Risk and Uncertainty
How to Think About What You Cannot Know
Introduction
You make decisions all day. Should you take the umbrella? Eat the leftovers? Sign the lease? Take the new job? Buy the index fund or the individual stock? Have the difficult conversation now or later? Almost every consequential decision involves outcomes you cannot fully predict, and almost every decision-making technique that humans have developed - from religious ritual to financial mathematics - is, in some sense, a way of dealing with what you cannot know.
Most people think they are good at handling uncertainty. The research is consistent that they are not, in specific predictable ways. The gap between how we feel about risk and what the math actually says is one of the most-studied subjects in modern psychology and behavioral economics. Understanding the gap does not make it disappear. But it does change which choices you are willing to make, which assumptions you are willing to challenge, and which advice you are willing to take from confident-sounding people who do not know what they think they know.
The page below tries to walk through the working version of risk and uncertainty. It covers what these terms mean technically, why the distinction matters, what the math says, where human intuition reliably fails, and what to do about it. The aim is the kind of practical understanding that helps with everyday decisions about money, career, family, and health, not just with abstract probability problems.
Risk vs Uncertainty: A Distinction That Matters
In 1921, the economist Frank Knight published a book that introduced one of the most important distinctions in modern decision-making. "Risk," in Knight's framework, refers to situations where you do not know the outcome but you do know the probabilities. Rolling a die is risk: you cannot predict what number will come up, but you know each face has a one-in-six chance. Insurance pricing is risk: a life insurer cannot predict which policyholder will die when, but the actuarial tables describe the probability distribution well enough to set premiums.
"Uncertainty," in Knight's framework, refers to situations where you do not even know the probabilities. Will artificial intelligence transform employment over the next ten years? Will this startup succeed? Will your country be at war in twenty years? Will the next pandemic emerge in 2027 or 2050? Will the climate transition produce more economic disruption than expected or less? These are not risk in the technical sense; they are uncertainty. There is no probability distribution to apply because the future of these processes cannot be statistically characterized from past data.
Why the distinction matters. The mathematical machinery of risk - expected value, variance, standard deviation, correlation, the central limit theorem - works only when the underlying probabilities are known or can be estimated from enough data. Applying it to uncertainty in Knight's sense is not just inaccurate; it is misleading. The financial-engineering models that contributed to the 2008 financial crisis assumed that mortgage default rates followed a known statistical distribution; the actual world produced correlations and tail behavior that the models had no way to characterize. The same pattern shows up in many other domains: when people apply risk-mathematics to uncertainty, they get falsely precise answers and often end up worse off than they would have been with cruder but more honest tools.
What to do under uncertainty. First, lower your confidence. Make smaller bets. Diversify across futures rather than betting on the most likely one. Build in slack so you can absorb surprises. Avoid commitments you cannot unwind if the world turns out differently than you expected. None of these is as satisfying as a confident probability estimate, but they are more useful than false precision when the underlying uncertainty is real. Most consequential life decisions involve some risk and some uncertainty mixed together; recognizing which is which is itself one of the most important moves.
Expected Value and Why It Is Not Enough
The basic mathematical tool for thinking about risky decisions is expected value: the average outcome of a decision, weighted by the probability of each outcome. A coin flip that pays $10 if heads and $0 if tails has an expected value of $5. A lottery ticket that pays $100 million with probability one-in-300-million has an expected value of about 33 cents - well below the price of the ticket, which is why playing the lottery is a losing proposition on average. The expected-value calculation works well for repeated decisions, where the law of large numbers can play out. An insurance company writing millions of policies will see its average loss per policy converge on the expected value over time.
But expected value is not enough for one-off decisions. Consider a coin flip that pays $1 million on heads and costs you everything you own on tails. The expected value depends on what "everything you own" is. For most people, the math says they should not take that bet even if the expected value is positive, because the downside (ruin) is much worse for them than the upside (a million dollars) is good. The principle is called "diminishing marginal utility of wealth," and it explains why most ordinary decisions about money are governed by the variance of outcomes, not just the expected value.
