Outlier Cells And Their Influence On Reserve Variance

Why an Unusual Cell Deserves a Sensitivity Check

A single large or small claims-triangle cell can affect more than the development factor selected from its age-to-age data. It can also change the residuals used to estimate process variance, and those changes feed into the projected reserve uncertainty. The key diagnostic question is therefore not simply whether a cell looks unusual, but how much the conclusions depend on it and whether there is a defensible reason to exclude or adjust it.

The calculations below use a simplified Mack chain-ladder setting. They compare results with all observations included against a sensitivity calculation that omits one cell. That second calculation is a diagnostic, not an automatic recommendation to delete the cell.

Model and notation: Let $C_{i,j}$ be cumulative claims for origin year $i$ at development age $j$, and let $R_{i,j}=C_{i,j+1}/C_{i,j}$ be the observed link ratio when $C_{i,j}>0$. At a fixed development age $j$, the Mack assumptions used here are $E[C_{i,j+1}\mid C_{i,j}]=f_jC_{i,j}$ and $\operatorname{Var}(C_{i,j+1}\mid C_{i,j})=\sigma_j^2C_{i,j}$. The common factor $f_j$ and variance parameter $\sigma_j^2$ apply across the origin years being compared; the observations used in the variance estimate are treated as independent across origin years. The formulas below require positive cumulative exposures and at least two observed link ratios for the stated variance estimator.

How a Cell Influences a Development Factor

At one development age, the chain-ladder factor is a cumulative-claims-weighted average of the observed link ratios. If the available origin years have exposures $C_{i,j}$, then

\hat f_j=\frac{\sum_i C_{i,j+1}}{\sum_i C_{i,j}} =\frac{\sum_i C_{i,j}R_{i,j}}{\sum_i C_{i,j}}.

Thus a cell’s effect depends on both its link ratio and its exposure. A striking ratio on a small exposure may have less effect on the factor than a less striking ratio on a large exposure. A useful first sensitivity measure is the change in the selected factor when the cell is omitted, reported alongside the original factor rather than in place of it.

The weighted-average identity explains the direction of the change. If a cell with exposure $w$ and ratio $r$ is added to other observations with total exposure $W$ and weighted factor $f$, the new factor is

f_{\mathrm{new}}=\frac{Wf+wr}{W+w}, \qquad f_{\mathrm{new}}-f=\frac{w}{W+w}(r-f).

The added cell moves the factor toward its own ratio. Its influence grows with its exposure relative to the existing total and with the distance between its ratio and the existing factor. This is a direct consequence of the weighted-average formula, not a separate adjustment rule.

Worked Example: Measuring a Cell’s Effect on a Selected Factor

A hypothetical triangle has three observations at one development age. Two ordinary cells have exposures of $60 thousand each and link ratios $17/15$ and $16/15$. A third cell has exposure of $480 thousand and a link ratio of $1.10$. A separate cell, flagged for investigation, has exposure of $400 thousand and a ratio of $1.50$. Calculate the factor with and without the flagged cell.

Method. Calculate each factor as total next-age cumulative claims divided by total current-age cumulative claims. The stated exposures and ratios are in thousands of dollars.

Without the flagged cell. The next-age amounts for the three retained observations are $60(17/15)=68$, $60(16/15)=64$, and $480(1.10)=528$, all in thousands. The total next-age amount is $68+64+528=660$; the total exposure is $60+60+480=600$. Therefore

\hat f_{\text{omit}}=\frac{660}{600}=1.10.

With the flagged cell. Its next-age amount is $400(1.50)=600$. The combined next-age amount is $660+600=1{,}260$, and the combined exposure is $600+400=1{,}000$. Hence

\hat f_{\text{all}}=\frac{1{,}260}{1{,}000}=1.26.

The flagged cell raises the factor by $1.26-1.10=0.16$, or about $14.5\%$ of the omit-cell factor. This agrees with the weighted-average check: $\frac{400}{1{,}000}(1.50-1.10)=0.16$. The arithmetic confirms the direction and size of the shift; it does not establish whether the observation should be excluded.

