8.1 Classical ruin theory and infinite time horizon

Classical ruin theory overview

Classical ruin theory gives insurance companies a mathematical framework for answering a fundamental question: what is the probability that we go bankrupt? It models an insurer's surplus over time, accounting for the randomness of incoming premiums and outgoing claims, and quantifies the likelihood that surplus drops below zero. For actuarial work, this is the foundation for setting capital requirements, choosing reinsurance, and stress-testing solvency.

The theory relies on stochastic processes (Poisson and compound Poisson) to model claim patterns, and produces key results like Lundberg's inequality (an upper bound on ruin probability) and the Pollaczek-Khinchine formula (connecting ruin quantities to the claim size distribution).

Surplus process in classical ruin theory

The surplus process $U(t)$ tracks the insurer's financial position at time $t$:

$U(t) = u + ct - S(t)$

• $u$ = initial surplus (starting capital)
• $c$ = premium income rate (constant, collected continuously)
• $S(t)$ = aggregate claims paid out up to time $t$

The insurer is ruined if the surplus ever becomes negative, meaning $U(t) < 0$ for some $t > 0$. Notice that ruin depends on the entire path of $U(t)$, not just its value at a single point.

Probability of ruin in infinite time

The ultimate ruin probability $\psi(u)$ is the chance that surplus ever drops below zero, given initial surplus $u$:

$\psi(u) = P\left(\inf_{t \geq 0} U(t) < 0 \;\middle|\; U(0) = u\right)$

This is an infinite time horizon measure, meaning you're asking whether ruin happens at any point in the future, no matter how far out. A higher initial surplus $u$ generally reduces $\psi(u)$, which is intuitive: more starting capital provides a bigger buffer against bad claim experience.

The practical goal is to keep $\psi(u)$ acceptably low through appropriate premium rates, reinsurance arrangements, and capital reserves.

Poisson process for claim arrivals

Claim arrivals are modeled as a Poisson process with rate $\lambda$:

• The number of claims in any interval of length $t$ follows a Poisson distribution with mean $\lambda t$
• The time between consecutive claims follows an exponential distribution with mean $1/\lambda$
• Claim arrivals in non-overlapping time intervals are independent

This is a memoryless model: knowing when the last claim arrived tells you nothing about when the next one will come. That's a simplification, but it makes the mathematics tractable and works well for large, homogeneous portfolios.

Compound Poisson process for aggregate claims

The aggregate claims process $S(t)$ combines the random number of claims with random claim sizes:

$S(t) = \sum_{i=1}^{N(t)} X_i$

• $N(t)$ = number of claims up to time $t$ (Poisson process with rate $\lambda$)
• $X_i$ = size of the $i$-th claim (i.i.d. random variables, independent of $N(t)$)

Claim sizes $X_i$ are assumed non-negative and typically continuous. Common choices for the claim size distribution include exponential, gamma, and Pareto. The choice matters a great deal: light-tailed distributions (exponential, gamma) versus heavy-tailed distributions (Pareto, lognormal) lead to very different ruin behavior.

Lundberg's inequality

Lundberg's inequality gives you a quick, conservative upper bound on the ruin probability without needing to compute the exact value. It's one of the most important results in classical ruin theory because it connects solvency directly to a single parameter: the adjustment coefficient.

Exponential bound for ruin probability

The inequality states:

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

where $R > 0$ is the adjustment coefficient (also called the Lundberg exponent). This tells you that ruin probability decays at least exponentially with initial surplus $u$.

A few things to note:

• The bound gets tighter as $u$ increases, so it's most useful for insurers with substantial capital
• For small $u$, the bound can be quite loose
• The entire bound hinges on finding $R$, which depends on the premium rate, claim frequency, and claim size distribution

Adjustment coefficient in Lundberg's inequality

The adjustment coefficient $R$ is the unique positive solution to Lundberg's fundamental equation:

$\lambda + cR = \lambda M_X(R)$

where $M_X(r) = E[e^{rX}]$ is the moment generating function (MGF) of the claim size distribution.

To find $R$, you need to solve this equation, which typically requires numerical methods unless the claim size distribution is exponential (where a closed-form solution exists).

A larger $R$ means a tighter bound on ruin probability and a more stable portfolio. What drives $R$ higher?

• Higher premium rate $c$ relative to expected claims
• Lower claim frequency $\lambda$
• Lighter-tailed claim size distributions

The adjustment coefficient only exists when the MGF $M_X(r)$ is finite for some $r > 0$. For heavy-tailed distributions like Pareto, the MGF doesn't exist, so Lundberg's inequality cannot be applied. This is a critical limitation.

Pollaczek-Khinchine formula

The Pollaczek-Khinchine formula provides a more detailed characterization of ruin than Lundberg's inequality. Rather than just bounding the ruin probability, it expresses the ruin probability (or its Laplace transform) in terms of the claim size distribution through a series representation.

