The Cramér-Lundberg Ruin Model

Picture a bucket sitting under a slow, steady hose — that's an insurer collecting premiums, or a trading desk earning a small steady edge every day. The bucket fills gradually and predictably. But every so often, without warning, someone dumps a pail of water back OUT of the bucket — a claim gets filed, or a bad trading day hits. Most of the time the pail is small and the bucket barely notices. Occasionally the pail is huge. The question the Cramér-Lundberg model answers precisely: starting from a given amount of water already in the bucket, what's the probability the bucket ever runs completely dry? That's "ruin," and this is the founding model of the entire field that studies it — a field that matters just as much to a prop trading desk sizing its capital cushion as it does to an insurance company setting its reserves.

The setup, one symbol at a time

The insurer's surplus at time $t$ — its capital cushion — is written $U(t)$, starting from an initial capital $u = U(0)$. Surplus grows steadily from premium income collected at a constant rate $c$ per unit time (dollars of premium collected per day, say), and drops each time a claim arrives. Claims arrive according to a Poisson process with rate $\lambda$ — random, memoryless arrival moments, averaging $\lambda$ claims per unit time — and each claim has a random size $X_i$, all independent and identically distributed with mean $E[X]$. Putting this together:

$$U(t) = u + ct - \sum_{i=1}^{N(t)} X_i$$

In words: surplus at time $t$ equals the starting capital, plus everything collected in premiums by time $t$, minus the sum of every claim that has arrived by time $t$ (where $N(t)$ is the random number of claims that have occurred so far). For this to make sense as an ongoing business at all, premiums must come in faster on average than claims go out — the safety loading (or net profit) condition:

$$c > \lambda \, E[X]$$

In words: the premium rate must exceed the average rate at which claims drain the bucket ($\lambda \, E[X]$ is claims-per-unit-time times average claim size — the average outflow rate). Without this, surplus drifts toward $-\infty$ with probability one no matter how large $u$ is.

Given the safety loading holds, the probability of eventual ruin starting from capital $u$, written $\psi(u) = P(U(t) < 0 \text{ for some } t > 0)$, satisfies the Lundberg inequality:

$$\psi(u) \leq e^{-Ru}$$

Here $R > 0$ is the adjustment coefficient (or Lundberg exponent), a number determined by $\lambda$, $c$, and the full distribution of claim sizes (via its moment generating function), not just its mean. In words: ruin probability is bounded above by an exponentially decaying function of starting capital — every extra unit of starting capital multiplies your maximum ruin probability by the same shrinking factor $e^{-R}$.

Adjust the rate parameter in the explorer above to see how claim-arrival randomness itself behaves — claims arrive according to this same Poisson shape, and it's the combination of Poisson arrival timing with random claim sizes that produces the surplus process shown earlier.

Worked example 1: checking the safety loading

An insurer collects premiums at $c = 120$ (thousand dollars per month) and expects $\lambda = 20$ claims per month, each averaging $E[X] = 5$ (thousand dollars). Average outflow rate is $\lambda \, E[X] = 20 \times 5 = 100$. Since $c = 120 > 100$, the safety loading condition holds, with a loading of $120/100 - 1 = 0.20$, i.e. premiums run 20% above expected claims — a viable, profitable book on average. If instead claims averaged $E[X] = 6.5$, outflow would be $20 \times 6.5 = 130 > 120 = c$, violating the condition: surplus drifts to $-\infty$ eventually regardless of starting capital, no matter how large $u$ is — ruin becomes a near-certainty in the long run, not just a tail risk.

Worked example 2: how starting capital changes the ruin bound

Suppose the adjustment coefficient works out to $R = 0.02$ per thousand dollars of surplus (a number that in practice comes from solving an equation involving $\lambda$, $c$, and the claim-size distribution — here just taken as given to focus on how the bound behaves). Compare two starting capital levels:

• $u = 50$: $\psi(50) \leq e^{-0.02 \times 50} = e^{-1.0} \approx 0.368$ — up to a 36.8% chance of ruin.
• $u = 150$: $\psi(150) \leq e^{-0.02 \times 150} = e^{-3.0} \approx 0.050$ — down to about 5%.
• $u = 250$: $\psi(250) \leq e^{-0.02 \times 250} = e^{-5.0} \approx 0.007$ — under 1%.

Tripling starting capital from 50 to 150 didn't cut the ruin bound by a third — it cut it by more than 7x. That's the exponential at work: each additional block of capital multiplies the remaining ruin probability by the same factor, the same way compound interest multiplies rather than adds.

What this means in practice

• This is a risk-of-ruin framework, not just an insurance model. A trading desk earning a steady edge (the analog of $c - \lambda E[X]$, its expected daily P&L) but exposed to occasional sharp drawdowns (the analog of claims) faces exactly this mathematics when sizing how much capital it needs to survive a bad stretch with high probability — this is the rigorous cousin of Kelly-criterion risk-of-ruin reasoning.
• The safety loading condition is a hard floor, not a soft target. If expected inflow doesn't exceed expected outflow, no amount of starting capital saves you in the long run — ruin becomes certain eventually, only the timing is uncertain.
• Reinsurance and capital retention decisions are literally about moving points on this curve — ceding some claims to a reinsurer lowers both $c$ (premium retained) and the claim-size distribution simultaneously, and the net effect on $R$ (and hence on ruin probability) is the actual quantity being optimized.
• The exponential bound is a worst case, not the exact answer — computing $\psi(u)$ exactly typically requires either an explicit formula for special claim distributions (e.g. exponential claims give a clean closed form) or simulation, but the Lundberg bound is often tight enough to be the number people actually quote.