Introduction
A chain-ladder projection gives an estimated ultimate loss and reserve, but a point estimate does not describe how uncertain that projection is. Mack’s model provides a framework for estimating the mean squared error of prediction (MSEP) from the claims triangle itself. Its calculation depends on two distinct sources of uncertainty: variation in future claims development and uncertainty in the estimated development factors.
This tutorial works through both sources using one hypothetical cumulative triangle. The key discipline is to use only observed links at each development age: the number of usable links decreases as development age increases.
The Triangle and the Development-Period Estimates
Consider the following hypothetical cumulative claims triangle, with amounts in thousands of dollars. The square-root terms are exact values chosen to make the age-2 variance calculation transparent; their approximate dollar amounts are also shown.
| Accident year | Age 1 | Age 2 | Age 3 |
|---|---|---|---|
| AY1 | 100 | 160 | \(200+\sqrt{112}\approx210.583\) |
| AY2 | 100 | 140 | \(175-\sqrt{112}\approx164.417\) |
| AY3 | 200 | 300 | 375 |
| AY4 | 100 | 150 | — |
| AY5 | 100 | — | — |
The age-1-to-age-2 factor uses four observed links, AY1 through AY4. The age-2-to-age-3 factor uses three observed links, AY1 through AY3. AY5 has no observed age-2 amount, so it contributes to neither factor at age 1 nor any calculation involving an observed age-2 link.
Worked Example: Estimating the Factors and Variance Parameters
Method. Calculate each factor from the sums over observed links, then calculate its variance estimate from the weighted squared deviations of those links from the factor. This method avoids treating an unobserved cell as a zero or as an available link.
Age 1 to age 2. The observed age-1 amounts are \(100,100,200,100\), totalling \(500\). Their corresponding age-2 amounts are \(160,140,300,150\), totalling \(750\). Thus $$ \hat f_1=\frac{750}{500}=1.5. $$ The four observed ratios are \(1.6,1.4,1.5,1.5\), so the weighted squared deviations are $$ 100(1.6-1.5)^2=1,\quad 100(1.4-1.5)^2=1,\quad 200(1.5-1.5)^2=0,\quad 100(1.5-1.5)^2=0. $$ Their sum is \(2\), and \(n_1=4\). Therefore $$ \hat\sigma_1^2=\frac{2}{4-1}=\frac{2}{3}. $$
Age 2 to age 3. Only AY1 through AY3 have both amounts observed. Their age-2 amounts total \(160+140+300=600\). Their age-3 amounts total $$ (200+\sqrt{112})+(175-\sqrt{112})+375=750. $$ Consequently, \(\hat f_2=750/600=1.25\). The deviations of the three link ratios from 1.25 can equivalently be calculated as the residual amounts \(C_{i,3}-1.25C_{i,2}\): these are \(\sqrt{112},-\sqrt{112},0\). The weighted squared deviations sum to $$ \frac{(\sqrt{112})^2}{160} +\frac{(-\sqrt{112})^2}{140} +\frac{0^2}{300} =\frac{112}{160}+\frac{112}{140} =0.7+0.8=1.5. $$ There are three observed links, so $$ \hat\sigma_2^2=\frac{1.5}{3-1}=0.75. $$
Check. The factor totals use four links at age 1 and three at age 2. The respective variance denominators are therefore \(4-1=3\) and \(3-1=2\). Both factors are weighted by cumulative amounts, not calculated as unweighted averages of the individual ratios.
Why the Mack MSEP Has Two Components
At one development step, a prediction can differ from the eventual observed cumulative amount because the next development itself is variable and because the factor used to predict it is estimated. The first is process uncertainty; the second is parameter uncertainty.
