Exposure rating is a tool for insurance pricing that allocates premium to bands of damage ratios or severity of losses. First ideas were published in (Salzmann 1963). It is often used to price non-proportional reinsurance contracts, such as excess of loss (XL) reinsurance.
Exposure rating uses the loss experience of a similar portfolio of policies to estimate the expected losses of the portfolio to be covered. The method is frequently used as a benchmark when there is no sufficient credible claims history from the client.
In this vignette, we first present the general notation and concepts of loss modelling. Secondly, we focus on destruction models implemented in the package. One popular model uses the MBBEFD distribution introduced by (Bernegger 1997). Finally, we provide an example of pricing of XL contract.
First, let us assume we have perfect information, i.e. we know the loss distribution for a certain risk.
To keep it simple, we assume the loss distribution is log-normal with a mean (M) of 65 and a coefficient of variation (CV) of 30%. The corresponding log-normal parameters μ and σ can be derived σ2 = log (1 + CV2), μ = log (M) − σ2/2.
The following chart shows the corresponding probability density curve f(x), cumulative distribution function F(x) and survival function S(x) = 1 − F(x).
In the insurance context, the survival function is
often called exceedance probability function, as it describes
the probability of exceeding a certain loss.
Let’s define X as the random variable that describes the ground up loss. The expected loss (in plain English the probability weighted sum of the losses) of a positive random variable is then E[X] = ∫0∞x f(x)dx.
Let’s further assume that losses are limited by α in the contract. Then the limited expected value (LEV), written as E[X ∧ α], is given as E[X ∧ α] = ∫0αx f(x)dx + ∫α∞α f(x)dx.
To evaluate this sum of integrals we recall that F(x) = ∫f(x)dx and F(∞) = 1. Hence ∫α∞α f(x)dx = α∫α∞f(x)dx = α − αF(α).
The integration by parts theorem helps with the first part of the integral ∫0αx f(x)dx = αF(α) − ∫0αF(x)dx.
Therefore, we retrieve the well-known identity E[X ∧ α] = α − ∫0αF(x)dx = ∫0α1 − F(x)dx = ∫0αS(x)dx.
The integral describes the area above the CDF up to α, or the area under the survival function up to α. In the example below the limit α was set to 100.
Suppose we want to insure the layer of claims between 80 and 100.
The expected loss cost would be the difference between the limited expected values of E[X ∧ 100] − E[X ∧ 80] and can be evaluated numerically
sigma <- sqrt(log(1 + 0.3^2))
mu <- log(65) - sigma^2/2
S <- function(x){ 1 - plnorm(x, mu, sigma) }
(lyr <- integrate(S, 80, 100)$value)
## [1] 2.22814
On the other hand the ratio of the original loss cost with a limit of 100 to a limit of 80 is called an increased limit factor (ILF):
## [1] 1.03592
Therefore we would expect the LEV to increase by 3.5% as the limit increases from 80 to 100.
Note ILFs are often used for pricing casualty business.
From the loss distribution we could read of the expected loss for any layer. However, often we will not have that level of information about the risk.
Instead, we will have to infer information from others risks that share similar loss characteristics as a function of the overall exposure, assuming that the relative loss size distribution is independent of the individual risk characteristic.
Hence, we will require a view on the expected full value loss cost for the overall exposure.
To make risks more comparable we look at ratio of losses to the underlying exposure, where the exposure is given as the sum insured (SI), or better the total insured value (TIV), or perhaps as the maximum probable loss (MPL).
Other definitions and measures are also popular, such as possible maximum loss (PML), estimate maximum loss (EML), or maximum foreseeable loss (MFL).
While the TIV and SI are straightforward to understand, the other metrics can be little more challenging to assess.
Suppose we insure a big production plant with of several buildings against fire. It is unlikely that a fire will destroy the whole facility, instead it is believed that the distance between the building will ensure that fires can be contained in a local area. Hence the MPL might be only the highest value of any of those buildings.
Shifting the metric from loss amounts to damage ratio or destruction rates (loss as a % of the exposure metric) allows us to compare and benchmarks risks.
Most insurance policies are written with a deductible, so that the insured will cover the first $X of the losses.
Thus, the reinsurer needs to understand, by how much the claims burden is reduced for a given deductible.
From the previous section we have learned how to calculate the limited expected value, which is the loss cost to the insured.
The exposure curve (or also called first loss curve or original loss cure ) is defined as the proportion of the LEV for a given deductible d compared to the overall expected claims cost E[X] $$ G(d) = \frac{E[X \wedge d]}{E[X]}. $$
For our example, using a log-normal distribution and a MPL of 200 we get:
MPL <- 200
ExpectedLoss <- 65
Deductible <- seq(0, MPL, 1)
G <- sapply(Deductible, function(x){
LEVx <- integrate(S, 0, x)$value
LEVx/ExpectedLoss
})
plot(Deductible/MPL, G,
t="l", bty="n", lwd=1.5,
main="Exposure Curve",
xlab="Deductible as % of MPL",
ylab="% of expected loss paid by insured",
xlim=c(0,1), ylim=c(0,1))
abline(a=0, b=1, lty=2)
The steepness of the curve is related to the severity of the loss distribution. The closer the curve is to the diagonal the greater the proportion of large loss.
