LLN and CLT in Aggregate Claim Models

Why Claim Averages and Aggregate Totals Behave Differently

An insurer may care about both the average amount paid per claim and the total amount paid across many claims. These quantities are connected, but they do not fluctuate on the same scale. As the number of claims grows, the average claim size tends to stabilize near its expected value, while the aggregate total grows roughly in proportion to the number of claims and continues to vary around its own, growing mean.

The law of large numbers (LLN) describes the stabilizing average. The central limit theorem (CLT) describes the approximate distribution of its fluctuations, and therefore also the fluctuations of the aggregate total. Both results require a clearly stated model. The examples below use a fixed number of independent, identically distributed claim amounts; no claim-count uncertainty is included.

Definition: Let $X_1,\ldots,X_n$ be independent and identically distributed claim amounts, each with finite mean $\mu=E[X_i]$ and finite, positive variance $\sigma^2=\operatorname{Var}(X_i)$. Define the aggregate claims and average claim amount by $$ S_n=\sum_{i=1}^{n}X_i,\qquad \bar{X}_n=\frac{S_n}{n}. $$ Then $E[S_n]=n\mu$, $\operatorname{Var}(S_n)=n\sigma^2$, $E[\bar{X}_n]=\mu$, and $\operatorname{Var}(\bar{X}_n)=\sigma^2/n$. The LLN states that $\bar{X}_n$ converges in probability to $\mu$. The iid CLT states that $(S_n-n\mu)/(\sigma\sqrt{n})$ converges in distribution to a standard normal random variable as $n$ grows.

The formulas for the mean and variance use a fixed number of claims. Independence gives $\operatorname{Var}(S_n)=\sum_i\operatorname{Var}(X_i)$; identical distributions make this sum $n\sigma^2$. If claim amounts are dependent, the variance generally includes covariance terms. If the number of claims is random, its variability must also be modeled; the fixed-$n$ formulas do not account for it.

How the LLN and CLT Translate to Claim Models

The LLN says that the average paid per claim becomes close to $\mu$ with high probability as the number of claims increases. Equivalently, aggregate claims per exposure, $S_n/n$, become close to $\mu$. It does not say that the total $S_n$ becomes close to a fixed number: its expected value is $n\mu$, which increases with $n$.

A short justification follows from the variance of the average. For any $\varepsilon>0$, Chebyshev’s inequality gives

P\left(\left|\bar{X}_n-\mu\right|\geq\varepsilon\right) \leq \frac{\operatorname{Var}(\bar{X}_n)}{\varepsilon^2} =\frac{\sigma^2}{n\varepsilon^2} \longrightarrow 0.

Thus the probability of an average claim amount differing from $\mu$ by at least a fixed positive amount tends to zero. The CLT adds a more detailed approximation for large $n$: the aggregate has approximate mean $n\mu$ and standard deviation $\sigma\sqrt{n}$, while the average has approximate mean $\mu$ and standard deviation $\sigma/\sqrt{n}$. These are approximations, not exact normal distributions in general.

Theorem: Under the iid assumptions above, for sufficiently large $n$, use $$ S_n\ \approx\ N(n\mu,n\sigma^2), \qquad \bar{X}_n\ \approx\ N\left(\mu,\frac{\sigma^2}{n}\right). $$ The second parameter in $N(\cdot,\cdot)$ is the variance. These approximations follow by rescaling the CLT: $S_n=n\mu+\sigma\sqrt{n}\,Z_n$, where $Z_n=(S_n-n\mu)/(\sigma\sqrt{n})$ is approximately standard normal. Dividing by $n$ gives $\bar{X}_n\approx\mu+(\sigma/\sqrt{n})Z_n$.

