Common Errors in Exposure and Model Inputs

Why Model Inputs Must Match

An aggregate loss model combines a frequency model for the number of losses with a severity model for the amount of each loss. The calculation can be internally correct and still answer the wrong question if its inputs do not match. A claim count recorded over policy-months cannot be used as though it were a count over policy-years, and a sample of capped insurance payments is not automatically a sample of the underlying losses.

A reliable input check asks three questions: What is being counted? What exposure period does that count represent? What loss amount does the severity data measure? These checks are central to integrated actuarial calculations aligned with the SOA Exam FAM (Fundamentals of Actuarial Mathematics) syllabus.

Definition: Exposure is the amount of risk-time represented by a frequency observation, expressed in a specified unit such as policy-years or vehicle-days. A frequency rate is meaningful only with its exposure unit attached. Ground-up loss, denoted $X$, is the loss amount before policy coverage modifications. Payment per loss, denoted $Y^L$, is the insurer’s payment associated with each loss, including zero when the loss produces no payment. Payment per payment, denoted $Y^P$, is the amount conditional on a payment being made.

Match Frequency Counts to Exposure

Suppose a portfolio is modeled by a homogeneous Poisson frequency process: claims occur at a constant rate $\lambda$ per unit of exposure, with independent counts over disjoint exposure amounts. If $N$ claims are observed over total exposure $E$, the likelihood is proportional to $\lambda^N e^{-\lambda E}$, so maximizing it gives $\widehat{\lambda}=N/E$. The units of this estimate are claims per unit of $E$, not simply “claims per period.”

L(\lambda)\propto \lambda^N e^{-\lambda E}, \qquad \ell(\lambda)=N\ln(\lambda)-\lambda E+\text{constant}, \qquad \frac{d\ell}{d\lambda}=\frac{N}{\lambda}-E=0 \quad\Longrightarrow\quad \widehat{\lambda}=\frac{N}{E}.

If a rate is estimated per policy-month but the model requires a rate per policy-year, multiply by 12. If the desired projection is for a different exposure amount, multiply the rate by that exposure. Do not convert the count and exposure in inconsistent ways. The count must also have the same definition in the data and model: for example, counts of all reported losses should not silently be substituted for counts of paid claims.

Worked Example: Converting Policy-Month Exposure

A hypothetical insurer records 24 claims over 600 policy-months. Estimate the Poisson claim rate per policy-year and the expected number of claims over 125 policy-years. If mean ground-up severity is $2,500 and frequency and severity are independent, calculate the expected aggregate ground-up loss over those 125 policy-years.

Method. Convert exposure to policy-years before estimating the annual rate. The observed exposure is $600/12=50$ policy-years. Therefore,

\widehat{\lambda}=\frac{24}{50}=0.48 \quad\text{claims per policy-year}.

Over 125 policy-years, the expected claim count is $125(0.48)=60$ claims. For a compound model with independent severities of mean $2,500, the expected aggregate loss is expected claim count times mean severity:

E[S]=60(\$2{,}500)=\$150{,}000.

Check. The rate times exposure is $0.48(125)=60$, and multiplying by $2,500 gives $150,000. A common mistake is to call $24/600=0.04$ an annual rate; it is claims per policy-month. Treating it as annual would predict $125(0.04)=5$ claims, underestimating the expected count by a factor of 12. The correction is to convert the exposure unit, giving $0.04(12)=0.48$ claims per policy-year.

Exposure totals also need consistent units across records. A vehicle-day rate cannot be applied directly to vehicle-years without conversion. When risks differ substantially, a single common rate may not be appropriate; the model must then reflect the intended risk groups or exposure-specific rates rather than treating unlike exposures as interchangeable.

Read Frequency Parameters From the Model Definition

A parameter name alone does not define a distribution. In particular, negative binomial conventions differ over whether $p$ denotes the probability of success or failure. State the probability mass function before calculating. One standard convention counts the number of failures $N$ before the $r$th success, with success probability $p$:

Definition: Under the failures-before-$r$-successes convention, $N$ has probability mass function $P(N=k)=\binom{k+r-1}{k}p^r(1-p)^k$, for $k=0,1,2,\ldots$. Its mean is $r(1-p)/p$, and its variance is $r(1-p)/p^2$. Here $p$ is the success probability; the factor $1-p$ is the failure probability.