If all losses were total losses then the exposure curve would be identical with the diagonal.
Different perils and exposures will have different exposure curves. Well known exposure curves are those used by Swiss Re and Lloyd’s.
As already mentioned, the exposure curve function of X is defined as the ratio of the limited expected value and the expectation. Since the loss X is a positive random variable, the exposure curve is $$ G_X(d) = \frac{\int_0^d (1-F_X(x))dx }{\int_0^1 (1-F_X(x))dx }. $$ Note that the exposure curve is a concave function for d ∈ ]0, 1[.
There is a direct link between the distribution function and the exposure curve. Since $$ F_X(x) = \left(1- \frac{G_X'(x)}{G_X'(0)}\right)1\!\!1_{[0,1[}(x) + 1\!\!1_{[1,+\infty[}(x), $$ defining the exposure curve or the distribution function is equivalent. The exposure curve is also a concave increasing function, see e.g. (Antal 2003).
The most trivial example of exposure curve is obtained for the uniform distribution on I. We consider FX(x) = x leading to GX(d) = d(2 − d).
A more interesting example is obtained for the Beta distribution on I (e.g.(Johnson, Kotz, and Balakrishnan 1994) for which the density is fX(x) = xa − 1(1 − x)b − 1/β(a, b) for x ∈ ]0, 1[ and a, b > 0 where β(., .) denotes the beta function, see e.g.(Olver et al. 2010). The distribution function is obtained in terms of the incomplete beta ratio function FX(x) = β(x; a, b)/β(a, b) = I(x; a, b). We get $$ G_X(d) \beta(a,b;x) - \frac{x^a(1-x)^b}{a\beta(a,b)} + x(1-\beta(a,b;x)) \frac{a+b}{a}, $$ where β(.; ., .) denotes the incomplete beta function.
Let us consider a continuous distribution function F0 of a random variable X0. The corresponding distribution function of the one-inflated random variable X1 is F1(x) = (1 − p1)F0(x) + p11 1[1, +∞[(x). There is no density but an improper density (1 − p)F0′(x) and a probability mass p1 at x = 1.
We consider the one-inflated beta distribution. Using Section , we obtain the following distribution function $$ F_X(x) = \left\{\begin{array}{ll} 0 & \text{if } x<0 \\ I(x;a,b)(1-p_1) & \text{if } 0\leq x <1 \\ 1 & \text{if } x\geq 1 \end{array}\right. $$ where I(x; a, b) denotes the incomplete beta ratio function. This leads to a non-null probability at x = 1, P(X = 1) = p1. The improper density function is $$ \tilde f_X(x) = (1-p_1) \frac{x^{a-1}(1-x)^{b-1}}{\beta(a,b)}. $$ The expectation is $$ E(X) = p_1 + (1-p_1)\frac{a}{a+b}. $$ The exposure curve is $$ G_X(d) = \frac{(1-p_1) \left(1 - I(d;a,b) \frac{b}{a+b} - \frac{d^a(1-d)^b}{(a+b)\beta(a,b)}\right) + p_1d }{p_1 + (1-p_1)\frac{a}{a+b}}. $$
We denote this first parametrization by MBBEFD(a, b). We define the parameter domain 𝒟a, b as Let us note that this domain includes two particular sets 𝒟a, 1 = {(a, 1), a + 1 > 0} and 𝒟0, b = {(0, b), b > 0}.
The MBBEFD distribution is defined by the following exposure curve for (a, b) ∈ 𝒟a, b The two special cases of a(1 − b) = 0 correspond to 𝒟a, 1 and 𝒟0, b. Note that the denominator is a normalizing constant to ensure the term belongs to [0, 1].
Differentiating GX, we obtain the following distribution function for (a, b) ∈ 𝒟a, b Note that the MBBEFD distribution is a mixed-type distribution with mass probability at x = 1 which equals to 1 when a(1 − b) = 0. In other words, for 𝒟a, 1 and 𝒟0, b, X has a Dirac distribution at x = 1. When a = +∞, the total loss probability is P(X = 1) = b.
For a(1 − b) > 0, the improper density function is The quantile function is
Using the definition of the exposure curve, we have E(X) = 1/GX′(0). The expectation for MBBEFD(a, b) is $$ E(X)=\frac{\ln(\frac{a+b}{a+1})}{\ln(b)} (a+1). $$ When a = 0 or b = 1, the expectation is simply E(X) = 1.
For fitting purposes and for verifying parameter constraints, (Bernegger 1997) proposed a second parametrization MBBEFD(g, b). Using the following parameter g = 1/pa, b, it is possible to reformulate the MBBEFD(a, b). That is $$ g= \frac{a+b}{(a+1)b} \Leftrightarrow a=\frac{(g-1)b}{1-gb}. $$ So g ≥ 1 guarantees that $\frac{a+b}{(a+1)b}\in [0,1]$, in addition to b > 0. The special case g = 1 leading to a Dirac distribution at x = 1 corresponds to a(1 − b) = 0 in the previous parametrization. The parameter domain is $$ \widetilde{\mathcal D}_{g,b}= \left\{ (g,b)\in\mathbb R^2, b>0, g\geq 1 \right\}. $$
The exposure curve is defined for x ∈ [0, 1] as Note that the case g > 1, bg = 1 implies g = 1/b, b < 1.
