Credibility-Adjusted Development Selections

Why Credibility-Adjust Development Factors?

An observed age-to-age factor reflects the claims experience available for a particular development interval. When that experience is limited, unusual claims or random fluctuation can make the observed factor a weak guide to future development. An external selection, based on broader or more stable experience, can provide a useful reference. Credibility adjustment blends the two: more supporting exposure gives the observed factor more influence, while less exposure leaves more weight on the external selection.

The aim is not to replace a carefully chosen external assumption with a formula. It is to make the role of the observed experience explicit, use a consistent exposure basis, and carry each adjusted factor through the projection without applying credibility a second time.

Definition: For development interval $j$, let $\widehat F_j$ be the observed age-to-age factor, $F_{0,j}$ the external selected factor, $E_j$ the exposure supporting the observed factor, and $Z_j$ the credibility assigned to that observation. A credibility-adjusted factor is $$ F_j^{\mathrm{sel}}=Z_j\widehat F_j+(1-Z_j)F_{0,j}, $$ where $0\leq Z_j\leq 1$. The selected factor is a weighted average: $Z_j=1$ gives full weight to observed experience, and $Z_j=0$ gives full weight to the external selection.

Observed Factors and Exposure Bases

Suppose $C_{i,j}$ denotes cumulative losses for origin period $i$ at development age $j$. A common observed age-to-age factor is the volume-weighted ratio $$ \widehat F_j=\frac{\sum_{i\in I_j}C_{i,j+1}}{\sum_{i\in I_j}C_{i,j}}, $$ where $I_j$ contains the origin periods with usable cumulative losses at both ages. This ratio gives more weight to origin periods with larger cumulative losses at age $j$. It is important to calculate it from aligned observations: the numerator and denominator must cover the same origin periods and use the same loss basis.

Exposure $E_j$ is a separate input to the credibility weight. It should measure the information supporting the observed factor, on the basis specified by the model or problem. Depending on the setting, that may be claim count, policy exposure, or another justified volume measure. Do not assume that the denominator of the volume-weighted factor is automatically the correct credibility exposure. Confirm that the exposure is comparable across intervals and that it corresponds to the experience being blended.

One useful exposure-based credibility model is $Z_j=E_j/(E_j+K)$, where $K$ is a positive credibility constant calibrated for the relevant process and exposure units. The following derivation explains this form and the assumptions it requires.

Why the Exposure Weight Has This Form

Consider one development interval and suppress its subscript. Suppose the underlying factor $\theta$ varies around the external selection $F_0$, with $\theta\sim N(F_0,\tau^2)$. Conditional on $\theta$, suppose the observed estimate satisfies $\widehat F\mid\theta\sim N(\theta,\sigma^2/E)$. Thus, more exposure reduces the sampling variance of the observed estimate. Under this model, the conditional distribution of $\theta$ given $\widehat F$ is normal, and its mean is a weighted average of the observed estimate and $F_0$.

Write the two precision contributions.
The prior variance is $\tau^2$, so its precision is $1/\tau^2$. The observed estimate has variance $\sigma^2/E$, so its precision is $E/\sigma^2$.

Form the posterior mean.
For independent normal prior and sampling models, the posterior mean weights each mean in proportion to its precision: $$ E[\theta\mid\widehat F] =\frac{(E/\sigma^2)\widehat F+(1/\tau^2)F_0}{E/\sigma^2+1/\tau^2}. $$

Identify the credibility weight.
The coefficient of $\widehat F$ is $$ Z=\frac{E/\sigma^2}{E/\sigma^2+1/\tau^2} =\frac{E}{E+\sigma^2/\tau^2}. $$ Setting $K=\sigma^2/\tau^2$ gives $Z=E/(E+K)$, as stated.

This derivation depends on the stated normal model and on the interpretation of $F_0$ as the prior mean. The expression $E/(E+K)$ is not a universal credibility rule: $K$ must be calibrated appropriately, and the model is only useful when the exposure measure and its units are meaningful. If a problem specifies a different credibility formula, use that formula rather than substituting this one.

Key result: With $Z_j=E_j/(E_j+K)$, the adjusted factor is $$ F_j^{\mathrm{sel}}=\frac{E_j}{E_j+K}\widehat F_j+\frac{K}{E_j+K}F_{0,j}. $$ For positive $K$, increasing $E_j$ increases the weight on observed experience. The weights sum to one, so the selected factor lies between the observed and external factors.

