Alternative Estimators Of Development-Period Variance

Why Estimate Development-Period Variance in More Than One Way?

In the Mack chain-ladder model, the variance of the next cumulative amount depends on both the development period and the current cumulative amount. A separate variance parameter for each period is flexible, but it may be very unstable when only a few origin years have observations at a late development age. Pooling or smoothing can reduce that instability, at the cost of adding assumptions about how variance behaves across development periods.

The choice is not merely computational. The selected variance estimates feed into estimates of reserve prediction error, so a noisy late-period estimate can materially affect reported uncertainty. The aim here is to distinguish what the standard period-by-period estimator assumes from what alternatives add.

Definition: Let $C_{i,j}$ be the cumulative claim amount for origin year $i$ at development age $j$, and let $f_j$ be the development factor from age $j$ to age $j+1$. Under the Mack assumptions, $$ E[C_{i,j+1}\mid C_{i,j}] = f_j C_{i,j}, \qquad \operatorname{Var}(C_{i,j+1}\mid C_{i,j}) = \sigma_j^2 C_{i,j}. $$ The parameter $\sigma_j^2$ is specific to development period $j$. Despite the superscript $2$, its units are the same monetary units as $C_{i,j}$, because it is multiplied by an amount to produce a variance.

For a given development period, use only origin years with both $C_{i,j}$ and $C_{i,j+1}$ observed. Denote their number by $n_j$. The usual chain-ladder estimate of the factor is the volume-weighted ratio $ \hat f_j = \sum_i C_{i,j+1}/\sum_i C_{i,j} $. The standard Mack variance estimate is calculated from the residuals around this fitted factor.

Definition: For $n_j\geq 2$ usable origin years with positive $C_{i,j}$, the standard period-specific Mack estimator is $$ \hat{\sigma}_j^2 = \frac{1}{n_j-1} \sum_i \frac{(C_{i,j+1}-\hat f_j C_{i,j})^2}{C_{i,j}} = \frac{1}{n_j-1} \sum_i C_{i,j} \left(\frac{C_{i,j+1}}{C_{i,j}}-\hat f_j\right)^2. $$ The sums are over the origin years observed at both ages. If $n_j=1$, this estimator cannot be calculated: there are no residual degrees of freedom.

Why the Residual Estimator Has This Form

The conditional variance assumption says that, at a fixed $C_{i,j}$, the residual $C_{i,j+1}-f_j C_{i,j}$ has variance $\sigma_j^2 C_{i,j}$. Dividing its square by $C_{i,j}$ therefore puts residuals from different-sized origin years on a common variance scale. Equivalently, the squared development-ratio residual is weighted by its starting cumulative amount.

The fitted factor is estimated from the same observations, so one degree of freedom is used in fitting it. The following justification makes the connection to weighted least squares explicit.

Fit the factor.
Minimizing $\sum_i (C_{i,j+1}-f C_{i,j})^2/C_{i,j}$ over $f$ gives $\hat f_j=\sum_i C_{i,j+1}/\sum_i C_{i,j}$.

Use the conditional variance scale.
Each squared residual is divided by its conditional variance multiplier $C_{i,j}$, giving $(C_{i,j+1}-\hat f_j C_{i,j})^2/C_{i,j}$.

Account for fitting.
With $n_j$ observations and one fitted factor, the residual degrees of freedom are $n_j-1$. Dividing the weighted residual sum by this number gives the standard Mack estimate.

This justification relies on the Mack mean and variance structure and on the usual conditional independence across origin years. It does not make a sparse estimate precise: with only two observations, the estimate has one residual degree of freedom and can be highly sensitive to either observation. The estimator is also not suitable when the required cumulative amounts are zero or negative, since the ratio and variance weighting are then not valid in this form.

Worked Example: A Period-Specific Mack Variance Estimate

A hypothetical insurer records amounts in thousands of dollars. At one development transition, two origin years have starting cumulative amounts of 100 and 300, with next-age amounts of 120 and 420, respectively. Calculate the development factor and the standard Mack variance estimate. State the units.

Method. Estimate the factor by the volume-weighted ratio, then calculate the weighted residual sum and divide by $n_j-1$.

The estimated factor is $$ \hat f_j=\frac{120+420}{100+300}=\frac{540}{400}=1.35. $$ The two development ratios are $120/100=1.20$ and $420/300=1.40$. Their weighted squared residual contributions are $$ 100(1.20-1.35)^2=100(0.0225)=2.25, \qquad 300(1.40-1.35)^2=300(0.0025)=0.75. $$ Thus the residual sum is $2.25+0.75=3.00$. There are $n_j=2$ observations, so the residual degrees of freedom are $2-1=1$, and $$ \hat{\sigma}_j^2=\frac{3.00}{1}=3.00. $$ The variance parameter is $3.00$ thousand dollars, not squared thousands of dollars. Multiplying it by a cumulative amount measured in thousands of dollars gives a conditional variance measured in squared thousands of dollars.