Daniel Bernoulli worked this out in 1738, in answer to the St. Petersburg paradox that his cousin Nicolaus had posed in 1713. The paradox is a hypothetical bet whose expected value is mathematically infinite but that almost no one would actually pay much for, because the probability of the very large outcomes is so low and the probability of small outcomes is so high. His answer was to value each outcome by its utility, which grows more slowly than wealth, rather than by the wealth itself - a principle that underlies most modern decision theory and is one of the deeper insights about how to think under risk.
The practical implication. For decisions you will make many times (small recurring bets, ordinary saving and spending, repeated business decisions), expected value is a useful guide. For one-off decisions with potentially large outcomes (whether to buy a house, take a job, start a business, undergo surgery, marry someone), you need to think about the full distribution of possible outcomes, not just the average. If losing what you cannot afford to lose is on the table, weight the bad outcomes more heavily than the good ones. This sounds obvious; it is constantly violated in actual decision-making, particularly when people are excited about possible upside.
Insurance as a Technology for Handling Risk
Insurance is one of the most useful inventions in human economic history, and one of the least appreciated as a technology rather than just a financial product. The basic idea is simple: many people each face a small probability of a large loss. Each person pays a small fee that, pooled across the group, covers the actual losses suffered by the unlucky few. Each individual gives up a little expected value (the insurance company has to make a profit) in exchange for not facing catastrophic risk on their own.
What makes insurance work mathematically. The law of large numbers says that as the pool grows, the average loss per policyholder converges on the expected loss. A small insurance company with 100 customers can be wiped out by an unlucky year. A large insurance company with 10 million customers cannot be, because the variation in any single year averages out across the pool. This is why insurance companies are large by design - their fundamental product depends on scale.
What can break insurance. Two things in particular. First, correlated risk: if everyone's loss happens at the same time (a massive earthquake hitting all the policyholders, a pandemic affecting all employers, a financial crisis correlating defaults), the law of large numbers no longer applies, because the variation does not average out. Some kinds of risk are uninsurable for this reason; flood insurance has often needed government backing because too many losses in a single event can wipe out private insurers. Second, asymmetric information: if customers know more about their risk than the insurer does, the insurer prices the average customer's risk while attracting disproportionately high-risk customers. This is "adverse selection" and is one of the central problems in healthcare insurance design.
The mental model of insurance applies far beyond financial products. Diversifying your investment portfolio is a form of insurance - you give up the chance that one concentrated bet pays off spectacularly, in exchange for protection against any single bad outcome. Maintaining multiple income streams is a form of insurance against losing one. Building professional relationships across industries is a form of insurance against your own industry contracting. Most reliable risk-management strategies in life are, in some way, insurance-shaped. Understanding the structure helps with applying it.
How Humans Perceive Risk
Human intuition about risk is reliably misleading in several specific directions, and most of them have been documented across decades of behavioral-economics research. Daniel Kahneman and Amos Tversky's work, published in many academic papers and summarized in Kahneman's 2011 book "Thinking, Fast and Slow," mapped the systematic gaps. A short tour of the most consequential ones.
Loss aversion. Losses weigh more heavily than equivalent gains. The familiar "about twice as much" figure comes from Kahneman and Tversky's estimates and is best treated as the upper end of a wide range rather than a constant. As the Decisions page sets out, later work found that the ratio varies widely with context, stakes, and how the choice is posed, and in some settings nearly vanishes. What holds up best is the direction and the reference-point structure. Offered a coin flip that pays $150 on heads and costs $100 on tails, many people refuse despite its positive expected value. That produces decisions which look mathematically irrational but feel emotionally normal. It helps explain why people hold on to losing investments rather than realizing the loss, why governments under-invest in long-term changes that produce short-term losses, and why people tolerate a worse known situation rather than risking an unknown improvement.
Availability heuristic. People estimate the probability of an event by how easily examples come to mind. Plane crashes are much more memorable than car accidents, so many people overestimate flying risk and underestimate driving risk - even though driving is far more dangerous per mile. Spectacular but rare events (terrorism, kidnapping, shark attacks) are systematically overestimated; common but quiet ones (heart disease, depression, falls in the bathroom) are systematically underestimated. Politics, marketing, and journalism all interact with the availability heuristic in ways that distort what people fear and what they prepare for.
