Comparing Mack And Overdispersed-Poisson Uncertainty

Why Compare the Two Approaches?

Chain ladder provides projected claims, but a reserve estimate also needs an uncertainty measure. The Mack model and the overdispersed-Poisson (ODP) model can produce the same chain-ladder projections while expressing variability differently. Comparing them helps distinguish variation in future claims from uncertainty in estimated development parameters, and reveals when shared parameters make an aggregate reserve less precise than a simple sum of individual standard errors suggests.

Both methods rely on a development triangle and assumptions about how claims develop. Their standard errors are not properties of the triangle alone: they depend on the model, variance estimation choices, and any adjustments made to the data. The comparison below focuses on the usual chain-ladder-compatible forms of the two models.

Definition: Write \(C_{i,k}\) for cumulative claims for origin period \(i\) through development age \(k\), and \(Y_{i,k}=C_{i,k}-C_{i,k-1}\) for the corresponding incremental claims (with \(Y_{i,1}=C_{i,1}\)). The Mack model specifies \(E[C_{i,k+1}\mid C_{i,1},\ldots,C_{i,k}]=f_k C_{i,k}\) and \(\operatorname{Var}(C_{i,k+1}\mid C_{i,1},\ldots,C_{i,k})=\sigma_k^2 C_{i,k}\). The ODP model specifies independent incremental observations with fitted means \(\mu_{i,k}=E[Y_{i,k}]\) and \(\operatorname{Var}(Y_{i,k})=\phi\mu_{i,k}\), where \(\phi\) is the overdispersion parameter.

What the Models Treat as Random

Under Mack, the conditional variance grows in proportion to the cumulative amount at the start of a development step. The \(f_k\) and \(\sigma_k^2\) values vary by development age. Origin periods are assumed independent in the model, but predictions for different origins can still be associated because they use the same estimated development factors.

Under ODP, the random observations are incremental amounts. A standard chain-ladder-compatible ODP model represents their means with origin-period and development-period effects, subject to an identifiability constraint. Its variance is proportional to each fitted incremental mean, with the common scale factor \(\phi\). The development pattern and origin effects are estimated from the triangle. Uncertainty in these fitted parameters contributes to reserve prediction error as well as the randomness of future increments.

A literal Poisson distribution permits only nonnegative integer observations. In reserving, ODP is commonly used as a quasi-likelihood mean-variance specification for nonnegative incremental amounts, which need not be integer counts. Neither formulation should be applied mechanically to negative increments: the variance and fitting assumptions need to be reconsidered or the data handled using an appropriate method.

Feature	Mack	Overdispersed Poisson
Random quantity	Next cumulative amount, conditional on current cumulative amount	Incremental amount
Variance structure	\(\sigma_k^2 C_{i,k}\)	\(\phi\mu_{i,k}\)
Development uncertainty	Development-age variance parameters and estimated factors	Dispersion and fitted model parameters
Common source of aggregate dependence	Shared estimated development factors	Shared fitted development parameters

Estimating Uncertainty From a Triangle

For a development age \(k\) with \(n_k\) observed origin periods, the usual chain-ladder factor estimate is the volume-weighted average of the observed age-to-age ratios. A common Mack estimate of the variance parameter is the weighted residual estimate shown below. This estimator requires at least two observed ratios at that age. Alternative estimators may be used, especially when only a few observations are available; the selected estimator should be stated because it can materially affect results.

\widehat f_k=\frac{\sum_i C_{i,k+1}}{\sum_i C_{i,k}}, \qquad \widehat{\sigma}_k^2= \frac{1}{n_k-1}\sum_i C_{i,k} \left(\frac{C_{i,k+1}}{C_{i,k}}-\widehat f_k\right)^2

For ODP, a common estimate of \(\phi\) is the Pearson chi-square statistic divided by residual degrees of freedom. Here \(\widehat\mu_{i,k}\) is the fitted mean for an observed incremental cell; the residual degrees of freedom are the number of fitted cells minus the number of independently estimated model parameters.

\widehat\phi= \frac{\sum_{(i,k)\ \mathrm{observed}} (Y_{i,k}-\widehat\mu_{i,k})^2/\widehat\mu_{i,k}} {\text{residual degrees of freedom}}

This Pearson estimate requires positive fitted means for the cells in the sum. A small residual degrees-of-freedom count makes it unstable. A different dispersion estimator, such as a deviance-based estimate, can give a different result, so comparisons should use a clearly stated and consistently applied convention.

Why the One-Step Variances Can Match

A simple one-step development model makes the relationship transparent. Suppose observed cumulative amounts \(T_i\) are known at the start of the step, and each future incremental amount has mean \(qT_i\), where \(q\) is the incremental-to-starting-cumulative ratio. Then the next cumulative amount is \((1+q)T_i\), so the age-to-age factor is \(f=1+q\). Assume the Mack variance model holds, the ratio is estimated from complete rows, and the ODP mean model uses the same exposures \(T_i\).

