A few years ago I had to decide whether to insure a laptop worth around £600. The premium was roughly £8 a month. My instinct was that it felt like poor value, but I hadn't actually worked through the numbers. Later, thinking it through properly: for the insurance to be positive expected value, the probability of losing the laptop in any given year would need to exceed about 16% — which felt high. But the expected-value framing also missed the real question. The point of insurance isn't to be a good expected-value bet — it almost never is, or the insurer would go out of business. The point is to convert a potentially painful loss into a manageable, known cost. Whether that conversion is worth £8 a month depends entirely on how painful a £600 loss would actually be. For me, it was manageable. For someone whose laptop was their primary income-generating tool, the same arithmetic might reach a different conclusion. That difference between quantifiable risk and personal tolerance for loss is what this article is about.
Risk: When Probabilities Are Known
Risk applies when you have enough information to assign a meaningful probability to an outcome. Car insurance is a classic risk product: actuaries have data on accident rates by age, driving experience, vehicle type, and postcode. They can calculate the expected cost of claims with enough precision to set profitable premiums. The probability is uncertain at the individual level — you don't know if you'll have an accident this year — but it's well-estimated at the population level.
In this environment, the useful mathematical tool is expected value: probability × consequence. A 10% chance of losing £1,000 has the same expected value (£100) as a 1% chance of losing £10,000. Whether you prefer one to the other depends on your risk tolerance, but the expected value calculation makes the comparison precise. Our probability calculator handles the probability component — the rest is judgement about how much the magnitude matters to you.
Uncertainty: When Probabilities Can't Be Estimated
Uncertainty is different. Starting a restaurant in a new neighbourhood, launching a genuinely novel product, making the first move in a new market — these involve outcomes for which there's no reliable comparable data. You can estimate the probability that the restaurant will succeed, but any number you assign is highly speculative. You genuinely don't know whether the probability is 5% or 40%.
In this environment, expected value calculations are less useful because the probability input is so uncertain that the output is spuriously precise. More robust approaches focus on: understanding the range of possible outcomes, ensuring that bad outcomes are survivable, maintaining optionality so you can adjust as you learn, and limiting downside exposure rather than optimising for expected return.
How Human Intuition Gets This Wrong
Humans are systematically biased in how we perceive risk. We overweight low-probability, high-salience events — dramatic plane crashes, rare diseases that appear in the news — and underweight routine risks that accumulate quietly, like the long-term effects of diet or the compounding probability of injury from a repeated dangerous behaviour. We also tend to treat uncertainty as risk, assigning confident-sounding probabilities to genuinely uncertain situations because a number feels more like information than an admission of ignorance.
Insurance as Risk Transfer
Insurance makes sense when the probability is low but the consequence of the event is large enough that it would be genuinely damaging — financially or otherwise. You pay a certain, manageable premium to transfer the risk of an uncertain but potentially severe loss. This is rational even if you expect, on average, to pay more in premiums than you receive in claims — the value is in converting a potentially unmanageable loss into a known, manageable cost.
The Role of Sample Size
Risk estimates improve as data accumulates. A medical treatment tested on 50 patients might show a 12% side-effect rate — but the confidence interval around that estimate is wide. The same treatment tested on 5,000 patients gives a far more reliable estimate of the true rate. This is why early-stage clinical trials are not the right basis for strong claims about safety or efficacy — the sample size is too small for the probabilities to be well-calibrated. Risk estimates from large, well-designed studies deserve more weight than those from small preliminary ones.
Practical Decision-Making Under Risk
For decisions involving genuine risk (estimable probabilities), calculate the expected outcomes, consider how the magnitude of different outcomes compares to your tolerance for loss, and identify whether there are ways to reduce the probability or the consequence independently. For decisions under uncertainty, prefer strategies that maintain flexibility, limit irreversible commitments, and allow you to gather information as you go rather than committing everything on an uncertain forecast upfront.
When the Model Runs Out of Road: Tail Risks and Black Swans
Risk models work by extrapolating from historical data. If a type of event has occurred with a certain frequency over the past 50 years, the model assumes it will occur at a similar rate going forward. This works well for routine risks — car accidents, equipment failures, seasonal demand variation — where the underlying conditions are relatively stable. It works much less well for tail risks: events in the extreme tails of the probability distribution that are rare, severe, and may not have occurred in the historical record at all.
Nassim Taleb coined the term "black swan" for consequential events that are outside the range of normal expectations — the 2008 financial crisis, the COVID pandemic, the 2011 Fukushima earthquake. These events were not literally unpredictable — some analysts did warn of them — but they fell outside the risk models that most organisations were using, which were calibrated on recent history in which such events had not occurred. The lesson is not that risk modelling is useless, but that it should be supplemented by explicit thinking about what would happen if conditions changed radically, and by ensuring that bad outcomes, however improbable, are survivable rather than catastrophic.