Why Examine Residuals Across Development Ages?
A reserve uncertainty estimate depends on more than the selected development factors: it also depends on how well the model describes the variation in observed claims. Standardized residual plots make it possible to inspect that fit across development ages. A plot may reveal that residuals are more variable at one age than the model allows, that signs persist in a pattern, or that a particular origin year follows a different development path.
The goal is not to declare a model invalid from one unusual point. Residuals are noisy, especially when few origin years are available at a development age. Instead, look for repeated structure and ask whether it is consistent with the assumptions underlying the uncertainty calculation.
A common estimate of the age-specific variance parameter, when \(n_j=|I_j|>1\), is $$ \hat\sigma_j^2= \frac{1}{n_j-1} \sum_{i\in I_j} \frac{(C_{i,j+1}-\hat f_jC_{i,j})^2}{C_{i,j}}. $$ The denominator \(n_j-1\) reflects estimation of one development factor at that age. If only one transition is available, this estimator cannot be calculated. A standardized residual plot therefore uses only observed transitions and depends on the selected factor and variance estimates.
Why the Leverage Adjustment Appears
The adjustment is a consequence of fitting the development factor to the same observations whose residuals are being plotted. Under the Mack model, write \(C_{i,j+1}=f_jC_{i,j}+\varepsilon_{i,j}\), where the conditional mean of \(\varepsilon_{i,j}\) is zero and its conditional variance is \(\sigma_j^2C_{i,j}\). The weighted chain-ladder factor estimate is $$ \hat f_j=\frac{\sum_{i\in I_j}C_{i,j+1}}{\sum_{i\in I_j}C_{i,j}}. $$
For a fixed age, abbreviate \(x_i=C_{i,j}\) and \(T=\sum_i x_i\). Then $$ \hat f_j-f_j=\frac{\sum_i\varepsilon_{i,j}}{T}, \qquad e_{i,j}=\varepsilon_{i,j}-\frac{x_i}{T}\sum_k\varepsilon_{k,j}. $$ Assuming errors from different origin years are uncorrelated at that age, the variance of this fitted residual is $$ \begin{aligned} \operatorname{Var}(e_{i,j}\mid C_{i,j}) &=\sigma_j^2x_i-2\frac{x_i}{T}\sigma_j^2x_i +\frac{x_i^2}{T^2}\sigma_j^2T\\ &=\sigma_j^2x_i\left(1-\frac{x_i}{T}\right). \end{aligned} $$ Thus the residual standard deviation is \(\sigma_j\sqrt{C_{i,j}(1-h_{i,j})}\), which motivates the leverage-adjusted denominator. This result relies on the stated variance model and uncorrelated errors across origin years at that age. It does not establish that residuals from different development ages are independent.
Worked Example: Calculating a Leverage-Adjusted Residual
In a hypothetical triangle, a cell has \(C_{i,j}=64\), the sum of prior cumulative amounts at age \(j\) is \(256\), and the raw residual is \(2.4\). The estimated process standard deviation parameter is \(\hat\sigma_j=0.15\). Calculate the standardized residual to four decimal places and state its direction.
Method. Calculate the leverage, then the fitted-residual standard deviation, and divide the raw residual by that standard deviation.
The leverage is $$ h_{i,j}=\frac{64}{256}=0.25. $$
The leverage-adjusted standard deviation is $$ 0.15\sqrt{64(1-0.25)} =0.15\sqrt{48} =0.15(6.928203) =1.039230. $$
Therefore, $$ r_{i,j}=\frac{2.4}{1.039230}=2.309401\ldots\approx 2.3094. $$ The residual is positive, so observed cumulative claims at the next age exceed the fitted amount for this cell.
Check and common mistake. The adjustment reduces the residual standard deviation below \(0.15\sqrt{64}=1.2\), because fitting the factor makes a cell’s fitted residual less variable than an unfitted error. Omitting leverage would give \(2.4/1.2=2.0000\), not the correct \(2.3094\). The corrected value is consistent with dividing by the smaller fitted-residual standard deviation.