The formula and parameter meanings must travel together. A software output or data note that says only “negative binomial with $r$ and $p$” is incomplete until the convention is identified. A second check is dimensional: $N$, $r$, and $p$ are dimensionless, while an exposure-based rate such as $\lambda$ has units. Do not treat a negative binomial parameter as an annual rate unless the model explicitly defines it that way.

Worked Example: Checking a Negative Binomial Parameterization

A hypothetical portfolio uses the failures-before-$r$-successes convention for annual claim count, with $r=4$ and success probability $p=0.8$. Calculate the mean, variance, probability of zero claims, and probability of exactly two claims. Then identify the error in using $p=0.8$ as the failure probability in the same formula.

Method. Apply the moments and probability mass function for the stated convention. The failure probability is $1-p=0.2$, so

E[N]=\frac{4(0.2)}{0.8}=1, \qquad \operatorname{Var}(N)=\frac{4(0.2)}{0.8^2} =\frac{0.8}{0.64}=1.25.

For zero claims, $k=0$, so the binomial coefficient and failure-probability factor are both 1:

$$ P(N=0)=0.8^4=0.4096. $$

For exactly two claims,

P(N=2) =\binom{2+4-1}{2}(0.8)^4(0.2)^2 =10(0.4096)(0.04) =0.16384.

Thus $P(N\geq 1)=1-0.4096=0.5904$. The values are consistent with a mean of one claim: the zero-claim probability is substantial, while positive counts remain possible. A common mistake is to insert $0.8$ as the failure probability and $0.2$ as the success probability. That reverses the stated convention; it would give $P(N=0)=0.2^4=0.0016$, not 0.4096. If another convention uses $q$ for failure probability, the same model must use $q=0.2$, with the formula adjusted accordingly.

Match the Severity Variable to the Question

Severity observations can represent different quantities. A policy may have deductible $d$ and maximum payment $u$. Under this coverage, the payment per loss is $Y^L=\min(\max(X-d,0),u)$. Losses at or below the deductible produce zero payment; losses above it are reduced by the deductible and the payment is capped at the limit. Therefore, the distribution and mean of $Y^L$ generally differ from those of ground-up loss $X$.

A sample containing only positive payments estimates a payment-per-payment quantity, not the average payment per loss. In particular, its sample mean omits losses that produced zero payments. Neither that conditional mean nor the mean of all modified payments can be used as the ground-up mean without an appropriate model and information about the coverage modification. The intended severity variable must be specified before estimating or projecting.

E[Y^L]=E\!\left[\min(\max(X-d,0),u)\right], \qquad E[Y^P]=E[Y^L\mid Y^L>0],

The distinction follows directly from what is averaged. $E[Y^L]$ averages the payment over every loss, including zero-payment losses. $E[Y^P]$ averages only over losses for which $Y^L>0$. Neither operation reverses the deductible or limit: the mapping from $X$ to $Y^L$ discards information about losses below the deductible and above the cap.

Worked Example: Ground-Up Losses Versus Capped Payments

A hypothetical loss has the following ground-up severity distribution: $X=\$1{,}000$ with probability 0.4, $X=\$5{,}000$ with probability 0.4, and $X=\$10{,}000$ with probability 0.2. The policy has deductible $d=\$1{,}000$ and payment limit $u=\$3{,}000$. Calculate the ground-up mean, payment-per-loss mean, and payment-per-payment mean. If annual claim count is independent Poisson with mean 2 and severities follow this distribution, calculate expected annual ground-up loss and expected annual insurer payment.

