A surplus that drifts up and lurches down
The earlier guides in this rung set the stage. You met the surplus process — the insurer's running balance through time — and you learned to ask the starkest question in the business: will it ever fall below zero, and with what probability of ruin? What you did *not* yet have was a concrete, solvable picture of how that balance actually moves. The Cramér-Lundberg model is that picture: the classical risk model, simple enough to yield real theorems, rich enough to capture the essential shape of an insurer's fortunes.
Picture the surplus as a single line drawn through time, and watch it do two things at once. Premiums flow in smoothly, every day, so the line *drifts upward* at a steady slope like water trickling into a tank. Then, at unpredictable moments, a claim arrives and the line *lurches straight down* by the size of that claim — a sudden vertical drop, a stone dropped into the tank. Between drops the line climbs again. Up gently, down sharply, up gently, down sharply. That sawtooth, climbing-then-plunging shape is the whole model in one image, and ruin is simply the first moment the line touches zero from above.
Three ingredients fix the whole drama. First, where the line starts: the initial surplus, call it u, the cushion of capital the insurer holds on day one. Second, how fast it climbs: the premium income rate c, the dollars of premium earned per unit of time. Third, how it plunges: claims arrive as a Poisson stream at some average rate, and each claim's size is an independent draw from a fixed severity distribution. The total claims piled up by any time t is therefore exactly the compound Poisson aggregate loss you built in the previous rung — now set in motion, watched as it accumulates rather than frozen at year-end.
The premium must lean the right way
Before any clever mathematics, one piece of plain arithmetic decides everything. Compare the premium pouring in each year to the claims expected to flow out each year. If the premium rate c is *less* than the expected claims per year, the surplus drifts downward on average — and over an infinite horizon ruin becomes certain, no matter how large the starting cushion u. You cannot out-cushion a leaking boat. So the model only earns its keep when premiums exceed expected claims, and we measure that surplus margin with the relative security loading θ.
The loading θ is just the fractional cushion built into the price. If expected claims are 1,000,000 a year and you charge a premium rate of 1,150,000, you have loaded the pure cost by 15%, so θ = 0.15. That extra 150,000 a year is the steady upward tilt of the surplus line — the slope that, over the long run, fights against the downward jumps. Set θ to zero and the line has no upward bias at all; set it negative and the line is doomed. The phrase actuaries use, the net profit condition, is simply: c must exceed expected claims, i.e. θ must be strictly positive.
One number that captures the danger: the adjustment coefficient
Here is the model's masterstroke. All the moving parts — the premium rate, the claim arrival rate, the entire shape of the severity distribution — can be boiled down into a single positive number that measures how dangerous this insurer's situation is. That number is the adjustment coefficient, written R and also called the Lundberg coefficient. Think of R as a risk thermometer: a *large* R means a safe, well-padded operation, while a *small* R means a fragile one, perilously close to the brink.
Where does R come from? It is defined through the moment-generating function of the claim severity — the mathematical object that packs every moment of the claim-size distribution, its mean, its spread, the full weight of its tail, into one function. R is the special positive value where the upward pull of the premium and the downward pull of the claims exactly balance out in that function. You will not be asked to solve that balance equation by hand here; the point to carry away is *what R responds to*. It is not a free dial — it is squeezed out of the real economics of the book.
Two intuitions make R feel concrete. First, a heavier loading θ pushes R *up*: charge more relative to expected claims and you become safer, exactly as common sense demands. Second — and this is the subtle, important part — a heavier-tailed severity pushes R *down*. Two books with identical average claims and identical premiums can have very different R if one is exposed to rare monster claims and the other is not. The thermometer feels the shape of the tail, not just the average. This is why R earns its keep: it senses danger that the mean alone is blind to.
Lundberg's inequality: ruin bounded by an exponential
Now the coefficient pays its dividend. The classic result, Lundberg's inequality, says something remarkably clean: the probability of ever being ruined, starting from initial surplus u, is no greater than e raised to the power minus R times u. In words — the chance of ruin falls off *exponentially* as you hold more starting capital, and the rate of that fall-off is exactly the adjustment coefficient R. A bigger cushion u helps; a bigger R helps; and the two multiply together inside the exponent.
Lundberg's inequality: P(ruin | start with u) <= e^(-R u) Suppose the adjustment coefficient works out to R = 0.001 and you ask how much capital caps ruin risk at 1%: e^(-0.001 * u) = 0.01 -0.001 * u = ln(0.01) = -4.605 (natural log) u = 4.605 / 0.001 = 4,605 Hold u = 4,605 of initial surplus -> ruin probability <= 1%. Double the cushion to u = 9,210 -> bound squares to <= 0.01%.
Be precise about what the inequality delivers and what it does not. It is an *upper bound*, not the exact probability — the true chance of ruin sits somewhere at or below the exponential, often comfortably below it. That is a feature, not a flaw: a guaranteed ceiling on the risk is exactly what a regulator or a board wants, because it lets you say 'ruin is *at most* this likely' without needing the full, often intractable, exact answer. For one special case — when claim sizes follow an exponential distribution — the bound turns out to be very nearly tight, and ruin theory hands you an exact closed-form probability as a bonus.
What the model honestly is — and isn't
The single most important honesty here: the exponential bound rests on the assumption that claims have a *light* tail — that the moment-generating function exists, that monster claims fade out fast enough. The instant you face a genuinely heavy-tailed risk — large fire losses, liability, catastrophe — that assumption snaps. For such claims the adjustment coefficient does not even exist, Lundberg's tidy exponential simply does not apply, and ruin probability decays far more slowly, dragged down by the rare giant claim. Honest practitioners know which world they are in before reaching for R.
There are gentler simplifications too, and it is worth naming them so you do not mistake the map for the territory. The classical model assumes a constant premium rate, claims that arrive independently at a steady Poisson pace, and no investment return on the surplus sitting in the tank — none of which is quite true of a real insurer, where premiums shift, claims cluster after storms, and reserves earn interest. The model also looks at ruin over an *infinite* horizon; in practice a finite window matters too, which is why the earlier guide drew the line between infinite-time and finite-time ruin. Cramér-Lundberg is the clean skeleton on which all the realistic muscle gets hung.