Partial Credibility for Aggregate Loss

Why Aggregate-Loss Credibility Needs Its Own Standard

For claim frequency, the quantity being judged is a count or rate. For aggregate loss, the quantity is a dollar total, or a dollar amount per unit of exposure. The variability of that amount depends on both how often claims occur and how large they are. A full-credibility standard for aggregate losses must therefore reflect the aggregate loss variance, not just the variability of claim counts.

Limited-fluctuation credibility first asks how much exposure is needed for experience to meet a specified precision and probability requirement. When observed exposure is below that full standard, the square-root rule assigns a partial-credibility factor. This factor measures the exposure relative to the full standard; it is not itself a guarantee that the full-credibility probability requirement has been met.

Definition: Let $S_e$ be the aggregate loss from $e$ units of homogeneous exposure. Write $E[S_e]=e\mu$ and $\operatorname{Var}(S_e)=e\tau^2$, where $\mu$ and $\tau^2$ are the mean and variance of aggregate loss per unit of exposure. A relative-error requirement with tolerance $r$ and two-sided confidence level $p$ is $\displaystyle P(|S_e-e\mu|\le r e\mu)\ge p.$ The full-credibility standard $e_0$ is the exposure required to meet this requirement under the stated model or approximation.

From Aggregate-Loss Variance to Full Credibility

For a compound Poisson model, suppose the number of claims per unit exposure is Poisson with mean $\lambda$, and individual claim amounts $X$ are independent and identically distributed, independent of the claim count, and have a finite second moment. The aggregate loss per unit exposure has mean $\lambda E[X]$ and variance $\lambda E[X^2]$. In particular, its variance uses the second raw moment of severity.

For a large enough exposure, use the normal approximation to $S_e$. Let $z=\Phi^{-1}((1+p)/2)$, where $\Phi$ is the standard normal distribution function. The approximation turns the two-sided requirement into a condition on the standard deviation:

P\left(\left|S_e-e\mu\right|\le re\mu\right) \approx 2\Phi\left(\frac{re\mu}{\sqrt{e\tau^2}}\right)-1 \ge p.

Since $z$ is the standard-normal quantile for the specified two-sided confidence level, the condition is equivalent to $re\mu/\sqrt{e\tau^2}\ge z$. Squaring both nonnegative sides and solving for $e$ gives the full standard. The square-root partial-credibility rule follows because the standard deviation of the aggregate grows as $\sqrt e$, while its mean grows as $e$.

e_0=\frac{z^2\tau^2}{r^2\mu^2}, \qquad Z=\min\left(1,\sqrt{\frac{e}{e_0}}\right).

Here $Z$ is the partial-credibility factor. The cap at one means that exposure exceeding the full standard is treated as full credibility, not as credibility greater than one. For compound Poisson aggregate losses, substitute $\mu=\lambda E[X]$ and $\tau^2=\lambda E[X^2]$:

e_0 = \frac{z^2\lambda E[X^2]} {r^2\left(\lambda E[X]\right)^2} = \frac{z^2E[X^2]} {r^2\lambda(E[X])^2}.

When this formula applies: The formula uses a normal approximation and assumes that aggregate loss has mean and variance proportional to exposure, as in a homogeneous compound Poisson model with independent claim amounts and finite second moment. For other models, use the appropriate per-unit mean $\mu$ and variance $\tau^2$ if they scale linearly with exposure. If the model does not have that variance structure, or if the normal approximation is unsuitable, this standard should not be used without adjustment. A one-sided confidence requirement also uses a different normal quantile.

Worked Applications

Worked Example: Aggregate Loss With Fixed Claim Severity

A hypothetical portfolio has a Poisson claim frequency of $\lambda=0.10$ claims per policy-year. Every claim is exactly $1,000. Determine the full-credibility standard for aggregate loss to be within 20% of its expected value with 90% two-sided confidence, then calculate the partial factor for 250 policy-years. Use the normal approximation.

Method. Calculate the compound Poisson mean and variance per policy-year, find $e_0$, and apply the capped square-root rule.

