Loss Distributions: Modeling Insurance Claims with Probability

Choosing a Loss Distribution

Insurance losses are typically non-negative, right-skewed, and may have heavy tails. The choice of distribution depends on the type of insurance and the tail behavior of the losses. Light-tailed distributions (exponential, gamma, Weibull) are appropriate for routine claims like auto physical damage or medical expenses. Heavy-tailed distributions (Pareto, lognormal) better model catastrophic or liability claims where extremely large losses are possible.

Key characteristics to consider when selecting a loss distribution include: the support (most loss distributions are defined on [0, infinity)), the tail behavior (how quickly P(X > x) decreases), the number of parameters (more parameters allow more flexible fitting), and whether the distribution has closed-form expressions for moments and insurance quantities like E[min(X, d)].

On Exam P, you do not need to justify a distributional choice, but you should be able to work efficiently with each of these distributions, computing expected values, variances, and insurance modifications like deductibles and policy limits.

The Pareto Distribution

The two-parameter Pareto distribution has survival function S(x) = (theta / (theta + x))^alpha for x > 0, where alpha > 0 is the shape parameter and theta > 0 is the scale parameter. The PDF is f(x) = alpha * theta^alpha / (theta + x)^(alpha + 1).

The moments of the Pareto exist only for sufficiently large alpha. The mean is theta / (alpha - 1) for alpha > 1, and the variance is alpha * theta^2 / ((alpha - 1)^2 * (alpha - 2)) for alpha > 2. The k-th moment exists only when alpha > k. This dependence of moment existence on the shape parameter is a hallmark of heavy-tailed distributions.

The Pareto is heavy-tailed because its survival function decreases polynomially: S(x) behaves like (theta/x)^alpha for large x. Compare this to the exponential, whose survival function S(x) = e^(-lambda*x) decreases exponentially fast. The polynomial decay means the Pareto assigns much more probability to extreme values.

For insurance calculations with deductibles: E[(X - d)+] = (theta + d) / (alpha - 1) * (theta / (theta + d))^alpha when alpha > 1. The limited expected value is E[min(X, d)] = theta/(alpha - 1) * [1 - (theta/(theta + d))^(alpha - 1)]. These formulas appear frequently on Exam P.

The mean excess function for the Pareto is e(d) = (theta + d) / (alpha - 1), which increases linearly with d. An increasing mean excess function means that the larger a loss is, the more you expect it to exceed its current value. This is a defining property of heavy-tailed distributions and contrasts with the constant mean excess function of the exponential.

The Lognormal Distribution

If ln(X) ~ Normal(mu, sigma^2), then X follows a lognormal distribution. The parameters mu and sigma are the mean and standard deviation of the log-transformed variable, not of X itself. The moments of X are E[X] = exp(mu + sigma^2/2) and Var(X) = exp(2*mu + sigma^2) * (exp(sigma^2) - 1).

Probabilities are computed by converting to the normal distribution: P(X <= x) = Phi((ln(x) - mu) / sigma), where Phi is the standard normal CDF. This conversion makes the lognormal easy to work with on Exam P, since you can use the standard normal table.

The lognormal is widely used for modeling insurance claim sizes because it is always positive, right-skewed, and has a heavier tail than the exponential (though not as heavy as the Pareto). It also arises naturally when losses are the product of many independent factors, by the multiplicative central limit theorem.

Example: If claim amounts follow a lognormal distribution with mu = 7 and sigma = 1.5, find P(X > 5000). Compute ln(5000) = 8.517. Then P(X > 5000) = P(Z > (8.517 - 7)/1.5) = P(Z > 1.011) = 1 - Phi(1.011) = 1 - 0.844 = 0.156.

The Weibull Distribution

The Weibull distribution has survival function S(x) = exp(-(x/theta)^tau) for x > 0, where theta > 0 is the scale parameter and tau > 0 is the shape parameter. When tau = 1, the Weibull reduces to the Exponential(1/theta).

The hazard function h(x) = (tau/theta) * (x/theta)^(tau - 1) is increasing when tau > 1 (wear-out failure) and decreasing when tau < 1 (infant mortality). This flexibility makes the Weibull popular for modeling lifetimes and durations.

The mean of the Weibull is theta * Gamma(1 + 1/tau), which involves the gamma function. For Exam P, you typically work with specific values of tau that simplify the calculations, such as tau = 2 (Rayleigh distribution) or tau = 1 (exponential).

The Gamma Distribution as a Loss Model

The Gamma(alpha, lambda) distribution with PDF f(x) = (lambda^alpha / Gamma(alpha)) * x^(alpha-1) * e^(-lambda*x) for x > 0 provides a flexible family for modeling losses. With alpha = 1 it reduces to the exponential; larger alpha values produce more bell-shaped (but still right-skewed) densities.

The gamma distribution is light-tailed (exponential decay), making it suitable for moderate-sized claims. Its mean is alpha/lambda and variance is alpha/lambda^2. The coefficient of variation is 1/sqrt(alpha), which decreases as alpha increases, meaning higher alpha values produce less variable (relative to the mean) claim sizes.

On Exam P, the gamma distribution often appears in compound models and as a mixing distribution. If individual claims follow a gamma distribution and claim counts follow a Poisson process, the aggregate loss model has tractable properties.

Comparing Loss Distributions

Understanding how distributions compare in their tail behavior helps you select and interpret models on Exam P:

• Exponential: constant hazard rate, light tail, memoryless property
• Gamma: flexible shape via alpha parameter, light tail, reduces to exponential when alpha = 1
• Weibull: flexible hazard rate (increasing or decreasing), light tail, reduces to exponential when tau = 1
• Lognormal: heavy tail, multiplicative structure, probabilities via normal tables
• Pareto: very heavy tail, polynomial decay, increasing mean excess function

The ordering from lightest to heaviest tail is roughly: exponential, gamma, Weibull (tau > 1), lognormal, Pareto. On Exam P, if a problem describes a loss distribution with a heavy tail or an increasing mean excess function, think Pareto first.

Study Tips

For Exam P, build a reference table with the PDF, CDF, survival function, mean, variance, and key insurance formulas (E[min(X, d)] and E[(X - d)+]) for each loss distribution. Focus on the exponential and Pareto distributions, as these appear most frequently. Practice computing probabilities for the lognormal by converting to the standard normal. Know the mean excess function for the exponential (constant) and Pareto (linearly increasing). When a problem gives you a distribution and asks for an insurance quantity, identify the distribution first, then apply the appropriate formula. These problems typically take three to five minutes on exam day.