To see the one-step formula, write the next amount as \(C_{i,k+1}=f_kC_{i,k}+\epsilon_{i,k}\), where the process deviation has conditional mean zero and variance \(\sigma_k^2C_{i,k}\). The estimated factor is a weighted ratio. Under the Mack variance model, its approximate variance is \(\sigma_k^2/\sum_i C_{i,k}\) over the observed links. Treating the future process deviation and factor estimation error as separate contributions gives the plug-in one-step prediction variance $$ \hat C_{i,k}\hat\sigma_k^2 +\hat C_{i,k}^{\,2}\frac{\hat\sigma_k^2}{\sum_i C_{i,k}}. $$ The first term is process variance: it grows with the current cumulative amount. The second is parameter variance: it reflects uncertainty in the shared estimated factor and therefore uses the total observed volume at that development age in its denominator.
The propagation rule follows from the variance of a scaled forecast: if the existing prediction error has variance \(V\), multiplying the forecast by a fixed factor \(\hat f\) multiplies that contribution by \(\hat f^2\). The next development introduces new process and parameter uncertainty, so those one-step terms are added. This recursion is a plug-in Mack estimate; it relies on the model’s variance structure and the selected factors, and it is not a distribution-free guarantee about the actual prediction error.
Worked Example: AY5 Reserve and Recursive MSEP
Method. Project AY5 through each remaining development age, calculate the first-step MSEP, carry it forward by the square of the next factor, and add the new age-2 uncertainty. All amounts are in thousands of dollars.
AY5 is observed at age 1 with \(C_{5,1}=100\). Its projected age-2 cumulative amount is $$ \hat C_{5,2}=100(1.5)=150. $$ The projected age-3 ultimate is $$ \hat C_{5,3}=150(1.25)=187.5. $$ The estimated reserve is ultimate less observed cumulative claims: $$ \hat R_5=187.5-100=87.5. $$
At age 1, the process contribution to the MSEP is $$ 100\left(\frac{2}{3}\right)=\frac{200}{3}=66.6667. $$ The parameter contribution uses the four observed age-1 amounts, whose sum is \(500\): $$ \frac{100^2(2/3)}{500}=\frac{200}{3}=66.6667? $$
The expression above must be evaluated carefully: \(100^2(2/3)/500=10{,}000(2/3)/500=40/3=13.3333\), not \(200/3\). Thus the correct age-1 MSEP is $$ \frac{200}{3}+\frac{40}{3}=80. $$
Carrying that uncertainty to age 3 multiplies it by \(\hat f_2^2=1.25^2=1.5625\): $$ 80(1.5625)=125. $$ At age 2, AY5’s projected cumulative amount is \(150\). The age-2 process contribution is $$ 150(0.75)=112.5, $$ and the parameter contribution, using the sum \(600\) of the three observed age-2 amounts, is $$ \frac{150^2(0.75)}{600} =\frac{16{,}875}{600}=28.125. $$ Therefore $$ \widehat{\operatorname{MSEP}}_5=125+112.5+28.125=265.625. $$ The estimated standard error is the square root of the MSEP: $$ \widehat{\operatorname{SE}}_5=\sqrt{265.625}\approx16.298. $$ Rounded to three decimals, the reserve standard error is \(16.298\) thousand dollars, or \(\$16{,}298.01\) when converted and rounded to cents.
Check. The MSEP components sum to \(265.625\), and \(16.298^2\) is approximately \(265.625\). The reserve is \(87.5\) thousand dollars; its standard error is not the standard error of the ultimate, since the observed age-1 amount is fixed in this conditional reserve calculation.
Applying the Calculation to AY4
AY4 is already observed at age 2, so it has only one remaining development step in this triangle. Its projected age-3 ultimate is \(150(1.25)=187.5\), and its reserve is \(187.5-150=37.5\) thousand dollars. The age-2 one-step MSEP uses the same age-2 variance estimate and denominator as the AY5 age-2 contribution:
Worked Example: AY4 One-Step Reserve Error
Method. Since AY4 is observed at age 2, use the age-2 one-step formula directly; there is no age-1 uncertainty to propagate for this accident year.
