Aggregating Reserve Prediction Error Across Origin Years

Why Total Reserve Error Is Not Usually a Sum of Variances

A reserve portfolio is often reported as one total, but its estimate is assembled from accident-year predictions. Those predictions can share estimated development factors. Consequently, their prediction errors need not be independent, even when the underlying claim processes for different accident years are assumed independent under the Mack framework. Ignoring the covariance can understate or, if covariance is negative, overstate the uncertainty in the total.

This tutorial focuses on aggregating accident-year reserve prediction errors. It does not change the selected chain-ladder estimates or the individual-year MSEPs; it shows how to combine those quantities consistently. The calculation is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Definition: For accident year $i$, let $\hat R_i$ be the estimated unpaid reserve and $R_i$ the eventual unpaid amount. The prediction error is $e_i=\hat R_i-R_i$. Its mean squared error of prediction is $\operatorname{MSEP}(\hat R_i)=E[e_i^2]$, and its prediction-error variance is $\operatorname{Var}(e_i)$. When the prediction error has mean zero, these are equal. For total reserve $\hat R_{\mathrm{tot}}=\sum_i\hat R_i$, the total prediction error is $e_{\mathrm{tot}}=\sum_i e_i$.

In Mack calculations, accident years can have distinct future claim process variation while still sharing uncertainty from fitted development factors. That shared estimation uncertainty creates cross-year covariance. The total MSEP therefore requires both the individual-year MSEPs and the covariances between pairs of years.

The Aggregation Identity

For $m$ accident years, expand the square of the total prediction error. This gives the exact aggregation identity:

\operatorname{MSEP}(\hat R_{\mathrm{tot}}) =E\!\left[\left(\sum_{i=1}^{m}e_i\right)^2\right] =\sum_{i=1}^{m}\operatorname{MSEP}(\hat R_i) +2\sum_{1\leq i<k\leq m} E[e_i e_k].

If each error has mean zero, $E[e_i e_k]=\operatorname{Cov}(e_i,e_k)$. The factor of 2 appears because the expansion contains both $e_i e_k$ and $e_k e_i$ for each distinct pair. The aggregate standard error is the square root of the aggregate MSEP, not the sum of the individual standard errors.

Key result: Under the usual centered-error convention, add all individual accident-year MSEPs and twice each distinct pairwise covariance. Then take the square root to obtain the total-reserve standard error. If prediction errors are not centered, use $E[e_i e_k]$ in the identity rather than substituting covariance without adjustment.

The following Mack setup explains why cross-year covariance arises. Write $C_{i,j}$ for cumulative claims for accident year $i$ at development age $j$. Under the Mack model, the conditional mean and variance of the next cumulative value are

E[C_{i,j+1}\mid C_{i,1},\ldots,C_{i,j}]=f_j C_{i,j}, \qquad \operatorname{Var}(C_{i,j+1}\mid C_{i,1},\ldots,C_{i,j}) =\sigma_j^2 C_{i,j}.

The model also assumes independent claim development across accident years, conditional on the development structure, and a common development factor $f_j$ at age $j$. In practice, $f_j$ is estimated from multiple accident years. Two reserve predictions that both use $\hat f_j$ are therefore affected by the same factor-estimation error. This is the source of the Mack cross-year covariance contribution used below. The formula is an estimate under the Mack framework, not a claim that all actual portfolio dependencies have been captured.

Shared-Factor Covariance

Suppose there are $n$ accident years and development ages. Let $k_i$ be the latest observed development age for accident year $i$, and let $\hat U_i$ be its chain-ladder ultimate estimate. The projected ultimate uses estimated factors from age $k_i$ through age $n-1$. Thus two accident years share only those future development ages included in both projections.

With the standard Mack variance estimate for a development factor, $\widehat{\operatorname{Var}}(\hat f_j)=\hat\sigma_j^2/\sum_{r=1}^{n-j}C_{r,j}$, a first-order factor-uncertainty covariance estimate is

\widehat{\operatorname{Cov}}(e_i,e_k) = \hat U_i\hat U_k \sum_{j=\max(k_i,k_k)}^{n-1} \frac{\hat\sigma_j^2} {\hat f_j^2\displaystyle\sum_{r=1}^{n-j}C_{r,j}}, \qquad i\ne k.

