Ruin Probabilities With Heavy-Tailed Claims

A trading desk sets aside capital as a buffer against losses and earns a steady stream of income (premium, spread capture, or carry) to replenish it. Losses arrive randomly over time. Will the buffer ever be wiped out — will accumulated losses ever exceed capital plus income collected so far? This is the classical "ruin probability" question from actuarial science, and it has a clean answer when individual losses are light-tailed (like normal or exponential-sized). But real trading losses — a single catastrophic drawdown, one enormous claim — are often heavy-tailed, and that changes not just the number but the entire character of the answer.

An analogy: leaks versus a burst pipe

Imagine a water tank with a slow steady inflow and random outflows representing losses. If outflows are all modest and similar in size (light-tailed), the tank only ever empties through the cumulative effect of an unusually long unlucky streak of many moderate outflows in a row — ruin is a slow grind. If instead outflows occasionally include a single catastrophic burst — a pipe bursting rather than a series of small leaks (heavy-tailed losses) — then the tank is far more likely to empty because of one single event than because of an accumulated bad streak. Knowing which picture is true completely changes how much buffer you actually need, and where to focus attention: smoothing many small leaks, or bracing for the one burst pipe.

The idea, one symbol at a time

Let $u$ be the initial capital (buffer), let losses arrive according to a process with claims $X_1, X_2, \ldots$ having mean $\mathbb{E}[X]$, and let income accumulate at rate $c$ per unit time with $c > \lambda \mathbb{E}[X]$ (income exceeds average loss rate — a "positive safety loading," or the buffer isn't doomed to shrink on average). Let $\psi(u)$ be the ruin probability: the probability that accumulated losses ever exceed capital plus income collected so far, starting from buffer $u$.

When claim sizes are light-tailed (their distribution has a moment generating function, e.g. exponential or normal), the classical Cramér–Lundberg result gives an exponential decay in $u$:

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

for constants $C$ and $R > 0$ (the "adjustment coefficient"). In plain English: doubling your buffer roughly squares how small the ruin probability becomes — buffer helps enormously, and fast.

When claim sizes are heavy-tailed (subexponential, e.g. Pareto-like), the ruin probability instead decays like the tail of a single claim:

$$\psi(u) \sim \frac{\lambda \mathbb{E}[X]}{c - \lambda \mathbb{E}[X]} \, \overline{F}(u), \qquad \overline{F}(u) = P(X > u) .$$

In plain English: ruin probability shrinks only as fast as the chance of seeing one loss bigger than the whole buffer — polynomially, not exponentially. Buffer still helps, but far more slowly, and ruin is driven overwhelmingly by that single worst loss, not by an accumulation of smaller ones (this is the same "single big jump" idea covered in subexponential distributions).

Worked example 1: light-tailed losses

Suppose losses are exponential with mean $1 (in millions), safety loading gives adjustment coefficient $R = 0.5$, and $C \approx 1$. For buffer u = \4$m:

$$\psi(4) \approx e^{-0.5 \times 4} = e^{-2} \approx 0.135 .$$

Doubling the buffer to $8m: $\psi(8) \approx e^{-4} \approx 0.018$ — an over 7-fold reduction in ruin probability from doubling capital.

Worked example 2: heavy-tailed losses

Now suppose losses follow a Pareto-like tail $\overline{F}(u) = u^{-1.5}$ (heavy-tailed, in the same $ millions units), with $\lambda \mathbb{E}[X]/(c - \lambda\mathbb{E}[X]) \approx 1$. For buffer $u = 4$:

$$\psi(4) \approx 1 \times 4^{-1.5} = \frac{1}{8} = 0.125 .$$

Doubling the buffer to $u=8$: $\psi(8) \approx 8^{-1.5} \approx 0.044$ — only about a 2.8-fold reduction, far less improvement than the light-tailed case got from the same doubling of capital, even though the two starting ruin probabilities (0.135 vs 0.125) looked almost identical at $u=4$.

What this means in practice

This distinction matters directly for sizing capital buffers, margin, and risk limits against strategies exposed to occasional catastrophic losses — short volatility positions, insurance-like carry trades, or concentrated single-name bets. If loss severity is heavy-tailed, the standard intuition that "more capital buys exponentially more safety" is wrong; capital still helps, but the marginal benefit of each extra unit decays much more slowly, and the dominant risk driver is capping the size of the single worst possible loss (position limits, stop-losses on tail scenarios) rather than simply holding more capital.