Changing Development Patterns And Mack Assumption Failure

Why Changing Development Patterns Matter

Chain-ladder projections use observed development to estimate how claims will emerge in the future. Mack’s model adds a framework for estimating prediction error, but that framework depends on development behavior being sufficiently stable across accident years. If recent cohorts develop differently from older cohorts, a pooled factor may describe no cohort particularly well. The resulting reserve can be biased, and the usual Mack standard error need not reflect the full uncertainty.

The diagnostic question is not simply whether accident-year factors differ. Some variation is expected. The question is whether the differences look like random scatter around one common development pattern or a persistent change associated with accident-year order. A shift is a warning that the model may be inappropriate; a small triangle cannot by itself establish why the pattern changed.

Definition: Let $C_{i,j}$ be cumulative claims for accident year $i$ at development age $j$, and let $R_{i,j}=C_{i,j+1}/C_{i,j}$ be the observed development ratio when $C_{i,j}>0$. Mack’s model assumes, for each development age $j$, that the conditional mean and variance are $$ E[C_{i,j+1}\mid\text{claims observed through age }j]=f_jC_{i,j},\qquad \operatorname{Var}(C_{i,j+1}\mid\text{claims observed through age }j)=\sigma_j^2C_{i,j}. $$ The factors $f_j$ and variance parameters $\sigma_j^2$ are common to accident years at that development age. The model also assumes independence of the relevant development outcomes across accident years. It specifies conditional moments; it does not require a particular probability distribution.

A changing development pattern violates the common-factor assumption when the conditional mean factor at a given age depends systematically on accident year. Changes in claim handling, coverage, settlement speed, or data practices are possible hypothetical explanations, but a triangle alone does not identify the cause. Calendar-year effects are another possible source of pattern; the key point here is that an apparent change by accident year makes a single common development factor questionable.

Factors, Residuals, and What They Can Show

For one development age, the chain-ladder factor is the exposure-weighted ratio of total claims at the next age to total claims at the current age. The Mack variance estimator measures weighted residual dispersion around that fitted factor. With $n$ usable accident years, the standard plug-in estimators are:

\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}}, \qquad \hat\sigma_j^2=\frac{1}{n-1}\sum_i C_{i,j}(R_{i,j}-\hat f_j)^2.

The residual $R_{i,j}-\hat f_j$ is a simple way to inspect whether individual accident years sit above or below the selected factor. A plot or table ordered by accident year is more informative than an unordered list: several older residuals on one side of zero followed by several newer residuals on the other may signal a changing pattern. Residuals are diagnostic evidence, not proof. With few accident years, even a striking-looking sequence can occur by chance.

Calculate the ratios by accident year.
For each usable cell, divide the cumulative amount at age $j+1$ by the cumulative amount at age $j$. Keep the accident-year order visible.

Calculate the pooled factor.
Use the ratio of totals, not an unweighted average unless all starting cumulative amounts are equal.

Inspect residuals for structure.
Compare their signs and sizes across accident years. Persistent runs or a change in level deserve investigation.

Assess the projection, not only the fit.
Ask whether the pooled factor is a reasonable estimate for the accident years still to develop. A good historical average need not represent a changed recent regime.

Why the Mack Estimators Have These Forms

The ratio-of-totals factor follows directly from the conditional mean assumption. Given the current cumulative amounts, each next-age value has conditional mean $f_jC_{i,j}$. Therefore the conditional mean of the sum of next-age values is $f_j\sum_i C_{i,j}$; dividing by the observed sum of current amounts gives the pooled factor as a natural estimate of $f_j$.

The variance estimator also reflects the model’s weighting. Since $R_{i,j}=C_{i,j+1}/C_{i,j}$, the conditional variance of a ratio is $\sigma_j^2/C_{i,j}$. Thus a squared ratio residual is weighted by $C_{i,j}$, giving equal model-scale variance after weighting. Fitting the common factor uses one degree of freedom, so the residual sum is divided by $n-1$, not $n$. Under the common-factor and independence assumptions, the conditional variance of the pooled factor estimate is correspondingly estimated by:

\widehat{\operatorname{Var}}(\hat f_j)=\frac{\hat\sigma_j^2}{\sum_i C_{i,j}}.

These justifications rely on the assumed common factor and variance. When factors change across accident years, residuals contain both ordinary variation and systematic differences between cohorts. The pooled estimate can still be calculated, but its interpretation as an estimate of one common factor is no longer warranted.

Worked Example: A Late Accident Year Sits Above the Pooled Factor

At one development transition, three hypothetical accident years have cumulative amounts of 50, 50, and 50 thousand dollars. Their next-age amounts are 60, 60, and 70 thousand dollars. Calculate the pooled factor and Mack variance estimate, then interpret the pattern.