Probability neglect. When people consider an outcome that produces a strong emotional reaction, the actual probability of the outcome often drops out of consideration. People worry about extremely rare but emotionally salient events (their child being kidnapped) more than about much more probable but less salient ones (their child being injured in a car accident). The decisions that follow from this neglect are often expensive (driving instead of flying after a crash, overinvesting in security against rare threats) without measurably reducing actual risk.
Base-rate neglect. When people are given specific information about a case, they often ignore the base rate at which the case appears in the broader population. The classic example: if a screening test for a rare disease is 99% accurate and the disease appears in 1 in 1000 people, a positive result is more likely to be a false positive than a true positive - because most positive results come from the much larger population of healthy people. Most people get this badly wrong on first encounter. The same logic applies to medical diagnoses, security screening, hiring decisions, and many other settings where base rates matter.
Optimism about the self, pessimism about others. Most people consistently overestimate their own driving ability, business prospects, marriage durability, and life expectancy relative to averages. The same people often correctly estimate (or even pessimistically estimate) the same outcomes for other people in similar situations. The asymmetry produces a systematic bias toward over-confident planning.
Black Swans and Fat Tails
Some kinds of risk are not well-described by the bell-shaped probability distributions that most introductory statistics courses cover. Many real-world processes have "fat tails" - extreme events are much more likely than a normal distribution would predict. Stock-market crashes, earthquakes, war casualties, financial panic spreading, and a long list of other phenomena follow distributions where the tail is much heavier than the bell curve suggests. The implication is that the rare-but-large events can dominate the long-run outcomes more than the average events do.
Nassim Taleb's "Black Swan" framing made this distinction widely known. A "black swan" is an event that is rare, has high impact, and is rationalized after the fact as if it had been predictable - even though predicting it before the fact would have required information that was not available. The September 11 attacks and the 2008 financial crisis had this character, as, arguably, did the rise of generative AI as a mass-market product. Few people saw them coming, yet in retrospect the signs looked clear. Taleb himself rejected the label for the COVID-19 pandemic, calling it a "white swan" because pandemics had long been predicted.
The strategic implication of fat tails. In domains where the tail dominates - financial markets, geopolitics, technology, careers in winner-take-all fields - the average outcome is a poor guide to what to expect. The few extreme events (the 2008 crisis, the dot-com crash, the rise of a single dominant company in a field) shape the long-run outcomes far more than the steady years. Strategies built on smoothing the average outcome miss the tails. Taleb argues that strategies which explicitly position for tails - through diversification, optionality, skin-in-the-game, and avoiding fragility - do better over multi-decade horizons, even when they look inferior in the steady years. Critics such as the fund manager Cliff Asness answer that protection has a price: his firm's research found that buying crash insurance through stock-index options lost money in each decade from 1985 to 2020, crashes included.
The practical move is what Taleb calls "antifragility" - building systems that benefit from volatility rather than just surviving it. A diversified portfolio with some exposure to assets that gain during crises (volatility, gold, certain non-correlated investments) can profit from a shock that would only damage an undiversified one. A career that builds skills usable across industries is antifragile against the decline of any single industry. A life that has multiple sources of meaning, relationship, and purpose is antifragile against the loss of any one of them. The principle generalizes beyond finance: the more parts of your life depend on a single specific outcome continuing, the more fragile you are to its disruption.
Base Rates and the Outside View
Base rates are the average rates at which things happen in similar situations. One of the most reliable findings in decision research is that people systematically underweight them in favor of specific information about the particular case in front of them. The result is consistent over-optimism about how long projects will take, how successful new businesses will be, and how durable new relationships will be.
The "planning fallacy" is the canonical example. When asked to estimate how long a project will take, people give estimates that are systematically optimistic. In a classic 1994 study by Roger Buehler and colleagues, students expected to finish their honors theses in 34 days on average; they took 56, and fewer than a third finished by their own best estimate. The pattern is robust across domains: software projects, construction projects, academic research, personal renovations, government infrastructure. The estimates do not improve much with experience. People who have learned that their previous projects took longer than expected still underestimate the next one in the same way.