Laplace transform of time to ruin

Define the time to ruin as $T = \inf\{t > 0 : U(t) < 0\}$, and its Laplace transform as:

$\phi(u) = E[e^{-sT} \mid U(0) = u]$

The Pollaczek-Khinchine formula connects $\phi(u)$ to the Laplace transform of the claim size distribution and the adjustment coefficient. Setting $s = 0$ recovers the ruin probability itself.

Pollaczek-Khinchine formula derivation

The derivation proceeds through these key ideas:

1. Ladder heights: Consider the successive record lows (descending ladder heights) of the surplus process. Each ladder height represents how much the surplus drops below its previous minimum.

2. Ladder height distribution: The distribution of these ladder heights is related to the integrated tail distribution of the claim sizes: $f_e(x) = \frac{1 - F_X(x)}{E[X]}$

3. Geometric compounding: Ruin occurs if and only if the surplus eventually crosses zero, which can be decomposed into a geometric number of ladder height steps.

4. Renewal equation: Connecting the Laplace transform of the time to ruin with the ladder height distribution yields the Pollaczek-Khinchine formula.

The resulting formula for the ruin probability takes the form of a geometric compound distribution:

$\psi(u) = \sum_{n=1}^{\infty} (1 - \rho)\rho^n \left(1 - F_e^{*n}(u)\right)$

where $\rho = \frac{\lambda E[X]}{c}$ is the claims-to-premium ratio and $F_e^{*n}$ is the $n$-fold convolution of the integrated tail distribution.

Lundberg's fundamental equation

This equation is the central algebraic relationship in classical ruin theory. Everything flows from it: the adjustment coefficient, Lundberg's inequality, and the structure of exact ruin probabilities.

Roots of Lundberg's fundamental equation

The equation is:

$\lambda + cr = \lambda M_X(r)$

or equivalently:

$\lambda(M_X(r) - 1) = cr$

To understand why a positive root exists, consider the behavior of both sides:

• At $r = 0$: both sides equal zero (trivial root)
• The left side $\lambda(M_X(r) - 1)$ is convex and grows faster than the linear right side $cr$, provided the MGF exists
• The net profit condition $c > \lambda E[X]$ ensures that the derivative of the left side at $r = 0$ is less than the slope of the right side, guaranteeing a second crossing at some $R > 0$

The net profit condition is essential. It says the insurer collects more in premiums than it expects to pay in claims. Without it, ruin is certain ($\psi(u) = 1$ for all $u$), and no positive root exists.

Lundberg coefficient vs. adjustment coefficient

These terms are sometimes used interchangeably in the literature, but they can refer to different quantities depending on the source:

• The adjustment coefficient $R$ is the positive root of $\lambda + cR = \lambda M_X(R)$, as described above
• Some texts define a Lundberg coefficient $\alpha$ as the positive root of $\lambda M_X(r) = \lambda + cr$ written in a slightly different parameterization

In most standard actuarial references, these are the same thing. If your course distinguishes them, pay close attention to which equation each one solves. The key point is that this coefficient directly controls the exponential decay rate of the ruin probability.

Exact ruin probabilities

Lundberg's inequality is useful but conservative. When you need the actual ruin probability, exact formulas are available for certain claim size distributions.

Ruin probability as a geometric compound

For exponential claim sizes with mean $1/\beta$, the exact ruin probability has a clean closed form:

$\psi(u) = \frac{\lambda}{\beta c} e^{-(\beta - \lambda/c)u}$

For mixtures of exponentials (hyper-exponential distributions), the ruin probability becomes a sum of exponential terms:

$\psi(u) = \sum_{i=1}^{n} A_i e^{-R_i u}$

where $R_i$ are the positive roots of Lundberg's fundamental equation and $A_i$ are constants determined by the mixing weights and parameters. This structure comes from partial fraction decomposition of the Laplace transform of the ruin probability.

Beekman's convolution formula

For general claim size distributions, the ruin probability satisfies this integral equation (also called a defective renewal equation):

$\psi(u) = \frac{\lambda}{c} \int_0^u \psi(u - x)\,\overline{F}_X(x)\,dx + \frac{\lambda}{c} \int_u^\infty \overline{F}_X(x)\,dx$

where $\overline{F}_X(x) = 1 - F_X(x)$ is the survival function of the claim size distribution.

Reading this equation:

• The second term is the probability of ruin on the first claim (a single claim large enough to wipe out surplus $u$)
• The first term accounts for ruin happening after the first claim, through a convolution with the remaining surplus

This can be solved numerically (e.g., by discretization or recursive methods) or analytically for special distributions.

