Deriving Process Variance For Incremental Claims

Why Incremental-Claim Variance Needs Care

A reserving triangle records cumulative claims by origin period and development age, while the claims expected to emerge in the future are incremental amounts. To quantify uncertainty in a future increment under the Mack chain-ladder model, it is essential to distinguish the variance assumption on the next cumulative amount from the variance of the increment itself.

The key result is direct: conditional on the claims observed so far for an origin period, the variance of its next incremental amount is proportional to the cumulative claims at the current development age. The proportionality constant can differ by development age. The result is a conditional process variance; it is not, by itself, a complete measure of reserve prediction error.

Definition: Let $C_{i,j}$ be the cumulative claims for origin period $i$ at development age $j$. Define the next incremental amount by $\Delta C_{i,j+1}=C_{i,j+1}-C_{i,j}$. Let $\mathcal H_{i,j}$ denote the information from the claims for origin period $i$ observed through development age $j$. The Mack model specifies, for development age $j$, $$ E[C_{i,j+1}\mid\mathcal H_{i,j}]=f_j C_{i,j}, \qquad \operatorname{Var}(C_{i,j+1}\mid\mathcal H_{i,j})=\sigma_j^2 C_{i,j}. $$ Here $f_j$ is the age-specific expected cumulative development factor and $\sigma_j^2$ is the age-specific process variance parameter.

These are conditional assumptions: $C_{i,j}$ is already known given $\mathcal H_{i,j}$, while $C_{i,j+1}$ is not yet known. The claims amounts and $\sigma_j^2$ must use consistent units. For example, if claims are expressed in thousands of dollars, the variance parameter must be expressed in the corresponding thousands-of-dollars scale.

The standard Mack setup also assumes that origin periods are independent for the process being modeled. The displayed conditional mean and variance assumptions are the ones needed for the single-increment derivation below. They do not assert that successive increments within one origin period are unconditionally independent: a later increment depends on the cumulative amount that has emerged by then.

Deriving the Incremental-Claim Variance

Given $\mathcal H_{i,j}$, the current cumulative amount $C_{i,j}$ is fixed. The next increment is the next cumulative amount minus this known amount. Subtracting a known constant changes the conditional mean but does not change the conditional variance.

\begin{aligned} E[\Delta C_{i,j+1}\mid\mathcal H_{i,j}] &=E[C_{i,j+1}-C_{i,j}\mid\mathcal H_{i,j}]\\ &=f_j C_{i,j}-C_{i,j}\\ &=(f_j-1)C_{i,j},\\[4pt] \operatorname{Var}(\Delta C_{i,j+1}\mid\mathcal H_{i,j}) &=\operatorname{Var}(C_{i,j+1}-C_{i,j}\mid\mathcal H_{i,j})\\ &=\operatorname{Var}(C_{i,j+1}\mid\mathcal H_{i,j})\\ &=\sigma_j^2 C_{i,j}. \end{aligned}

The first line gives the expected increment; the last line gives its conditional variance. A development factor therefore determines the conditional mean increment, but it does not appear directly in the Mack formula for the conditional increment variance. The variance depends on the current cumulative amount and the variance parameter for that development age.

Key result: Under the Mack conditional variance assumption, $$ \operatorname{Var}(\Delta C_{i,j+1}\mid\mathcal H_{i,j})=\sigma_j^2 C_{i,j}. $$ This is a process variance conditional on the observed history and the model parameters being used. It is not an unconditional variance that includes uncertainty in estimated development factors.

How Cumulative Claims and Development Age Enter

At a fixed development age $j$, the parameter $\sigma_j^2$ is held constant across origin periods in the Mack model. Consequently, if one origin period has twice the current cumulative claims of another, its conditional process variance for the next increment is twice as large, provided both are being evaluated at the same development age. Its conditional standard deviation is multiplied by $\sqrt{2}$, not by $2$.

Across different development ages, the process variance parameter can change. The model does not require variance to increase or decrease with age. A larger cumulative amount at a later age does not guarantee a larger incremental variance: the value of $\sigma_j^2$ at that later age matters too.

An equivalent way to understand the assumption is through the cumulative development ratio $R_{i,j+1}=C_{i,j+1}/C_{i,j}$, when $C_{i,j}>0$. Since $C_{i,j}$ is known conditionally, the Mack assumptions imply:

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

Multiplying the ratio by the known current cumulative amount gives $\Delta C_{i,j+1}=C_{i,j}(R_{i,j+1}-1)$. Its conditional variance is therefore $C_{i,j}^2(\sigma_j^2/C_{i,j})=\sigma_j^2 C_{i,j}$, agreeing with the direct derivation. This ratio form also shows that the assumed variance of the development ratio decreases as current cumulative claims increase, at a fixed age and fixed $\sigma_j^2$. If $C_{i,j}=0$, the ratio is undefined; use the cumulative-amount variance formula rather than the ratio representation.

