Reconciling Component Estimates to Aggregate Losses

Why Reconciliation Matters

A frequency-severity estimate builds an aggregate loss estimate from two components: expected claims per exposure and expected loss per claim. That structure is useful for analysis, but multiplying two displayed component estimates does not always reproduce a separately calculated aggregate loss. The difference may be only rounding, or it may reveal that the calculations use different claims, loss measures, weights, or time periods.

Reconciliation is the process of making those bases explicit and quantifying the difference. It is not a rule that the components must be forced to match an aggregate benchmark. First determine whether the figures are intended to describe the same experience. If they are, the arithmetic should be explainable; if they are not, the discrepancy may be appropriate but should be documented.

This tutorial uses deterministic data reconciliation: no interest rate, mortality assumption, or probability distribution is required. The methods are aligned with the CAS Exam 5 (Basic Ratemaking and Estimated Claim Liabilities) syllabus.

Definition: Let $E$ be the number of exposure units, $N$ the number of claims on a specified count basis, and $A$ the aggregate loss on a specified loss basis. Frequency is claims per exposure, $f=N/E$, and severity is loss per claim, $s=A/N$, when $N>0$. The aggregate implied by frequency and severity is $Efs$. For estimated components, write $\hat f$ and $\hat s$, so the component estimate is $E\hat f\hat s$.

The phrase “specified basis” is essential. A claim count might include all reported claims, while a severity calculation might include only closed claims. Aggregate loss might mean paid loss, incurred loss, or an ultimate estimate. Those figures cannot be assumed interchangeable. A reconciliation begins by naming the exposure period, claim-count definition, loss measure, and valuation date for every quantity.

The Exact Identity and the Reconciliation Residual

When frequency and severity come from the same claims and the same aggregate loss, their product reproduces that aggregate exactly. The result follows directly from their definitions.

Efs =E\left(\frac{N}{E}\right)\left(\frac{A}{N}\right) =A

This identity requires $E>0$, $N>0$, and consistent definitions throughout. It is an arithmetic identity, not a statistical assumption. It applies to a data set when the same $N$ appears in frequency and severity and the same $A$ defines severity and the aggregate benchmark. It does not guarantee that independently selected or adjusted estimates $\hat f$ and $\hat s$ will multiply to an observed aggregate.

For a benchmark aggregate $A_{\text{bench}}$, define the dollar residual as the benchmark minus the component estimate. A positive residual means the component estimate is below the benchmark; a negative residual means it is above. The percentage residual uses the benchmark as its denominator when the benchmark is positive.

D=A_{\text{bench}}-E\hat f\hat s, \qquad D_{\%}=\frac{A_{\text{bench}}-E\hat f\hat s}{A_{\text{bench}}}

The percentage residual is undefined when the benchmark is zero. In that case, report the dollar difference and explain the zero-loss context. A residual is a diagnostic, not a conclusion: it does not by itself identify which input is wrong or whether any input is wrong.

A Practical Reconciliation Sequence

A reliable review checks arithmetic and basis before revising assumptions. Keep unrounded component values available where possible, then compare the final displayed result with the aggregate benchmark.

Fix the target.
State the exposure period, valuation date, claim-count definition, and loss measure for the aggregate benchmark.

Recompute the component product.
Multiply exposure by frequency and severity, retaining precision until the final reporting step.

Check definitions and weights.
Confirm that frequency, severity, and aggregate loss cover compatible claims, and that group values use appropriate aggregate weights.

Check period alignment.
Determine whether the components and benchmark refer to the same historical or projected period and have the same adjustments.

Quantify and explain what remains.
Report the dollar and, where meaningful, percentage residual. Attribute it only to causes supported by the calculation records.

This order prevents a common analytical error: adjusting a component to eliminate a residual before checking whether the compared quantities have the same basis. It also distinguishes a presentation difference caused by rounding from a substantive difference caused by data definitions or period selection.

Worked Example: Isolating a Rounding Difference

A hypothetical book has 12,000 exposure units and 1,000 claims. Aggregate losses on the stated claim basis are $4,320,000. A summary displays frequency as 0.083 claims per exposure and severity as $4,320 per claim. Reconcile the displayed component estimate to the aggregate. Assume the frequency display was rounded to three decimal places; the source count and loss total are unrounded. This is a deterministic calculation, with no interest, mortality, or probability distribution assumption.

Method. First calculate the exact data frequency and severity. Then compare their exact product with the product using the displayed frequency.