What works to correct this. Kahneman and others have called it "taking the outside view" - looking at the base rate of outcomes across similar situations, then asking what would make your specific situation different. If about four in ten new US restaurants close within five years, your specific business plan needs to explain what is unusual about it; the default expectation should be the base rate. If most venture-backed startups never return their investors' capital, the question for any specific startup is what puts it in the minority. Shikhar Ghosh of Harvard Business School put that share at about three in four for US venture-backed companies - lower than the popular "90% of startups fail," and based on a stated definition of failure. If 40-50% of US first marriages end in separation or divorce, the question is what distinguishes your relationship from the average, not an assumption that yours will be the exception.
This discipline is uncomfortable. Most people, asked to take the outside view of their own plans, find that the base rates are much less favorable than they have been assuming. The right response is rarely to abandon the plan; it is to plan more carefully for the most common failure modes, build in fallbacks for the steps most likely to go wrong, and not over-commit on the assumption that everything will go right. Projects that survive contact with the base rate, with appropriate adjustments, tend to do much better than projects that ignore it.
Variance, Ruin, and the Asymmetry of Survival
Some risks have a special property: they cannot be undone. Going broke once means you cannot make any more bets. Dying once ends the sequence. Destroying a critical relationship, a critical professional reputation, a critical environmental resource has consequences that do not reverse on the same timescale as ordinary fluctuations. Risks of this kind are sometimes called "ruin risks" - risks that, if they materialize, end the game rather than producing a setback within it.
The asymmetry matters because expected-value reasoning fails for ruin risks. A bet that has positive expected value over many repetitions can still be a terrible idea if the outcome of any single bet includes a chance of ruin. The "Kelly criterion" in betting theory makes this precise: the optimal fraction of your wealth to bet, even on a positive-expected-value bet, is much smaller than people intuit, because the path-dependence of repeated bets means that a streak of bad outcomes can wipe out everything you have. Most professional gamblers and investors understand this in their bones; most amateurs do not.
The practical implication is that some kinds of decisions are not just decisions about averages. They are decisions about whether to expose yourself to ruin. Most drunk drivers arrive home safely most of the time, but the small probability of a catastrophic outcome makes drunk driving a bad bet over a lifetime of similar choices. Concentrating all of your savings in a single investment has positive expected value if the investment is reasonable, but the variance can produce ruin. Taking on credit-card debt at 20% interest for short-term spending has positive utility for a few weeks but can compound into ruin over years. The rule is something like: avoid bets that include any meaningful probability of irreversible harm, even if the expected value is positive.
What this looks like in practice. Maintain emergency reserves. Diversify your sources of income. Maintain multiple close relationships rather than centering all emotional reliance on one. Avoid health behaviors that have catastrophic-failure modes (drinking and driving, heroin, ignoring serious symptoms). Build optionality so that bad outcomes in any one domain do not destroy the others. The pattern across all of these is the same: protect yourself against ruin even at the cost of some expected value, because a wider distribution that includes ruin is much worse than a narrower one without it.
Working Practically With Risk and Uncertainty
The same disciplines repeat across every domain, which is the real payoff of thinking in base rates and tails. Diversify - across investments, income sources, skills, and sources of meaning - so that no single bad outcome is ruinous. Buy optionality, like an emergency fund or transferable skills, because the freedom to walk away or wait is one of the most valuable returns in life. Take the outside view on your own plans, health, and relationships: assume the base rate applies to you unless you can name what genuinely makes your case different, then plan for the most common failure modes rather than the optimistic case. And lower your confidence in confident predictions - be willing to say "I don't know", build in slack, and avoid commitments you cannot reverse if the world turns out differently than you expected.
Most people are confidently wrong about risk and uncertainty in specific predictable ways. The biases are not failures of intelligence; they are how the human mind reliably operates. What changes outcomes is the discipline to notice when the situation is governed by base rates the inside view is ignoring, when the math says diversify even though the heart says concentrate, when an attractive bet includes ruin in its tails, and when "I don't know" is the most accurate answer available. None of this is glamorous. The compounding return of practical risk discipline is one of the largest under-recognized advantages in finance, career, health, and life over a span of decades.