Identify the current age and amount.
Use $C_{i,j}$, the cumulative claims already observed at the age immediately before the increment being forecast.

Use the matching age-specific inputs.
Use $f_j$ for the expected development and $\sigma_j^2$ for the process variance from age $j$ to $j+1$.

Calculate the conditional moments.
The mean increment is $(f_j-1)C_{i,j}$; the variance is $\sigma_j^2 C_{i,j}$. Take the square root of the variance if a standard deviation is requested.

Worked Calculations

Worked Example: A Future Increment at One Development Age

A hypothetical insurer has cumulative claims of $800$ thousand dollars for an origin period at age $j=3$. For development from age 3 to age 4, take $f_3=1.25$ and $\sigma_3^2=0.04$, with the variance parameter expressed in thousands of dollars. Find the conditional mean, variance, and standard deviation of the next incremental amount.

Method. Apply the Mack conditional mean and variance assumptions at age 3. The observed amount is $C_{i,3}=800$, so the expected next cumulative amount is $f_3C_{i,3}=1.25(800)=1{,}000$ thousand dollars.

The conditional mean increment is the difference between the expected next cumulative amount and the known current cumulative amount: $1{,}000-800=200$ thousand dollars. Equivalently, $(1.25-1)(800)=200$.

The conditional variance of the increment is $\sigma_3^2C_{i,3}=0.04(800)=32$ in squared-thousands-of-dollars units. Its conditional standard deviation is $\sqrt{32}=5.656854\ldots$ thousand dollars, or $\$5{,}656.85$ to the nearest cent.

Check. The increment mean is positive because the development factor exceeds 1. The standard deviation is the square root of 32, not 32 thousand dollars. The calculation uses the current cumulative amount, 800, rather than the expected next cumulative amount, 1,000.

Common mistake avoided: Multiplying the variance parameter by the next expected cumulative amount would give $0.04(1{,}000)=40$, which is not the Mack conditional variance for this transition. The formula uses $C_{i,3}$, known at the time of conditioning.

Worked Example: Comparing Two Development Ages

For two hypothetical origin periods, the current cumulative amounts and age-specific Mack inputs are as follows. All claim amounts are in thousands of dollars, and the variance parameters are in the same amount scale.

Transition	Current cumulative amount	$f_j$	$\sigma_j^2$
Age 1 to 2	500	1.30	18
Age 2 to 3	1,200	1.10	5

For age 1 to 2, the conditional mean increment is $(1.30-1)(500)=0.30(500)=150$ thousand dollars. Its conditional variance is $18(500)=9{,}000$ squared-thousands of dollars, giving a standard deviation of $\sqrt{9{,}000}=94.8683\ldots$ thousand dollars.

For age 2 to 3, the conditional mean increment is $(1.10-1)(1{,}200)=0.10(1{,}200)=120$ thousand dollars. Its conditional variance is $5(1{,}200)=6{,}000$ squared-thousands of dollars, giving a standard deviation of $\sqrt{6{,}000}=77.4597\ldots$ thousand dollars.

Thus, the first transition has the larger conditional variance: $9{,}000$ compared with $6{,}000$, despite having the smaller current cumulative amount. Its age-specific variance parameter is much larger. The standard deviations, rounded to two decimal places in thousands of dollars, are $94.87$ and $77.46$, respectively.

Check. Each variance uses the parameter for its own development age. Comparing cumulative amounts alone would incorrectly suggest that the second transition must have the larger variance.

Common mistake avoided: Applying one age’s $\sigma_j^2$ to both transitions ignores that the process variance parameter is development-age-specific.

Worked Example: Converting Units Without Changing the Model

A hypothetical portfolio has $C_{i,4}=\$2{,}400{,}000$. For development from age 4 to age 5, $f_4=1.10$ and $\sigma_4^2=\$12{,}000$, with the parameter stated in dollar units. Find the expected increment and conditional standard deviation in dollars.

The conditional mean increment is $(1.10-1)(\$2{,}400{,}000)=0.10(\$2{,}400{,}000)=\$240{,}000$. The conditional variance is $\$12{,}000\times \$2{,}400{,}000=28{,}800{,}000{,}000$ dollars squared.

