Overdispersed Poisson Model For Reserve Variability

Why Model Incremental Claims?

A claims development triangle records amounts at successive development ages. Chain ladder uses observed cumulative amounts to project future development. To describe variability around those projections, an actuary can model the incremental amounts directly. The overdispersed Poisson (ODP) model gives a multiplicative structure for their expected values and lets their variances exceed their means by a common scale factor.

The model is useful for connecting a chain-ladder projection to an explicit mean-and-variance framework. Its assumptions also matter: a good match to a chain-ladder estimate does not establish that the triangle’s variability, independence, or development pattern is adequately represented.

Definition: Let $X_{ij}$ be the incremental claim amount for origin period $i$ and development period $j$, for cells observed in a development triangle. The ODP model specifies $$ E[X_{ij}]=\mu_{ij}=\alpha_i\beta_j,\qquad \operatorname{Var}(X_{ij})=\phi\mu_{ij}, $$ where $\alpha_i>0$ is an origin-period parameter, $\beta_j>0$ is a development-period parameter, and $\phi>0$ is a common dispersion parameter. In the usual model, distinct incremental cells are independent. The parameters must be normalized in some way because multiplying all $\alpha_i$ by a constant and dividing all $\beta_j$ by that constant leaves $\mu_{ij}$ unchanged.

Mean, Variance, and Dispersion

The product $\alpha_i\beta_j$ says that the expected increment separates into an origin-period level and a development-period pattern. Two origin periods may have different overall expected claim amounts through $\alpha_i$, while the $\beta_j$ terms describe how expected incremental amounts vary by development age. The model does not require the development parameters to sum to one.

The variance is proportional to the expected increment. If $\phi=1$, it equals the mean, as in the Poisson variance formula. If $\phi>1$, the model allows greater variability than a Poisson model with the same mean; this is the usual overdispersed case. If $0<\phi<1$, the variance is less than the mean. The coefficient of variation for a cell with positive mean is

\frac{\sqrt{\operatorname{Var}(X_{ij})}}{E[X_{ij}]} =\sqrt{\frac{\phi}{\mu_{ij}}}.

Thus, under a common $\phi$, cells with smaller expected amounts have larger relative variability. The model’s mean and variance specifications do not, by themselves, define a complete probability distribution. “Overdispersed Poisson” in reserving is often a quasi-likelihood model: it uses a Poisson-like mean structure and a scaled variance without asserting that each observed payment is an integer-valued Poisson random variable.

Worked Example: Interpreting a Cell’s Dispersion

A hypothetical insurer models an incremental cell with expected amount $\mu_{ij}=\alpha_i\beta_j=\$240$. The estimated common dispersion parameter is $\phi=1.8$. Calculate the model variance, standard deviation, and coefficient of variation. State how the variance compares with the Poisson variance.

Method. Apply the ODP variance formula, then take its square root and divide by the mean for the coefficient of variation.

The mean is $\mu_{ij}=\$240$. Therefore,

\operatorname{Var}(X_{ij})=\phi\mu_{ij} =1.8(240)=432

Because the amount is in dollars, the variance is in squared dollars. The standard deviation is

\sqrt{432}=\$20.7846\ldots

The coefficient of variation is

\frac{\sqrt{432}}{240}=0.08660\ldots

To four decimal places, the coefficient of variation is $0.0866$, or $8.660\%$. A Poisson model with this mean would have variance $240$, so the ODP variance is $432/240=1.8$ times as large. Check: $1.8(240)=432$, and $20.7846^2$ is approximately $432$. The common mistake is to treat $\phi$ as an additional mean multiplier; it scales the variance, not the expected amount.

How the Mean Structure Relates to Chain Ladder

Write $C_{ij}=\sum_{k=1}^{j}X_{ik}$ for cumulative claims through development period $j$. Under the ODP mean structure, linearity of expectation gives

E[C_{ij}] =\sum_{k=1}^{j}E[X_{ik}] =\alpha_i\sum_{k=1}^{j}\beta_k.

For a fixed origin period, the ratio of expected cumulative amounts at successive ages is therefore

\frac{E[C_{i,j+1}]}{E[C_{ij}]} =\frac{\sum_{k=1}^{j+1}\beta_k}{\sum_{k=1}^{j}\beta_k}.