Finite mean and variance support the iid LLN and CLT as asymptotic results, but do not guarantee that a normal approximation is accurate for every particular sample size. Strong skewness, heavy tails, or a sample too small for the claim distribution can make the approximation less reliable. For independent claims that are not identically distributed, the aggregate mean and variance are sums of the individual means and variances; a normal approximation also requires suitable conditions preventing a small number of terms from dominating the total variation.

Worked Aggregate-Claim Calculations

Worked Example: An Exponential Claim Model

In a hypothetical one-year portfolio, exactly 100 independent claims are observed. Assume each claim amount $X_i$ has an exponential distribution with mean $\$1{,}000$, so its rate is $\lambda=1/1{,}000$ per dollar and its variance is $\sigma^2=\$1{,}000{,}000$. The model includes no discounting ($i=0$) and no mortality decrement. Approximate the probability that aggregate claims exceed $\$110{,}000$.

Method. Use the iid CLT for the aggregate. For this exponential distribution, $\mu=\$1{,}000$ and $\sigma=\$1{,}000$.

The aggregate mean and variance are

E[S_{100}]=100(1{,}000)=\$100{,}000,\qquad \operatorname{Var}(S_{100})=100(1{,}000{,}000)=100{,}000{,}000.

Therefore the aggregate standard deviation is $\sqrt{100{,}000{,}000}=\$10{,}000$. Standardizing the threshold gives

z=\frac{110{,}000-100{,}000}{10{,}000}=1.

Using the standard normal distribution, $P(S_{100}>\$110{,}000)\approx P(Z>1)=1-\Phi(1)=0.1587$. The answer is an approximation, rounded to four decimal places. It is consistent with the average-claim formulation: the same event is $\bar{X}_{100}>\$1{,}100$, whose approximate standard deviation is $\$1{,}000/\sqrt{100}=\$100$, giving the same standardized value $100/100=1$.

Check and common mistake avoided. The aggregate standard deviation is $\$10{,}000$, not $100(\$1{,}000)=\$100{,}000$. Standard deviations combine by the square root of the number of independent claims, not by multiplying by $n$. Using $\$100{,}000$ would give $z=0.1$ and an incorrect upper-tail probability of $0.4602$.

Worked Example: A Uniform Claim Model and an Average Threshold

For a hypothetical one-year block, exactly 64 independent claims have amounts $X_i\sim\operatorname{Uniform}(0,6{,}000)$ dollars. There is no discounting ($i=0$) or mortality decrement. Approximate the probability that the average claim amount exceeds $\$3{,}375$.

Method. Calculate the uniform distribution’s mean and variance, then apply the CLT to $\bar{X}_{64}$. For a uniform distribution on $[a,b]$, the mean is $(a+b)/2$ and the variance is $(b-a)^2/12$. Thus

\mu=\frac{0+6{,}000}{2}=\$3{,}000,\qquad \sigma^2=\frac{(6{,}000-0)^2}{12}=3{,}000{,}000.

The average has approximate mean $\$3{,}000$ and standard deviation

\frac{\sigma}{\sqrt{64}} =\frac{\sqrt{3{,}000{,}000}}{8} =\sqrt{46{,}875} \approx \$216.5064.

The threshold is $\$375$ above the mean, so its standardized value is

z=\frac{3{,}375-3{,}000}{216.5064}\approx 1.7321.

Consequently, $P(\bar{X}_{64}>\$3{,}375)\approx 1-\Phi(1.7321)=0.0416$, rounded to four decimal places. Equivalently, the aggregate threshold is $64(3{,}375)=\$216{,}000$; the aggregate mean is $64(3{,}000)=\$192{,}000$, and its standard deviation is $\sqrt{64(3{,}000{,}000)}\approx\$13{,}856.4065$. This gives the same standardized value.

Check and common mistake avoided. The average’s variance is $\sigma^2/64$, not $64\sigma^2$. The latter is the variance of the aggregate, not the average. The two formulations agree only when the mean, variance, and threshold are all scaled consistently.