The resulting distribution function is As in the previous parametrization, there is a non-null probability at x = 1, The improper density function is for x ∈ ]0, 1[ The quantile function is
Let us compute the first two moments. $$ E(X) = 1/G_X'(0) = \left\{\begin{array}{ll} \frac{\ln(gb)(1-b)}{\ln(b)(1-gb)} & \text{ if } g>1, b\neq 1, b\neq 1/g \\ \frac{\ln(g)}{g-1} & \text{ if } g>1, b= 1 \\ \frac{b-1}{\ln(b)} & \text{ if } g>1, bg= 1 \\ 1 & \text{ if } g=1 \text{ or } b= 0 \end{array}\right. $$
We consider two fitting methods, namely the maximum likelihood
estimation (MLE) and the total-loss-moment matching estimation. Both
methods are implemented in the fitDR
function of the
package.
As its name suggest, the maximum likelihood function consists in maximizing the likelihood function defined as $$ \mathcal L(\theta, x_1,\dots,x_n)= \prod_{i=1}^n f_X(x_i), $$ where fX denotes the density of the loss random variable X and θ is the parameter vector. When X is a continuous non-inflated distribution, fX is the usual density function (obtained by differentiating the distribution function) when X is a mixed-type distribution with a non-null probability P(X = 1), fX is the product of P(X = x)1 1x = 1 and FX′(x)1 1x ≠ 1.
For example when considering the beta distribution, the likelihood function to be maximized is $$ \mathcal L((a,b), x_1,\dots,x_n)= \prod_{i=1}^n \frac{x_i^{a-1}(1-x_i)^{b-1}}{\beta(a,b)}. $$
The total-loss-moment matching estimation consists in matching the total loss and the first moments of a distribution. That is we solve the system $$ \left\{\begin{array}{ll} P(X = 1;\theta) = tl_n \\ E(X; \theta) = \bar x_n\\ \vdots \\ E(X^k; \theta) = m_{n,k}\\ \end{array}\right. \text{ where } \left\{\begin{array}{ll} tl_n = \frac{1}{n}\sum_{i=1}^n 1\!\!1_{x_i =1},\\ \bar x_n = \frac{1}{n}\sum_{i=1}^n x_i,\\ m_{n,k} = \frac{1}{n}\sum_{i=1}^n (x_i)^k. \end{array}\right. $$ The number of equation k + 1 equals the number of parameters. This methods is relevant only for one-inflated distributions as for classic continuous distributions such as the beta distribution we know that P(X = 1) = 0.
The function of the package returns an object of class inheriting from the class for which generic functions are available . There is also a bootstrap procedure implemented in the function.
Let us fit some distributions with a simulated dataset based on a mixture of two distributions (regular beta and mbbefd).
set.seed(123456)
x <- c(rbeta(50, 3, 1/2), rmbbefd(50, 1/2, 1/10))
f1 <- fitDR(x, "mbbefd", method="mle")
summary(f1)
## Fitting of the distribution ' mbbefd ' by maximum likelihood
## Parameters :
## estimate Std. Error
## a 0.035909278 NA
## b 0.009872438 NA
## Loglikelihood: -39.69145 AIC: 83.3829 BIC: 88.59324
## Correlation matrix:
## [1] NA
## Parametric bootstrap medians and 95% percentile CI
## Median 2.5% 97.5%
## a 0.03603436 0.012784198 0.06813441
## b 0.00930872 0.002569712 0.02273344
The graph correspond respectively to the histogram (and empirical density) against the fitted density and a two-dimensional level curve and scatter plot for bootstrap parameters.
As in the package, we can compare multiple fits on the same graph or with the same statistics. We just have to pass a list of objects.
f2 <- fitDR(x, "oibeta", method="mle")
f3 <- fitDR(x, "oiunif", method="mle")
gofstat(list(f1, f2, f3))
## Goodness-of-fit statistics
## 1-mle-mbbefd 2-mle-oibeta 3-mle-oiunif
## Kolmogorov-Smirnov statistic 0.1400000 0.1400000 0.2154661
## Cramer-von Mises statistic 0.2715687 0.1414871 1.7856663
## Anderson-Darling statistic Inf Inf Inf
##
## Goodness-of-fit criteria
## 1-mle-mbbefd 2-mle-oibeta 3-mle-oiunif
## Akaike's Information Criterion 83.38290 41.84037 82.99270
## Bayesian Information Criterion 88.59324 49.65588 85.59787
par(mfrow=c(1, 2))
cdfcomp(list(f1, f2, f3), leg=c("mbbefd", "oibeta", "oiunif"))
denscomp(list(f1, f2, f3), leg=c("mbbefd", "oibeta", "oiunif"),
ylim=c(0,4), xleg="topleft")
Of course, we can also compare the different fitted exposure curves using the function . That is