Why Development-Factor Selection Affects Uncertainty
A chain-ladder projection depends on selected development factors. Changing a factor can change the projected ultimate loss and the variance contribution associated with that development period. The direction of the reserve change does not, by itself, determine the direction of the standard-error change: the ultimate estimate and the relative prediction error both matter.
This tutorial examines sensitivity by holding the observed data and the stated variance inputs fixed, changing one selected factor, and recalculating the projection. The comparison isolates the arithmetic effect of factor judgment; it does not establish that either selection is better supported by the data. A sensitivity analysis is most informative when it shows the changed factor, the affected uncertainty terms, and the resulting reserve and prediction error side by side.
Variance Contributions Under a Selected Factor
For a development period, write \(C\) for the cumulative amount entering that period, \(f\) for its selected development factor, and \(\sigma^2\) for the process-variance parameter. Under the Mack framework, the conditional mean of the next cumulative amount is \(fC\), and its conditional process variance is \(\sigma^2C\). Let \(\tau^2\) denote the variance assigned to the selected factor estimate. The relative contribution for this period can be written as
The first term represents process uncertainty; the second represents uncertainty in estimating the development factor. In the usual Mack calculation, factor-estimation variance is obtained from the development data under the model. In a sensitivity calculation, use the factor-variance input appropriate to the selection being evaluated. For several projected development periods, let \(q_k\) be the relative contribution at period \(k\). Under the standard Mack MSEP calculation, the contributions are accumulated and scaled by the squared projected ultimate.
These expressions assume the Mack model is appropriate: development factors describe a common pattern, conditional process variance is proportional to the cumulative amount, and the relevant development-period and accident-year errors satisfy the model’s uncorrelatedness conditions. A factor sensitivity calculation does not repair violations of those assumptions. In particular, results require care when cumulative amounts are zero or negative, or when dependence or development-pattern changes undermine the model.
Why the Relative Contribution Has This Form
A one-period derivation explains the two terms. Suppose the future cumulative amount has process variance \(\sigma^2C\), while uncertainty in the factor estimate has variance \(\tau^2\). Treat the process error and factor-estimation error as uncorrelated, as in this decomposition. The variance attributable to process error is \(\sigma^2C\); the variance attributable to factor uncertainty is \(C^2\tau^2\), since the factor multiplies \(C\). Thus the combined variance is \(\sigma^2C+C^2\tau^2\). Dividing by the squared projected amount \((fC)^2\) gives
This is a relative contribution. It cannot be compared with a dollar standard error without also using the projected ultimate. In a multi-period Mack calculation, the period contributions are formed using the appropriate projected amounts and factor-uncertainty terms, then combined as shown above.
Worked Example: Recalculating a Selected Late Factor
A hypothetical accident year has a current cumulative amount of $1,000. Two development periods remain. The first selected factor is 2.00, and the late factor is initially 1.50. For the first period, \(\sigma_1^2=400\) and \(\tau_1^2=0.0100\). At the second period, the cumulative amount entering development is $2,000, with \(\sigma_2^2=400\) and \(\tau_2^2=0.2525\). Compare the initial selection with a late factor of 1.40, holding those variance inputs fixed.
Method. Compute each period’s process and factor-estimation contributions, sum them to obtain relative MSEP, then multiply by the squared ultimate. This comparison avoids treating a lower projected ultimate as automatically implying a lower standard error.
Initial ultimate. The projected ultimate is \(1{,}000(2.00)(1.50)=\$3{,}000\).
First-period contribution. The process term is \(400/(2.00^2\times1{,}000)=400/4{,}000=0.1000\). The factor-estimation term is \(0.0100/2.00^2=0.0025\). The contribution is \(0.1025\).
Second-period contribution under 1.50. The process term is \(400/(1.50^2\times2{,}000)=400/4{,}500=0.0888889\). The factor-estimation term is \(0.2525/1.50^2=0.2525/2.25=0.1122222\). The contribution is \(0.2011111\). Therefore, total relative MSEP is \(0.1025+0.2011111=0.3036111\).
The MSEP is \(3{,}000^2(0.3036111111)=\$2{,}732{,}500\). The standard error is \(\sqrt{2{,}732{,}500}=\$1{,}653.03\), rounded to the nearest cent.