Applying Selected Factors to an Ultimate Projection

For an origin period observed at development age $k$, the projected ultimate is the latest cumulative loss multiplied by the selected factors for the remaining intervals. If development is projected through age $m$, then $$ \widehat U_i=C_{i,k}\prod_{j=k}^{m-1}F_j^{\mathrm{sel}}. $$ The estimated unpaid amount on the same cumulative-loss basis is $\widehat U_i-C_{i,k}$. This projection assumes the selected factors are appropriate for that origin period and that the cumulative loss definitions are consistent across ages. If the projection includes an additional tail factor, it must be stated and applied once, after the specified development intervals.

Worked Example: Credibility-Adjusting One Development Interval

A hypothetical insurer has an observed age-to-age factor of $1.50$ for an interval. The external selection is $1.30$. The exposure supporting the observed factor is $30$ units, and the calibrated constant is $K=20$ exposure units. A recent origin period has cumulative losses of $\$800{,}000$ at the earlier age. Calculate its projected cumulative losses at the next age and the development amount.

Method and credibility. Use $Z=E/(E+K)$. Here, $$ Z=\frac{30}{30+20}=\frac{30}{50}=0.60. $$ The external-selection weight is $1-Z=0.40$.

Adjusted factor. $$ F^{\mathrm{sel}}=(0.60)(1.50)+(0.40)(1.30)=0.90+0.52=1.42. $$ The observed factor exceeds the external selection, so the blend is between $1.30$ and $1.50$, as it should be.

Projection. Applying the selected factor once gives $$ \widehat U=\$800{,}000(1.42)=\$1{,}136{,}000. $$ The projected development amount is $$ \$1{,}136{,}000-\$800{,}000=\$336{,}000. $$ Thus the projected next-age cumulative loss is $\$1{,}136{,}000$, with $\$336{,}000$ of additional development. The calculation avoids treating the external selection as a factor to multiply by the observed factor; the two are blended, not compounded.

Worked Example: Credible Selections Across Two Intervals

A hypothetical cumulative-loss triangle has the following usable observations, in thousands of dollars. At ages 1 and 2, the available pairs are $100,140$, $200,280$, and $300,420$. For the age-2-to-3 interval, the available pairs are $140,196$ and $280,392$. The youngest origin period is observed at age 1 with cumulative losses of $\$300{,}000$. For age 1 to 2, use $E_1=40$, external factor $1.25$, and $K=20$. For age 2 to 3, use $E_2=20$, external factor $1.20$, and $K=20$. Find the selected factors and projected ultimate for the youngest origin period.

Method and observed age-1-to-2 factor. Use the volume-weighted ratio on the three aligned pairs: $$ \widehat F_1=\frac{140+280+420}{100+200+300} =\frac{840}{600}=1.40. $$ The credibility is $$ Z_1=\frac{40}{40+20}=\frac{2}{3}. $$ Therefore, $$ F_1^{\mathrm{sel}}=\left(\frac{2}{3}\right)(1.40)+\left(\frac{1}{3}\right)(1.25) =0.933333\ldots+0.416666\ldots=1.35. $$

Observed age-2-to-3 factor. Use the two aligned pairs: $$ \widehat F_2=\frac{196+392}{140+280} =\frac{588}{420}=1.40. $$ Here, $$ Z_2=\frac{20}{20+20}=0.50, $$ so $$ F_2^{\mathrm{sel}}=(0.50)(1.40)+(0.50)(1.20)=0.70+0.60=1.30. $$

Project to ultimate. The youngest period is at age 1, so apply both remaining factors in order: $$ \widehat U= \$300{,}000(1.35)(1.30) =\$300{,}000(1.755)=\$526{,}500. $$ The implied development from the age-1 cumulative amount is $$ \$526{,}500-\$300{,}000=\$226{,}500. $$ The projected ultimate is $\$526{,}500$. Each interval has its own observed factor, external selection, and exposure weight; using one interval's credibility for both would ignore the stated evidence for the second interval.

Worked Example: Keeping the Credibility Exposure Basis Straight

For a hypothetical portfolio, an observed age-1-to-2 factor is calculated from aligned cumulative losses of $\$400{,}000$ at age 1 and $\$600{,}000$ at age 2. An observed age-2-to-3 factor is $1.10$. The relevant exposure amounts are $80$ units for the first interval and $20$ units for the second. The external selections are $1.40$ and $1.30$, respectively, and $K=20$ exposure units for each interval. A period has $\$500{,}000$ of cumulative losses at age 1. Calculate its ultimate and projected development.