Exact frequency is $1{,}000/12{,}000=0.083333\ldots$. Exact severity is $4{,}320{,}000/1{,}000=\$4{,}320$. Therefore the exact component calculation is:

12{,}000\left(\frac{1{,}000}{12{,}000}\right)(4{,}320) =1{,}000(4{,}320) =\$4{,}320{,}000

Using the displayed frequency instead gives $12{,}000(0.083)=996$ implied claims. The displayed component estimate is $996(\$4{,}320)=\$4{,}302{,}720$. The benchmark-minus-estimate residual is $\$4{,}320{,}000-\$4{,}302{,}720=\$17{,}280$, or $17{,}280/4{,}320{,}000=0.004=0.400\%$ of the benchmark.

Check. The frequency rounding reduced the implied count by $1{,}000-996=4$ claims. Four claims at $4,320 each equal $17,280, exactly the residual. The discrepancy is fully explained by rounding; it is not evidence of a different loss basis. The common mistake is to treat 0.083 as exact and alter severity to force agreement.

Definition Mismatches and Weighting

A component calculation can be internally correct and still fail to match an aggregate calculated on another basis. For example, frequency based on reported claims multiplied by severity based only on paid, closed claims does not generally equal incurred losses that include estimates for open claims. Reconcile the claim populations and loss measures before concluding that a component estimate is inaccurate.

Group aggregation also requires care. For groups $g=1,\ldots,G$, let $E_g$, $N_g$, and $A_g$ be exposure, claim count, and aggregate loss. The all-group frequency is total claims divided by total exposures; it is an exposure-weighted average of group frequencies. The all-group severity is total loss divided by total claims; it is a claim-count-weighted average of group severities.

f_{\text{all}}=\frac{\sum_g N_g}{\sum_g E_g} =\frac{\sum_g E_g f_g}{\sum_g E_g}, \qquad s_{\text{all}}=\frac{\sum_g A_g}{\sum_g N_g} =\frac{\sum_g N_g s_g}{\sum_g N_g}

Consequently, multiplying simple, unweighted averages of group frequencies and group severities by total exposure generally does not reproduce the total loss. The product of the correctly weighted overall frequency and severity does reproduce the total, provided each group’s components use the same definitions and periods.

Worked Example: Correcting an Unweighted Group Calculation

A hypothetical insurer reviews two territories. Territory A has 1,000 exposures, 100 claims, and $100,000 in losses. Territory B has 9,000 exposures, 180 claims, and $360,000 in losses. An analyst takes the simple average of the two territory frequencies and the simple average of the two severities, then applies their product to all 10,000 exposures. Find the resulting estimate, compare it with the correct aggregate, and show the weighting correction. This is deterministic; no interest, mortality, or probability distribution assumption is used.

Method. Compute the group rates, reproduce the analyst’s unweighted calculation, then calculate overall rates from total claims, exposures, and losses.

For A, frequency is $100/1{,}000=0.10$, and severity is $\$100{,}000/100=\$1{,}000$. For B, frequency is $180/9{,}000=0.02$, and severity is $\$360{,}000/180=\$2{,}000$. The simple averages are $ (0.10+0.02)/2=0.06$ and $(\$1{,}000+\$2{,}000)/2=\$1{,}500$. Thus the analyst estimates $10{,}000(0.06)(\$1{,}500)=\$900{,}000$.

The observed totals are 280 claims and $460,000 loss. Overall frequency is $280/10{,}000=0.028$; overall severity is $\$460{,}000/280=\$1{,}642.857142\ldots$. Their product is:

10{,}000(0.028)(\$1{,}642.857142\ldots) =280(\$1{,}642.857142\ldots) =\$460{,}000

The unweighted estimate exceeds the aggregate by $\$900{,}000-\$460{,}000=\$440{,}000$, or $440{,}000/460{,}000=95.6522\%$ of the aggregate. The exposure-weighted frequency is $0.10(1{,}000/10{,}000)+0.02(9{,}000/10{,}000)=0.028$. The claim-count-weighted severity is $\$1{,}000(100/280)+\$2{,}000(180/280)=\$1{,}642.857142\ldots$. These weights recover the same overall rates.

Check. Applying the weighted overall rates produces $460,000, equal to $100{,}000+360{,}000$. The common mistake is to give a small territory with 1,000 exposures the same influence on overall frequency as a territory with 9,000 exposures, and then to average severity without regard to claim counts.

Worked Example: Reconciling Paid-Only Severity to Incurred Losses

A hypothetical portfolio has 2,400 exposures and 120 reported claims. Of these, 100 claims are closed with total paid losses of $450,000. The other 20 claims remain open; their estimated incurred loss, in addition to the paid amounts already included, is $150,000. A component calculation uses the 120 reported claims for frequency but calculates severity as paid losses divided by the 100 closed claims. Compare that calculation with total incurred loss on all 120 claims. Assume all amounts refer to the same accident period and valuation date. This deterministic example uses no interest, mortality, or probability distribution assumption.

Method. Identify the loss definition represented by each input, calculate the mixed-basis component estimate, and then calculate the consistent incurred estimate.