With \(C_{4,2}=150\), \(\hat\sigma_2^2=0.75\), and \(\sum_jC_{j,2}=600\) over the three observed age-2-to-age-3 links, the process contribution is $$ 150(0.75)=112.5. $$ The parameter contribution is $$ \frac{150^2(0.75)}{600}=28.125. $$ Thus $$ \widehat{\operatorname{MSEP}}_4=112.5+28.125=140.625, \qquad \widehat{\operatorname{SE}}_4=\sqrt{140.625}\approx11.859. $$ The reserve estimate is \(37.5\) thousand dollars, and the standard error is \(11.859\) thousand dollars, rounded to three decimals. As a check, the AY4 MSEP equals the two age-2 contributions that are added for AY5; AY5 additionally has its earlier age-1 uncertainty, propagated to age 3 as \(125\).
Common Exam Traps and Practical Checks
The most consequential error in a small triangle is counting a missing cell as if it were an observed link. For the age-1 factor here, the correct calculation is \(750/500=1.5\) from AY1 through AY4. AY5 has an age-1 amount but no age-2 amount, so it contributes neither a link ratio nor a term to the age-1 variance estimate. The variance estimate’s denominator is the number of observed links minus one, not the number of accident years in the triangle minus one.
For example, including AY5 in the age-1 denominator would incorrectly use five links and produce \(2/(5-1)=0.5\). That is not the Mack estimate for this triangle. The correct four-link estimate is \(2/(4-1)=2/3\). This change affects both the first-step MSEP and its propagated contribution at the final age.
- Do not confuse cumulative and incremental amounts. The Mack formulas above use cumulative amounts \(C_{i,k}\) and observed cumulative link ratios.
- Do not omit the factor square in recursion. An existing MSEP carried through a factor \(1.25\) is multiplied by \(1.25^2\), not by \(1.25\).
- Do not use the wrong volume denominator. The age-1 parameter term uses \(500\), while the age-2 term uses \(600\), each summed over links observed at that age.
- Do not report MSEP as standard error. MSEP is in squared units; take its square root to obtain a standard error in the original monetary units.
- Keep the scale consistent. These calculations use thousands of dollars throughout. A standard error of \(16.298\) is therefore \(16.298\) thousand dollars, not \(\$16.298\).
The calculation is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus. In an exam-style solution, show the observed-link count at each age, state the factor and variance estimates, identify the process and parameter terms, and verify that the final standard error is the square root of the stated MSEP.
Check Your Understanding
For each question, use only the amounts specified as observed and keep track of the development age represented by each quantity.
- A hypothetical triangle has three observed links from age 1 to age 2: \((C_{i,1},C_{i,2})=(100,140),(200,280),(100,130)\). Calculate the volume-weighted factor. (Answer: \(\hat f_1=550/400=1.375\).)
- At one development age, three observed link ratios are \(1.2,1.4,1.5\), with corresponding starting cumulative amounts \(100,100,200\). Calculate the Mack variance estimate using the weighted squared-deviation estimator. (Answer: \(\hat f=1.4\), weighted squared deviations \(=100(0.2)^2+100(0)^2+200(0.1)^2=6\), so \(\hat\sigma^2=6/(3-1)=3\).)
- An accident year has \(C_{i,k}=80\), and the selected factor for its sole remaining development age is \(1.25\). Calculate its projected next cumulative amount and reserve if \(80\) is the currently observed cumulative amount. (Answer: projected amount \(100\); reserve \(20\).)
- For a one-step prediction, let \(\hat C_{i,k}=120\), \(\hat\sigma_k^2=0.5\), and \(\sum_j C_{j,k}=600\) over observed links. Calculate the process term, parameter term, and MSEP. (Answer: process \(=60\), parameter \(=120^2(0.5)/600=12\), MSEP \(=72\).)
- An accident year has an age-1 MSEP of \(50\), then develops through a factor of \(1.2\). At age 2, its projected cumulative amount is \(120\), with \(\hat\sigma_2^2=0.4\) and observed-link denominator sum \(480\). Calculate its age-3 MSEP and standard error. (Answer: propagated age-1 MSEP \(=50(1.2^2)=72\); age-2 process \(=120(0.4)=48\); parameter \(=120^2(0.4)/480=12\); total MSEP \(=132\); standard error \(=\sqrt{132}\approx11.489\).)