Method and first interval. The observed factor is $$ \widehat F_1=\frac{\$600{,}000}{\$400{,}000}=1.50. $$ Its credibility uses the stated exposure of $80$, not the loss amount: $$ Z_1=\frac{80}{80+20}=0.80. $$ Thus $$ F_1^{\mathrm{sel}}=(0.80)(1.50)+(0.20)(1.40)=1.20+0.28=1.48. $$

Second interval. Its credibility is separately calculated from its own exposure: $$ Z_2=\frac{20}{20+20}=0.50. $$ The selected factor is $$ F_2^{\mathrm{sel}}=(0.50)(1.10)+(0.50)(1.30)=0.55+0.65=1.20. $$

Projection and check. $$ \widehat U=\$500{,}000(1.48)(1.20) =\$500{,}000(1.776)=\$888{,}000. $$ Therefore the projected development is $$ \$888{,}000-\$500{,}000=\$388{,}000. $$ The ultimate exceeds the latest cumulative losses, and both selected factors lie between their corresponding observed and external factors. The calculation avoids using the first interval's exposure for the second interval merely because both factors appear in the same projection.

Common Traps and Practical Checks

Confusing the factor with its credibility weight. $Z$ is a weight between zero and one, not a development factor. Blend $Z\widehat F+(1-Z)F_0$; do not multiply $Z$ by the external factor alone.
Using an inconsistent exposure basis. If $K$ is calibrated in policy-year units, an exposure measure in claim counts cannot be substituted without a justified conversion. Use the basis specified for the credibility model.
Credibility-weighting twice. Once $F_j^{\mathrm{sel}}$ has been calculated, use that selected factor in the projection. Do not blend the projected losses with the external projection again unless a separate, explicitly defined method calls for it.
Misaligning the factor data. A volume-weighted ratio requires the same origin periods in its numerator and denominator. Adding a newer period's age-$j+1$ losses when that period has no matching age-$j$ amount changes the factor basis.
Applying factors in the wrong direction. An age-to-age factor moves cumulative losses from an earlier age to a later age. It is applied to cumulative losses at the earlier age, not reversed or applied to incremental losses as if they were cumulative.

A final reasonableness check is to verify that every selected factor lies between its two inputs, that greater exposure produces greater observed-data weight under the specified formula, and that the ultimate is obtained by multiplying only the factors for development ages still outstanding.

Key takeaway: Calculate each observed age-to-age factor on aligned data, assign credibility using the exposure basis and formula appropriate to that interval, blend with the external selection, and then project with the adjusted factors once in chronological order.

Check Your Understanding

For each problem, use $Z=E/(E+K)$ and the stated exposure units. Keep monetary amounts in dollars.

An observed factor is $1.50$, the external selection is $1.30$, $E=15$, and $K=45$. Find the credibility-adjusted factor. (Answer: $1.35$)
An observed factor is $1.28$, the external selection is $1.20$, $E=60$, and $K=20$. Find the selected factor. (Answer: $1.26$)
A period has $\$250$ of cumulative losses at age 1, where amounts are in thousands of dollars. The observed and external factors for age 1 to 2 are $1.50$ and $1.30$, with $E=30$, $K=30$. For age 2 to 3 they are $1.20$ and $1.40$, with $E=10$, $K=30$. Find the ultimate and development amount, in thousands of dollars. (Answer: Ultimate $472.5$; development $222.5$)
Aligned age-1 and age-2 cumulative losses sum to $\$750{,}000$ and $\$900{,}000$, respectively. The external factor is $1.30$, $E=40$, and $K=10$. A period has $\$600{,}000$ at age 1. Find its projected age-2 cumulative losses and development. (Answer: Cumulative losses $\$732{,}000$; development $\$132{,}000$)
A period has $\$400$ thousand at age 1. For age 1 to 2, the observed and external factors are $1.45$ and $1.25$, with $E=20$, $K=20$. For age 2 to 3, they are $1.10$ and $1.30$, with $E=60$, $K=20$. Find the ultimate and the development amount, in thousands. (Answer: Ultimate $621$; development $221$)

What You'll Learn