The estimated ratio is \(\widehat q=\sum_i Y_i/\sum_i T_i\), where \(Y_i\) is the observed increment. Under the Mack assumptions, the variance of the estimated age-to-age factor is approximately \(\sigma^2/\sum_iT_i\). Since \(q=f-1\), the estimated ratio has the same variance. The future process variance for a new row is \(\sigma^2T_{\mathrm{new}}\). Adding parameter uncertainty gives the following one-step mean squared prediction error (MSEP):

\operatorname{MSEP}_{\mathrm{Mack}} \approx \sigma^2T_{\mathrm{new}} +T_{\mathrm{new}}^2\frac{\sigma^2}{\sum_i T_i}

For the corresponding ODP model, \(\operatorname{Var}(Y_i)=\phi qT_i\). The ratio estimate has variance \(\phi q/\sum_i T_i\), so the new increment’s process variance and parameter variance give:

\operatorname{MSEP}_{\mathrm{ODP}} \approx \phi qT_{\mathrm{new}} +T_{\mathrm{new}}^2\frac{\phi q}{\sum_i T_i}

If the fitted models imply \(\sigma^2=\phi q\), these one-step MSEPs agree. This is a useful special-case check, not a claim that every Mack and ODP uncertainty calculation must agree. In a multi-age triangle, model fitting, variance estimation, tail treatment, and parameter covariance can all differ.

Worked Example: Comparing One-Step Prediction Errors

A hypothetical triangle has two complete origins with starting cumulative amounts \(T_1=100\) and \(T_2=200\), and observed next increments \(Y_1=20\) and \(Y_2=30\). A third origin has starting cumulative amount \(T_3=300\), with its next increment unobserved. Calculate the one-step Mack and ODP prediction errors using the weighted Mack variance estimate and Pearson ODP dispersion estimate. Use the Mack estimated variance of the fitted factor, \(\widehat\sigma^2/\sum_iT_i\), and the corresponding ODP ratio-estimation variance.

Method and projection. The fitted increment ratio is \(\widehat q=(20+30)/(100+200)=50/300=1/6\). The fitted age-to-age factor is \(1+\widehat q=7/6\). Therefore the third origin’s predicted increment is \(300(1/6)=50\).

Mack variance estimate. The observed age-to-age factors are \(120/100=1.2\) and \(230/200=1.15\), with weighted mean \(7/6\). The weighted squared residual sum is \(100(1.2-7/6)^2+200(1.15-7/6)^2 =100(1/30)^2+200(-1/60)^2 =1/9+1/18=1/6\). There are two ratios, so the denominator is \(2-1=1\), giving \(\widehat\sigma^2=1/6\).

Mack prediction error. The process component is \((1/6)(300)=50\). The factor-estimation variance is \((1/6)/(100+200)=1/1800\), so its contribution to the prediction variance is \(300^2(1/1800)=50\). Thus the MSEP is \(50+50=100\), and the prediction standard error is \(\sqrt{100}=10\).

ODP calculation. The fitted increments for the two complete origins are \(100/6=16.6667\) and \(200/6=33.3333\). The Pearson statistic is \((20-16.6667)^2/16.6667+(30-33.3333)^2/33.3333 =0.6667+0.3333=1.0000\). There are two fitted observations and one fitted ratio, giving one residual degree of freedom; hence \(\widehat\phi=1\). The future process variance is \(1(50)=50\). The estimated ratio variance is \(1(1/6)/300=1/1800\), contributing \(300^2/1800=50\). The ODP MSEP is also \(100\), and its standard error is \(10\).

Check and trap avoided. Both methods predict an increment of \(50\) and give a standard error of \(10\). The common trap is to treat the estimated mean \(50\) as the entire prediction variance. That omits the additional \(50\) from estimating the shared development ratio.

Aggregating Reserves: Shared Factors Matter

For multiple future origin periods, the aggregate reserve is the sum of their projected unpaid amounts. Its prediction variance is the sum of the individual variances plus twice the sum of pairwise covariances. If process errors are independent across origins conditional on the fitted development parameters, their process covariance is zero. But parameter uncertainty is shared: a high estimate of a development factor raises several origin-period projections together. Consequently, the reserve estimates can have positive covariance even when the underlying process errors are independent.

For the same one-step ratio setup, let \(T_a\) and \(T_b\) be starting cumulative amounts for two future origins. Their covariance due to the shared estimated ratio is approximately \(T_aT_b\operatorname{Var}(\widehat q)\). The aggregate MSEP therefore includes the cross-product term; summing the individual MSEPs alone omits it.

Worked Example: Aggregating Two Future Origins

Use the fitted variance information from the preceding example. Two hypothetical future origins have starting cumulative amounts \(T_a=300\) and \(T_b=150\). The estimated ratio variance is \(1/1800\), and the Mack process variance parameter is \(1/6\). Calculate the aggregate prediction variance and standard error. Assume the process errors are independent across origins.