Reported-claim frequency is $120/2{,}400=0.05$ claims per exposure. Paid-only severity for the 100 closed claims is $\$450{,}000/100=\$4{,}500$. Multiplying this severity by all 120 reported claims gives an estimated aggregate of $120(\$4{,}500)=\$540{,}000$.

The incurred aggregate is paid loss plus the stated additional incurred amount for open claims: $\$450{,}000+\$150{,}000=\$600{,}000$. Consistent incurred severity across all 120 claims is $\$600{,}000/120=\$5{,}000$. Therefore the frequency-severity calculation on the incurred basis is $2{,}400(0.05)(\$5{,}000)=\$600{,}000$.

The mixed-basis estimate is below the incurred benchmark by $\$600{,}000-\$540{,}000=\$60{,}000$, or 10.000% of the benchmark. That difference is not a rounding effect: the severity omitted the open-claim incurred amount and was calculated over a different claim subset. The common mistake is to multiply a closed-claim paid severity by all reported claims and label the result an incurred aggregate.

Check. The incurred severity is $500 higher per claim than the mixed-basis severity; across 120 claims, $120(\$500)=\$60{,}000$, exactly the residual.

Period Mismatches and Interpretation

A benchmark and its components can use different periods even when their labels look similar. For example, a historical aggregate may be compared with frequency projected to a future period and severity held at a historical level. That comparison is not an arithmetic reconciliation of one period’s experience. The appropriate response is to label the periods and adjustments, then either put the quantities on a common basis or report why they differ.

When a projected estimate is intended to represent a new period, its components should be interpreted as rates for that target period and applied to the corresponding exposure. A historical aggregate is then a reference point, not necessarily a value the projected components should reproduce. Conversely, if the purpose is to reproduce historical experience, introducing a future-period adjustment into only one component creates a mismatch.

A useful diagnostic table records each quantity’s definition alongside its value. This makes hidden differences visible before any recalibration is considered.

Quantity	Questions to document	Typical reconciliation issue
Exposure	Which policy or exposure units, and which period?	Components applied to a different exposure base
Frequency	Which claims count, and at what development or reporting status?	Reported count compared with closed-claim severity
Severity	Paid, incurred, or ultimate; which claims and valuation date?	Paid-only amount compared with incurred aggregate
Aggregate benchmark	Which loss definition, period, and adjustments?	Historical benchmark compared with projected components
Displayed estimates	How many digits were retained in each input?	Rounded rate creates a small apparent residual

Common Mistakes and Key Takeaway

Rounding too early. A rounded frequency can change the implied claim count. Retain the unrounded rate for calculations and round the final reported result.
Mixing claim populations. A severity from closed claims cannot automatically be applied to a count of all reported claims.
Using the wrong group weights. Overall frequency is exposure-weighted; overall severity is claim-count-weighted, not generally a simple average of group rates.
Comparing different periods. A historical aggregate need not equal a projection whose components have been adjusted to another period.
Forcing agreement without diagnosis. Changing a component solely to eliminate a residual can conceal a data-definition or period mismatch.

Key takeaway: For a consistent data set, $E(N/E)(A/N)=A$. When estimated frequency and severity do not reproduce an aggregate, first test rounding, claim and loss definitions, aggregation weights, and period alignment. Interpret the residual only after those bases are clear.

Check Your Understanding

For each problem, identify the relevant basis before calculating. Amounts are hypothetical and deterministic.

A portfolio has 5,000 exposures, 175 claims, and aggregate loss of $1,050,000. Calculate frequency and severity, then verify the aggregate component estimate. (Answer: Frequency $=0.035$; severity $=\$6{,}000$; implied aggregate $=\$1{,}050{,}000$.)
A summary uses 20,000 exposures, a displayed frequency of 0.047, and severity of $3,200. The benchmark based on source data is $3,200,000. Calculate the displayed component estimate and the benchmark-minus-estimate residual. (Answer: Component estimate $=\$3{,}008{,}000$; residual $=\$192{,}000$, or 6.000% of the benchmark.)
Group A has 2,000 exposures, 80 claims, and $240,000 in losses. Group B has 8,000 exposures, 120 claims, and $480,000 in losses. Calculate the properly weighted overall frequency, severity, and aggregate. (Answer: Frequency $=0.020$; severity $=\$3{,}600$; aggregate $=\$720{,}000$.)
A historical period has 40,000 exposures and $2,000,000 in losses. For a target period, exposure is 42,000, frequency is assumed unchanged, and severity is projected to increase by 5%. Calculate the target-period component estimate and explain why it need not equal the historical aggregate. (Answer: Historical pure premium $=\$50$ per exposure; target pure premium $=\$52.50$; target estimate $=42{,}000(\$52.50)=\$2{,}205{,}000$, which differs because it uses target exposure and projected severity.)

What You'll Learn