Method. First calculate the policy payment for each possible ground-up loss:

\begin{aligned} X=\$1{,}000 &: \quad Y^L=\min(\max(1{,}000-1{,}000,0),3{,}000)=\$0,\\ X=\$5{,}000 &: \quad Y^L=\min(5{,}000-1{,}000,3{,}000)=\$3{,}000,\\ X=\$10{,}000 &: \quad Y^L=\min(10{,}000-1{,}000,3{,}000)=\$3{,}000. \end{aligned}

The ground-up mean weights each loss amount by its probability:

E[X]=0.4(\$1{,}000)+0.4(\$5{,}000)+0.2(\$10{,}000) =\$400+\$2{,}000+\$2{,}000 =\$4{,}400.

The payment-per-loss mean includes the zero payment:

E[Y^L]=0.4(\$0)+0.4(\$3{,}000)+0.2(\$3{,}000) =\$1{,}200+\$600 =\$1{,}800.

A payment is made with probability $0.4+0.2=0.6$. Conditional on a payment, it is always $3,000, so $E[Y^P]=\$3{,}000$. These are three different means: $4,400 for ground-up loss, $1,800 per loss, and $3,000 per payment.

With independent Poisson frequency of mean 2, expected annual ground-up loss is $2(\$4{,}400)=\$8{,}800$, while expected annual insurer payment is $2(\$1{,}800)=\$3{,}600$. Check. The payment-per-loss mean cannot exceed the ground-up mean here: $1,800 is below $4,400, as expected under this deductible and limit. A common mistake is to use the observed payment-per-payment mean of $3,000 as though it were the mean ground-up severity; this would give $2(\$3{,}000)=\$6{,}000$, which is neither expected ground-up loss nor expected insurer payment. The correction is to match the severity mean to the target aggregate quantity.

A Practical Input Audit

Before combining frequency and severity, write down the meaning and units of each input. This short audit catches many errors before they propagate into an expected loss or a more complex aggregate calculation.

Define the counted event.
State whether the frequency data count reported losses, claims with positive payments, or another clearly defined event.

Write the exposure unit.
Record the total exposure in the same unit used by the rate, such as policy-years. Convert units before estimating or projecting.

State the distribution convention.
For a frequency model, identify its probability formula and the meaning of every parameter before calculating moments or probabilities.

Name the severity variable.
Specify whether the amount is ground-up loss $X$, payment per loss $Y^L$, or payment per payment $Y^P$. Apply coverage modifications only when they match the target quantity.

Check the aggregate interpretation.
For independent claim severities and frequency, the expected aggregate amount is expected count times the mean of the matching severity variable.

The central discipline is consistency: frequency, exposure, and severity must describe the same event population and the same time basis, while the severity variable must measure the loss amount the calculation is intended to estimate. Labeling units and variables explicitly makes a factor-of-12 error, a reversed probability convention, or an inappropriate covered-payment mean much easier to detect.

Check Your Understanding

For each problem, check the exposure unit, model convention, and target severity before calculating.

A hypothetical book records 27 claims over 900 policy-months. Estimate the annual claim rate and the expected number of claims over 80 policy-years. (Answer: $0.36$ claims per policy-year; $28.8$ claims.)
A fleet records 15 claims over 6,000 vehicle-days. Assuming a constant Poisson rate, estimate the rate per vehicle-day and the expected count over 100 vehicles insured for 365 days each. (Answer: $0.0025$ claims per vehicle-day; $91.25$ claims.)
Under the failures-before-$r$-successes convention, $N$ has $r=3$ and $p=0.75$. Find its mean, variance, and probability of zero claims. (Answer: $E[N]=1$, $\operatorname{Var}(N)=1.3333$, $P(N=0)=0.4219$.)
Ground-up loss is $200, $1,200, or $4,000 with probabilities 0.5, 0.3, and 0.2, respectively. A policy has deductible $500 and payment limit $1,000. Find $E[X]$, $E[Y^L]$, and $E[Y^P]$. (Answer: $E[X]=\$1{,}260$, $E[Y^L]=\$410$, $E[Y^P]=\$820$.)
A hypothetical portfolio has 36 claims over 2,400 policy-months. Estimate the annual rate and expected claims over 75 policy-years. State the unit conversion used. (Answer: $0.18$ claims per policy-year; $13.5$ claims; $2{,}400$ policy-months equals $200$ policy-years.)

What You'll Learn