Individual process variances. For origin \(a\), the process variance is \((1/6)(300)=50\). For origin \(b\), it is \((1/6)(150)=25\). The aggregate process variance is \(50+25=75\).

Parameter contribution. The common ratio estimate contributes \((T_a+T_b)^2(1/1800)=450^2/1800=112.5\). Equivalently, the two individual parameter contributions are \(300^2/1800=50\) and \(150^2/1800=12.5\), and their covariance is \(300(150)/1800=25\); the aggregate includes twice that covariance, so \(50+12.5+2(25)=112.5\).

Aggregate result. The aggregate prediction variance is \(75+112.5=187.5\), so its standard error is \(\sqrt{187.5}=13.6931\), or \(13.69\) to two decimals. If the individual MSEPs were simply added, the result would be \(50+50+25+12.5=137.5\), which is too small by \(2(25)=50\). This calculation illustrates why shared-factor covariance must not be discarded.

Reading the Comparison in Practice

A useful comparison asks not merely which reported standard error is larger, but which assumptions and estimates caused the difference. Mack’s development-age variances can be sensitive to the observed link ratios and to the estimator selected for a sparse age. ODP’s estimate of \(\phi\) is sensitive to its fitted means, residual degrees of freedom, and the dispersion estimator. Both methods can produce unstable results when only a few informative observations support a development age.

Triangle diagnostics are therefore part of the comparison, not an optional afterthought. Earlier material on standardized chain-ladder residuals, residual patterns, calendar-year effects, influential outliers, and sparse late development identifies features that can contradict a stable link-ratio or incremental mean-variance structure. For example, an outlier can increase a Mack variance estimate and the ODP Pearson statistic, while a systematic calendar-year pattern indicates that a model with development effects alone may not capture the dependence in the data.

The estimates need not react identically. Mack variance estimates are organized by development age and depend on the chosen link-ratio variance estimator. ODP uses a common dispersion parameter in its standard form, so a lack of fit concentrated in one age may be spread across the residual-based estimate rather than represented as a separate variance parameter for that age. A comparison should report any selected tail factor or variance adjustment and explain whether both approaches use compatible assumptions and cells.

Key takeaway: Mack and ODP can agree when their mean structures and implied variances align, but agreement is not guaranteed. Distinguish future process variation from fitted-parameter uncertainty, include covariance from shared factors when aggregating, and interpret each estimate alongside the triangle diagnostics.

Common Exam Traps

Comparing unlike quantities. A reserve estimate is an expected amount; a prediction error is a standard deviation or MSEP. Check whether a reported figure is a variance, MSEP, or square-root standard error.
Omitting parameter uncertainty. A future process variance alone does not account for estimating the development pattern from the same triangle.
Adding standard errors. Aggregate variances require covariance terms. Do not add standard errors, and do not assume total variance equals the sum of individual variances when factors are shared.
Equating the two dispersion estimates. Mack’s \(\sigma_k^2\) is age-specific and has the scale of cumulative claims; ODP’s \(\phi\) scales incremental variance relative to its fitted mean. They are not directly interchangeable.
Ignoring diagnostics. A numerical standard error does not establish that the underlying development assumptions are credible. Outliers, changing patterns, calendar-year effects, and unstable tail development can undermine either model.

Check Your Understanding

For each question, state any assumptions used in the calculation. Answers are provided for checking.

A one-step Mack model has \(\widehat\sigma^2=0.24\), complete-row starting cumulative total \(400\), and a future origin starting cumulative amount \(200\). Using \(\operatorname{Var}(\widehat f)=\widehat\sigma^2/400\), calculate the factor-estimation contribution to the future prediction variance. (Answer: \(200^2(0.24/400)=24\).)
For an ODP fit, the Pearson statistic is \(12\) and the residual degrees of freedom are \(6\). What is the estimated dispersion parameter? (Answer: \(\widehat\phi=12/6=2\).)
Two future origins have exposures \(100\) and \(200\), and a common estimated ratio has variance \(0.002\). Calculate their covariance due to ratio estimation and the aggregate variance contribution from this parameter uncertainty. (Answer: covariance \(100(200)(0.002)=40\); aggregate parameter contribution \((100+200)^2(0.002)=180\).)
In a one-step ODP model, \(q=0.2\), \(\phi=1.5\), the future exposure is \(250\), and the total complete-row exposure is \(500\). Using the one-step formulas, calculate process variance, parameter variance, and MSEP. (Answer: process variance \(1.5(0.2)(250)=75\); parameter variance \(250^2(1.5)(0.2)/500=37.5\); MSEP \(112.5\).)
A triangle’s ODP Pearson residuals show persistent calendar-year structure, while one Mack link-ratio residual is highly influential. Is it sufficient to choose whichever method gives the smaller standard error? Explain briefly. (Answer: No. The calendar-year pattern challenges the ODP development-only structure, and the influential residual can distort Mack variance estimates; investigate model fit and sensitivity before interpreting either standard error.)

What You'll Learn