8.2 Lundberg's inequality and adjustment coefficients

Lundberg's inequality

Lundberg's inequality gives you a practical upper bound on the probability that an insurance company goes broke. Instead of computing the exact ruin probability (which is often intractable), this inequality lets you bound it with a clean exponential expression tied to the company's initial surplus and the riskiness of its portfolio.

The core idea: model the insurer's surplus as a random walk with a positive drift (premiums come in faster than claims go out, on average). Even with that positive drift, random fluctuations in claims can still push the surplus below zero. Lundberg's inequality quantifies the worst-case likelihood of that happening.

Definition of Lundberg's inequality

Lundberg's inequality states that the probability of ruin $\psi(u)$ satisfies:

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

where $u$ is the initial surplus and $R$ is the adjustment coefficient (also called the Lundberg exponent). The bound decreases exponentially as initial surplus increases, which matches intuition: the more capital you start with, the harder it is for claims to wipe you out.

Assumptions

For the classical Lundberg's inequality to hold, you need four conditions:

1. Premium income arrives at a constant rate $c$ per unit time.
2. Claim sizes $X_1, X_2, \ldots$ are independent and identically distributed with common distribution $F$ and mean $\mu$.
3. Claims arrive according to a Poisson process with rate $\lambda$.
4. The net profit condition holds: $c > \lambda\mu$. Premiums must exceed expected claims per unit time, otherwise ruin is certain.

The net profit condition is sometimes written using the safety loading $\theta$, defined by $c = (1 + \theta)\lambda\mu$ with $\theta > 0$.

The exponential bound

The bound $e^{-Ru}$ comes from a martingale argument applied to the surplus process (sometimes presented via moment generating function techniques related to large deviations). The adjustment coefficient $R$ is the unique positive solution to:

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

where $M_X(r) = E[e^{rX}]$ is the moment generating function (MGF) of the claim size distribution. A larger $R$ means the bound decays faster with surplus, so the portfolio is less risky.

Applications

• Capital requirements: Given a target ruin probability threshold, solve $e^{-Ru} \leq \epsilon$ for $u$ to find the minimum initial surplus needed.
• Premium setting: Adjusting $c$ changes $R$, so you can see how premium levels affect solvency risk.
• Portfolio comparison: Comparing $R$ values across portfolios gives a quick relative risk assessment.
• Reinsurance optimization: Evaluate how ceding part of the risk changes the adjustment coefficient and tightens the ruin bound.

Adjustment coefficients

The adjustment coefficient $R$ is the single most important parameter in Lundberg's inequality. It captures how the balance between premium income and claim risk translates into an exponential decay rate for the ruin probability.

Definition

The adjustment coefficient $R$ is defined as the unique positive root of the Lundberg equation:

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

where:

• $\lambda$ = Poisson claim arrival rate
• $M_X(r) = E[e^{rX}]$ = moment generating function of the claim size distribution
• $c$ = premium income per unit time

An equivalent form you'll sometimes see is $M_X(r) = 1 + \frac{cr}{\lambda}$, which is just a rearrangement.

Why does a unique positive root exist? At $r = 0$, both sides equal zero. The left side is convex (since $M_X(r)$ is convex) and grows faster than the linear right side for large $r$, but the net profit condition $c > \lambda\mu$ ensures the linear term dominates near the origin. This creates exactly one positive crossing point.

Calculating the adjustment coefficient

For most claim distributions, you can't solve the Lundberg equation in closed form. Here's the general approach:

1. Write down the MGF $M_X(r)$ for your claim size distribution.

2. Substitute into $\lambda(M_X(r) - 1) = cr$.

3. Rearrange to get $g(r) = \lambda(M_X(r) - 1) - cr = 0$.

4. Use a numerical root-finding method (Newton-Raphson or bisection) to find the positive root.

Exponential claims example (closed form): If claim sizes follow an exponential distribution with mean $\mu = 1/\alpha$, then $M_X(r) = \frac{\alpha}{\alpha - r}$ for $r < \alpha$. Substituting and solving:

$\lambda\left(\frac{\alpha}{\alpha - r} - 1\right) = cr$

$\frac{\lambda r}{\alpha - r} = cr$

Dividing both sides by $r$ (since $r > 0$):

$\frac{\lambda}{\alpha - r} = c$

$R = \alpha - \frac{\lambda}{c}$

With $c = (1+\theta)\lambda\mu = (1+\theta)\frac{\lambda}{\alpha}$, this simplifies to $R = \frac{\theta\alpha}{1+\theta}$.

Interpretation

• A higher $R$ means the ruin bound $e^{-Ru}$ decays faster, so the portfolio is safer.
• A lower $R$ means even large initial surpluses don't reduce the ruin bound much, signaling higher risk.
• $R$ increases when you raise the safety loading $\theta$ (charge higher premiums relative to expected claims).
• $R$ decreases when the claim size distribution becomes heavier-tailed or more variable.

Ruin probability

Definition

The ruin probability $\psi(u)$ is the probability that the surplus process 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-horizon ruin probability. It asks whether ruin ever occurs, not whether it occurs by a specific time. Finite-time ruin probabilities $\psi(u, T)$ are harder to compute and typically require simulation or numerical methods.

