Reading the question backwards
Every guide in this rung so far has pointed the same way: you fix the initial cushion *u*, the loading, and the claim process, and you ask what the probability of ruin comes out to be. That is the natural direction for *analysis* — you have the company, you want to know how safe it is. But it is the wrong direction for *management*. A board does not start by choosing how much capital to lock up; it starts by deciding how much risk of failure it is willing to tolerate, and only then asks what that decision costs.
So we run the arrow the other way. Name a ruin probability you can live with — say one chance in two hundred over the year, a 0.5% target — and solve for the *smallest* initial surplus that delivers it. That inverted quantity is the required capital: the cushion the company must actually hold so that the chance of being driven through zero stays below the line it drew in the sand. The whole of risk theory, turned from a descriptive instrument into a prescriptive one, lives in this single inversion.
Solving for the cushion: a worked inversion
In the light-tailed classical model you already met, the Lundberg inequality gives the inversion an almost embarrassingly clean form. It says the ruin probability starting from surplus *u* is bounded above by e raised to the power of −R·u, where R is the adjustment coefficient — that single number, set by the loading and the claim distribution, which measures how strongly the surplus leans uphill. Ruin probability falls *exponentially* as the cushion grows. Double the danger you can tolerate? You do not need to halve the capital — you only shave a little off the top.
Run it as numbers. Suppose the adjustment coefficient works out to R = 0.01 per unit of surplus, and you want the (infinite-horizon) ruin probability held at 0.5%. Set e^(−0.01·u) = 0.005 and solve: u = −ln(0.005) / 0.01 ≈ 530. Now tighten the target tenfold, to 0.05%. The required cushion rises only to about 760 — not ten times larger, just roughly 50% more. That is the signature of an exponential bound: each extra factor-of-ten of safety costs the *same fixed slab* of capital (here about ln(10)/R ≈ 230), no matter how deep into the tail you are already standing.
Light-tailed (Lundberg) inversion: psi(u) <= e^(-R*u) Target psi Required u (R = 0.01) ---------- --------------------- 0.5% -ln(0.005)/0.01 ~= 530 0.05% -ln(0.0005)/0.01 ~= 760 0.005% -ln(0.00005)/0.01 ~= 990 Each 10x more safety -> + ln(10)/R ~= 230 of capital (a fixed slab, the same every time you go deeper)
When the tail is heavy, the answer changes character
Everything above quietly assumed the claim sizes were *light-tailed* — the kind whose chance of a giant value dies off exponentially, so the adjustment coefficient R even exists. Many real claim distributions are not like that. A heavy-tailed severity — liability, catastrophe, large-fire, cyber — has a tail that fades only as a *power* of size, not exponentially. For these, R does not exist at all, the tidy exponential bound simply does not apply, and the relationship between ruin probability and required capital changes its entire character.
The deep result of tail weight and ruin theory is this: under a heavy tail, ruin is overwhelmingly caused by *one single colossal claim*, not by a slow drift of ordinary bad luck. And the ruin probability decays only like a *power* of the cushion, not exponentially. So now, to cut ruin probability tenfold, you may have to make the cushion several times larger — and to cut it tenfold again, larger still by an even bigger multiple. The fixed-slab bargain of the light-tailed world is gone. Safety against catastrophe gets ruinously expensive precisely where you need it most.
The bridge to economic capital and ERM
Step out of the idealised surplus process and look at how a real company actually fixes its number, and you will find the same inversion wearing modern clothes. Instead of a continuous-time ruin probability, the firm picks a *one-year* horizon and a confidence level — Solvency II famously uses 99.5%, i.e. a 1-in-200 chance of insolvency in the year. It then computes the loss at that confidence point and holds capital to cover it. The amount it needs is its economic capital, and the intuition behind it is exactly the cushion-for-a-bad-year story you built from ruin theory.
But notice the subtle, important shift in *what is being measured*. The 99.5% loss point is a value at risk: it tells you the threshold the loss will not exceed with 99.5% probability. And here lies a famous, honest limit you must carry with you — VaR tells you *where* the cliff edge is, but says nothing about how far you fall once you go over it. Two companies can share an identical VaR while one faces a survivable shortfall beyond it and the other faces a bottomless one. For heavy-tailed, catastrophe-exposed risk, that blind spot is not academic; it is the whole danger.
This is why required capital does not live alone. It is the quantitative heart of enterprise risk management — the discipline of choosing, monitoring, and pricing risk across the whole firm. ERM takes the board's stated appetite ("a 1-in-200 year, no worse"), turns it into a capital target by exactly the inversion of this guide, and then asks the further questions a single ruin probability cannot: which lines of business consume the most capital, how much does diversification across them give back, and is the *cost* of carrying that capital — the return investors demand on it — earned back in the price charged to policyholders? Required capital is where survival, strategy, and price finally meet.
Honest about the limits
Before you trust any of these numbers, hold them at arm's length. The classical surplus process ignores investment returns, inflation, expenses, new business, management reaction, and the brute fact that a company in trouble does not sit passively waiting for ruin — it raises rates, buys reinsurance, stops writing, or is rescued. Every required-capital figure is the output of a *model*, and a model is a deliberately simplified story, not the world. The number's precision — "hold 537.2" — is almost entirely spurious; its order of magnitude is what deserves your respect.
The sharpest danger is model risk on the tail. The entire required-capital answer is dominated by the rarest, largest, least-observed claims — exactly the region where you have the *fewest* data points and the *most* freedom to choose a distribution. Pick a heavy tail and your capital triples; pick a light one and it does not, and nothing in your historical data may yet tell you which was right. The 1-in-200 events have, almost by definition, mostly not happened to you yet. Required capital is therefore not a fact you measure; it is a judgement you defend, leaning on extreme-value theory, stress tests, and professional humility as much as on any formula.
None of this makes the exercise pointless — quite the opposite. A model you know to be wrong in stated ways is far more useful than a vague feeling of safety, because it tells you precisely which assumptions to argue about, stress, and disclose. That is the mature actuarial stance you carry out of this rung: compute the required capital, then spend at least as much effort understanding *why it might be wrong* as you spent computing it. Survival is the question; the capital is the model's best honest answer; and the limits of that answer are themselves part of what you are paid to know.