The origin parameter cancels. Consequently, the model implies a common expected cumulative development pattern across origin periods, even though their expected claim levels differ. This is the structural reason the ODP mean model is associated with chain ladder: both use a development pattern that is common across origin periods.

A standard chain-ladder age-to-age factor is estimated as the ratio of the aggregate observed cumulative amounts at the later and earlier ages, using origin periods observed at both ages. Under the usual triangular-data setup, fitting the multiplicative ODP mean model with a log link gives the same fitted incremental values and projections as volume-weighted chain ladder. This equivalence concerns the fitted means. It does not mean that the variance assumptions are automatically validated or that every form of chain ladder is equivalent to every ODP fit.

A short derivation shows why the model has the relevant structure. Since $E[X_{ij}]=\alpha_i\beta_j$, the expected cumulative amount at age $j$ is $\alpha_i$ times the partial sum of development parameters. Advancing one age adds $\alpha_i\beta_{j+1}$. Thus, the expected cumulative age-to-age ratio depends on the development parameters but not on the origin-period level. Pooling observed cumulative amounts across origins estimates this shared development pattern; the chain-ladder factors are empirical versions of those ratios.

Worked Example: Projecting Means in a Small Triangle

A hypothetical triangle has observed incremental amounts, in dollars, as follows: origin period 1 has $100, 50, 25$; origin period 2 has $120, 60$; and origin period 3 has $80$. Use volume-weighted chain ladder to project the missing increments. Then interpret the projections as ODP fitted means under the standard equivalence conditions. Assume $\phi=2$, and calculate the ODP variance for the projected origin-period-2, development-period-3 cell.

Method. First form cumulative amounts and estimate the age-to-age factors from aggregate cumulative amounts. Then apply each factor to the latest cumulative amount for an origin period. Under the standard ODP and chain-ladder equivalence, these projected increments are fitted means.

The observed cumulative amounts are:

Origin period 1: $100,\ 100+50=150,\ 150+25=175$.
Origin period 2: $120,\ 120+60=180$.
Origin period 3: $80$.

For development periods 1 to 2, the eligible origins are 1 and 2. The first age-to-age factor is

\widehat f_1=\frac{150+180}{100+120} =\frac{330}{220}=1.5.

For development periods 2 to 3, only origin period 1 is observed at both ages. Thus,

\widehat f_2=\frac{175}{150}=\frac{7}{6}.

The projected origin-period-2 cumulative amount at development period 3 is $180(7/6)=210$, so its projected incremental amount is $210-180=\$30$. For origin period 3, the projected cumulative amount at development period 2 is $80(1.5)=120$, giving an incremental amount of $120-80=\$40$. Its projected cumulative amount at development period 3 is $120(7/6)=140$, giving an incremental amount of $140-120=\$20$.

Under the stated equivalence conditions, the ODP fitted mean for the origin-period-2, development-period-3 cell is $\mu_{23}=\$30$. Its variance is

\operatorname{Var}(X_{23})=\phi\mu_{23} =2(30)=60.

The variance is $60$ squared dollars, and its standard deviation is $\sqrt{60}=\$7.7460\ldots$. Check: the projected cumulative amounts reconcile: $180+30=210$, and $120+20=140$. A common mistake is to treat an age-to-age cumulative factor as an incremental factor. The $7/6$ factor applies to the cumulative amount $180$; the projected increment is the increase, $210-180=30$.

Assumptions and Limits of the Model

The ODP model is more than a formula for variance. Its assumptions determine what conclusions an actuary may draw from the fit.

Multiplicative expected increments.
Each cell’s expected amount is $\alpha_i\beta_j$. The development pattern is shared across origins; an origin-specific development pattern or an unmodeled calendar-year effect can violate this structure.

Variance proportional to the mean.
Each cell has variance $\phi\mu_{ij}$, with the same dispersion parameter across cells. Systematic changes in dispersion by origin or development age are not captured by a single $\phi$.

Independent incremental cells.
Distinct cells are assumed independent in the standard model. Shared shocks, claim-level dependence across periods, or other dependence can make uncertainty calculations based on independence unreliable.

Nonnegative amounts and positive fitted means.
The multiplicative specification uses positive means. A standard Poisson-type variance model is not directly appropriate for negative incremental amounts; such data require careful treatment rather than silently applying the formula.