Taking the square root gives a conditional standard deviation of $\sqrt{28{,}800{,}000{,}000}=\$169{,}705.627\ldots$, or $\$169{,}705.63$ to the nearest cent.

As a unit check, express claims in thousands: $C_{i,4}=2{,}400$ and $\sigma_4^2=12$. Then the variance is $12(2{,}400)=28{,}800$ squared-thousands of dollars. Its standard deviation is $\sqrt{28{,}800}=169.7056\ldots$ thousand dollars, consistent with the dollar calculation.

Check. The expected increment is 10% of the current cumulative claims, and the two unit calculations produce the same standard deviation. Mixing $C_{i,4}=2{,}400$ in thousands with $\sigma_4^2=\$12{,}000$ in dollars would not be a valid multiplication.

Common mistake avoided: Changing the scale of cumulative claims without changing the scale of the variance parameter can produce a numerically plausible but dimensionally inconsistent result.

Interpretation and Exam Traps

The Mack variance assumption describes how the conditional spread of a future increment scales with the cumulative amount already observed. It does not specify a full probability distribution for the increment. In particular, the formula alone does not establish that increments are normally distributed or that a calculated mean and standard deviation determine all tail probabilities.

The conditional qualification matters. At the valuation point, $C_{i,j}$ is observed and treated as fixed when calculating this one-step process variance. If instead one asks for an overall reserve prediction error, additional sources of uncertainty and dependencies may need to be considered. The conditional process variance here should not be mistaken for that broader quantity.

Using the wrong development age: The transition from age $j$ to $j+1$ uses $\sigma_j^2$, not the parameter for the destination age.
Using the wrong cumulative amount: The variance is $\sigma_j^2C_{i,j}$, not $\sigma_j^2C_{i,j+1}$ or the expected next cumulative amount.
Adding variance for a known amount: Conditional on the history, $C_{i,j}$ is fixed. Subtracting it from $C_{i,j+1}$ does not add a second variance term.
Confusing variance and standard deviation: First compute $\sigma_j^2C_{i,j}$; take its square root only when a standard deviation is requested.
Mixing units: Keep the cumulative amount and variance parameter on compatible scales, and remember that the resulting variance has squared monetary units.

For a quick consistency check, verify the development age, confirm that the formula uses the current cumulative amount, and check the units before taking a square root. The result should be nonnegative when cumulative claims and the process variance parameter are nonnegative.

Check Your Understanding

For each question, calculate the requested conditional quantities using the Mack assumptions. Keep amounts and variance parameters on the stated scale.

A hypothetical origin period has $C_{i,2}=900$ thousand dollars. For age 2 to 3, $f_2=1.08$ and $\sigma_2^2=0.03$ in the stated thousands-of-dollars scale. Find the conditional mean increment and conditional standard deviation in thousands of dollars. (Answer: mean $72$; standard deviation $\sqrt{27}\approx5.1962$)
At one development age, $C_{i,j}=400$ thousand dollars, $f_j=1.50$, and $\sigma_j^2=16$ in the same scale. Find the mean increment and variance. (Answer: mean $200$ thousand dollars; variance $6{,}400$ squared-thousands of dollars)
For one transition, $C_{i,j}=750$ and $\sigma_j^2=8$; for another, $C_{k,r}=1{,}500$ and $\sigma_r^2=3$. All values use thousands of dollars. Which transition has the larger conditional variance, and what are the two variances? (Answer: the first, with variance $6{,}000$, compared with $4{,}500$ squared-thousands of dollars)
Claims are measured in thousands of dollars, with $C_{i,3}=2{,}000$, $f_3=1.12$, and $\sigma_3^2=10$ in the same scale. Find the conditional mean, variance, and standard deviation of the next increment. (Answer: mean $240$ thousand dollars; variance $20{,}000$ squared-thousands of dollars; standard deviation $\sqrt{20{,}000}\approx141.4214$ thousand dollars)
A hypothetical origin period has $C_{i,1}=1{,}000$ thousand dollars, $f_1=1.15$, and $\sigma_1^2=0.02$. A calculation uses $0.02(1{,}150)=23$ as the variance. Identify the error and give the correct variance and standard deviation. (Answer: it uses the expected next cumulative amount rather than the current amount; the correct variance is $0.02(1{,}000)=20$ squared-thousands of dollars and the standard deviation is $\sqrt{20}\approx4.4721$ thousand dollars)

What You'll Learn