Applying Credibility to Indicated Rate Changes

Why Apply Credibility to a Rate Change?

An experience-based rate indication reflects the losses and exposures observed for a particular risk or group. A manual indication reflects a broader benchmark or pricing basis. The experience may be highly informative, but a small or volatile portfolio can produce an unstable indication. Credibility provides a systematic way to temper that indication with the manual view.

This tutorial focuses on a single blending decision: how to combine one experience-based rate change with one manual rate change. The result depends on both indications and on the credibility weight assigned to the experience. Credibility methods are not interchangeable: the weight should be selected using an appropriate method and assumptions, not chosen simply because it produces a preferred rate. The treatment is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Definition: Let $c_E$ be the experience-based indicated percentage change and $c_M$ the manual indicated percentage change, each measured from the same current rate level and for the same coverage, period, and basis. For an experience credibility weight $Z$, where $0 \leq Z \leq 1$, the blended indicated change is $$ c_B=Zc_E+(1-Z)c_M. $$ The weight $Z$ is the share assigned to experience; $1-Z$ is the share assigned to the manual indication.

A percentage change is relative to the current rate. If the current rate is $R_0$, the indicated rate after blending is $R_B=R_0(1+c_B)$, where changes are written as decimals in the calculation. For example, a 12% increase is represented by $c_B=0.12$. The blend is a weighted average: if $Z=0$, the manual indication controls; if $Z=1$, the experience indication controls. When the two indications agree, the blended indication is the same regardless of $Z$.

Blend Changes or Rate Factors Consistently

The same blend can be performed using rate factors. Define $r_E=1+c_E$ and $r_M=1+c_M$. Since the weights sum to one, $$ Zr_E+(1-Z)r_M =Z(1+c_E)+(1-Z)(1+c_M) =1+Zc_E+(1-Z)c_M =1+c_B. $$ Thus, blending percentage changes and then adding one gives the same result as blending factors directly, provided the same weights and base are used.

This equivalence does not mean that percentages should be added to the current rate without applying the base. A 10% change to a current rate of $500 means an increase of $50, not $10. It also does not justify blending indications that use different definitions of rate, coverage, or time period: those bases must first be made consistent.

Worked Example: Blending Two Indications into a Rate

A hypothetical coverage has a current annual rate of $400. Its experience indication is a 14% increase, while its manual indication is a 6% increase. The selected experience credibility weight is $Z=0.65$. Find the blended percentage change and indicated annual rate.

Method. Apply the specified credibility weight to the two changes, then apply the resulting factor to the current rate. This is a single-period rate calculation; no interest rate or mortality assumption is involved. No probability distribution is needed because the credibility weight is given.

Step 1: Weight the experience change. $0.65(14\%)=9.10\%$.

Step 2: Weight the manual change. The manual weight is $1-0.65=0.35$, so $0.35(6\%)=2.10\%$.

Step 3: Add the weighted changes. $c_B=9.10\%+2.10\%=11.20\%$.

Step 4: Apply the change to the current rate. The rate factor is $1+0.112=1.112$, giving $R_B=\$400(1.112)=\$444.80$.

Check. The blended change, 11.20%, lies between the two indications, 6% and 14%, as it must for weights between zero and one. The result is an $44.80 increase over $400. The common mistake avoided is applying the percentage change as a dollar amount rather than applying its rate factor.

Limited Fluctuation Credibility and Exposure Volume

One way to set $Z$ is limited fluctuation credibility. It starts with a full credibility standard: enough independent, comparable exposure that the aggregate experience is sufficiently close to its expected value with a specified probability. Under the usual normal approximation, let $X_j$ be the loss for exposure unit $j$, with independent, identically distributed losses, mean $\mu$, and coefficient of variation $C=\sqrt{\operatorname{Var}(X_j)}/\mu$. For $n$ exposure units, aggregate loss $S_n=\sum_{j=1}^n X_j$ has mean $n\mu$ and variance $n\operatorname{Var}(X_j)$.

Result: For a two-sided relative tolerance $k$ and a normal critical value $z_p$, the approximate full credibility standard is $$ N_0=\left(\frac{z_p C}{k}\right)^2. $$ A commonly used partial credibility weight based on exposure volume $n$ is $$ Z=\min\left(1,\sqrt{\frac{n}{N_0}}\right). $$ This rule assumes comparable independent exposure units and variability that scales with exposure. The normal approximation and the selected tolerance and probability must be suitable for the application; the formula is not a universal weight for every data structure.

The square root reflects how sampling variability declines with exposure. Under the assumptions above, the standard deviation of aggregate loss grows in proportion to $\sqrt{n}$, while its mean grows in proportion to $n$. Therefore, relative fluctuation is proportional to $1/\sqrt{n}$. If exposure is one quarter of the full standard, the square-root rule assigns $Z=0.5$, not $Z=0.25$.

