From two pieces to one total
Everything in this rung so far has lived on one side or the other of the great frequency–severity split. You learned to model how *often* claims arrive — the frequency distribution, typically a Poisson count — and, quite separately, how *big* each claim is — the severity distribution, a positive, heavy-tailed amount. Two clean little models, sitting in two clean little boxes. But no insurer ever wrote a cheque for a 'frequency.' What it pays is the aggregate loss: the grand total of every claim in a year, added up. That single random number is what this guide is about.
Write the total as S. In words: S is the sum of N individual claim amounts X1, X2, …, where N is itself random (you do not know in advance how many claims this year will bring) and each Xi is random too. Both dice are rolling at once — the number of claims *and* the size of each. This double randomness is what makes S richer, and harder, than anything you have handled yet. It is the object that every property-and-casualty (P&C) premium, capital figure, and reinsurance decision ultimately rests on.
Two ways to add it up: individual vs collective
There are two honest ways to build the distribution of S, and they answer slightly different questions. The individual risk model walks policy by policy. For each of the n contracts in the portfolio it asks: will *this* policy have a claim, and if so how large? Then it adds those n possibly-zero amounts together. It is the natural picture for a fixed book where most policies file nothing — group life insurance is the classic case: a known list of people, each of whom either dies in the year or does not.
The collective risk model stops looking at individual policies and watches the whole portfolio as one claim-producing machine. It does not care which policy fired; it only asks: how many claims N did the *book* produce, and what was the size of each? This is the collective view, and S = X1 + … + XN with a random N is its signature form. For a large, churning portfolio — cars insured and cancelled every day — this is far more natural than tracking each policy. You rarely know exactly which car will crash; you can model very well how many crashes the fleet will produce.
Notice the trade. The individual model is exact about the portfolio's composition but clumsy to compute — n separate yes/no-and-how-much questions. The collective model throws away the policy labels and gains enormous mathematical convenience in return; that convenience is the whole reason the next section exists. Crucially, for a big book the two give almost the same answer for S, because the law of large numbers smooths the difference. The collective model is an approximation you choose with open eyes, not a corner you cut.
The compound Poisson model and its two clean answers
The most-loved version of the collective model takes the count N to be Poisson. This is the compound Poisson model: a Poisson number of claims, each an independent draw from the same severity distribution. 'Compound' simply means a random number of random pieces stacked together — the resulting distribution of S is a compound distribution. It is the default starting point for almost all of property-and-casualty risk theory, and it earns that place by being both realistic for many lines and astonishingly easy to summarise.
Here is the payoff, and it is one of the prettiest results in the whole field. You do not need the full distribution to get the two numbers you most want. The mean total loss is just the average count times the average claim size — intuitively obvious: expected number of claims, each costing the expected amount. The variance is the subtle one, and it is where the magic of the double randomness shows itself.
Compound Poisson, with lambda = expected # claims:
E[S] = lambda * E[X] (mean count x mean severity)
Var[S] = lambda * E[X^2] (count x second moment of severity)
Example: lambda = 100 claims/year
severity mean E[X] = 2,000
severity E[X^2] = 20,000,000 (so Var[X]=16,000,000)
E[S] = 100 x 2,000 = 200,000
Var[S] = 100 x 20,000,000 = 2,000,000,000
SD[S] = sqrt(2,000,000,000) = 44,721Two honest cautions before you fall in love with those formulas. First, the mean and variance are *not* the distribution. In our example E[S] is 200,000 with a standard deviation of about 44,700 — but S is right-skewed, so the bad years stretch much further above the mean than the good years fall below it. Pricing and capital live in that right tail, not at the average. Second, the clean formulas lean on independence: claims are assumed not to clump. A single hailstorm that triggers a thousand correlated roof claims at once breaks that assumption hard — which is exactly why catastrophe risk gets modelled separately rather than crammed into a tidy compound Poisson.
Getting the whole distribution: the Panjer recursion in words
Mean and variance are not enough when the questions you face live in the tail: 'What total loss will we exceed only one year in two hundred?' For that you need the full distribution of S, and that is genuinely hard — adding up a *random* number of random claims is not a sum any single formula resolves. Before computers, actuaries leaned on the central limit theorem to pretend S followed a normal bell curve, but that approximation lies badly in the right tail, precisely where it matters most.
The **Panjer recursion** is the elegant escape. It is an exact, computer-friendly way to build the distribution of S one rung at a time — the probability that the total is 0, then 1 unit, then 2 units, and so on — each step bootstrapping off the answers already found. You climb the ladder of possible total losses, and the recursion tells you exactly how to get the next rung's probability from the ones below it. No simulation, no normal-curve fudge: just an honest, mechanical climb.
There is a catch worth naming. The recursion only works when the frequency distribution belongs to a special, well-behaved family — the (a,b,0) class, whose members are exactly the Poisson, the negative binomial, and the binomial, with a slightly larger (a,b,1) class that also handles a tweaked probability of zero claims. That is not a narrow cage; those are the very count distributions actuaries actually use. The recursion also needs the severity put onto a discrete grid first — claim sizes rounded to the nearest unit, say nearest 100 — which introduces a small, controllable discretisation error you trade for the exact recursion.
Where this distribution carries you next
Once you can produce the distribution of S, you hold the master key to most of risk theory, and three doors swing open. The first is pricing: the expected aggregate loss per policy is the pure premium, the actuarially fair cost of the cover before any loading for expenses or profit. The second is capital: read a high quantile off the right tail — the loss you would exceed only 1 year in 200 — and you have a Value-at-Risk. Average everything beyond that point and you get the gentler, fuller-tail Tail Value-at-Risk, the basis of modern economic capital.
Mind a famous misconception at this door: VaR tells you the *threshold* you breach once in two hundred years, but it says nothing about how bad things get *once you breach it*. A storm that just clears the threshold and a storm ten times larger look identical to VaR — it ignores the far tail. That blind spot is exactly why TVaR, which averages the severity of the breaches themselves, has steadily replaced it in solvency regimes. And a reminder that runs through this whole rung: a reserve set from S is not idle cash waiting to be spent — it is a measured promise about losses that have not yet fully revealed themselves.
The third door leads straight into the next rungs of the ladder. Track the running balance of premiums-in minus losses-out over time and S turns dynamic: you get the surplus process of ruin theory, asking how likely it is the insurer's surplus ever falls below zero. And the moment you ask 'how much should I trust *this* portfolio's own loss experience versus the wider market's?' you have stepped into the credibility problem — blending a group's data with broader information, the very next idea you will meet. The collective risk model is not a destination; it is the junction where pricing, capital, ruin, and credibility all meet.