Method. Calculate the three ratios, fit the exposure-weighted factor, and calculate the weighted residual sum of squares. This checks the common mistake of dividing by $n$ rather than the residual degrees of freedom $n-1$.

The ratios are $60/50=1.20$, $60/50=1.20$, and $70/50=1.40$. The pooled factor is:

\hat f=\frac{60+60+70}{50+50+50} =\frac{190}{150} =\frac{19}{15} =1.2666667.

The residuals are $1.20-19/15=-1/15$, $-1/15$, and $1.40-19/15=2/15$. The weighted squared-residual sum is:

50\left[\left(\frac{-1}{15}\right)^2+ \left(\frac{-1}{15}\right)^2+ \left(\frac{2}{15}\right)^2\right] =50\left(\frac{6}{225}\right) =\frac{4}{3}.

There are $n=3$ observations, so:

\hat\sigma^2=\frac{4/3}{3-1}=\frac{2}{3}=0.6667, \qquad \widehat{\operatorname{Var}}(\hat f) =\frac{2/3}{150}=\frac{1}{225}=0.0044444.

The pooled factor is $1.2667$, and the variance estimate is $0.6667$. The two older ratios are both $1.20$, while the newest is $1.40$. That ordering is a warning worth investigating, though three observations alone cannot establish a persistent shift. Common mistake avoided: dividing the weighted squared-residual sum by 3 would give $4/9$; fitting the factor has used one degree of freedom, so the specified estimator divides by $3-1=2$.

A Shift Can Leave Bias Even When Residuals Are Included

A shift may increase the estimated residual dispersion, but that does not guarantee that the Mack prediction error captures the error in projecting a changed cohort. The pooled factor averages the historical experience according to its starting cumulative amounts. If the future accident year belongs to a different regime, its expected factor can be systematically above or below that average. The resulting forecast error has a bias component, which is not the same as random variation around the pooled model.

Worked Example: A Recent Regime Changes a One-Step Reserve Projection

Suppose six hypothetical accident years have a cumulative amount of 100 thousand dollars at a particular age. Three older years each develop by a factor of $1.20$, and three more recent years each develop by $1.40$. A current accident year has 200 thousand dollars at that age. For illustration, suppose its expected factor under the recent regime is $1.40$. Calculate the pooled Mack factor, its variance estimate, the conventional one-step MSEP, and the difference between the pooled projection and the recent-regime expected value.

Method. Fit the pooled factor and Mack variance using all six observations, then apply the one-step Mack plug-in formula. The calculation illustrates why a model-based standard error should not be read as a guarantee against structural bias.

The historical next-age amounts total $3(100)(1.20)+3(100)(1.40)=360+420=780$ thousand dollars. The current amounts total $600$ thousand dollars. Hence:

\hat f=\frac{780}{600}=1.30.

Each ratio differs from $1.30$ by $0.10$ in magnitude. The weighted squared-residual sum is $600(0.10)^2=6$, so:

\hat\sigma^2=\frac{6}{6-1}=1.20, \qquad \widehat{\operatorname{Var}}(\hat f)=\frac{1.20}{600}=0.002.

For the current amount $C=200$ thousand dollars, the pooled projection is $200(1.30)=260$ thousand dollars. The conventional one-step Mack plug-in MSEP has process and parameter components:

\widehat{\operatorname{MSEP}} =\hat\sigma^2 C+C^2\widehat{\operatorname{Var}}(\hat f) =(1.20)(200)+(200)^2(0.002) =240+80=320

The corresponding estimated standard error is $\sqrt{320}=17.8885$ thousand dollars, or 17.89 thousand dollars to two decimal places. Under the stated recent-regime expectation, the next-age expected amount is $200(1.40)=280$ thousand dollars, so the pooled projection is 20 thousand dollars lower. The conventional MSEP is calculated under the pooled Mack model; the difference of 20 thousand dollars is a systematic discrepancy from using the pooled factor for a changed regime. The calculation does not make the pooled MSEP a complete measure of uncertainty under that changed regime.

Changes Across Multiple Development Ages

A cohort shift can affect more than one development transition. Since chain-ladder projections multiply factors, differences at several ages can compound. For this reason, an overall ultimate comparison should consider the sequence of age-specific factors, not just one isolated transition.

Worked Example: Pooled Factors Compound a Recent Pattern Change

A hypothetical triangle has two older accident years with cumulative amounts $100,120,132$ thousand dollars at ages 1, 2, and 3, respectively. Two more recent years each have cumulative amounts $100,140,168$ thousand dollars at those ages. A current accident year has 200 thousand dollars at age 1. Compare its ultimate projection using pooled factors with a projection using the recent cohort’s factors.

