Covariance Of Ultimate Estimates Across Accident Years

Why Accident-Year Prediction Errors Can Be Related

In chain-ladder reserving, an accident-year ultimate estimate is a product of its observed cumulative claims and estimated development factors. Two accident years can therefore depend on the same estimated factors. Even if their future claim processes are independent, uncertainty in a factor used by both estimates can move both predictions in the same direction. Their prediction errors may then have positive covariance.

This distinction matters when describing uncertainty for more than one accident year. A covariance of zero is justified by independent process errors alone only if shared parameter uncertainty is absent as well. This tutorial uses the Mack chain-ladder framework and a first-order sensitivity calculation to quantify the shared-factor contribution. The method is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Definition: Let \(C_{i,k}\) denote cumulative claims for accident year \(i\) through development age \(k\), and let \(\hat f_k\) be the estimated development factor from age \(k\) to \(k+1\). If accident year \(i\) is observed through age \(j_i\), its chain-ladder ultimate estimate is \(\hat U_i=C_{i,j_i}\prod_{k=j_i}^{n-1}\hat f_k\). The prediction error is \(e_i=U_i-\hat U_i\), where \(U_i\) is the future ultimate amount. Covariance between \(e_i\) and \(e_m\) describes how their prediction errors vary together.

Under Mack’s model, conditional on the cumulative amount at age \(k\), the next cumulative amount has mean \(f_k C_{i,k}\) and variance \(\sigma_k^2 C_{i,k}\). The model assumes the relevant claim development processes for different accident years are independent, conditional on the model structure. A standard estimated variance for the link-ratio factor is \(\widehat{\operatorname{Var}}(\hat f_k)=\hat\sigma_k^2/S_k\), where \(S_k\) is the sum of the eligible cumulative amounts at age \(k\). Here \(\hat\sigma_k^2\) is the estimated Mack process variance parameter for development age \(k\).

The variance of a factor estimate represents parameter uncertainty, not the future process variance for one accident year. When the same factor estimate appears in two ultimate estimates, this uncertainty is a common source of variation. The examples below state \(S_k\) and \(\hat\sigma_k^2\) directly, so no particular estimator for the variance parameter needs to be assumed.

Deriving the Shared-Factor Contribution

For a fixed accident year, changing one factor changes the ultimate estimate in proportion to the product of the other factors. This gives a direct way to measure how strongly each ultimate estimate responds to uncertainty in a shared factor.

\frac{\partial \hat U_i}{\partial \hat f_k}=\frac{\hat U_i}{\hat f_k}

To justify the covariance formula, approximate each ultimate estimate near the selected development factors by its first-order expansion. The part of the change in \(\hat U_i\) due to factor \(k\) is approximately \((\hat U_i/\hat f_k)(\hat f_k-\operatorname{E}[\hat f_k])\). For two accident years, the covariance contribution from that factor is therefore the product of their sensitivities times the factor variance. If development-factor estimates are treated as uncorrelated across ages, summing over common factors gives the following plug-in approximation.

\widehat{\operatorname{Cov}}(e_i,e_m) \approx \hat U_i\hat U_m \sum_{k\in K_i\cap K_m} \frac{\hat\sigma_k^2}{\hat f_k^2 S_k}

Here \(K_i\) is the set of future development ages used in \(\hat U_i\), and \(K_i\cap K_m\) contains the ages whose estimated factors are used in both ultimates. The same shared-parameter contribution applies whether the error is written as actual ultimate minus estimate or estimate minus actual ultimate, because changing both error signs leaves covariance unchanged. Under independent future process errors across accident years, their process-error covariance is zero; the displayed expression captures the additional first-order covariance from shared factor estimation.

The diagonal sum is an assumption or approximation, not a universal identity. If development-factor estimates at different ages have estimated nonzero covariances, the first-order calculation must also include cross terms:

\sum_{k\in K_i}\sum_{\ell\in K_m} \frac{\hat U_i}{\hat f_k} \frac{\hat U_m}{\hat f_\ell} \widehat{\operatorname{Cov}}(\hat f_k,\hat f_\ell)

In particular, use the appropriate covariance structure if it is supplied. Do not add a factor that is used by only one accident year to the diagonal shared-factor sum.