Alternative late factor, 1.40. The ultimate becomes \(1{,}000(2.00)(1.40)=\$2{,}800\). The first-period contribution is unchanged at \(0.1025\). For the late period, the process term is \(400/(1.40^2\times2{,}000)=400/3{,}920=0.1020408\), and the factor-estimation term is \(0.2525/1.40^2=0.2525/1.96=0.1288265\). The late contribution is \(0.2308673\), so total relative MSEP is \(0.1025+0.2308673=0.3333673\).
The alternative MSEP is \(2{,}800^2(0.3333673469)=\$2{,}613{,}600\), and its standard error is \(\sqrt{2{,}613{,}600}=\$1{,}616.66\), rounded to the nearest cent. The selected late factor reduces the ultimate by $200 and the MSEP by $118,900. The standard-error reduction from the unrounded results is about $36.36. Because the two standard errors are rounded separately to cents, subtracting the displayed figures gives $36.37; the difference of the unrounded figures rounds to $36.36.
Check. The alternative has higher relative MSEP, \(0.3333673\) rather than \(0.3036111\), but a lower dollar MSEP because it is applied to a smaller squared ultimate. Both comparisons are needed to describe the result.
Comparing a Reserve Change With a Relative-MSEP Change
A factor selection can lower the projected reserve while increasing relative uncertainty. The following example uses period contributions already calculated from the two selections’ development and variance inputs. This focuses attention on the final comparison: each set of contributions must be summed and scaled by its own ultimate.
Worked Example: A Smaller Ultimate With Higher Relative Uncertainty
A hypothetical two-period development analysis gives a projected ultimate of $1,500 under the baseline selection and $1,350 under an alternative selection. The period relative-MSEP contributions are \(0.0600000000\) and \(0.0936117378\) under the baseline, and \(0.0600000000\) and \(0.1170370370\) under the alternative. Calculate and compare the MSEP and standard error.
Baseline. Total relative MSEP is \(0.0600000000+0.0936117378=0.1536117378\). The MSEP is \(1{,}500^2(0.1536117378)=2{,}250{,}000(0.1536117378)=\$345{,}626.41\), to the nearest cent. The standard error is \(\sqrt{345{,}626.41}=\$587.90\).
Alternative. Total relative MSEP is \(0.0600000000+0.1170370370=0.1770370370\). The MSEP is \(1{,}350^2(0.1770370370)=1{,}822{,}500(0.1770370370)=\$322{,}650.00\). The standard error is \(\sqrt{322{,}650}=\$568.02\), rounded to the nearest cent.
The alternative lowers the ultimate by $150 and the MSEP by $22,976.41. The unrounded standard errors differ by approximately $19.88. Relative MSEP increases, even though dollar MSEP decreases: the relative measure is larger, but it is multiplied by the smaller squared ultimate. Do not substitute one of these comparisons for the other.
Tracing the Effect Through More Than One Development Period
When an earlier factor changes, it can alter the cumulative amount entering later periods as well as the ultimate. When only a late factor changes, earlier contributions may remain fixed, but the late-period relative contribution can change because the factor appears in its denominator. The next example shows the calculation period by period.
Worked Example: Changing the Final Factor in a Two-Period Projection
A hypothetical accident year has a current cumulative amount of $1,000. The first selected factor is 1.50, with \(\sigma_1^2=100\) and \(\tau_1^2=0.0125\). The second period begins at cumulative amount $1,500, with \(\sigma_2^2=180\) and \(\tau_2^2=0.0400\). Compare a final factor of 2.00 with 1.80, holding all stated variance inputs fixed.
Baseline projection. The ultimate is \(1{,}000(1.50)(2.00)=\$3{,}000\). For the first period, the process contribution is \(100/(1.50^2\times1{,}000)=100/2{,}250=0.0444444\), and the factor-estimation contribution is \(0.0125/2.25=0.0055556\). Thus \(q_1=0.0500000\).
For the second period, the process contribution is \(180/(2.00^2\times1{,}500)=180/6{,}000=0.0300000\), and the factor-estimation contribution is \(0.0400/4=0.0100000\). Thus \(q_2=0.0400000\), and total relative MSEP is \(0.0900000\). MSEP is \(3{,}000^2(0.09)=\$810{,}000\); the standard error is \(\sqrt{810{,}000}=\$900.00\).