Reading Spread and Location Across Ages
A plot of \(r_{i,j}\) against development age should be read in two ways. First, compare the spread at different ages. The Mack model already allows \(\sigma_j\) to vary by age, so raw residuals need not have equal spread. After standardization, however, the spread should be broadly comparable across ages if the age-specific variance estimates are appropriate. Second, inspect location and shape: repeated signs, a trend, or a cluster can signal structure that a single age-specific variance parameter does not capture.
A simple descriptive check is the root mean square (RMS) of the standardized residuals at an age: $$ \operatorname{RMS}_j=\sqrt{\frac{1}{n_j}\sum_{i\in I_j}r_{i,j}^2}. $$ An RMS near one is a rough calibration reference, not a formal pass/fail rule. Small samples can produce substantial variation, and an RMS can be large because of one extreme residual rather than broad overdispersion.
Worked Example: Comparing Standardized Spread by Age
A hypothetical insurer plots four standardized residuals at an early development age, \((-1,1,-1,1)\), and four at a late age, \((-2,2,-2,2)\). Compare their means and RMS values, then interpret the result for a Mack uncertainty assessment.
Early age. The sum is \(-1+1-1+1=0\), so the mean is \(0/4=0\). The sum of squares is \(1+1+1+1=4\), giving $$ \operatorname{RMS}_{\text{early}}=\sqrt{\frac{4}{4}}=1. $$
Late age. The sum is \(-2+2-2+2=0\), so the mean is also zero. The sum of squares is \(4+4+4+4=16\), giving $$ \operatorname{RMS}_{\text{late}}=\sqrt{\frac{16}{4}}=2. $$
The late-age RMS is twice the early-age RMS; its mean squared standardized residual is \(2^2=4\) times as large. This pattern is a warning that variation at the late age may exceed the model’s calibration. With only four residuals per age, it is not conclusive evidence that the late-age variance parameter is wrong. Check the underlying cells and whether one small group is driving the pattern before changing an uncertainty estimate.
Common mistake. Comparing unstandardized residual spreads across ages would ignore the Mack model’s age-specific scale. The comparison here is meaningful because the values have already been standardized using their age-specific estimates.
Runs and Development-Pattern Departures
A residual plot should preserve enough structure to show which observations belong to the same origin year. Connecting residuals for one origin year across successive development ages can reveal a pattern hidden by age-by-age summaries. For example, a sequence that moves from negative to positive means that observed development is below the fitted amount at earlier transitions and above it at later transitions. Repeated sequences of that kind suggest a development-shape departure, rather than merely a single unusual cell.
The following hypothetical standardized residuals illustrate why both views matter. Each column represents one development age, and each row follows an origin year across ages.
| Origin year | Age 1 | Age 2 | Age 3 | Age 4 |
|---|---|---|---|---|
| A | -1.2 | -0.8 | 0.8 | 1.2 |
| B | 1.2 | 0.8 | -0.8 | -1.2 |
| C | -0.6 | 0.4 | 0.6 | 0.4 |
| D | 0.6 | -0.4 | -0.6 | -0.4 |
Each column sums to zero, so its mean is zero. Yet origin year A moves from two negative residuals to two positive residuals. This is a directional change across development ages that an age-wise mean alone would miss. It is a diagnostic clue, not proof: with few observations, a sequence may occur by chance. If similar transitions recur across multiple origin years, investigate whether the development pattern is adequately represented before relying on its reserve uncertainty estimate.
Worked Example: Interpreting a Residual Sequence
For one hypothetical origin year, standardized residuals at four successive development ages are \((-1.4,-0.6,0.7,1.3)\). The corresponding fitted-residual standard deviations are \((2,2,2,2)\), and the raw residuals are therefore \((-2.8,-1.2,1.4,2.6)\). Interpret the sequence and calculate the raw-residual total.