Heavy-tailed claim size distributions

Heavy-tailed distributions like Pareto and lognormal have tail probabilities that decay slower than any exponential. This fundamentally changes the ruin picture: Lundberg's inequality no longer applies (the MGF doesn't exist for positive arguments), and ruin probabilities are much higher for the same initial surplus.

Subexponential distributions in ruin theory

A distribution $F$ is subexponential if:

$\lim_{x \to \infty} \frac{\overline{F^{*2}}(x)}{\overline{F}(x)} = 2$

This condition means that the sum of two independent claims exceeding a large threshold is almost entirely driven by one of them being large, not both. The "single big jump" principle captures this: ruin with heavy-tailed claims is typically caused by one catastrophic claim, not an accumulation of many moderate ones.

Common subexponential distributions:

• Pareto: $\overline{F}(x) = (x_0/x)^\alpha$ for $x > x_0$
• Lognormal: $\ln(X) \sim N(\mu, \sigma^2)$
• Weibull with shape parameter $< 1$

Ruin probabilities for heavy-tailed claims

For subexponential claim sizes, the ruin probability has the asymptotic form:

$\psi(u) \sim \frac{\lambda}{c - \lambda E[X]} \int_u^\infty \overline{F}_X(x)\,dx \quad \text{as } u \to \infty$

This is sometimes written as $\psi(u) \sim \frac{\rho}{1 - \rho} \overline{F}_e(u)$, where $\rho = \lambda E[X]/c$ and $\overline{F}_e$ is the integrated tail distribution.

The practical consequence: ruin probability decreases much more slowly with initial surplus $u$ compared to light-tailed cases. Doubling your capital might cut ruin probability in half (polynomial decay) rather than reducing it by orders of magnitude (exponential decay).

Because exact computation is difficult, Monte Carlo simulation and the single big jump approximation are the standard tools for estimating ruin probabilities with heavy-tailed claims.

Surplus process modifications

The classical model assumes constant premium income, no investment returns, and no dividends. Real insurers don't operate this way. Several modifications make the model more realistic while preserving analytical tractability.

Brownian motion approximation

When individual claims are small relative to the surplus and premium income, the surplus process can be approximated by a Brownian motion with drift:

$U(t) \approx u + (c - \lambda E[X])t + \sqrt{\lambda E[X^2]}\, B(t)$

where $B(t)$ is a standard Brownian motion. The drift term $c - \lambda E[X]$ is the expected profit rate, and the volatility term $\sqrt{\lambda E[X^2]}$ captures the randomness of claims.

Under this approximation, the ruin probability has a closed-form expression:

$\psi(u) = \exp\left(-\frac{2(c - \lambda E[X])}{\lambda E[X^2]} u\right)$

This is a useful quick estimate, but it tends to underestimate ruin probability for heavy-tailed claims because Brownian motion has light (Gaussian) tails.

Diffusion approximation for ruin probability

The diffusion approximation generalizes the Brownian motion approach using the functional central limit theorem. It approximates the surplus process by a general diffusion process, which is valid when:

• Claims arrive frequently (large $\lambda$)
• Individual claim sizes are small relative to the surplus
• The portfolio is well-diversified

The ruin probability under the diffusion approximation is found by solving a boundary value problem for the associated ordinary differential equation (in the time-homogeneous case). This approach extends naturally to models with state-dependent premium rates or investment income, where the classical compound Poisson results don't have closed forms.

Ruin theory applications

The mathematical results above aren't just theoretical exercises. They directly inform how insurers manage capital, set prices, and structure risk transfer.

Optimal surplus and dividend problems

A central question in corporate finance for insurers: how should you balance holding capital (which reduces ruin probability) against paying dividends (which shareholders want)?

The barrier strategy is a classic approach:

1. Choose a dividend barrier $b$
2. Whenever surplus exceeds $b$, pay the excess as dividends
3. When surplus is below $b$, retain all income

The optimization problem is to find the barrier $b^*$ that maximizes the expected present value of total dividends paid until ruin. This is solved using Hamilton-Jacobi-Bellman (HJB) equations from stochastic control theory. The optimal barrier depends on the discount rate, claim distribution, and premium rate.

Reinsurance and ruin probability impact

Reinsurance reduces the insurer's net claim exposure at the cost of ceding some premium. Two main types:

• Proportional (quota-share): The reinsurer pays a fixed fraction $\alpha$ of every claim. This scales down both the mean and variance of net claims proportionally.
• Non-proportional (excess-of-loss): The reinsurer pays the portion of each claim exceeding a retention level $d$. This primarily reduces the tail risk.

Ruin theory helps determine the optimal retention level by analyzing how different reinsurance arrangements affect the adjustment coefficient $R$. A reinsurance contract that increases $R$ (tightening Lundberg's bound) improves long-term solvency. The trade-off is that ceding more risk means ceding more premium, which reduces the profit loading and can actually increase ruin probability if taken too far.