Worked Calculations

Worked Example: One Shared Development Factor

Two hypothetical accident years have estimated ultimate claims of 1,250 and 2,500 thousand dollars. Both estimates use a single shared factor with \(\hat f_1=1.25\), \(\hat\sigma_1^2=0.04\), and \(S_1=4{,}000\), with amounts measured in thousands of dollars. Future process errors are independent, and treat this factor as the only parameter source of covariance. Calculate the covariance contribution.

Method and model. Use the Mack factor-variance estimate and the first-order shared-factor formula. Interest and mortality assumptions do not apply to this claims development calculation; the stated assumptions are the Mack model and the supplied factor variance inputs.

Step 1: Estimate factor variance. The estimated variance of the shared factor is \(\hat\sigma_1^2/S_1=0.04/4{,}000=0.00001\).

Step 2: Calculate sensitivities. The first ultimate’s sensitivity is \(1{,}250/1.25=1{,}000\). The second is \(2{,}500/1.25=2{,}000\).

Step 3: Multiply sensitivities by factor variance. The covariance contribution is \(1{,}000(2{,}000)(0.00001)=20\) squared thousands of dollars. Equivalently, \(1{,}250(2{,}500)(0.04)/(1.25^2(4{,}000))=125{,}000/6{,}250=20\).

Check. The direct factor-variance calculation and the sensitivity calculation agree. The covariance is positive: an upward movement in the shared factor raises both ultimate estimates. The result is \(20.00\) squared thousands of dollars, rounded to two decimals. The common mistake avoided here is losing a factor of 1,000 when multiplying the sensitivities by the factor variance.

Worked Example: Two Shared Factors

Two hypothetical accident-year ultimate estimates are 1,500 and 2,000 thousand dollars. Both use two estimated development factors. For the first factor, \(\hat f_1=1.25\), \(\hat\sigma_1^2=0.05\), and \(S_1=5{,}000\). For the second, \(\hat f_2=1.10\), \(\hat\sigma_2^2=0.0225\), and \(S_2=9{,}000\). Assume the two factor estimates are uncorrelated, and assume independent future process errors across accident years. Find the approximate covariance from shared factor estimation.

Method and model. Apply the diagonal shared-factor formula to both common factors. All monetary values are in thousands of dollars; no interest or mortality model is involved.

Step 1: Contribution from factor 1. The contribution is \(1{,}500(2{,}000)(0.05)/(1.25^2(5{,}000))\). The numerator is \(150{,}000\), and the denominator is \(1.5625(5{,}000)=7{,}812.5\). Thus the contribution is \(150{,}000/7{,}812.5=19.2\) squared thousands of dollars.

Step 2: Contribution from factor 2. The numerator is \(1{,}500(2{,}000)(0.0225)=67{,}500\). The denominator is \(1.10^2(9{,}000)=1.21(9{,}000)=10{,}890\). The contribution is \(67{,}500/10{,}890\approx6.19835\) squared thousands of dollars.

Step 3: Add the contributions. The total is \(19.2+6.19835=25.39835\) squared thousands of dollars, or \(25.40\) squared thousands of dollars to two decimals.

Check. Both factor contributions are positive, and the sum exceeds either individual contribution. The common mistake avoided is rounding each contribution too early or using \(\hat f_k\) rather than \(\hat f_k^2\) in the denominator.

Worked Example: Include Only Factors Shared by Both Accident Years

Two hypothetical accident-year estimates are 900 and 1,200 thousand dollars. They share factor 2, with \(\hat f_2=1.20\), \(\hat\sigma_2^2=0.09\), and \(S_2=6{,}000\). The first estimate also uses an estimated factor at age 1, but the second estimate does not. Assume factor estimates at different ages are uncorrelated and future process errors are independent. Calculate the shared-parameter covariance contribution.