The severity is constant at $1,000, so $E[X]=1{,}000$ and $E[X^2]=1{,}000{,}000$. Thus the mean aggregate loss per policy-year is $\mu=0.10(1{,}000)=\$100$, and the variance per policy-year is $\tau^2=0.10(1{,}000{,}000)=100{,}000$ dollars squared.

For 90% two-sided confidence, $z=\Phi^{-1}(0.95)=1.6449$. With $r=0.20$,

e_0 = \frac{(1.6449)^2(100{,}000)} {(0.20)^2(100)^2} = \frac{2.70569601(100{,}000)}{400} = 676.4240\text{ policy-years}.

The observed exposure is $e=250$ policy-years, below the full standard. Therefore, $\displaystyle Z=\sqrt{250/676.4240}=\sqrt{0.3696}=0.6079.$

The partial-credibility factor is 0.6079. The calculation is dimensionally consistent: both $e$ and $e_0$ are measured in policy-years, so their ratio is dimensionless. The result is capped at one if exposure reaches or exceeds the full standard.

Common mistake avoided: Using the claim count standard alone would ignore the $1,000 loss attached to each claim. Here the requested quantity is aggregate dollars, so the relevant variance includes claim severity.

Worked Example: A Variable Severity Distribution

A hypothetical coverage has a Poisson frequency of $0.04$ claims per exposure unit. Each claim is $100 or $900, with probability $0.5$ for each amount. Find the exposure required for aggregate loss to be within 25% of its mean with 90% two-sided confidence. Then find the partial factor at 400 exposure units.

Method. Compute the first and second raw severity moments, then use the compound Poisson full standard.

The severity mean and second moment are $\displaystyle E[X]=0.5(100)+0.5(900)=500$ and $\displaystyle E[X^2]=0.5(100^2)+0.5(900^2)=0.5(10{,}000)+0.5(810{,}000)=410{,}000.$ Thus $\mu=0.04(500)=20$ dollars per exposure unit, and $\tau^2=0.04(410{,}000)=16{,}400$ dollars squared per exposure unit.

For 90% two-sided confidence, $z=1.6449$. Substituting $r=0.25$ gives

e_0 = \frac{(1.6449)^2(16{,}400)} {(0.25)^2(20)^2} = \frac{2.70569601(16{,}400)}{25} = \frac{44{,}373.4146}{25} = 1{,}774.9366\text{ exposure units}.

At $e=400$, the partial factor is $\displaystyle Z=\sqrt{400/1{,}774.9366}=\sqrt{0.2254}=0.4747.$ The full standard is about 1,774.94 exposure units, and the factor at 400 units is 0.4747.

Check. Since $400<1{,}774.9366$, the factor must be below one, as it is. The units are exposure units in both numerator and denominator of the ratio.

Worked Example: Partial Credibility for a Pure Premium

For a hypothetical portfolio, claim frequency is Poisson with mean $0.10$ claims per policy-year. Claim severity is exponential with mean $600, so its variance is $(600)^2$. Determine the full standard for estimating the expected pure premium within 15% with 95% two-sided confidence. Find the partial factor at 1,500 policy-years. If observed aggregate claims are $103,200 over that exposure, also calculate the observed pure premium.

Method. A pure premium is aggregate loss per policy-year. Divide the aggregate mean and variance by the appropriate exposure scaling, then apply the same relative-error standard.

For exponential severity, $E[X]=600$ and $\displaystyle E[X^2]=\operatorname{Var}(X)+(E[X])^2=600^2+600^2=720{,}000.$ The aggregate mean per policy-year is $\mu=0.10(600)=\$60$; its variance per policy-year is $\tau^2=0.10(720{,}000)=72{,}000$ dollars squared. For 95% two-sided confidence, $z=\Phi^{-1}(0.975)=1.9600$.

Therefore,

e_0 = \frac{(1.9600)^2(72{,}000)} {(0.15)^2(60)^2} = \frac{3.8416(72{,}000)}{81} = \frac{276{,}595.2}{81} = 3{,}414.7556\text{ policy-years}.