How the Cell Affects Process-Variance Estimation

Under the Mack assumptions, an observed link ratio should fluctuate around the common factor. A commonly used estimator of the development-period variance parameter, when $n_j>1$, is the exposure-weighted residual sum of squares divided by $n_j-1$:

\hat{\sigma}_j^2 =\frac{1}{n_j-1} \sum_i C_{i,j}(R_{i,j}-\hat f_j)^2.

Here $n_j$ is the number of observed link ratios used in the calculation. Each residual is measured from the factor estimated from that same set of observations. Consequently, omitting a cell changes both the factor at the centre of the residuals and the residual sum of squares; it also changes $n_j$ and the degrees-of-freedom denominator. Reusing the full-data factor or denominator after omission is not a leave-one-out variance estimate.

In the previous example, the three retained ratios have factor $1.10$. Their residuals are $1/30$, $-1/30$, and $0$. Their weighted squared residual sum is

60(1/30)^2+60(-1/30)^2+480(0)^2 =\frac{2}{15}=0.133333\ldots.

There are three retained ratios, so the estimated variance is $(2/15)/(3-1)=1/15$. Including the flagged cell gives a different result; that calculation is shown in the next example.

Worked Example: Comparing Process-Variance Estimates

Use the four cells from the factor example. Calculate the weighted residual sum of squares and the Mack variance estimate with all four cells, then compare it with the estimate from the three-cell sensitivity calculation.

All cells included. The factor is $1.26$, as calculated above. The four weighted squared residuals are:

For exposure $60$ and ratio $17/15$, the residual is $17/15-1.26=-19/150$, giving $60(19/150)^2=0.962667$.
For exposure $60$ and ratio $16/15$, the residual is $16/15-1.26=-29/150$, giving $60(29/150)^2=2.242667$.
For exposure $480$ and ratio $1.10$, the residual is $-0.16$, giving $480(0.16)^2=12.288$.
For exposure $400$ and ratio $1.50$, the residual is $0.24$, giving $400(0.24)^2=23.040$.

Their sum is $0.962667+2.242667+12.288+23.040=38.533334$, with the displayed values rounded. With four ratios, the denominator is $4-1=3$, so

\hat{\sigma}_{\text{all}}^2 =\frac{38.533333\ldots}{3} =12.844444\ldots.

Flagged cell omitted. The weighted residual sum is $2/15=0.133333\ldots$ around the refitted factor $1.10$. There are three ratios, so the denominator is $3-1=2$:

\hat{\sigma}_{\text{omit}}^2 =\frac{0.133333\ldots}{2} =\frac{1}{15}=0.066666\ldots.

The full-data estimate is much larger in this example, because the flagged cell has both a large exposure and a substantial residual from the full-data factor. These estimates are in the scale implied by the stated thousands-of-dollars exposures and the Mack variance model. The comparison indicates sensitivity; it does not prove that the high estimate is inappropriate.

Carrying the Sensitivity Through to Reserve Uncertainty

For a one-step projection from an observed cumulative amount $C$, the conditional Mack prediction variance has a process component and a factor-parameter component. With the factor estimated from historical exposures whose total is $W=\sum_i C_{i,j}$, the one-step approximation used here is

\widehat{\operatorname{MSEP}} =\underbrace{\hat{\sigma}_j^2 C}_{\text{process}} +\underbrace{C^2\frac{\hat{\sigma}_j^2}{W}}_{\text{factor parameter}}.

The first term represents variation in the future claim amount around its conditional mean. The second represents uncertainty in the estimated development factor, using the approximate variance $\hat{\sigma}_j^2/W$ of that factor. This is a one-step result under the stated Mack model; it is not a general formula for every reserving model or a substitute for the multi-step recursion when more than one development period remains.

Worked Example: Correcting the Exposure in Reserve MSEP

A current cumulative claim amount is $C=500$ thousand dollars. Use the leave-one-out factor $1.10$, variance estimate $1/15$, and remaining historical exposure $W=600$ thousand dollars from the earlier calculations. Calculate the projected reserve, the two MSEP components, and the standard error. Report amounts in thousands and round the standard error to three decimals.

Projected amount and reserve. The projected cumulative amount is $500(1.10)=550$. The estimated unpaid reserve is $550-500=50$ thousand dollars.

Process component. Using the observed current amount $C=500$,

\frac{1}{15}(500)=33.333333\ldots

in squared-thousands units.

