Why Aggregate Loss Is a Random Sum
A portfolio’s total loss depends on two different sources of uncertainty. The number of claims can vary from period to period, and the amount of each claim can also vary. Aggregate loss combines these two sources: it is the sum of the severities of however many claims occur.
Keeping these pieces separate is essential. A frequency model describes the claim count; a severity model describes the amount of one claim. Neither model alone specifies the aggregate loss. A standard compound model joins them using a random count and a sequence of claim severities.
The units and time period must be stated for both components. For example, \(N\) might count claims per policy-year, while \(X\) measures dollars per claim. The aggregate \(S\) is then dollars of loss over a policy-year. A claim count measured per month cannot be combined unadjusted with a severity description intended for a different coverage period if the underlying risk changes over time.
The notation \(X_i\) labels the severity of the \(i\)th claim, not the loss for the \(i\)th policy. In the standard model, claims share a common severity distribution and their severities are mutually independent. In addition, claim count is independent of those severities. These assumptions are modeling choices, not consequences of the notation.
Separating Frequency and Severity
Frequency and severity answer different questions. A frequency model might be Poisson, binomial, or negative binomial, depending on the setting and assumptions. It assigns probabilities to possible claim counts. A severity model assigns probabilities or a density to possible amounts for one claim. The aggregate model uses both, but its definition does not require that a particular frequency family be chosen.
| Component | Random quantity | What it describes | Example of a parameter or specification |
|---|---|---|---|
| Frequency | \(N\) | Number of claims in the defined period and exposure | Poisson claim rate per policy-year |
| Severity | \(X_i\) | Dollar amount of one claim | A discrete distribution for claim amounts |
| Aggregate | \(S\) | Total dollar loss from all claims in the period | The sum of the realized severities |
If the number of claims happens to be \(n\), the realized total is the sum of the first \(n\) claim amounts. This gives a direct construction of the random sum. For each possible claim index \(j\), define \(I(N\geq j)\) to be 1 if at least \(j\) claims occur and 0 otherwise. Then
To see why, fix a realization \(N=n\). For \(j=1,\ldots,n\), the indicator is 1, so those terms contribute \(X_1+\cdots+X_n\). For every \(j>n\), the indicator is 0, so no further severity contributes. Thus the expression is exactly the total of the claims that occurred. This identity explains the random-sum construction without imposing a particular count or severity distribution.
Worked Example: Constructing an Aggregate From a Count Model
A hypothetical service plan has a claim frequency modeled as Poisson with rate \(0.6\) claims per policy-year. The modeled severity \(X\), in dollars per claim, is \(320\) with probability \(0.5\) and \(880\) with probability \(0.5\). Assume claims are independent of the count and their severities are independent and identically distributed. For a two-policy-year exposure, the observed count is 3, and the three severities are \(320\), \(880\), and \(320\). Specify the count model for this exposure and calculate the observed aggregate loss.
Method. Scale the frequency rate by exposure, then add the severities for the observed claims. The rate is \(0.6\) claims per policy-year and the exposure is 2 policy-years, so the count model is $$ N\sim\operatorname{Poisson}(0.6\times2)=\operatorname{Poisson}(1.2). $$ The severity distribution remains a per-claim distribution: \(P(X=320)=0.5\) and \(P(X=880)=0.5\). The frequency parameter \(1.2\) is a claim-count parameter, not a dollar amount.
For the observed count of 3, substitute the three observed severities into the random sum: $$ S=320+880+320=1{,}520. $$ The aggregate loss is therefore $1,520.00. The calculation is consistent with the stated severity support: each of the three observed claim amounts is either $320 or $880.
Check and common mistake avoided. There are three claims, so exactly three severity amounts are included. The two-policy-year exposure changes the count model’s parameter from \(0.6\) to \(1.2\); it does not multiply each claim amount by 2. Multiplying the observed total by the exposure again would incorrectly give $3,040.
When One Common Severity Distribution Is Not Appropriate
The identically distributed assumption deserves particular attention in a mixed portfolio. Claim severities might differ systematically by product, territory, or coverage type. If the claim class is known and classes have different severity distributions, the claim amounts across all classes are generally not identically distributed. A single \(X\) then does not accurately describe every claim without further modeling.
One approach is to keep separate counts and severity sequences by class, and add the class-specific aggregate losses. Each class can have its own frequency and severity assumptions. This is still a sum of losses, but it is not automatically the single-frequency, single-iid-severity model in the definition above.
Worked Example: Checking the Identical-Severity Assumption Across Classes
A hypothetical insurer tracks two types of equipment. In one period, class A has 2 claims with severities $500 and $900. Its claim amounts are modeled as independent copies of \(X_A\), where \(P(X_A=500)=0.75\) and \(P(X_A=900)=0.25\). Class B has 1 claim with severity $2,400; its claim amounts are modeled as independent copies of \(X_B\), where \(P(X_B=1{,}800)=0.5\) and \(P(X_B=2{,}400)=0.5\). The counts are modeled as \(N_A\sim\operatorname{Poisson}(0.8)\) and \(N_B\sim\operatorname{Poisson}(0.3)\), independent of the applicable severity sequences. Find the observed combined loss and assess whether all three claims use one common iid severity distribution.