Step 1: Check the signs. The first two standardized residuals are negative, so cumulative claims at those next ages were below the fitted amounts. The last two are positive, so they were above the fitted amounts.
Step 2: Check the raw values. The supplied values agree with the standardization: \(-2.8/2=-1.4\), \(-1.2/2=-0.6\), \(1.4/2=0.7\), and \(2.6/2=1.3\).
Step 3: Calculate the total. $$ -2.8-1.2+1.4+2.6=0. $$ A zero total does not remove the pattern: the first two transitions are below fit and the last two above fit. The sequence is consistent with a possible departure in the shape of development for this origin year, but one sequence alone is not enough to establish a model failure.
Common mistake. Summing residuals and finding zero does not mean the residuals contain no information. The ordering by development age matters; a plot can display a systematic sequence even when positive and negative errors cancel.
What Residual Patterns Mean for Uncertainty
The pattern matters because the reserve uncertainty calculation uses estimates of process variation by development age and assumes a development structure for the expected cumulative claims. When standardized residuals at one age are persistently more dispersed than expected, the corresponding variance estimate may not represent future variation well. When residuals show repeated directional sequences, the fitted factors may not capture the observed development pattern. In either case, the reported prediction error can give a misleading impression of precision.
Use a disciplined review rather than reacting to a single point:
Confirm that the factor, cell, residual sign, leverage, and age-specific variance estimate were used consistently.
Look for broad changes in standardized spread, and note how many observed transitions support each age.
Follow residuals by origin year across development ages; repeated runs or trends deserve more attention than isolated signs.
Treat a pattern as evidence to investigate the assumptions and uncertainty estimate, not as automatic proof that a model or reserve is wrong.
There are important limits. Standardization uses estimated quantities, so residuals do not have exactly known unit variance in a finite triangle. Late development ages often contain fewer observed transitions, making plots less stable. Also, a pattern involving calendar time, an influential cell, or dependence across cells calls for additional diagnostics; a plot against development age by itself does not identify the cause.
Exam trap: A large raw residual at a high-volume development cell is not automatically the most unusual one. Compare residuals on the model’s scale, including leverage, and then assess whether unusual values form a repeated pattern. Conversely, a plot with residuals centered near zero can still show age-specific overdispersion or a trend across development ages.
Check Your Understanding
For each problem, calculate or interpret the standardized residual information before drawing a conclusion about the uncertainty model.
- A cell has \(C_{i,j}=81\), the prior-amount sum at that age is \(324\), raw residual \(-3\), and \(\hat\sigma_j=0.20\). Calculate the leverage-adjusted standardized residual to four decimal places. (Answer: \(-1.9245\))
- At one age, standardized residuals are \((-1,1,-1,1)\); at another, they are \((-1.5,1.5,-1.5,1.5)\). Calculate both RMS values and the ratio of the second age’s mean squared residual to the first age’s. (Answer: RMS values \(1.0000\) and \(1.5000\); mean-square ratio \(2.25\))
- For one origin year, residuals across successive ages are \((-0.9,-0.2,0.6,1.1)\). State what the sign sequence suggests and one reason it is not conclusive by itself. (Answer: A shift from below-fit to above-fit development suggests a possible development-pattern departure; one origin-year sequence may be chance variation.)
- A cell has \(C_{i,j}=100\), the prior-amount sum is \(200\), raw residual \(4\), and \(\hat\sigma_j=0.10\). Calculate the standardized residual to four decimals and compare it with the value obtained by omitting leverage. (Answer: With leverage, \(5.6569\); without leverage, \(4.0000\).)
- Four standardized residuals at a late age are \((-2,0,1,1)\). Calculate their mean and RMS to four decimals, then state the diagnostic implication. (Answer: Mean \(0.0000\), RMS \(1.2247\); this is a modest spread above one and a small-sample warning to investigate, not proof of miscalibration.)