Why the Sign of a Claim Amount Matters
A zero or negative incremental claim amount does not automatically make a development calculation impossible. An increment can be zero because no new claims emerged, or negative because recoveries, re-estimates, or other adjustments reduced cumulative claims. Mack’s variance calculations, however, use cumulative claims at the start of a development period as the predictor. The sign of that cumulative amount—not merely the sign of the increment—determines whether the standard calculation is defined and meaningful.
This distinction matters when estimating development-factor variance and assessing standardized residuals. Treating every negative increment as unusable can discard valid information; applying the formula to a negative cumulative predictor can produce an impossible negative variance. The calculations and diagnostic rules below are aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.
The Mack Factor and Variance Calculations
For a development age \(j\), suppose there are \(m\) observations with positive starting cumulative claims. The chain-ladder development-factor estimate is the ratio of total next cumulative claims to total starting cumulative claims. The usual Mack estimator of the development-period variance is based on the squared residuals, each scaled by its starting cumulative amount:
Equivalently, when \(C_{i,j}>0\), each term in the variance sum can be written as \(C_{i,j}(C_{i,j+1}/C_{i,j}-\widehat f_j)^2\). The residual in dollars is \(e_{i,j}=C_{i,j+1}-\widehat f_j C_{i,j}\). A commonly used standardized residual is
This standardized residual is defined only when \(C_{i,j}>0\) and \(\widehat{\sigma}_j^2>0\). It is an estimated diagnostic quantity, not a guarantee that residuals are independent or normally distributed.
Why These Formulas Have This Form
The factor estimate is the weighted average of observed development ratios, with starting cumulative claims as weights. To see why, write each ratio as \(R_i=C_{i,j+1}/C_{i,j}\). The weighted average is
The weights reflect the variance assumption: a larger \(C_{i,j}\) has a larger modeled variance, but a smaller relative variance. The variance estimator then measures residual dispersion on the scale implied by that assumption. Each squared residual is divided by \(C_{i,j}\), and one degree of freedom is used because the factor is estimated from the same observations. This is why the divisor is \(m-1\), where \(m\) counts usable positive predictors—not automatically every row shown in the triangle.
Worked Example: A Zero Increment and a Negative Increment
For one development age, a hypothetical triangle has the following three transitions, in dollars: \(100\) to \(120\), \(200\) to \(200\), and \(100\) to \(80\). The second increment is zero and the third is negative. Calculate the Mack factor and variance estimate. Use the standard Mack formulas and identify the mistake of excluding the negative increment automatically.
Method. All three starting cumulative amounts are positive, so all three observations can be included. The increments are \(120-100=20\), \(200-200=0\), and \(80-100=-20\).
Factor. The total next cumulative amount is \(120+200+80=400\), and the total starting amount is \(100+200+100=400\). Therefore,
Residuals. The fitted next cumulative amounts are \(1(100)=100\), \(1(200)=200\), and \(1(100)=100\). Thus the residuals are \(20\), \(0\), and \(-20\).
Variance. The scaled squared residual terms are \(20^2/100=4\), \(0^2/200=0\), and \((-20)^2/100=4\). There are \(m=3\) usable observations, so
Check. The estimate is nonnegative, as a variance estimate must be. The zero increment contributes a zero residual term, while the negative increment contributes a positive squared-residual term. Excluding that row solely because its increment is negative would discard a valid transition under the stated predictor condition and change the estimate. The common mistake is to confuse a negative increment with a negative starting cumulative amount.
Zero Starting Cumulative Amounts
If \(C_{i,j}=0\), the ratio \(C_{i,j+1}/C_{i,j}\) is undefined, and the variance contribution divides by zero. If the next cumulative amount is also zero, that transition contributes no information about the factor in the weighted-average formula: it adds zero to both totals. It should not be counted as an ordinary usable observation when applying the \(m-1\) divisor. If the next cumulative amount is nonzero, the transition conflicts with the Mack conditional mean and variance at a zero predictor: the model specifies both a zero mean and zero variance. Investigate the data and model rather than forcing the cell into the standard calculation.
Worked Example: A Zero Predictor and the Degrees of Freedom
A hypothetical development age has transitions \(0\) to \(0\), \(100\) to \(120\), and \(200\) to \(200\), in dollars. Calculate the factor and Mack variance estimate using the usable observations. Then show what goes wrong if the zero-predictor row is counted in the degrees of freedom.
Factor. The zero-to-zero row adds nothing to either total and has an undefined ratio. The two positive-predictor observations give
Residuals and variance. The fitted next amounts for the usable rows are \(100(1.0666667)=106.6667\) and \(200(1.0666667)=213.3333\). The residuals are \(13.3333\) and \(-13.3333\). Their scaled squared terms are \(13.3333^2/100=1.7778\) and \(13.3333^2/200=0.8889\), with unrounded values summing to \(2.6667\). There are two usable observations, so
Check and trap. The estimated process variance for a later transition from a cumulative amount of \(250\) would be \(2.6667(250)=666.6667\) dollars squared, with standard deviation \(\sqrt{666.6667}=\$25.82\). If the zero-predictor row were incorrectly counted, the divisor would be \(3-1=2\) and the reported variance would be \(1.3333\), half the correct estimate. The zero row is not an extra residual observation for this variance calculation.
