How Claim Frequency Changes Required Exposure

Why Frequency Changes Required Exposure

A credibility standard may say that an aggregate-loss result needs a specified number of expected claims to meet a fluctuation criterion. That number alone does not tell an actuary how much exposure is required. The conversion depends on how often claims occur per exposure unit: a low-frequency risk needs more exposure units to produce the same expected claim count.

The key is to keep the units explicit. If frequency is measured in claims per policy-year, the resulting exposure is in policy-years. If frequency is measured in claims per vehicle-year, the result is in vehicle-years. The arithmetic is simple, but a mismatch in time period or exposure base can make the answer materially wrong.

Definition: Let $e$ denote exposure measured in a stated unit, and let $\lambda$ be the expected number of claims per exposure unit. Under a constant-frequency model, the expected claim count at exposure $e$ is $\lambda e$. If $N_0$ is the required expected claim-count threshold, the exposure required to reach that threshold is $e_0=N_0/\lambda$.

This is a conversion of a threshold into exposure, not a new credibility criterion. The threshold $N_0$ is taken as given, or obtained from the applicable aggregate-loss standard. To isolate the effect of frequency, hold that threshold and all other assumptions fixed. Then doubling frequency halves the required exposure; halving frequency doubles it.

From an Aggregate-Loss Criterion to a Claim-Count Threshold

The frequency conversion is especially clear when the threshold comes from a compound Poisson aggregate-loss model. Assume claims occur according to a Poisson count with mean $\lambda e$. Claim severities are independent and identically distributed, independent of the count, with finite mean $\mu=E[X]$ and variance $\sigma^2=\operatorname{Var}(X)$. Aggregate loss is the sum of the severities for the claims that occur.

For a relative-error standard, suppose the aggregate loss is required to be within a fraction $r$ of its expected value with approximate two-sided confidence $p$. Let $z_p$ be the corresponding standard normal quantile, so that the normal approximation uses $P(|Z|\le z_p)=p$. The limited-fluctuation condition is $z_p\sqrt{\operatorname{Var}(S)}\le rE[S]$, where $S$ is aggregate loss. This is an approximate standard: it relies on the normal approximation being suitable and on the stated model assumptions.

E[S]=\lambda e\mu,\qquad \operatorname{Var}(S)=\lambda e(\sigma^2+\mu^2)

The variance expression follows from conditioning on the claim count $N$. Given $N$, the aggregate has mean $N\mu$ and variance $N\sigma^2$. Therefore, by the law of total variance, its unconditional variance is the expected conditional variance plus the variance of the conditional mean:

\operatorname{Var}(S) =E[N\sigma^2]+\operatorname{Var}(N\mu) =\lambda e\sigma^2+\lambda e\mu^2 =\lambda e(\sigma^2+\mu^2).

Substitute these moments into the relative-error condition. Squaring both nonnegative sides and dividing by positive quantities gives a requirement on the expected claim count $\lambda e$:

z_p\sqrt{\lambda e(\sigma^2+\mu^2)} \le r\lambda e\mu \quad\Longleftrightarrow\quad \lambda e\ge \frac{z_p^2(\sigma^2+\mu^2)}{r^2\mu^2} = \frac{z_p^2(1+\operatorname{CV}_X^2)}{r^2}, \qquad \operatorname{CV}_X=\frac{\sigma}{\mu}.

The right-hand side is the required expected claim count $N_0$ under this model and criterion. Once it is established, frequency enters through $\lambda e$, so the exposure requirement is $e_0=N_0/\lambda$. If the prescribed threshold is an integer number of expected claims, round the threshold up as required before dividing by frequency. Do not round the exposure down if doing so would leave expected claims below the threshold.

Identify the threshold.
Use the required expected claim count $N_0$ for the applicable criterion. If deriving it from an aggregate-loss standard, state the loss model and the criterion's assumptions.

State the frequency unit.
Write $\lambda$ as expected claims per exposure unit, such as claims per policy-year.

Divide and check units.
Calculate $e_0=N_0/\lambda$. The exposure unit in the answer must match the denominator unit in $\lambda$.

Worked Applications

Worked Example: Comparing Two Claim Frequencies

A hypothetical aggregate-loss standard requires 600 expected claims. For a fixed coverage and severity assumption, one risk class has an expected frequency of 0.12 claims per policy-year and another has 0.30 claims per policy-year. Find the required policy-years for each class and compare them.

Method. Convert the same expected claim-count threshold separately using $e_0=N_0/\lambda$. The threshold remains 600 for both classes; only frequency changes.

For the first class:

e_1=\frac{600\text{ expected claims}} {0.12\text{ claims per policy-year}} =5{,}000\text{ policy-years}.

Check: $5{,}000\times0.12=600$ expected claims.

For the second class:

e_2=\frac{600\text{ expected claims}} {0.30\text{ claims per policy-year}} =2{,}000\text{ policy-years}.

Check: $2{,}000\times0.30=600$ expected claims. The first class requires $5{,}000/2{,}000=2.5$ times as many policy-years because its frequency is $0.30/0.12=2.5$ times lower. Common mistake avoided: applying a lower-frequency class's rate to the higher-frequency class's exposure requirement, rather than using each class's own frequency.

Worked Example: Deriving a Threshold, Then Converting It

A hypothetical portfolio has claim severity with mean $\mu=\$4{,}000$ and standard deviation $\sigma=\$6{,}000$. Assume a compound Poisson model, a 10% relative-error tolerance, and an approximate two-sided 95% standard with $z_p=1.96$. Claim frequency is 0.25 claims per policy-year. Find the required policy-years. Use the normal limited-fluctuation approximation and round the expected claim threshold up to a whole claim.