The summation starts at the later of the two latest observed ages: only factors still to be applied to both years contribute. If there are no common future development ages, the sum is zero under this shared-factor expression. Since the observed cumulative amount for each year is fixed when calculating its reserve, subtracting that amount to obtain a reserve does not change the covariance contribution.

Find the individual-year results.
Use the Mack calculation for each accident year to obtain its reserve estimate and MSEP, including the required process and parameter components.

Identify common future factors.
For every distinct pair, compare the future development ages used in its two projections.

Calculate each pairwise covariance.
Use the Mack covariance estimate for shared factors, or use the pairwise covariances supplied with the individual-year results.

Aggregate and take the square root.
Add the individual MSEPs and twice the sum of distinct pairwise covariances; the square root is the total standard error.

Worked Calculations

Worked Example: Aggregating Three Accident-Year Results

A hypothetical portfolio has three accident years. Individual Mack calculations give reserve MSEPs of $1.44$, $2.25$, and $4.00$ million dollars squared. The estimated pairwise covariances are $0.30$, $0.40$, and $0.50$ million dollars squared, for pairs $1,2$, $1,3$, and $2,3$, respectively. Calculate the total-reserve MSEP and standard error.

Method. Apply the aggregation identity, counting each distinct covariance twice.

Step 1: Sum the individual MSEPs.

1.44+2.25+4.00=7.69 \quad\text{million dollars squared}.

Step 2: Sum the distinct pairwise covariances.

0.30+0.40+0.50=1.20 \quad\text{million dollars squared}.

Step 3: Add twice that covariance sum.

\operatorname{MSEP}(\hat R_{\mathrm{tot}}) =7.69+2(1.20)=10.09 \quad\text{million dollars squared}.

Step 4: Take the square root. Rounded to four decimal places, the total standard error is $\sqrt{10.09}=3.1765$ million dollars. The covariance terms increase the MSEP above the $7.69$ million dollars squared that would result from incorrectly assuming independence.

Check. The variance is positive, and the standard error squared returns the aggregate MSEP to rounding. The common mistake avoided is adding only the three individual MSEPs, or counting each covariance once rather than twice.

Worked Example: Calculating Covariance From a Shared Mack Factor

A hypothetical cumulative claims triangle, measured in thousands of dollars, has observed values $C_{1,1}=100$, $C_{1,2}=200$, $C_{1,3}=240$, $C_{2,1}=100$, $C_{2,2}=200$, and $C_{3,1}=100$. The fitted Mack factors are estimated by volume-weighted averages. For development age 2, take $\hat\sigma_2^2=8$. Calculate the shared-factor covariance contribution between accident years 2 and 3.

Step 1: Calculate the fitted factors. At age 1, the observed transitions are $100$ to $200$ for each of the first two years. At age 2, the observed transition is $200$ to $240$ for the first year. Therefore,

\hat f_1=\frac{200+200}{100+100}=2, \qquad \hat f_2=\frac{240}{200}=1.2.

Step 2: Find the estimated variance of $\hat f_2$. There is one observed age-2 cumulative amount, $C_{1,2}=200$, so the Mack factor-variance estimate is

\widehat{\operatorname{Var}}(\hat f_2) =\frac{\hat\sigma_2^2}{\sum_r C_{r,2}} =\frac{8}{200}=0.04.

Step 3: Calculate the two projected ultimates. Accident year 2 is observed through age 2, so $\hat U_2=200(1.2)=240$. Accident year 3 is observed through age 1, so $\hat U_3=100(2)(1.2)=240$. Their only common future factor is the age-2 factor.

Step 4: Apply the shared-factor covariance expression. The relative factor-variance contribution is $0.04/(1.2)^2=0.04/1.44=0.027777\ldots$. Hence

\widehat{\operatorname{Cov}}(e_2,e_3) =240(240)\frac{0.04}{1.44} =1{,}600 \quad\text{(thousands of dollars)}^2.

This pair contributes $2(1{,}600)=3{,}200$ thousands of dollars squared to the total MSEP. The individual MSEPs must still be added to obtain the full portfolio MSEP. The common mistake avoided is including age 1 in the pairwise covariance sum: it is not a future factor shared by both projections.