Negative Cumulative Predictors and Invalid Variances
A negative increment is compatible with a positive starting cumulative amount. For example, a transition from \(100\) to \(80\) has a negative increment, but its starting predictor is positive and the ratio is defined. A negative starting cumulative amount is different: the Mack model would assign conditional variance \(\sigma_j^2 C_{i,j}<0\), which is impossible for a variance. The standard variance estimator also divides a squared residual by that negative amount, potentially producing a negative “variance estimate.”
Worked Example: A Negative Cumulative Predictor
A hypothetical development age has transitions \(100\) to \(120\), \(50\) to \(45\), and \(-20\) to \(-10\), in dollars. Show why a mechanical Mack variance calculation is invalid, even though an algebraic factor can be calculated.
Algebraic factor. The totals give \(\widehat f_j=(120+45-10)/(100+50-20)=155/130=31/26\), approximately \(1.1923\). This arithmetic does not establish that the Mack assumptions are appropriate.
Variance terms. The residuals are \(120-100(31/26)=10/13\), \(45-50(31/26)=-190/13\), and \(-10-(-20)(31/26)=180/13\). The scaled squared terms are
Their sum is \(-897/169\), so dividing by \(3-1\) gives \(-897/338\), approximately \(-2.6538\). That cannot be a variance. The negative contribution arises from the negative predictor, not from the fact that one increment is negative.
Defensible treatment. Do not report this as a Mack variance estimate or repair it by taking an absolute value. Check the cumulative data and the treatment of recoveries or adjustments; document the issue and do not apply the standard Mack calculation to this development age unless the data and model are made appropriate. A negative cumulative amount may also become a predictor at the next development age, where the same problem recurs.
Residual Diagnostics and Logarithm Traps
For positive starting cumulative claims, a zero increment gives \(C_{i,j+1}/C_{i,j}=1\), so the logarithm of the cumulative development ratio is \(\log(1)=0\). A negative increment can also leave the next cumulative amount positive, in which case that cumulative ratio remains positive and its logarithm is defined. But taking the logarithm of the increment itself is invalid for a zero or negative increment. Likewise, a nonpositive next cumulative amount makes the cumulative ratio nonpositive, so its real logarithm is undefined.
Do not manufacture a log-based residual by adding a small positive constant, taking an absolute value, or dropping cells without a stated rationale. Such changes alter the data or diagnostic. Instead, inspect the original amounts, calculate Mack residuals only where their denominators are valid, and report excluded or incompatible observations explicitly. If \(\widehat{\sigma}_j^2=0\), standardized residuals at that development age are also undefined because their denominator is zero; do not label them zero.
| Observed condition | Standard Mack calculation | Diagnostic implication |
|---|---|---|
| \(C_{i,j}>0\), increment is zero | Defined; the ratio is 1 and the residual term may be zero | Do not discard merely because the increment is zero |
| \(C_{i,j}>0\), increment is negative, next cumulative remains positive | Defined for the transition | Use the residual formula; do not take a logarithm of the increment |
| \(C_{i,j}=0\) | Ratio and variance contribution are undefined | Do not count as an ordinary usable observation |
| \(C_{i,j}<0\) | Conditional variance and scaled residual term are invalid | Do not report a standard Mack variance estimate from that row |
A Practical Decision Sequence
When an unusual increment appears, work from the cumulative data and the model assumptions rather than deciding from the increment’s sign alone.
Record the starting amount \(C_{i,j}\), the next amount \(C_{i,j+1}\), and their difference. Confirm whether the unusual sign belongs to the increment or to a cumulative predictor.
If \(C_{i,j}>0\), a zero or negative increment can still be used in the factor and residual calculations. If \(C_{i,j}=0\) or \(C_{i,j}<0\), the standard variance term is not usable.
Use only valid observations, apply the \(m-1\) divisor to their count, and verify that the resulting variance estimate is nonnegative. A negative result signals an invalid application or data/model problem.
Standardize residuals only with positive estimated variance and positive starting cumulative claims. Do not take logarithms of zero or negative amounts, and state any limitation rather than masking it.
Check Your Understanding
For each problem, decide whether the stated Mack calculation is valid and show the relevant factor, variance, or diagnostic conclusion.
- At one development age, the transitions are \(50\) to \(60\) and \(150\) to \(150\). Calculate \(\widehat f_j\) and \(\widehat{\sigma}_j^2\). (Answer: \(\widehat f_j=1.05\), \(\widehat{\sigma}_j^2=1.5\).)
- A transition is \(80\) to \(80\), so its increment is zero. Is the logarithm of the cumulative development ratio defined, and what is its value? Is the logarithm of the increment defined? (Answer: The ratio is \(1\), so its logarithm is \(0\); the logarithm of the zero increment is undefined.)
- At a development age, three transitions are \(0\) to \(0\), \(100\) to \(110\), and \(200\) to \(190\). Using only usable predictors, calculate \(\widehat f_j\). (Answer: \(\widehat f_j=(110+190)/(100+200)=1.00\); the zero-to-zero row is not a usable ratio observation.)
- If \(\widehat{\sigma}_j^2=0.8\) and a future starting cumulative amount is \(400\), calculate the Mack conditional variance and standard deviation for the next cumulative amount. (Answer: Variance \(=0.8(400)=320\) dollars squared; standard deviation \(=\sqrt{320}=\$17.89\).)
- A transition has starting cumulative claims \(-25\) and next cumulative claims \(-20\). Should its Mack scaled squared residual term be included even though the observed increment is positive? (Answer: No. The starting cumulative predictor is negative, making the model’s conditional variance and the standard Mack scaled term invalid.)