Check. The weighted factor reproduces the total next-age amount: $1.35(100+300)=540=120+420$. The residuals have weighted sum $100(-0.15)+300(0.05)=0$, as required by the fitted factor.

Common mistake avoided: Dividing the residual sum by two would ignore that one factor was fitted. The standard denominator is $n_j-1=1$, not $n_j=2$.

Pooling Variance Information Across Periods

A pooled estimator combines residual information from several development periods. The simplest version assumes that the selected periods share a common Mack variance parameter, $\sigma^2$. Each period may still have its own development factor; pooling variance does not mean forcing the factors to be equal.

Definition: For a selected set of development periods $J$, a common-variance pooled estimator is $$ \hat{\sigma}_{\mathrm{pool}}^2 = \frac{\displaystyle\sum_{j\in J}\sum_i (C_{i,j+1}-\hat f_j C_{i,j})^2/C_{i,j}} {\displaystyle\sum_{j\in J}(n_j-1)}. $$ It pools weighted residual sums and their residual degrees of freedom. This is appropriate only if the selected periods can reasonably be treated as having a common variance parameter under the Mack model.

Pooling is often helpful when separate late-development estimates have very few degrees of freedom. But it trades period-specific flexibility for a shared-variance assumption. If variance genuinely changes with development age, pooling unlike periods can conceal that pattern and bias the estimates for individual ages.

Worked Example: Pooling Two Sparse Development Periods

For the first of two development periods, suppose the weighted residual sum is 3.00 with one degree of freedom. For a second period, two origin years have starting amounts of 200 and 200, measured in thousands of dollars, and development ratios 1.10 and 1.50. Calculate each period’s estimate and the pooled estimate if the two periods are assumed to share a variance parameter.

Method. Keep each period’s factor fit separate, calculate its residual sum, and pool residual sums over total degrees of freedom.

For the second period, the fitted factor is $$ \hat f=\frac{200(1.10)+200(1.50)}{200+200} =\frac{220+300}{400}=1.30. $$ The weighted residual contributions are $$ 200(1.10-1.30)^2=200(0.04)=8.00, \qquad 200(1.50-1.30)^2=200(0.04)=8.00. $$ The residual sum is $16.00$, and its degrees of freedom are $2-1=1$. Therefore the separate estimates are $3.00/1=3.00$ for the first period and $16.00/1=16.00$ for the second.

Under the common-variance assumption, the pooled estimate is $$ \hat{\sigma}_{\mathrm{pool}}^2 =\frac{3.00+16.00}{1+1} =\frac{19.00}{2}=9.50. $$ It is $9.50$ thousand dollars. Pooling moves the first estimate up from 3.00 and the second down from 16.00, reflecting the shared-variance restriction.

Check. The pooled estimate lies between the two separate estimates, and multiplying it by the pooled two degrees of freedom recovers the total residual sum: $9.50(2)=19.00$.

Common mistake avoided: A simple average of period estimates is only appropriate here because both periods have one degree of freedom. In general, pool residual sums and degrees of freedom, rather than averaging estimates equally.

Smoothing Instead of Assuming One Common Variance

Smoothing is a middle ground between estimating every period independently and imposing one common variance across all periods. A smoothing rule allows the variance to change with development age, but assumes the estimates should vary in a structured or gradual way. There is no single universal smoothed estimator: the chosen method and its assumptions should be stated.

One transparent local method is to average nearby period-specific estimates with prespecified weights. For example, a symmetric weighted average of the previous, current, and next estimates can estimate the current variance if the true variance follows a locally linear pattern across equally spaced ages. The symmetric weights cancel a linear slope. If the underlying pattern is curved, however, the smoothed estimate may be biased for the current period.

Worked Example: A Local Linear Smoother for a Sparse Period

Suppose three consecutive development periods have weighted residual sums of 3.00, 18.00, and 45.00, with respective residual degrees of freedom 1, 2, and 3. Their separate variance estimates are noisy. Estimate the middle-period variance using weights $1/4, 1/2, 1/4$ on the three period estimates. Assume equally spaced ages and a locally linear true variance pattern.

Method. First calculate the separate period estimates, then apply the stated symmetric smoother.

The separate estimates are $$ \hat{\sigma}_1^2=\frac{3.00}{1}=3.00,\qquad \hat{\sigma}_2^2=\frac{18.00}{2}=9.00,\qquad \hat{\sigma}_3^2=\frac{45.00}{3}=15.00. $$ The smoothed estimate for the middle period is $$ \tilde{\sigma}_2^2 =\frac14(3.00)+\frac12(9.00)+\frac14(15.00) =0.75+4.50+3.75=9.00. $$ The estimate is $9.00$ thousand dollars.