Worked Example: Aggregating a Four-Year Covariance Table

For a hypothetical portfolio, the four individual reserve MSEPs are $100$, $144$, $196$, and $256$ thousands of dollars squared. The pairwise covariances, also in thousands of dollars squared, are $10$ for years 1 and 2, $8$ for years 1 and 3, $0$ for years 1 and 4, $12$ for years 2 and 3, $6$ for years 2 and 4, and $15$ for years 3 and 4. Find the total MSEP and standard error.

Step 1: Add the diagonal (individual MSEP) terms.

$$ 100+144+196+256=696. $$

Step 2: Add all six distinct covariance terms.

$$ 10+8+0+12+6+15=51. $$

Step 3: Double the covariance sum and add it to the diagonal sum.

\operatorname{MSEP}(\hat R_{\mathrm{tot}}) =696+2(51)=798 \quad\text{(thousands of dollars)}^2.

Step 4: Take the square root. Rounded to four decimal places, the total standard error is $\sqrt{798}=28.2489$ thousand dollars, or approximately $\$28{,}248.89$. The conversion is $28.2489\times \$1{,}000$, with rounding applied to the dollar amount at the final step.

Check. The zero covariance for years 1 and 4 contributes nothing, but it remains one of the six possible pairs. The common mistake avoided is omitting a pair from the covariance table or treating the covariance entries as already doubled.

Interpretation and Exam Traps

A covariance is not another accident-year MSEP. It describes how two prediction errors vary together. Positive covariance means that shared factor estimation tends to move both errors in the same direction, increasing total uncertainty. Negative covariance, when present in an applicable model, reduces the aggregate MSEP. The Mack shared-factor expression above is nonnegative when its inputs are nonnegative, but one should use the covariance estimates appropriate to the model rather than impose that sign on every possible reserving setting.

A frequent error is to calculate the total standard error as the sum of individual standard errors. For example, even under independence the correct result is the square root of the sum of individual variances, not the sum of their square roots. With dependence, the pairwise covariance terms are needed as well.

Exam check: Keep units consistent throughout. If MSEPs and covariances are in thousands of dollars squared, the square root is in thousands of dollars. Also check that the covariance list contains each unordered pair once before multiplying its sum by 2.

Aggregation is an essential step in reporting reserve uncertainty for a portfolio rather than for isolated accident years. The distinction between individual and aggregate MSEP, the shared-factor source of covariance, and the factor of 2 in the aggregation identity are all central to applying Mack results correctly.

Check Your Understanding

For each problem, aggregate the stated Mack prediction-error quantities and check the units before taking a square root.

Two accident years have MSEPs of $9$ and $16$ thousands of dollars squared and covariance $3$ thousands of dollars squared. Find the total MSEP and standard error. (Answer: $31$ thousands of dollars squared; $5.5678$ thousand dollars.)
Three years have MSEPs $25$, $36$, and $49$ millions of dollars squared. Their covariances for pairs $1,2$, $1,3$, and $2,3$ are $2$, $4$, and $3$ millions of dollars squared. Find the total MSEP. (Answer: $128$ millions of dollars squared.)
Two projected ultimates are $300$ and $250$ thousand dollars. Their only common future factor has $\hat f=1.25$, $\hat\sigma^2=5$, and $\sum_r C_{r}=500$, with amounts measured in thousands. Their individual MSEPs are $900$ and $625$ thousands of dollars squared. Calculate the factor covariance contribution and aggregate MSEP. (Answer: covariance $480$ thousands of dollars squared; aggregate MSEP $2{,}485$ thousands of dollars squared.)
Three accident years have MSEPs $4$, $9$, and $16$ thousands of dollars squared. The only nonzero pairwise covariance is $2$ between years 1 and 3. Find the aggregate standard error. (Answer: $\sqrt{33}=5.7446$ thousand dollars.)
A four-year portfolio has individual MSEPs $16$, $25$, $36$, and $49$ thousands of dollars squared. Its six pairwise covariances sum to $12$ thousands of dollars squared. Find the aggregate MSEP and standard error. (Answer: $150$ thousands of dollars squared; $\sqrt{150}=12.2474$ thousand dollars.)

What You'll Learn