Changing Risk Means Across Periods

Why a Constant Risk Mean Matters

Bühlmann-Straub credibility uses observations from multiple periods to estimate a risk’s expected rate. Its interpretation depends on a substantive assumption: after allowing for exposure, a given risk has the same expected rate in each period. This is not merely a convenient formula condition. If the expected rate changes over time, a weighted average of historical rates may describe the historical mix of conditions rather than the risk’s expected rate on the basis needed for a future decision.

Trend, changes in the composition of exposure, and changes in the risk itself can all create apparent or actual changes in a risk’s mean. The appropriate response depends on which change occurred. A known common trend may be removed by adjusting rates to a common time basis. A known mix change may call for standardization to a fixed mix. A change in the risk’s underlying characteristics may require a separate adjustment or a new classification. If the change is unknown, a credibility estimate alone cannot identify and correct it.

Definition: In the Bühlmann-Straub model, let $Y_{ij}$ be the observed rate for risk $i$ in period $j$, with exposure $w_{ij}>0$. Conditional on the risk parameter $\Theta_i$, the standard model assumes $E[Y_{ij}\mid\Theta_i]=\mu(\Theta_i)$ in every period and $\operatorname{Var}(Y_{ij}\mid\Theta_i)=\sigma^2(\Theta_i)/w_{ij}$. The conditional mean $\mu(\Theta_i)$ is constant across periods. The usual credibility calculation also assumes that period observations are conditionally uncorrelated given $\Theta_i$.

Here, a “rate” might be loss cost per exposure unit. The model allows rates to vary from period to period because of process variation; it does not allow the conditional expected rate for the same risk to drift systematically without changing the model. The earlier material on dependence between periods concerns the separate issue of correlated observations. Even if observations are conditionally uncorrelated, a changing conditional mean can still invalidate a constant-mean interpretation.

What a Changing Mean Does to a Credibility Estimate

For a risk with total exposure $W_i=\sum_j w_{ij}$, the exposure-weighted observed rate is

\bar{Y}_i=\frac{\sum_j w_{ij}Y_{ij}}{W_i}.

Under the constant-mean assumption, its conditional expectation is $\mu(\Theta_i)$. If instead the period-specific conditional mean is $\mu_{ij}$, then

E[\bar{Y}_i\mid\Theta_i] =\frac{\sum_j w_{ij}\mu_{ij}}{W_i}.

Thus, the weighted average estimates an exposure-weighted average of the period means, not necessarily the mean for the next period. The distinction is important: more exposure gives a historical period more weight, but it does not make that period’s conditions representative of a different future basis.

When all observations have been placed on a comparable basis and the model assumptions hold, the Bühlmann-Straub estimate has the form

\widehat{\mu}_i=Z_i\bar{Y}_i+(1-Z_i)\mu,\qquad Z_i=\frac{W_i}{W_i+K},\qquad K=\frac{\text{EPV}}{\text{VHM}}.

Here $\mu$ is the collective mean on the same basis as $\bar{Y}_i$, EPV is the expected process variance per unit exposure, and VHM is the variance of hypothetical means. A conventional application also requires $K>0$ and a suitable stable collective structure. The formula does not itself adjust for trend or a change in risk composition.

Adjusting for Trend and Exposure Mix

A known trend can make rates from different periods comparable if its effect is appropriately specified. For example, if rates are expected to rise by a known dollar amount between two periods, add that amount to the earlier rate to express it on the later-period basis. For a multiplicative trend, convert each rate using the applicable trend factor. Then calculate the exposure-weighted average of the adjusted rates and use a collective mean defined on that same basis. An adjustment should not be applied twice, and the adjustment direction must match the chosen basis.

Mix changes require a different comparison. Suppose a portfolio contains two classes with different expected rates. The total rate can change solely because the proportions of exposure in the classes change, even if neither class’s rate changes. A comparison intended to measure underlying rate change should therefore use a consistent class mix, or should model the class-specific rates and exposure composition separately.

A useful way to see why the credibility coefficient has its usual form is to consider the adjusted weighted average under the model. If the conditional process variance per unit exposure is represented by EPV, then the variance of the weighted average due to process variation is $\text{EPV}/W_i$. Combining an estimate with this variance and a collective estimate whose between-risk variance is VHM gives a credibility weight equal to the between-risk variance divided by the sum of between-risk and process variances:

Z_i=\frac{\text{VHM}}{\text{VHM}+\text{EPV}/W_i} =\frac{W_i}{W_i+\text{EPV}/\text{VHM}} =\frac{W_i}{W_i+K}.

This derivation also shows the limitation: the variance calculation presumes a common conditional mean across periods. If period means change, reducing process variance by adding exposure does not remove systematic basis differences.

Worked Example: Put Two Periods on a Common Trend Basis

A hypothetical account has rates of $100 and $140 per exposure unit, with exposures 3 and 2, respectively. A known trend adjustment adds $20 to the first-period rate to express it on the second-period basis. Assume the adjusted experience is suitable for the Bühlmann-Straub model, $K=5$, and the collective mean on the second-period basis is $110. Find the adjusted observed rate and credibility estimate.