Why the linear-pattern assumption matters. If the three true variances are $a-b$, $a$, and $a+b$, the expected weighted estimate is $\frac14(a-b)+\frac12a+\frac14(a+b)=a$. The two slope terms cancel, so the smoother targets the middle variance under a linear pattern. A sharp change or pronounced curvature would not generally cancel.

Check. The symmetric weighted average of 3.00 and 15.00 is 9.00, equal to the middle estimate in this example. The smoothing rule does not change the result here, but in uneven data it can pull a noisy estimate toward its neighbors.

Common mistake avoided: Smoothing does not remove the need to state a model for change across ages. A weighted average is not automatically an unbiased estimate of the current period’s variance under every possible variance pattern.

Units, Sparse Data, and Consequences for Reserve Error

Because $\operatorname{Var}(C_{i,j+1}\mid C_{i,j})=\sigma_j^2 C_{i,j}$, the parameter’s unit follows by division: variance has squared monetary units and $C_{i,j}$ has monetary units, so $\sigma_j^2$ has monetary units. For example, if amounts are in thousands of dollars, the parameter is reported in thousands of dollars. This is also why the numerical parameter changes linearly, not quadratically, when the amount unit changes.

To check the scaling, let dollar amounts be converted to thousands by $C'=C/1000$. Then $$ \operatorname{Var}(C'_{i,j+1}\mid C'_{i,j}) =\frac{\sigma_j^2 C_{i,j}}{1000^2} =\frac{\sigma_j^2}{1000}C'_{i,j}. $$ Thus the parameter in thousands-of-dollars units is $\sigma_j^2/1000$. For example, a parameter of $2{,}500{,}000$ dollars becomes $2{,}500$ thousand dollars. The corresponding variance of an amount in thousands is, as expected, expressed in squared thousands.

An estimated conditional process variance for a future cell uses the form $\hat{\sigma}_j^2 C_{i,j}$. Consequently, replacing a separate estimate with a pooled or smoothed one changes the estimated process uncertainty for that period, and may affect reserve prediction error. Pooling or smoothing should therefore be motivated by the data and a defensible variance pattern, not chosen merely because it produces a more convenient number.

Choosing and Reporting an Estimator

A sound analysis reports the estimator, the observations or periods used, and the assumptions that justify combining information. Separate estimates are the direct choice when each period has adequate data and period-specific variation matters. Pooling is more stable when the periods plausibly share one variance. Smoothing is useful when adjacent periods are expected to have related but not identical variances and a trend model is plausible.

Check the information available. State $n_j$ and the residual degrees of freedom. A period with one usable origin year has no separate residual estimate.
Do not pool development factors by accident. Variance pooling can retain a distinct $\hat f_j$ at each age.
Examine whether periods are comparable. A common variance assumption is questionable if development behavior changes sharply across ages.
Keep units consistent. Calculate factors from consistent amounts and label $\hat{\sigma}_j^2$ in the same monetary unit as cumulative claims.
Interpret precision cautiously. A pooled or smoothed estimate may be less variable, but its validity depends on the additional assumption being reasonable.

Key takeaway: The standard Mack estimator uses a period’s weighted squared residuals divided by $n_j-1$. Pooling combines residual sums under a common-variance assumption; smoothing allows structured variation across ages. Neither alternative creates information without assumptions, and the Mack variance parameter is measured in the same monetary unit as cumulative claims.

Check Your Understanding

Work each problem by identifying the estimator and its assumptions before calculating.

A development period has starting amounts of 50 and 150 thousand dollars, with development ratios 1.20 and 1.60. Find the weighted factor and standard Mack variance estimate. (Answer: $\hat f=1.5000$, $\hat{\sigma}^2=6.0000$ thousand dollars.)
Two periods have weighted residual sums 8 and 10, with one and two residual degrees of freedom, respectively. If a common variance is assumed, find the pooled estimate. (Answer: $18/3=6.0000$ in the monetary unit used for cumulative claims.)
A Mack variance parameter is estimated as 720,000 dollars. Express it when cumulative amounts are recorded in thousands of dollars. (Answer: 720 thousand dollars.)
Three equally spaced period estimates are 4, 10, and 16. Apply weights $1/4,1/2,1/4$ to estimate the middle variance, and state the pattern assumption that supports targeting the middle age. (Answer: 10; a locally linear variance pattern.)
One development period has one observed origin year, with no usable separate variance estimate. A comparable period has residual sum 12 and one residual degree of freedom; another has residual sum 8 and one degree of freedom. Under a common-variance assumption, calculate a pooled estimate. (Answer: $(12+8)/(1+1)=10$ in the monetary unit used for cumulative claims.)

What You'll Learn