Worked Example: Planning the Number of Claims

A hypothetical one-year portfolio is modeled using independent claim amounts with an exponential distribution of mean $\$800$, rate $\lambda=1/800$ per dollar, and variance $\$640{,}000$. There is no discounting ($i=0$) or mortality decrement. Use a CLT design approximation to find the smallest integer number $n$ for which the average claim amount is within $\$80$ of its mean with approximate probability at least $0.95$.

Method. The standard deviation of the average is $800/\sqrt{n}$. Under the normal approximation, the centered average divided by this standard deviation is approximately standard normal. A central probability of $0.95$ uses the standard normal critical value $1.96$, so the target requires

\frac{80}{800/\sqrt{n}}\geq 1.96.

Solving step by step,

0.1\sqrt{n}\geq1.96,\qquad \sqrt{n}\geq19.6,\qquad n\geq19.6^2=384.16.

Since $n$ must be an integer, the smallest candidate is $n=385$. At $n=385$, the standardized tolerance is $80\sqrt{385}/800\approx1.9621$, which exceeds $1.96$. At $n=384$, it is $0.1\sqrt{384}\approx1.9596$, which is below $1.96$. Thus 385 is the smallest count meeting the stated CLT design criterion. The implied central probability is approximately $0.9503$, rounded to four decimal places.

Check and common mistake avoided. The standard error of the average is $800/\sqrt{n}$, not $800/n$. This is a normal-approximation planning result, not a guarantee of the exact probability for 385 exponential claims.

Using the Results Responsibly

For a fixed claim count, the LLN gives a useful scale interpretation: the average claim amount settles toward $\mu$, and the aggregate divided by the number of claims settles toward the same value. The CLT gives a complementary risk interpretation: aggregate standard deviation grows like $\sqrt{n}$, while average-claim standard deviation shrinks like $1/\sqrt{n}$. Thus increasing the number of independent claims makes the average more stable even though the total amount remains variable.

In an exam calculation, first identify whether the event concerns $S_n$ or $\bar{X}_n$, and use the matching mean and standard deviation. State the claim distribution and its parameters, verify that the claims are independent under the model, and distinguish an exact moment calculation from a normal approximation. A finite variance is essential to the standard iid CLT stated here; if it is absent, these conclusions cannot be used as written.

Key takeaway: With iid claims of mean $\mu$ and variance $\sigma^2$, the aggregate has mean $n\mu$ and standard deviation $\sigma\sqrt{n}$; the average has mean $\mu$ and standard deviation $\sigma/\sqrt{n}$. The LLN describes stabilization, while the CLT approximates fluctuations around that stable level.

Check Your Understanding

Use the stated fixed-count iid models. For probability questions, use the CLT and report probabilities to four decimal places.

A hypothetical one-year block has 36 independent claims, each with mean $\$2{,}000$ and variance $\$4{,}000{,}000$. With no discounting or mortality decrement, find the aggregate mean and standard deviation. (Answer: mean $\$72{,}000$; standard deviation $\$12{,}000$.)
Claim amounts are independent and uniform on $[0,\$4{,}000]$, and there are 100 claims. Approximate the probability that the average claim exceeds $\$2{,}200$. (Answer: $0.0416$.)
Each of 144 independent claims has an exponential distribution with mean $\$600$ and variance $\$360{,}000$. Approximate the probability that aggregate claims lie between $\$82{,}800$ and $\$90{,}000$. (Answer: $0.3829$.)
For 64 independent exponential claims with mean $\$1{,}200$ and rate $1/1{,}200$ per dollar, approximate the probability that the aggregate exceeds $\$88{,}800$. (Answer: $0.1056$.)
Independent exponential claim amounts have mean $\$600$ and variance $\$360{,}000$. Using a central CLT design approximation with critical value $1.645$, find the smallest integer $n$ for which the average is within $\$60$ of its mean with approximate probability at least $0.90$. (Answer: $271$.)

What You'll Learn