis is only as good as the probabilities you feed it, and those probabilities are often optimistic (startups routinely model a 30 percent chance of success on campaigns that have a 5 percent base rate). Garbage probabilities produce a number that looks precise and is mostly noise. The discipline of EV thinking therefore comes bundled with the discipline of calibration — the practice of assigning probabilities that match historical base rates, not wishful thinking.
The law of large numbers is doing the heavy lifting
EV is meaningful because of the law of large numbers: over many trials, the average result converges to the expected value. Run the campaign a hundred times and the average gross profit per run approaches $80,000. The math holds.
The problem is that “many trials” does real work in that sentence. A startup running its first-ever marketing campaign doesn’t have the luxury of averaging over many trials. A once-in-a-decade infrastructure investment isn’t a repeated experiment. When you’re making a one-shot decision, the distribution matters — not just the mean, but the spread.
This is why professional risk managers care about variance (how spread out the outcomes are) and not just expected return. Two bets can have identical EV and wildly different risk profiles. Consider:
| Bet | Outcome A | Probability A | Outcome B | Probability B | EV |
|---|---|---|---|---|---|
| Alpha | +$100,000 | 0.5 | -$100,000 | 0.5 | $0 |
| Beta | +$10,000 | 0.5 | -$10,000 | 0.5 | $0 |
Both have EV of zero, but Bet Alpha has one hundred times the variance. For a small company where a $100,000 loss triggers a liquidity crisis, Bet Beta is vastly preferable despite identical EV. The variance of Alpha could end the game before the law of large numbers has any chance to operate.
This is the caveat that professionals internalize and novices skip: EV is the right objective when you can make the bet repeatedly. When a single loss means ruin, you must also manage variance.
Decision trees: making EV mechanical
Once you accept EV as the right framework, decision trees are the natural tool for applying it to multi-stage problems. A decision tree maps out a sequence of choices (squares, by convention) and chance events (circles), assigning probabilities to each branch of a chance event and a terminal value to each leaf.
The key operation is “folding back” or “rolling back” — you start from the right-hand side, compute the EV at each chance node, and work leftward. At each decision node, you pick the branch with the highest EV. The process is mechanical once the tree is built, which is the point: it removes the intuition bias from multi-step problems.
A product team deciding whether to run a $20,000 pilot before committing to a $200,000 full launch can draw this out explicitly. The pilot has some probability of providing a positive signal (which updates the probability of the full launch succeeding) and some probability of a negative signal (which avoids a bad investment). Fold back: what’s the EV with the pilot vs. without it? The pilot’s cost is justified if it provides enough information to shift the probability estimate enough to change the decision.
The technique makes one thing brutally clear: information has economic value. Any data you can gather before committing to a decision has value equal to the probability of it changing your decision times the size of the gain from making the right choice. This is why A/B testing, pilots, and staged rollouts aren’t risk aversion — they’re EV maximization.
Where EV thinking actually lives in industry
Poker players formalized EV thinking before most business schools did. In poker, “pot odds” is the ratio of the pot size to the call amount — a compact representation of EV applied to whether to continue with a hand. Professional players don’t think in terms of “I think I’m ahead”; they think in terms of “my equity in this pot at this probability justifies this call.” It’s mechanical EV.
Insurance pricing is EV from the insurer’s side. The premium is set above the expected payout (plus overhead and profit margin) so that the insurer’s EV across thousands of policies is positive. The policyholder’s EV is technically negative — they’re paying more than the expected claim — but they’re buying variance reduction, not positive EV. That’s rational too: a homeowner who can’t absorb a $400,000 house fire should pay $2,000 a year to transfer that variance to someone who can. EV and variance are both legitimate objectives; the trick is knowing which one your situation calls for.
Pharmaceutical R&D is run on EV thinking under a different name: portfolio management. A drug in Phase 1 trials has a roughly 10 percent chance of reaching market approval. A Phase 3 drug has roughly a 65 percent chance. The probability-adjusted net present value of a pipeline (a metric called rNPV, for risk-adjusted NPV) tells you what the portfolio is worth in expectation. That number drives capital allocation decisions — which programs to fund, which to kill, which to acquire.
The common thread: industries where the outcomes are uncertain and the sample sizes are large enough for the law of large numbers to operate have converged on EV as the organizing principle.
The ruin problem is the most important caveat you’ll ever hear in this domain
Kelly’s criterion — developed by John Kelly at Bell Labs in 1956 and later popularized by gamblers and traders — formalizes the ruin problem. The Kelly criterion (the optimal fraction of your bankroll to bet on any given opportunity) equals edge / odds, where edge is how much you’re ahead in expectation and odds is the payoff ratio. Crucially, Kelly bets a fraction, never the full bankroll, even on a positive-EV bet.
Why? Because variance compounds. A 50 percent loss requires a 100 percent gain to recover. Two 50 percent losses require a 300 percent gain. The arithmetic of consecutive losses is asymmetric and brutal. If you bet 100 percent of your bankroll on a positive-EV bet and lose, the law of large numbers never gets to help you — you have nothing left to run the next trial.
This is why professional gamblers, hedge funds, and venture firms think about position sizing — how much of their total capital to deploy on any single bet — as seriously as they think about which bets to take. A bet can be good and still be run at the wrong size. The expected-value framework doesn’t tell you how much to risk; Kelly does.
For mortals making business decisions, the practical version of this insight is: never make a bet where the downside scenario threatens the existence of the enterprise, no matter how favorable the EV. The asymmetry of ruin means you don’t get to average across many trials if trial one kills you.
How to actually do this
Applying EV thinking doesn’t require a graduate probability course. It requires three habits.
First, enumerate the scenarios. Not just the base case — the optimistic case and the pessimistic case, with rough probability weights on each. Forcing yourself to name probabilities is uncomfortable, but the discomfort is the point. It exposes assumptions that were previously hidden.
Second, multiply and sum. Take each scenario’s outcome, multiply by its probability, and add them up. Do this for net cash, not gross revenue — the unit has to be comparable across scenarios. Write the number down. Is it positive? By how much? Compared to the cost of the bet?
Third, check your ruin exposure. What does the downside scenario actually do to you? If the worst case is painful but survivable, EV dominates. If the worst case ends the company or your career in a way that forecloses future bets, cut the bet size or don’t take it regardless of EV.
The magic isn’t in the arithmetic. It’s in the forced explicitness. Most bad decisions are made bad not because of math errors but because the decider never clearly stated what they thought the probabilities were. Once you write them down, other people can push back, and you can track whether your probability estimates are calibrated over time.
What the VC story actually teaches
The fund manager at the start of this essay wasn’t making a virtue of being wrong. She was making a claim about the shape of the distribution in her asset class. Venture returns follow a power law: a small number of investments return 10x to 100x, and most return nothing. In that environment, optimizing for win rate is actively harmful — it biases you toward safer companies with lower upside. You want positive-EV bets, not high-probability bets.
Most business decisions don’t have power-law payoff structures. But some do — new product lines, international expansion, acquisitions — and those are precisely the decisions where organizations most often revert to modal thinking and kill bets that would have paid out.
The discipline of EV thinking is, at root, a discipline of intellectual honesty about uncertainty. It doesn’t eliminate uncertainty. It prevents you from pretending it doesn’t exist, which is the thing that actually gets you killed.
The 60 percent chance the campaign returns nothing is real. So is the 40 percent chance it returns four times its cost. Holding both in mind at once, multiplying each by its probability, and making the call anyway — that’s not optimism. That’s arithmetic.