Two promises hidden inside a premium
Everything in this rung has been building two tools. The expected value gave you a single fair price — the pure premium — and the variance told you how wildly any one year might swing around it. Both are honest descriptions of a single risk. But an insurer never sells one policy; it sells thousands. The magic of insurance is not in any single contract — it is in what happens when you add a great many of them together.
When an insurer hands you a premium, it is quietly making two promises to its own shareholders. First: if we pool enough similar risks, our average claim will land close to the pure premium we charged, so the price is roughly right. Second: even though we can't be certain, we can pin down how far the total could plausibly drift, so we know exactly how much spare money — capital — to keep on hand. Those two promises are not wishful thinking. They are two of the most celebrated theorems in all of probability, and this guide is about meeting them.
The Law of Large Numbers: averages settle down
Flip a fair coin ten times and you might see seven heads — a wild 70%. Flip it ten thousand times and you will see something achingly close to 50%. That is the whole idea of the law of large numbers (LLN): as you average more and more independent observations of the same random thing, the average drifts ever closer to its expected value and stops jumping around. Importantly, the LLN is a statement about the *average*, not the *total* — the total number of heads actually wanders further from 5,000 in absolute terms; it is the average per flip that calms down.
Recall from the last guide that the variance of an average of n independent copies shrinks by a factor of n: the standard deviation of the mean falls like one over the square root of n. That square-root law is the LLN with its sleeves rolled up. Pool 100 identical independent policies and the per-policy average is ten times more stable than a single policy; pool 10,000 and it is a hundred times more stable. This is exactly the pooling mechanism from the Foundations rung, now stated as a theorem rather than an analogy. It is why an insurer can quote the pure premium and trust that, across the book, reality will cooperate.
The Central Limit Theorem: sums look bell-shaped
The LLN tells you the average lands near the right place — but it says nothing about the *shape* of the leftover uncertainty. That shape is the gift of the central limit theorem (CLT), and it is genuinely astonishing. Add up many independent risks — no matter how lopsided, skewed, or strange each one is on its own — and the distribution of the *sum* (or the average) drifts toward a single, universal shape: the normal distribution, the famous bell curve. The individual risks can be coin flips, dice, dental claims, or lopsided lottery tickets; the total forgets their personalities and remembers only their mean and variance.
Why should you care that the total is bell-shaped? Because a normal distribution is fully described by just two numbers — its mean and its standard deviation — and we already know how to compute both. Once an actuary can argue that the aggregate loss of a large book is approximately normal, the whole tail becomes calculable. The familiar rule of thumb falls out: about 95% of the time the total lands within roughly two standard deviations of the mean, and the rare years beyond that have known, quantifiable odds. The CLT is what turns a vague "we might lose a lot" into a number a regulator and a board can actually plan against.
Putting them to work: price, then capital
Let's make this concrete with the tiny policy from the last guide. Each policy has a 5% chance of a 2,000 claim, so its expected loss (pure premium) is 100 and its standard deviation is about 436 — more than four times the mean. One such policy is terrifyingly volatile. Now write 10,000 of them, independently. The two theorems take over.
One policy: E = 100 SD = 436 (SD/mean = 4.36, scary) 10,000 policies (independent): Total mean = 10,000 * 100 = 1,000,000 Total SD = 436 * sqrt(10,000) = 43,600 SD as % of mean = 43,600 / 1,000,000 = 4.36% (LLN: relative wobble shrank ~100x) CLT => total loss ~ Normal(1,000,000, 43,600) ~95% of years land within +/- 2 SD: 1,000,000 +/- 87,200 Capital to be ~99.5% safe (~2.58 SD): about 112,000 above the mean
Read that small table slowly, because it is the whole logic of an insurance company in eight lines. The LLN is the first miracle: pooling 10,000 independent policies leaves the *average* claim hugging 100, and the relative wobble collapses from a ruinous 436% to a manageable 4.36%. That stability is why charging the average premium is sane. The CLT is the second: it certifies the total is approximately normal, so the insurer can say "to be 99.5% safe we must hold about 112,000 of capital above expected losses" — a precise, defensible number rather than a guess. Price comes from the first theorem; capital comes from the second.
The honest catch: both lean hard on independence
Now the warning that separates an actuary from a tourist. Every clean number in that table rested on one word: independent. The total standard deviation grew like the square root of n only because each policy's outcome had nothing to do with its neighbour's. That is the precise role of independence you met earlier in this rung. Lose it, and the arithmetic that made insurance look effortless quietly falls apart.
Real risks are often not independent at all. An earthquake, a hurricane, a pandemic, or a market crash strikes thousands of policies at the same instant — their losses move together. When risks are positively correlated, the variance of the total no longer divides by n; the comforting square-root shrinkage weakens or vanishes, and in the extreme of perfect correlation it disappears entirely, leaving you no better off than holding one giant policy. The LLN stops calming the average, and the CLT's tidy bell grows a fat, dangerous tail the normal approximation badly underestimates.
There is a subtler trap too. Even with genuine independence, the CLT is a statement about the *centre* of the distribution; the bell-curve approximation is at its worst precisely in the far tail, where catastrophes live. A normal curve has thin, fast-fading tails, but real insurance losses are often heavy-tailed — extreme events happen more often than the bell predicts. So an actuary leans on these theorems where they are trustworthy (the bulk, with diversifiable risks) and reaches for sterner tools — explicit catastrophe models, correlation assumptions, and heavy-tailed distributions — exactly where they fail. The theorems are a foundation, not a fortress.
What you carry forward
You now hold the two theorems that make a stack of frightening individual risks into a calm, priceable, capital-backed business — and, just as importantly, you know the single assumption they both lean on and what goes wrong when it breaks. This is the keystone of the whole probability rung. From here the ladder turns from describing one distribution to fitting distributions to real claims data, estimating their parameters, and quantifying how much we trust those estimates — the statistics that lets an actuary learn the numbers the world refuses to hand over directly.