advanced 12 min read
Finance & economics · Topic
Loss Models
probability theory · random processes · numerical methods
Loss modeling sits at the heart of non-life (property and casualty) insurance mathematics. Actuaries model claim counts with frequency distributions and individual claim sizes with severity distributions, then combine them into an aggregate loss distribution for the portfolio. Techniques range from the elegant Panjer recursion for exact computation to stop-loss premiums and risk measures for capital adequacy. Understanding these tools is essential for pricing, reserving, and Solvency II internal model work.

Severity Distributions

The severity (individual loss size) $X$ is modeled by heavy-tailed distributions that capture the possibility of large claims.

Lognormal: If $\ln X \sim N(\mu, \sigma^2)$ then:

\[f_X(x) = \frac{1}{x\sigma\sqrt{2\pi}}\exp\!\left(-\frac{(\ln x - \mu)^2}{2\sigma^2}\right), \quad x > 0\] \[E[X] = e^{\mu + \sigma^2/2}, \quad \text{Var}(X) = e^{2\mu+\sigma^2}(e^{\sigma^2}-1)\]

Pareto (single-parameter): $F(x) = 1 - (\theta/(\theta+x))^\alpha$. The mean excess loss $e(d) = E[X-d \mid X>d]$ is linear in $d$:

\[e(d) = \frac{\theta + d}{\alpha - 1}, \quad \alpha > 1\]

making the Pareto the canonical heavy-tailed severity model.

Gamma: Shape $\alpha$, rate $\beta$. Closed under convolution for equal-rate gammas: sum of $n$ independent $\text{Gamma}(\alpha_i, \beta)$ is $\text{Gamma}(\sum \alpha_i, \beta)$.

Weibull: $F(x) = 1 - e^{-(x/\theta)^\tau}$. For $\tau < 1$ the hazard decreases (light-tailed); for $\tau > 1$ it increases. At $\tau = 1$, exponential.

Distribution Tail behavior Mean excess Typical use
Lognormal Moderate heavy Increasing General property
Pareto Heavy (power) Linear in $d$ Large liability
Gamma Light-medium Decreasing Workers’ comp
Weibull Flexible Varies Reliability, WC

Frequency Distributions and the (a, b, 0) Class

The frequency $N$ (number of claims) follows a counting distribution. The $(a, b, 0)$ class satisfies the recursion:

\[\frac{p_k}{p_{k-1}} = a + \frac{b}{k}, \quad k = 1, 2, 3, \ldots\]

This single recursion generates three canonical distributions:

Distribution $a$ $b$ $p_0$
Poisson($\lambda$) $0$ $\lambda$ $e^{-\lambda}$
Negative Binomial($r, p$) $1-p$ $(r-1)(1-p)$ $p^r$
Binomial($m, q$) $-q/(1-q)$ $(m+1)q/(1-q)$ $(1-q)^m$

The Poisson is the default model ($\text{Var}(N) = E[N]$). The Negative Binomial handles overdispersion ($\text{Var}(N) > E[N]$), while the Binomial handles underdispersion. Zero-modified and zero-truncated variants extend coverage to portfolios with excess zeros.

Compound Distributions and the Panjer Recursion

The aggregate loss for a portfolio is:

\[S = X_1 + X_2 + \cdots + X_N = \sum_{i=1}^{N} X_i\]

where $N$ is the claim count (frequency) and $X_i$ are i.i.d. severities independent of $N$. The mean and variance are:

\[E[S] = E[N]\,E[X], \quad \text{Var}(S) = E[N]\,\text{Var}(X) + \text{Var}(N)\,(E[X])^2\]

For discrete severities (or after discretization), the Panjer recursion computes the aggregate distribution exactly:

\[f_S(s) = \frac{1}{1 - a\,f_X(0)} \sum_{x=1}^{s} \left(a + \frac{bx}{s}\right) f_X(x)\, f_S(s-x), \quad s = 1, 2, \ldots\]

with $f_S(0) = p_N(0)$ if $f_X(0) = 0$, or more generally:

\[f_S(0) = P_N(f_X(0))\]

where $P_N$ is the probability generating function of $N$. The Panjer recursion is $O(n^2)$ in the number of discretization points — far more efficient than direct convolution or Fast Fourier Transform for moderate-sized loss ranges, though FFT wins for very large grids.

Stop-Loss Premiums and the Limited Expected Value

The stop-loss premium with retention $d$ (also called the net stop-loss premium) is the pure premium for excess-of-loss reinsurance:

\[\pi(d) = E[\max(S - d, 0)] = E[(S-d)_+] = \int_d^{\infty} (s-d)\,f_S(s)\,ds\]

It can be rewritten as:

\[\pi(d) = E[S] - d + \int_0^d F_S(s)\,ds = E[S] - d + \int_0^d F_S(s)\,ds\]

or equivalently $\pi(d) = E[S] - E[\min(S, d)]$ where $E[\min(S,d)]$ is the limited expected value:

\[E[X \wedge d] = \int_0^d [1 - F_X(x)]\,dx = E[X] - \pi(d)\]

For the Pareto($\alpha$, $\theta$) severity:

\[E[X \wedge d] = \frac{\theta}{\alpha-1}\left[1 - \left(\frac{\theta}{\theta+d}\right)^{\alpha-1}\right], \quad \alpha > 1\]

The excess loss variable (residual life) given the loss exceeds $d$ is:

\[Y = X - d \mid X > d, \quad E[Y] = \frac{E[(X-d)_+]}{1 - F_X(d)} = e(d)\]

For exponential $X \sim \text{Exp}(\theta)$, $e(d) = \theta$ for all $d$ (memoryless property); for Pareto, $e(d) = (\theta+d)/(\alpha-1)$, linear and increasing.

Risk Measures for Insurance

The Value-at-Risk at confidence level $p$ is the $p$-quantile of aggregate losses:

\[\text{VaR}_p(S) = F_S^{-1}(p) = \inf\{s : F_S(s) \ge p\}\]

For Solvency II, $p = 99.5\%$ over a one-year horizon defines the Solvency Capital Requirement (SCR).

The Conditional Value-at-Risk (CVaR, also called TVaR or ES) averages losses above VaR:

\[\text{CVaR}_p(S) = E[S \mid S > \text{VaR}_p(S)] = \frac{1}{1-p}\int_p^1 \text{VaR}_u(S)\,du\]

For a continuous distribution:

\[\text{CVaR}_p(S) = \text{VaR}_p(S) + \frac{\pi(\text{VaR}_p(S))}{1-p}\]

where $\pi(\cdot)$ is the stop-loss premium. CVaR is coherent (satisfies subadditivity, monotonicity, translation invariance, positive homogeneity) while VaR is not in general.

Risk measure Formula Coherent? Solvency II use
VaR$_{99.5\%}$ $F_S^{-1}(0.995)$ No SCR standard formula
CVaR$_{99\%}$ $E[S\mid S > \text{VaR}_{99\%}]$ Yes Internal models
Standard deviation $E[S] + \lambda\,\text{SD}(S)$ No Simple pricing loads