Factor-parameter component. The factor variance estimate is $(1/15)/600=1/9{,}000$. Thus

500^2\left(\frac{1/15}{600}\right) =250{,}000\left(\frac{1}{9{,}000}\right) =27.777777\ldots.

The total MSEP is $33.333333\ldots+27.777777\ldots=61.111111\ldots$ squared-thousands units. The standard error is $\sqrt{61.111111\ldots}=7.817$ thousand dollars, rounded to three decimals. The calculation is internally consistent: the parameter term uses the retained historical exposure of 600, not the full-data exposure of 1,000.

Common Trap: Omitting a Cell but Keeping Its Exposure

A leave-one-out calculation must refit the quantities affected by omission. A common error is to omit a cell from the factor and residual sum but retain its exposure in the factor-variance denominator. In the preceding example, that error would give

500^2\left(\frac{1/15}{1{,}000}\right)=16.666667,

instead of the correct parameter component

500^2\left(\frac{1/15}{600}\right)=27.777778.

The process component remains $500/15=33.333333$, so the incorrect total would be $50.000000$, with standard error $7.071$. The corrected total is $61.111111$, with standard error $7.817$. The denominator must be the exposure in the same data set used to estimate the factor and its variance.

Another trap is to treat a large residual as sufficient evidence for deletion. A cell may be unusual because of data error, a large claim, a settlement, a change in claim handling, or another feature of the observed experience. Those explanations have different implications. Check the underlying data and context, then present a transparent sensitivity comparison. If an adjustment or exclusion is justified, document the reason and show its effect; otherwise, retain the observation and acknowledge the resulting uncertainty.

A Practical Interpretation

An influential cell can shift the selected mean and the estimated uncertainty in different ways. In the worked triangle, including the flagged cell raised the factor from $1.10$ to $1.26$, and the estimated variance from $1/15$ to $12.844444\ldots$. These results are linked: the cell changes the factor, which changes the residuals used to estimate variance. A sensitivity analysis should therefore recalculate the factor, residual variance, relevant exposures, and projected uncertainty consistently.

The calculation does not decide which result is more credible. Consider the size and direction of the cell’s influence, its data quality, and whether there is a substantive explanation for its development. The earlier material on standardized chain-ladder residuals and calendar-year effects can help frame such investigation. A sensitivity result is most useful when reported as a comparison with a clear description of what was changed—not as a hidden replacement for the selected method.

Key takeaway: To assess an unusual cell, refit the weighted factor and the residual-based variance estimate with and without it, then carry each internally consistent set of estimates through the reserve calculation. Influence calls for investigation and disclosure; it does not, by itself, justify deletion.

Check Your Understanding

For each problem, calculate the requested sensitivity or variance quantity before checking the answer.

At one development age, two cells have exposures 100 and 300 and link ratios 1.05 and 1.15. Calculate the exposure-weighted factor. (Answer: 1.1250)
Two retained cells have exposures 200 and 200 with ratios 1.08 and 1.12. A flagged cell has exposure 100 and ratio 1.50. Find the factor with the flagged cell omitted and included. (Answer: omitted $1.1000$; included $1.1800$)
Three cells have exposures 100, 200, and 300, and link ratios 1.00, 1.10, and 1.20. Calculate the weighted factor and the variance estimate using the denominator $n-1$. (Answer: factor $17/15=1.1333$; variance estimate $5/3=1.6667$)
A one-step Mack projection has $C=800$, $\hat f=1.25$, $\hat\sigma^2=0.04$, and historical exposure $W=400$, with amounts in thousands. Calculate the reserve, process component, parameter component, total MSEP, and standard error. (Answer: reserve $160$ thousand; process $32$; parameter $64$; MSEP $96$ squared-thousands; standard error $9.798$ thousand)
For a one-step projection, $C=1{,}000$ and $\hat\sigma^2=0.09$. A leave-one-out calculation has historical exposure $500$, while the full-data exposure is $900$. Holding $C$ and $\hat\sigma^2$ fixed solely to isolate the denominator effect, calculate the process component and each parameter component and MSEP. (Answer: process $90$; leave-one-out parameter $180$, MSEP $270$; full-exposure parameter $100$, MSEP $190$)

What You'll Learn