Worked Example: Finding Credibility from a Full Credibility Standard

A hypothetical portfolio has 25 independent, comparable exposure units. The coefficient of variation of loss per unit is $C=0.5$. The actuary uses a two-sided normal critical value $z_p=2$ and a relative tolerance $k=10\%$; the critical value of 2 corresponds to approximately 95.45% central probability under a normal approximation. The experience indication is an 18% increase, the manual indication is a 6% increase, and the current rate is $240. Find the limited fluctuation blend and indicated rate.

Model assumptions. Losses per exposure unit are identically distributed and independent, with coefficient of variation 0.5. Aggregate loss is approximated by a normal distribution for the stated full credibility criterion. This is a one-period pricing example: no interest or mortality assumption is used.

Step 1: Calculate full credibility exposure. Substitute the specified inputs: $N_0=(2(0.5)/0.10)^2=(1/0.10)^2=10^2=100$ exposure units.

Step 2: Calculate partial credibility. $Z=\sqrt{25/100}=\sqrt{0.25}=0.5$. The manual weight is $1-0.5=0.5$.

Step 3: Blend the changes. The experience contribution is $0.5(18\%)=9\%$; the manual contribution is $0.5(6\%)=3\%$. Thus $c_B=9\%+3\%=12\%$.

Step 4: Calculate the rate. The factor is $1.12$, so $R_B=\$240(1.12)=\$268.80$.

Check. With 25 exposure units against a full standard of 100, credibility is one half under the square-root rule. The 12% blend falls between the 6% and 18% indications, and the dollar increase is $28.80. The method avoids treating 25% of the full exposure standard as 25% credibility.

Bühlmann Credibility as a Different Basis for the Weight

A different approach estimates credibility from the variation within risks and the variation between risks. In the Bühlmann model, let $\Theta$ represent an unobserved risk parameter. Conditional on $\Theta$, observations $X_1,\ldots,X_n$ are independent and identically distributed. Define the overall mean $m=E[E[X_j\mid\Theta]]$, the expected process variance $\mathrm{EPV}=E[\operatorname{Var}(X_j\mid\Theta)]$, and the variance of hypothetical means $\mathrm{VHM}=\operatorname{Var}(E[X_j\mid\Theta])$. If $\mathrm{VHM}>0$, set $K=\mathrm{EPV}/\mathrm{VHM}$.

Result: The Bühlmann credibility estimate of the risk's expected outcome is $$ \widehat{\mu}=Z\bar{X}+(1-Z)m,\qquad Z=\frac{n}{n+K},\qquad K=\frac{\mathrm{EPV}}{\mathrm{VHM}}, $$ where $\bar{X}$ is the observed mean for that risk. The estimate is the best linear predictor of the risk's conditional mean under the Bühlmann assumptions. In a pricing application, a rate indication derived from the experience can be credibility-weighted against the manual indication using the corresponding $Z$, provided both indications refer to the same target and basis.

The weight can be justified by minimizing mean squared prediction error among linear estimates $a\bar X+(1-a)m$. The model gives $\operatorname{Var}(\bar X)=\mathrm{VHM}+\mathrm{EPV}/n$, and $\operatorname{Cov}(\bar X,\mu(\Theta))=\mathrm{VHM}$, where $\mu(\Theta)=E[X_j\mid\Theta]$. Hence the error variance is $$ \operatorname{Var}\big(a\bar X+(1-a)m-\mu(\Theta)\big) =a^2\left(\mathrm{VHM}+\frac{\mathrm{EPV}}{n}\right)+(1-2a)\mathrm{VHM}. $$ Differentiating with respect to $a$ and setting the derivative to zero gives $$ 2a\left(\mathrm{VHM}+\frac{\mathrm{EPV}}{n}\right)-2\mathrm{VHM}=0, \qquad a=\frac{\mathrm{VHM}}{\mathrm{VHM}+\mathrm{EPV}/n} =\frac{n}{n+\mathrm{EPV}/\mathrm{VHM}}. $$ Thus $a=Z=n/(n+K)$. The result explains why more observations increase experience weight, while greater process variation relative to between-risk variation increases $K$ and reduces credibility. It is a linear credibility result, not a claim that every blended indication is the full Bayesian posterior expectation.

Worked Example: Comparing Two Credibility Methods

A hypothetical book has nine comparable exposure units. Its experience indication is a 17% increase; its manual indication is a 5% increase; the current rate is $1,000. For Bühlmann credibility, the applicable model estimates $\mathrm{EPV}=900$ and $\mathrm{VHM}=100$, in squared units of the modeled outcome. A separately specified limited fluctuation standard is $N_0=16$ exposure units. Compare the two blends.

