Insurance pooling
A Socratic walk-through of insurance pooling — reasoned out one step at a time, not lectured.
The question we started with
THE QUESTION #How can an insurer safely take on risks that would ruin any one of its customers?
A house fire would ruin its owner. An insurer accepts that same risk from a million owners and treats the result as a budget line. The intuitive explanation is that risk gets "spread" until it is thin enough not to hurt.
That phrasing hides something worth pressing on. The insurer holding a million policies is exposed to more total risk than any customer, not less. Nothing has been made to disappear. So the question is not how risk is destroyed — it cannot be — but what changes when many risks are held together, and what has to be true of them for the change to occur.
Reasoning it through
REASONING #Take a thousand houses, each with the same small chance of burning. What happens as we add more? Two things move at different speeds, and the whole subject lives in the gap. The expected loss grows in proportion to the number of houses: double the pool, double the claims you expect, and double the premium income. The uncertainty around that total grows much more slowly — with the square root of the number. So the total swing gets larger in absolute terms while getting smaller relative to the money coming in. Per-policy uncertainty falls as one over the square root of the pool size: a hundred policies cut it to a tenth, ten thousand to a hundredth.
Notice what was quietly assumed. That square-root behaviour requires the fires to be independent — one house burning must not make another more likely. Independence is not a footnote here; it is the load-bearing beam. Remove it and see what happens.
Suppose every pair of risks shares a small common element — the same storm, the same drought, the same legal ruling. Per-policy uncertainty still falls as the pool grows, but it does not fall to zero. It descends to a floor set by the correlation, and after that, adding policies buys nothing at all. That is the single most important fact about pooling: a pool of correlated risks has a size beyond which growth stops helping, no matter how large the insurer.
So which risks should be hard to insure privately? Exactly the ones where a single cause strikes everyone in the pool at once. A flood does not visit a random four percent of the floodplain; it visits the floodplain. Earthquakes, windstorms and pandemics are all in this family, and their correlation is why they need machinery that motor insurance does not.
What machinery? First, capital: if the average is not reliable, reserves must be sized against the bad tail rather than the expected loss — European solvency rules, for instance, require capital sufficient to survive the worst year in two hundred. Second, reinsurance — an insurer cedes part of its book to a reinsurer who pools across insurers, geographies and perils uncorrelated with each other. Japanese earthquakes and Florida hurricanes are largely independent even though neither is independent within itself, so pooling the pools restores some of the averaging that failed a level down. Third, where even that fails, the state: flood and terrorism cover in several countries exists only behind a public backstop.
One subtler failure defeats even genuinely independent risks. The square-root argument needs individual losses to have a finite, moderate spread. Some loss distributions are so heavy-tailed that a single claim can dominate the pool's whole total, and the averaging then converges so slowly that a realistic pool never becomes predictable.
Which answers "how big must a pool be?" There is no universal number of policies, and looking for one is the wrong move. What matters is the ratio of capital to the volatility that cannot be diversified away — so a thousand independent, bounded risks may be perfectly viable where a million highly correlated ones are not.
The analogy
THE ANALOGY #Think of a bridge carrying a crowd. Pedestrians arriving and leaving independently average out into a steady load the deck was designed for, and the more of them there are, the steadier that load becomes as a fraction of the total. But let them fall into step, as crowds on a swaying bridge involuntarily do, and the same number of people becomes one synchronised force. Nothing about the individuals changed. Their correlation did, and that is what the structure has to survive.
a bridge fails suddenly at a threshold, whereas an insurer degrades gradually — raising premiums, cutting cover or ceding more to reinsurers long before insolvency — and the crowd's synchrony is a physical feedback loop, while correlated claims usually share an external cause rather than influencing one another.
Clarifying the model
THE MODEL #The most common misstatement is that pooling reduces risk. It does not. It reduces risk per policy while increasing it in total, converting a distribution the individual cannot survive into one an institution can plan around. That matters because it locates the residual danger: no longer with the customer, but concentrated in one balance sheet — which is why insurer solvency is regulated so heavily.
A second refinement: independence is rarely perfect, so the useful question is never "are these risks independent?" but "how much correlation, from what cause, and does it appear only in the tail?" Many risks look near-independent in normal conditions and become tightly coupled in exactly the circumstances where a payout is needed — the correlation switches on with the disaster. Modelling that honestly is one of the harder open problems in the field.
Independence also fails in two ways that are not about weather at all. Claims can share a legal cause: a court reinterpreting standard policy wording triggers thousands at once, as happened with asbestos and with pandemic business-interruption cover. And behaviour can couple them — if being insured makes everyone slightly less careful, losses drift together in a way the original pricing did not assume.
A picture of it
THE PICTURE #How to readBoth lines start at 1, meaning a pool of one is exactly as uncertain as the risk itself. The lower line is independent risks: uncertainty per policy falls as one over the square root of the pool, heading towards zero — the ordinary insurance story. The upper line is the same risks with a mild common element, a shared storm or ruling affecting one part in ten; it drops quickly at first, then flattens near 0.32 and stays there. That plateau is the point: past a few hundred policies, growth buys the correlated pool almost nothing, so the insurer must reach for capital and reinsurance rather than for more customers.
What became clearer
WHAT CLEARED #Pooling does not destroy risk; it exploits independence to make the average predictable while the total grows, which is why an institution can carry what an individual cannot. Independence is the assumption doing all the work, and where it fails — floods, quakes, pandemics, a courtroom reinterpreting every contract at once — the averaging hits a floor no amount of growth clears. The answer is then not a bigger pool but a different structure: capital sized to the tail, reinsurance across perils that genuinely are unrelated, and a public backstop where even that runs out.
Where to go next
ONWARD #- How catastrophe bonds pool insurance risk with investors whose other holdings are unrelated to it.
- Why correlations that look small in ordinary years rise sharply in exactly the events that cause claims.
Key terms
TERMS #| Term | What it means |
|---|---|
| Risk pooling | combining many risks so that the average outcome becomes predictable, even though the total exposure grows. |
| Correlated risk | risks driven by a shared cause, so they occur together and defeat the averaging that pooling relies on. |
| Reinsurance | insurance bought by insurers, pooling books across perils and regions that are uncorrelated with each other. |
| Solvency capital | reserves held against the bad tail rather than the average, sized by regulation to a stated survival probability. |
Every term the collection defines is gathered in the glossary.