Method and model. Use only the common factor in the diagonal covariance formula. The age-1 factor is not common, and is assumed uncorrelated with factor 2, so it contributes no covariance here. The claims amounts are in thousands of dollars; interest and mortality assumptions are not relevant.

Step 1: Find the factor-2 variance estimate. \(\hat\sigma_2^2/S_2=0.09/6{,}000=0.000015\).

Step 2: Find each sensitivity. The sensitivities are \(900/1.20=750\) and \(1{,}200/1.20=1{,}000\).

Step 3: Calculate covariance. \(750(1{,}000)(0.000015)=11.25\) squared thousands of dollars. The equivalent direct calculation is \(900(1{,}200)(0.09)/(1.20^2(6{,}000))=97{,}200/8{,}640=11.25\).

Check. The direct calculation matches the sensitivity calculation, and the noncommon age-1 factor has not been included. The covariance contribution is \(11.25\) squared thousands of dollars, or \(11.25\) to two decimals.

Interpreting the Result and Avoiding the Main Trap

The shared-factor term is parameter risk: if the common estimated factor is higher than its underlying value, both accident-year estimates tend to be too high; if it is lower, both tend to be too low. Independent process errors do not cancel this dependence. Thus “the future claims processes are independent” does not imply “the total prediction errors are independent.”

A common incorrect calculation sets covariance to zero because the future process errors are independent. For the first worked example, that argument would give \(\widehat{\operatorname{Cov}}(e_1,e_2)=0\). It omits the shared factor-estimation term. The correction is \(1{,}000(2{,}000)(0.00001)=20\) squared thousands of dollars. Zero would be appropriate for the process-error covariance under the stated independence assumption, but not for the total covariance that includes uncertainty in the common estimated factor.

A second trap is to use the factor variance \(\hat\sigma_k^2\) directly in place of the variance of its estimate. In Mack’s calculation, the supplied factor-variance parameter is divided by \(S_k\): \(\widehat{\operatorname{Var}}(\hat f_k)=\hat\sigma_k^2/S_k\). Omitting \(S_k\) can greatly overstate the covariance. Another is mixing dollars with thousands of dollars. If ultimates and \(S_k\) are expressed in thousands, the covariance is in squared thousands; convert units consistently before reporting it in dollars squared.

Key takeaway: Under the first-order, uncorrelated-factor approximation, multiply the two ultimate estimates, then sum \(\hat\sigma_k^2/(\hat f_k^2S_k)\) over factors used by both accident years. Keep process covariance and shared parameter covariance conceptually separate.

Check Your Understanding

For each question, assume the stated factor estimates are uncorrelated across development ages and that future process errors are independent across accident years. Monetary amounts are in thousands of dollars. Report covariance in squared thousands of dollars, to two decimals.

Two ultimate estimates are 800 and 1,600. They share one factor with \(\hat f=1.10\), \(\hat\sigma^2=0.025\), and \(S=2{,}500\). Find their shared-factor covariance contribution. (Answer: \(10.58\) squared thousands of dollars.)
Two ultimate estimates are 1,000 and 1,500 and share two factors. For factor 1, \(\hat f_1=1.20\), \(\hat\sigma_1^2=0.036\), and \(S_1=5{,}000\). For factor 2, \(\hat f_2=1.25\), \(\hat\sigma_2^2=0.05\), and \(S_2=4{,}000\). Find the total contribution. (Answer: \(19.50\) squared thousands of dollars.)
Two estimates are 600 and 900. They share factor 3, for which \(\hat f_3=1.50\), \(\hat\sigma_3^2=0.09\), and \(S_3=3{,}000\). The first estimate also uses a nonshared factor, uncorrelated with factor 3. Find the shared-factor contribution. (Answer: \(7.20\) squared thousands of dollars.)
Two estimates are 2,000 and 3,000. They share factor 1, with \(\hat f_1=1.25\), \(\hat\sigma_1^2=0.04\), and \(S_1=5{,}000\). The second estimate also uses a factor that the first does not use, and this nonshared factor is uncorrelated with factor 1. Find their covariance contribution. (Answer: \(30.72\) squared thousands of dollars.)

What You'll Learn