Method. First add the observed amounts by class: $$ S_A=500+900=1{,}400,\qquad S_B=2{,}400. $$ The combined observed aggregate is $$ S=S_A+S_B=1{,}400+2{,}400=3{,}800. $$ Thus the combined loss is $3,800.00.
Within each class, the stated model uses independent, identically distributed severities: class A uses \(X_A\), and class B uses \(X_B\). Across the combined set of claims, however, a common identical severity distribution has not been specified. In fact, the supports differ: class A severities are $500 or $900, whereas class B severities are $1,800 or $2,400. The standard model with one shared \(X\) should therefore not be asserted for these claims without an additional justified mixture model.
Check and common mistake avoided. The count models describe how many claims occur in each class; they do not make the class severity distributions identical. Treating all three claims as iid from \(X_A\), for example, would make the observed $2,400 class B claim impossible under that severity model. The class-specific representation preserves the stated assumptions.
Exposure Units and Independence
A frequency parameter is meaningful only together with its exposure basis. Suppose a rate is specified per unit-year. For several units observed over part of a year, the exposure must be expressed in unit-years before the count parameter is set. This adjustment belongs to the frequency component. The severity distribution continues to describe one claim in dollars, unless the claim amounts themselves change with the exposure period or risk conditions.
Independence is a separate issue from the choice of distributions. The standard compound model assumes that the severity sequence is independent of \(N\). For example, if unusually severe claims are more likely to be reported, or if claim count and claim size share an unmodeled risk factor, that assumption may fail. The random-sum definition can still describe a total, but formulas and interpretations that rely on the standard independence assumptions may no longer apply.
Worked Example: Matching Frequency Exposure to the Claim Period
A hypothetical device portfolio has a Poisson claim rate of \(0.4\) claims per device-year. Five devices are covered for half a year each. The claim severity \(X\), in dollars, is $1,000 with probability \(0.8\) and $3,000 with probability \(0.2\), independently and identically across claims and independently of the count. For the observed period, two claims occur, with severities $1,000 and $3,000. Determine the count model for the period and its observed aggregate.
Method. First calculate exposure in device-years: $$ 5\times 0.5=2.5\text{ device-years}. $$ Multiply the rate per device-year by this exposure to obtain the Poisson parameter: $$ \lambda=0.4\times2.5=1.0. $$ Thus the frequency model for the period is \(N\sim\operatorname{Poisson}(1.0)\). The severity model remains per claim, with \(P(X=1{,}000)=0.8\) and \(P(X=3{,}000)=0.2\).
For the two observed claims, the realized aggregate is $$ S=1{,}000+3{,}000=4{,}000. $$ The observed total is $4,000.00, and both claim amounts are in the stated severity support.
Check and common mistake avoided. The exposure is 2.5 device-years, not 5 device-years, because each device is covered for only half a year. Using \(0.4\times5=2.0\) would overstate the count-model parameter. The severity amounts are summed once each; they are not multiplied by the number of covered devices.
Common Modeling Traps
- Confusing a claim count with a claim amount. \(N\) is an integer count; \(X_i\) is a per-claim severity; \(S\) is a monetary total. A frequency parameter does not have dollar units.
- Using the wrong exposure period. A rate per policy-year must be paired with policy-year exposure, not an unadjusted number of policies or months.
- Assuming identical severities across unlike classes. Separate risk classes may require separate severity distributions. A shared \(X\) needs justification.
- Assuming independence automatically. Specifying a frequency distribution and a severity distribution does not, by itself, establish independence between count and severity.
- Summing more or fewer severities than claims. For a realized count \(N=n\), the aggregate contains exactly \(n\) claim amounts.
Check Your Understanding
For each problem, keep the count model, severity model, and aggregate-loss calculation distinct.
- A hypothetical equipment plan has 2 claims in a period with severities $460 and $1,140. What is the observed aggregate loss? (Answer: $1,600.00)
- A frequency rate is 0.25 claims per machine-year. Three machines are each covered for 4 months. What is the total exposure in machine-years and the Poisson parameter for the period? (Answer: 1 machine-year; Poisson parameter 0.25)
- In one risk class, claim severities are iid as \(X_A\), taking $700 or $1,300. In another class, they are iid as \(X_B\), taking $2,000 or $3,500. Does this specification establish one common iid severity distribution across both classes? (Answer: No; it specifies different class-level severity distributions.)
- A policy-year count is modeled as Poisson with rate 0.7 per policy-year. A portfolio has 4 policies, each observed for 3 months. Find the Poisson parameter for the combined period. (Answer: \(0.7\times4\times(3/12)=0.7\))
- A hypothetical portfolio has 3 observed claims with severities $250, $250, and $1,600. Its count model is Poisson with parameter 0.9. What is the observed aggregate, and does the count parameter change the sum of these observed severities? (Answer: $2,100.00; no, the count parameter does not change the observed sum.)