Method. Calculate each factor as a ratio of totals. Then multiply the selected factors through the remaining development ages and compare the results. This avoids treating age-specific factors as additive.

At ages 1 to 2, the total at age 1 is $4(100)=400$ thousand dollars. The total at age 2 is $2(120)+2(140)=520$ thousand dollars, giving:

\hat f_1=\frac{520}{400}=1.30.

At ages 2 to 3, the total at age 2 is 520 thousand dollars and the total at age 3 is $2(132)+2(168)=264+336=600$ thousand dollars. Therefore:

\hat f_2=\frac{600}{520}=\frac{15}{13}=1.1538462.

The pooled cumulative factor is $1.30(15/13)=1.50$. The pooled ultimate projection for the current accident year is $200(1.50)=300$ thousand dollars, implying a reserve of $300-200=100$ thousand dollars. In the recent cohort, the observed age-specific factors are $140/100=1.40$ and $168/140=1.20$. Applying those factors gives an illustrative recent-pattern projection of $200(1.40)(1.20)=336$ thousand dollars and a reserve of $336-200=136$ thousand dollars. The difference in ultimate projections is $336-300=36$ thousand dollars. This is a sensitivity comparison, not proof that the recent factors will persist; its purpose is to show how a change at multiple ages can alter the projection materially.

Common Mistakes and Practical Interpretation

Treating the pooled factor as automatically appropriate for the latest year. It is a weighted historical average. If development behavior changed, that average may not represent the cohort being projected.
Using the wrong weights. The chain-ladder factor is a ratio of totals, equivalently a weighted average of ratios with weights $C_{i,j}$. An unweighted average is generally different when starting cumulative amounts differ.
Dividing the variance estimate by $n$. The common factor is estimated from the same observations, so the stated Mack estimator uses $n-1$ degrees of freedom.
Assuming that a larger residual variance solves model failure. A shift may add dispersion around the pooled factor, but the residual calculation does not automatically correct the projected mean for the newest cohort.
Declaring a trend from a small sample. Ordered residuals or factors can flag instability; they do not establish a cause or prove that a different factor will continue.

A practical conclusion should separate what the triangle shows from what is assumed about future development. Report the pooled estimate and its Mack uncertainty as conditional on the model, inspect factors and residuals in accident-year order, and test how the reserve changes under plausible recent-pattern selections. A systematic pattern calls for explicit judgment about model adequacy; it should not be hidden by presenting a standard error alone.

Key takeaway: Mack’s standard errors rely on common development behavior across accident years. A pattern of factors that shifts with accident-year order can make a pooled projection biased for recent cohorts. Residual dispersion may reveal the shift, but it does not guarantee that the reported MSEP includes the resulting model risk.

Check Your Understanding

For each problem, calculate the requested quantities before checking the answer.

At one development age, two hypothetical accident years have starting cumulative amounts of 80 and 120 thousand dollars and ratios of 1.10 and 1.30. Calculate the ratio-of-totals factor. (Answer: $1.2200$.)
Three accident years have starting amounts $50,100,150$ thousand dollars and ratios $1.20,1.30,1.40$. Calculate the pooled factor and Mack variance estimate using $n-1$ in the denominator. (Answer: $\hat f=1.3333$; $\hat\sigma^2=0.8333$.)
Two older years each have starting amount 100 and ratio 1.15; two recent years each have starting amount 200 and ratio 1.35, with amounts in thousands. Calculate the pooled factor and Mack variance estimate. (Answer: $\hat f=1.2833$; $\hat\sigma^2=1.7778$.)
Historical data give $\hat f=1.25$, $\hat\sigma^2=1.5$, and total starting exposure 500, all in thousands where applicable. A current amount is 300, and a recent-regime expected factor is 1.35. Calculate the pooled projection, the recent-regime expected amount, and the conventional one-step Mack MSEP and standard error. (Answer: pooled projection $375$; recent-regime expectation $405$; MSEP $720$ thousand-dollars-squared; standard error $26.8328$ thousand dollars.)
A current accident year has 250 thousand dollars at age 1. Pooled factors for ages 1 to 2 and 2 to 3 are 1.28 and 1.15; plausible recent-regime factors are 1.35 and 1.20. Calculate both ultimate projections and their difference. (Answer: pooled ultimate $368$ thousand dollars; recent-pattern ultimate $405$ thousand dollars; difference $37$ thousand dollars.)

What You'll Learn