Assumptions. For Bühlmann, the observations are conditionally independent and identically distributed given the risk parameter, and the stated EPV and VHM apply to the modeled experience measure. For limited fluctuation, exposure units are comparable and independent, and the square-root rule is appropriate. No interest rate or mortality assumption is relevant to this one-period rate comparison; no particular probability distribution is required for the Bühlmann formula.

Limited fluctuation calculation. $Z=\sqrt{9/16}=\sqrt{0.5625}=0.75$, leaving manual weight $0.25$. The weighted changes are $0.75(17\%)=12.75\%$ and $0.25(5\%)=1.25\%$. Their sum is $14.00\%$, giving a factor of $1.14$ and a rate of $\$1{,}000(1.14)=\$1{,}140$.

Bühlmann calculation. $K=900/100=9$, so $Z=9/(9+9)=0.5$. The weighted changes are $0.5(17\%)=8.5\%$ and $0.5(5\%)=2.5\%$. Their sum is $11.0\%$, giving a factor of $1.11$ and a rate of $\$1{,}000(1.11)=\$1{,}110$.

Check and interpretation. Both indications fall between the manual 5% and experience 17% changes. The methods differ because their weights are based on different information: one uses a stated full exposure standard, while the other uses estimated within-risk and between-risk variation. The difference in indicated rates is $30. The common mistake avoided is to treat a credibility weight from one method as if it were automatically the weight prescribed by another.

Common Exam Traps and Practical Checks

A blend is only as reliable as its inputs. Before calculating, confirm that both changes use the same current rate, coverage, exposure basis, and prospective period. In particular, a rate change should not be blended with a loss ratio or a pure premium unless it has first been converted to a comparable rate change. The indicated experience should also be developed on an appropriate basis before credibility is applied; credibility does not repair inconsistent trend, development, or exposure treatment.

Misweighting the manual indication. If experience has weight $Z$, the manual weight is $1-Z$, not $Z$. Check that the two weights sum to one.
Confusing a change with a factor. A 12% increase is a factor of 1.12. Apply 1.12 to the current rate.
Dropping a weighted contribution. For equal weights on +16% and +4%, the blend is $0.50(16\%)+0.50(4\%)=8\%+2\%=10\%$, not 8%. The second contribution is still part of the weighted average.
Using exposure count without checking comparability. The limited fluctuation square-root rule requires a suitable exposure measure and assumptions about independent, comparable observations. Raw policy count may not be appropriate if policy sizes or coverage periods differ materially.
Selecting a method to obtain a desired answer. The method and its parameters should be justified by the model and available data, not by which result is more attractive.

Key takeaway: First put the manual and experience indications on a common basis. Then determine $Z$ using a justified credibility method, calculate $Zc_E+(1-Z)c_M$, and apply the resulting factor to the current rate. Check that the result lies between the two changes and that its arithmetic is consistent in both percentage and factor form.

Check Your Understanding

For each problem, determine the indicated rate or rates requested. Unless stated otherwise, the examples are hypothetical one-period pricing calculations with no interest or mortality assumptions.

A current annual rate is $500. The experience indication is a 12% increase, the manual indication is a 2% increase, and $Z=0.40$. Find the blended change and rate. (Answer: 6.0% increase; $530.00)
A portfolio has 36 comparable independent exposure units, with full credibility standard $N_0=144$. The experience indication is +22%, the manual indication is +10%, and the current rate is $250. Using the limited fluctuation square-root rule, find the blended change and rate. (Answer: $Z=0.50$; 16.0% increase; $290.00)
For a Bühlmann model, $n=6$, $\mathrm{EPV}=480$, and $\mathrm{VHM}=120$. The experience indication is +15%, the manual indication is +5%, and the current rate is $800. Find $K$, $Z$, the blended change, and the indicated rate. (Answer: $K=4$; $Z=0.60$; 11.0% increase; $888.00)
A current rate is $1,250. The experience indication is an 8% decrease, the manual indication is a 4% increase, and $Z=0.70$. Calculate the blended change and rate, retaining the sign of each indication. (Answer: 4.4% decrease; $1,195.00)
A limited fluctuation standard is $N_0=64$ exposure units. At $n=16$, the experience indication is +20%, the manual indication is +8%, and the current rate is $150. Calculate the indicated rate at $n=16$, then calculate it if exposure increases to $n=64$ while the two indications and current rate remain unchanged. (Answer: at $n=16$, $Z=0.50$, blended change 14%, rate $171.00; at $n=64$, $Z=1$, blended change 20%, rate $180.00)

What You'll Learn