Method. First calculate the severity coefficient of variation and the expected claim-count threshold. Then divide that threshold by frequency.

The severity coefficient of variation is:

\operatorname{CV}_X=\frac{\sigma}{\mu} =\frac{\$6{,}000}{\$4{,}000}=1.5, \qquad \operatorname{CV}_X^2=2.25.

Under the stated compound Poisson and normal-approximation assumptions, the threshold is:

N_0= \frac{z_p^2(1+\operatorname{CV}_X^2)}{r^2} = \frac{1.96^2(1+2.25)}{0.10^2}.

Compute each component: $1.96^2=3.8416$, $1+2.25=3.25$, and $0.10^2=0.01$. Thus:

N_0=\frac{3.8416(3.25)}{0.01} =\frac{12.4852}{0.01} =1{,}248.52.

Rounding up gives a required threshold of 1,249 expected claims. Convert that to policy-years:

e_0=\frac{1{,}249}{0.25} =4{,}996\text{ policy-years}.

Check: $4{,}996(0.25)=1{,}249$ expected claims. Common mistake avoided: using $\sigma/\mu=1.5$ where the formula requires its square, $2.25$. That would understate the aggregate-loss fluctuation threshold and the required exposure.

Worked Example: Keeping the Exposure Period Consistent

A hypothetical book has a required expected claim count of 900. Its frequency is 0.06 claims per policy-year. A proposed data extract contains 2,500 policies, each exposed for an average of six months. Determine (a) the required exposure in policy-years, (b) the extract's exposure and expected claim count, and (c) the additional policy-years needed.

Method. Calculate the required exposure in the same unit as the frequency, then convert the extract's six-month policy terms into earned policy-years.

(a) Required policy-years:

e_0=\frac{900}{0.06}=15{,}000\text{ policy-years}.

(b) Six months is $6/12=0.5$ year per policy. Therefore, the extract contains:

e_{\text{extract}}=2{,}500(0.5)=1{,}250\text{ policy-years}.

At 0.06 claims per policy-year, its expected claim count is:

E[N_{\text{extract}}]=1{,}250(0.06)=75\text{ claims}.

(c) The additional exposure needed is $15{,}000-1{,}250=13{,}750$ policy-years. Check: the total exposure of 15,000 policy-years yields $15{,}000(0.06)=900$ expected claims. Common mistake avoided: treating each six-month policy as a full policy-year. Doing so would incorrectly report 2,500 policy-years and 150 expected claims for the extract.

What the Inverse Relationship Does—and Does Not—Say

For a fixed $N_0$, the exposure formula $e_0=N_0/\lambda$ gives an exact inverse relationship between required exposure and frequency. For example, if frequency falls by 20%, exposure must be multiplied by $1/0.8=1.25$, or increased by 25%, to preserve the same expected count. This comparison is valid only when the threshold, frequency definition, and exposure unit remain consistent.

Frequency may also be part of the assumptions used to determine the threshold itself. In the compound Poisson relative-error derivation above, the required expected claim count depends on the severity coefficient of variation, tolerance, and confidence level; frequency then converts that count into exposure. If a change in risk class also changes severity, dependence, or the count model, it is not enough to change $\lambda$ while assuming the rest of the standard is unchanged.

Most importantly, $e_0=N_0/\lambda$ makes the expected count equal to the threshold. It does not guarantee that the observed count will be at least $N_0$. Under a Poisson model, the observed count is random even when its expectation equals $N_0$. A separate probability requirement on the observed count would need its own calculation; it should not be silently substituted for an expected-count standard.

Key takeaway: First establish the required expected claim count under the applicable standard. Then divide by expected claims per exposure unit. Lower frequency means more exposure is needed for the same expected number of claims, provided the other assumptions and the exposure basis are held fixed.

Check Your Understanding

For each problem, keep the frequency denominator and the requested exposure unit consistent.

A hypothetical risk has a required expected count of 480 claims and frequency of 0.16 claims per account-year. How many account-years are required? (Answer: 3,000 account-years)
Two hypothetical classes have the same 750-claim threshold. Their frequencies are 0.25 and 0.10 claims per vehicle-year. Find each required exposure and the ratio of lower-frequency to higher-frequency exposure. (Answer: 3,000 and 7,500 vehicle-years, respectively; the ratio is 2.5)
A standard requires 1,100 expected claims. A hypothetical portfolio has 4,000 policies exposed for nine months each, with frequency 0.08 claims per policy-year. Find the portfolio's exposure, expected count, and shortfall in exposure relative to the requirement. (Answer: 3,000 policy-years; 240 expected claims; 10,750 additional policy-years required)
Under a hypothetical compound Poisson standard, severity has mean $\$2{,}000$ and standard deviation $\$2{,}000$. The relative tolerance is 20%, the approximate two-sided confidence is 95% with $z_p=1.96$, and frequency is 0.40 claims per policy-year. Find the required expected claim threshold, rounded up, and the required policy-years. (Answer: 193 expected claims; 482.5 policy-years)
A frequency is reported as 0.03 claims per month of coverage, while a required threshold is 360 expected claims. Convert the frequency to claims per coverage-year and find the required coverage-years. (Answer: 0.36 claims per coverage-year; 1,000 coverage-years)

What You'll Learn