Alternative projection. The ultimate is \(1{,}000(1.50)(1.80)=\$2{,}700\). The first contribution remains \(0.0500000\). For the second period, the process term is \(180/(1.80^2\times1{,}500)=180/4{,}860=0.0370370\), and the factor-estimation term is \(0.0400/3.24=0.0123457\). Thus \(q_2=0.0493827\) and total relative MSEP is \(0.0993827\).
The alternative MSEP is \(2{,}700^2(0.0993827160)=\$724{,}500\), and its standard error is \(\sqrt{724{,}500}=\$851.18\), rounded to the nearest cent. The ultimate is lower by $300, relative MSEP is higher, and dollar MSEP is lower by $85,500. The standard error falls by about $48.82.
Check. The alternative late-period contribution is higher because both late-period terms are divided by the smaller factor squared. Yet the absolute MSEP falls after scaling by the smaller squared ultimate. This is a direct example of why relative and absolute uncertainty should be reported separately.
Common Traps and a Practical Comparison Sequence
One arithmetic error can materially distort the conclusion. For the first period in the first example, the process term is
not \(0.0500\). The denominator is \(2.00^2\times1{,}000=4{,}000\); the numerator remains 400. Using \(0.0500\) understates relative MSEP by \(0.0500\), which in that example understates MSEP by \(3{,}000^2(0.0500)=\$450{,}000\).
A disciplined comparison helps prevent this and related errors. Recalculate each selection independently rather than carrying forward a result that depended on the old factor.
Record the alternative factor and identify which projected amounts and variance terms depend on it.
Apply the selected factors in development order. Do not use the old ultimate to scale a new relative MSEP.
Separate process and factor-estimation terms and check every squared factor and cumulative amount in the denominator.
Sum the relative contributions, multiply by the squared ultimate to obtain MSEP, and take the square root for standard error.
Report the changes in reserve, relative MSEP, dollar MSEP, and standard error separately; state the variance assumptions held fixed.
- Do not confuse reserve with ultimate. If the latest observed cumulative amount is unchanged, an ultimate reduction of $200 also reduces the reserve by $200. Otherwise, recalculate reserve as ultimate minus the latest observed amount.
- Do not compare only relative MSEP. Dollar MSEP depends on the squared ultimate, so the two measures can move in opposite directions.
- Do not use a rounded contribution too early. Retain adequate precision through the sum, MSEP, and square root; round only the reported final money amounts.
- Do not assume a selected factor is certain. Omitting its factor-estimation term understates predictive uncertainty when that term belongs in the model.
Check Your Understanding
For each problem, calculate the requested quantities before checking the answer. Treat all amounts and model inputs as hypothetical.
- A one-period projection starts from cumulative claims of $800. The selected factor is 2.00, with \(\sigma^2=320\) and \(\tau^2=0.0400\). Calculate the relative contribution, projected ultimate, and MSEP. (Answer: relative contribution \(0.1100\); ultimate $1,600; MSEP $281,600.)
- A two-period projection starts at $500. The first factor is 1.20, with \(\sigma_1^2=72\) and \(\tau_1^2=0.0144\). The second period starts at $600, with factor 2.00, \(\sigma_2^2=180\), and \(\tau_2^2=0.0400\). Calculate each contribution, total relative MSEP, ultimate, and MSEP. (Answer: \(q_1=0.1100\), \(q_2=0.0850\), total relative MSEP \(0.1950\); ultimate $1,200; MSEP $280,800.)
- A hypothetical two-period projection begins at $2,000. Its baseline ultimate is $4,000, with relative contributions \(0.0400\) and \(0.0600\). An alternative late-factor selection gives ultimate $3,750 and contributions \(0.0400\) and \(0.0750\). Calculate both MSEPs and state which is higher. (Answer: baseline MSEP $1,600,000; alternative MSEP $1,617,187.50; the alternative is higher by $17,187.50 despite its lower ultimate.)
- A projection starts at $1,000. The first factor is 1.40 and its relative contribution is \(0.0500\). Under a baseline final factor of 2.00, the final-period contribution is \(0.0800\); under an alternative final factor of 1.80, it is \(0.1000\). Calculate each ultimate and MSEP, then compare relative and dollar MSEP. (Answer: baseline ultimate $2,800 and MSEP $1,019,200; alternative ultimate $2,520 and MSEP $952,560; relative MSEP rises from \(0.1300\) to \(0.1500\), while dollar MSEP falls by $66,640.)