At 1,500 policy-years, $\displaystyle Z=\sqrt{1{,}500/3{,}414.7556}=\sqrt{0.4393}=0.6628.$ The observed pure premium is $\displaystyle \$103{,}200/1{,}500=\$68.80$ per policy-year. Thus the full standard is 3,414.76 policy-years, the partial factor is 0.6628, and the observed pure premium is $68.80.

Check. The observed amount is aggregate claims divided by policy-years, giving dollars per policy-year. The partial factor is based on exposure and the variability standard; it is not obtained by multiplying the observed pure premium by $Z$.

Exam Traps and Model Checks

The most consequential calculation error is substituting severity variance for its second raw moment. Under a compound Poisson model, the number of claims is random as well as claim amounts. The resulting aggregate variance is $\lambda E[X^2]$, not $\lambda\operatorname{Var}(X)$. Since $E[X^2]=\operatorname{Var}(X)+(E[X])^2$, using severity variance alone generally understates aggregate variability and therefore understates the exposure required.

Common mistake: Suppose $E[X]=500$, $\operatorname{Var}(X)=160{,}000$, and $\lambda=0.04$. The incorrect variance calculation is $0.04(160{,}000)=6{,}400$. The correct second moment is $E[X^2]=160{,}000+500^2=410{,}000$, so the aggregate variance per exposure unit is $0.04(410{,}000)=16{,}400$. Use the correct variance in $e_0=z^2\tau^2/(r^2\mu^2)$.

Other checks help catch errors before they affect the credibility factor:

Use the quantile that matches the stated confidence convention. A two-sided confidence level $p$ uses $\Phi^{-1}((1+p)/2)$; do not silently substitute a one-sided quantile.
Keep the exposure unit consistent. If $e_0$ is in policy-years, $e$ must also be in policy-years.
Use the mean aggregate loss as the basis for relative error. For a pure premium, the estimate is aggregate loss divided by exposure, but the relative-error standard has the same exposure ratio.
Apply the cap: if $e\ge e_0$, set $Z=1$. A value above one is not a partial-credibility factor.
Remember that the normal calculation is an approximation. A small expected claim count or a highly skewed aggregate distribution can make it less reliable.

Key takeaway: For aggregate-loss limited-fluctuation credibility, identify the mean and variance per exposure unit, calculate the full standard $e_0=z^2\tau^2/(r^2\mu^2)$, and then use $Z=\min(1,\sqrt{e/e_0})$. Under a compound Poisson model, the variance calculation requires the severity second moment $E[X^2]$.

Check Your Understanding

For each problem, use the stated normal-approximation convention. Keep monetary and exposure units explicit.

A hypothetical risk has Poisson frequency $0.20$ claims per policy-year and fixed severity of $500. Find the full standard for aggregate loss to be within 10% of its expected value with 90% two-sided confidence. (Answer: 1,352.8 policy-years.)
Claim frequency is Poisson with mean $0.05$ per exposure unit, and severity is exponential with mean $1,000. Find $e_0$ for 20% relative error and 90% two-sided confidence, then find $Z$ at 200 exposure units. (Answer: $e_0=2{,}705.5$ exposure units; $Z=0.2719$.)
A hypothetical portfolio has Poisson frequency $0.05$ claims per policy-year and fixed severity $800. Determine the full standard for 25% relative error and 95% two-sided confidence, then find $Z$ at 600 policy-years. (Answer: $e_0=1{,}229.3$ policy-years; $Z=0.6987$.)
Aggregate loss per exposure unit has mean $120 and variance 72,000 dollars squared. Find $e_0$ for 15% relative error and 90% two-sided confidence, and the capped partial factor at 3,000 exposure units. (Answer: $e_0=601.2$ exposure units; $Z=1.0000$.)
Claim frequency is Poisson with mean $0.20$ per exposure unit. Severity has mean $400 and variance 90,000 dollars squared. Find $e_0$ for 20% relative error and 90% two-sided confidence, then find $Z$ at 100 exposure units. (Answer: $e_0=528.4$ exposure units; $Z=0.4350$.)

What You'll Learn