Exact results vs. Lundberg's bound

For exponentially distributed claims, there's an exact formula:

$\psi(u) = \frac{1}{1+\theta} e^{-\frac{\theta\alpha}{1+\theta}u}$

Notice this matches the Lundberg bound's exponential form but with a prefactor $\frac{1}{1+\theta} < 1$. So Lundberg's inequality is indeed an upper bound here, and the gap between the bound and the true value is the factor $\frac{1}{1+\theta}$.

For general claim distributions, exact formulas rarely exist, which is why the Lundberg bound is so useful in practice.

The Cramér-Lundberg approximation

A refinement of Lundberg's inequality is the Cramér-Lundberg approximation, which says that for large $u$:

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

where $C = \frac{\lambda\mu}{c \cdot M_X'(R)/\lambda - 1}$ (the exact form of $C$ depends on the claim distribution). This isn't a bound but an asymptotic approximation. It's more accurate than the bare Lundberg bound for large initial surpluses.

Other approximations include the De Vylder approximation, which replaces the original surplus process with one having exponential claims matched on the first three moments. This tends to work well even for moderate values of $u$.

Surplus process

Definition

The classical surplus process (also called the Cramér-Lundberg model) is:

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

where:

• $u$ = initial surplus (starting capital)
• $c$ = premium rate (constant income per unit time)
• $N(t)$ = number of claims by time $t$ (Poisson process with rate $\lambda$)
• $X_i$ = size of the $i$-th claim (i.i.d. random variables)

The surplus starts at $u$, grows linearly from premiums, and drops randomly each time a claim arrives. The aggregate claims $S(t) = \sum_{i=1}^{N(t)} X_i$ form a compound Poisson process.

Connection to Lundberg's inequality

Lundberg's inequality bounds the probability that this surplus process ever goes negative. The positive drift condition $c > \lambda\mu$ ensures the surplus trends upward on average, but random claim clusters can still cause ruin. The adjustment coefficient $R$ is derived directly from the parameters of this process ($c$, $\lambda$, and the claim size distribution).

Modeling extensions

The classical model can be generalized in several ways:

• Sparre Andersen model: Replace the Poisson arrival process with a general renewal process, allowing non-exponential inter-claim times.
• Diffusion approximation: Add a Brownian motion term to capture premium variability or investment fluctuations.
• Dependent claims: Allow claim sizes to be correlated, relevant for catastrophic events.
• Variable premiums: Let $c$ depend on the current surplus level or past claim experience.

Each extension changes how you compute (or bound) the adjustment coefficient and ruin probability.

Premiums and claims

Premium calculation principles

The premium rate $c$ is typically set using a premium principle that adds a safety loading on top of expected claims. Common principles include:

• Expected value principle: $c = (1 + \theta)\lambda\mu$, where $\theta > 0$ is the safety loading factor. Simple and widely used.
• Variance principle: $c = \lambda\mu + \alpha\lambda E[X^2]$, adding a loading proportional to the variance of aggregate claims.
• Standard deviation principle: $c = \lambda\mu + \alpha\sqrt{\lambda E[X^2]}$, loading proportional to the standard deviation.

The safety loading $\theta$ directly affects $R$: higher loading means larger $R$ and a tighter ruin bound.

Claim size distributions

The choice of claim size distribution has a major impact on the adjustment coefficient. Common distributions and their tail behavior:

Impact on the adjustment coefficient

The adjustment coefficient is sensitive to both the premium level and the claim distribution:

• Increasing the mean claim size $\mu$ (with $c$ fixed) shrinks $R$ because the net profit margin narrows.
• Increasing claim size variance (even with the same mean) shrinks $R$ because heavier tails make the MGF grow faster.
• Increasing the safety loading $\theta$ grows $R$ because there's more buffer against adverse claim experience.

If the net profit condition $c > \lambda\mu$ is violated, no positive $R$ exists and ruin is certain ($\psi(u) = 1$ for all $u$).

Generalized Lundberg's inequality

Extensions beyond the classical model

The classical Lundberg's inequality assumes Poisson arrivals, i.i.d. claims, and constant premiums. Several extensions relax these:

• Sparre Andersen model: Allows general inter-claim time distributions. A Lundberg-type bound still holds, but the equation defining $R$ changes to involve the joint transform of the inter-claim time and claim size.
• Markov-modulated models: The claim arrival rate and/or claim size distribution depend on an underlying Markov chain (e.g., modeling economic regimes). The adjustment coefficient becomes the solution to a matrix equation.
• Dependent claims: When claim sizes are correlated, the classical MGF-based equation no longer applies. Bounds can still be derived using supermartingale arguments, but they're typically less tight.

Implications for ruin probability

Relaxing assumptions generally makes the analysis harder but more realistic:

• Generalized models may yield tighter or looser bounds depending on the specific dependence structure.
• In some cases (e.g., positively correlated claims), the true ruin probability is higher than what the classical model predicts, so using the classical bound could be dangerously optimistic.
• Numerical methods (matrix-analytic methods, recursive algorithms) and Monte Carlo simulation become essential tools when closed-form bounds aren't available.

Practical applications

• Reinsurance design: Generalized models help evaluate how different reinsurance treaties (quota share, excess-of-loss) affect the adjustment coefficient under realistic claim dependencies.
• Regulatory capital: Solvency frameworks like Solvency II require risk assessments that go beyond classical assumptions, making generalized ruin theory directly relevant.
• Product design: Understanding how claim characteristics affect $R$ helps actuaries design products with manageable ruin risk profiles.