Process variation is not parameter uncertainty.
The variance formula describes cell-level variation conditional on the model parameters. Estimates of $\alpha_i$, $\beta_j$, and $\phi$ are themselves uncertain; that estimation uncertainty is a separate issue.

In practice, diagnostics such as residual patterns can help assess whether the mean structure is plausible. A visually smooth chain-ladder fit is not enough: calendar-year patterns, outliers, sparse late development, and dependence may still matter. Also, because the ODP specification gives only a mean and variance unless a full distribution is imposed, those two moments alone do not justify a particular prediction-interval shape.

Worked Example: Interpreting a Dispersion Estimate

For a hypothetical portfolio, a variance study estimates that cells with fitted mean $\$200$ have process variance $700$ squared dollars. A modeller proposes using an ODP model with a common dispersion parameter. Calculate the implied dispersion parameter and standard deviation, then state what must be checked before using that parameter for all cells.

Method. Rearrange $\operatorname{Var}(X_{ij})=\phi\mu_{ij}$ to solve for $\phi$.

\widehat\phi=\frac{700}{200}=3.5.

The corresponding standard deviation is

\sqrt{700}=\$26.4575\ldots

To the nearest cent, it is $\$26.46$. The estimate $\widehat\phi=3.5$ indicates variance 3.5 times the mean at the stated fitted mean. Before applying it to all cells, the modeller should assess whether a common variance-to-mean ratio is reasonable across origin and development periods, whether incremental cells can reasonably be treated as independent, and whether the fitted mean structure is adequate. Check: $3.5(200)=700$. The common mistake is to conclude that one cell’s estimated dispersion proves the same dispersion applies throughout the triangle.

Common Exam Traps

Confusing mean and variance. The mean is $\mu_{ij}$; the variance is $\phi\mu_{ij}$. Multiplying the mean by $\phi$ changes the model rather than describing overdispersion.
Assuming literal Poisson data. Incremental claim amounts may be continuous or measured in dollars. The ODP mean-variance relationship does not require every amount to be an integer count.
Equating fitted means with validated uncertainty. Chain-ladder and ODP fitted values can agree even when independence or the common-dispersion assumption is implausible.
Applying cumulative factors directly to increments. Age-to-age factors act on cumulative amounts. The projected increment is the change in cumulative amount.
Ignoring the data structure. Negative increments, zeros, sparse late development, or calendar-year effects can make a standard multiplicative ODP specification unsuitable or require additional modelling choices.

Key takeaway: The ODP model specifies $E[X_{ij}]=\alpha_i\beta_j$ and $\operatorname{Var}(X_{ij})=\phi\alpha_i\beta_j$. Its common multiplicative development structure explains its connection to chain-ladder fitted means. Its independence and common-dispersion assumptions are separate claims that must be assessed, not inferred from that connection.

Check Your Understanding

For each problem, use the ODP model only where its assumptions are stated or are reasonable for the question.

A cell has fitted mean $\$320$ and $\phi=1.25$. Calculate its variance and standard deviation. (Answer: variance $400$ squared dollars; standard deviation $\$20.00$)
A cell has mean $\$90$ and variance $270$ squared dollars. Find $\phi$, and say whether the cell is overdispersed relative to a Poisson variance. (Answer: $\phi=3$; yes, its variance is three times its mean.)
For one origin period, $\beta_1=1,\ \beta_2=0.4,\ \beta_3=0.2$, with $\alpha_i=150$. Calculate the expected cumulative amount at development period 2 and the expected cumulative age-to-age factor from period 2 to 3. (Answer: expected cumulative amount $150(1+0.4)=\$210$; factor $(1+0.4+0.2)/(1+0.4)=8/7\approx1.1429$)
A hypothetical triangle has aggregate cumulative amounts of $240$ at development period 1 and $336$ at development period 2 among eligible origins. The latest cumulative amount for another origin is $150$. Under volume-weighted chain ladder, calculate the projected next cumulative amount and incremental amount for that origin. (Answer: factor $336/240=1.4$; projected cumulative amount $150(1.4)=210$; projected increment $210-150=60$)
An analyst reports that a fitted ODP model matches chain-ladder projections, then concludes that future cells are independent and that a normal prediction interval is justified. Identify the unsupported conclusions. (Answer: Matching fitted means does not establish independence; the mean-variance specification alone does not establish a normal distribution or justify a normal prediction interval.)

What You'll Learn