Method and assumptions. Use the given additive trend adjustment, exposure-weight the rates after adjustment, and then apply the credibility formula. No interest or mortality assumption is relevant. The example treats the observations as conditionally uncorrelated, with a stable risk mean after the adjustment.

Step 1: Adjust the first-period rate. Its second-period-basis value is $100+20=120$. The second-period rate is already $140 on that basis.

Step 2: Calculate the exposure-weighted rate. The weighted losses per unit-rate scale are $3(120)=360$ and $2(140)=280$. Total exposure is $3+2=5$, so

\bar{Y}=\frac{360+280}{5}=\frac{640}{5}=\$128.00.

Step 3: Calculate credibility. $Z=5/(5+5)=0.5$.

Step 4: Blend with the collective mean. The estimate is $0.5(128)+0.5(110)=64+55=\$119.00$.

Check. The estimate of $119.00 is halfway between $128.00 and $110.00, as $Z=0.5$ requires. The common mistake is to compute the adjusted weighted numerator as 660, producing a false mean of $132.00. The correct numerator is $3(120)+2(140)=360+280=640$; consequently the correct credibility estimate is $119.00, not $121.00.

Worked Example: Separate Mix Change from Rate Change

A hypothetical insurer tracks two equipment classes. Their expected annual rates are $80 for class A and $160 for class B, and neither class’s rate changes between periods. In the first period, exposures are 4 units of A and 1 unit of B. In the second period, exposures are 1 unit of A and 4 units of B. Calculate each period’s aggregate rate and explain the change.

Method and assumptions. Calculate each aggregate rate using that period’s exposure mix. Assume the class rates are stable and that there is no trend or change within either class. No interest, mortality, or claim distribution assumption is needed for this weighted-average calculation.

Step 1: First-period aggregate. The expected amount on the rate scale is $4(80)+1(160)=320+160=480$. Total exposure is 5, so the aggregate rate is $480/5=\$96.00$.

Step 2: Second-period aggregate. The expected amount is $1(80)+4(160)=80+640=720$. Total exposure is again 5, so the aggregate rate is $720/5=\$144.00$.

Step 3: Interpret the difference. The aggregate rate rises by $144-96=\$48.00$, or 50% relative to $96.00, even though neither class-specific rate changes. The cause is the shift toward class B, which has the higher rate.

Check and common mistake. Apply a fixed mix of 80% A and 20% B to either period: $0.8(80)+0.2(160)=64+32=\$96.00$. This confirms that the first-period rate is the fixed-mix benchmark; applying the second-period mix instead gives $0.2(80)+0.8(160)=16+128=\$144.00$. Treating the $48.00 aggregate increase as an underlying increase in both classes would confuse mix change with rate change.

When Risk Characteristics Change

A risk can change in ways that are not captured by a common trend factor or by reweighting known classes. Examples include a material change in operations, coverage, safety controls, or the set of locations insured. If the change alters the risk’s expected rate, the risk parameter is no longer adequately represented by one constant $\Theta_i$. Historical experience then reflects the former risk as well as random process variation.

If the effect is known and can be expressed on the same rating basis, adjustment may restore comparability. If it is not known, the analyst should not describe an unadjusted credibility-weighted average as an estimate of the current expected rate without qualification. Additional information, a changed classification, or a model with time-varying risk characteristics may be needed. A larger credibility weight can make the estimate more responsive to experience, but it cannot determine whether an observed change is trend, mix, a genuine shift in the risk, or random variation.

Worked Example: Adjust for a Known Change in Risk Characteristics

A hypothetical service fleet has observed rates of $90 and $120 in two periods, with exposures 2 and 3. Between periods, a documented operational change is expected to add $30 to the rate; for this example, assume the adjustment applies uniformly to the earlier period when expressing it on the later-period basis. The later-period rate is already on that basis. Let $K=5$, and let the collective mean on the later-period basis be $100. Calculate the adjusted credibility estimate and compare it with the estimate that would result from ignoring the change.

Method and assumptions. Add the known $30 change to the earlier rate, then exposure-weight both rates and use the credibility formula. Treat the adjustment as exact and assume the adjusted observations satisfy the model’s stable-mean and conditional-independence assumptions. No interest or mortality assumption is relevant.

Step 1: Adjust the earlier observation. Its later-basis rate is $90+30=\$120.00$. The second observation remains $120.00.

Step 2: Find the adjusted weighted rate. The weighted amount is $2(120)+3(120)=240+360=600$, and total exposure is $2+3=5$. Therefore, $\bar{Y}=600/5=\$120.00$.

Step 3: Find the credibility factor and estimate. $Z=5/(5+5)=0.5$, so the estimate is $0.5(120)+0.5(100)=60+50=\$110.00$.

Step 4: Compare with ignoring the change. The unadjusted weighted rate is $[2(90)+3(120)]/5=(180+360)/5=\$108.00$. Its credibility estimate would be $0.5(108)+0.5(100)=\$104.00$. The difference, $6.00, arises from failing to put the earlier experience on the later basis.

Check. After adjustment both observed rates equal $120.00, so their weighted average must also be $120.00. The adjusted estimate, $110.00, lies halfway between $120.00 and the $100.00 collective mean, consistent with $Z=0.5$. This adjustment is justified only because the example supplies a known, common $30 effect; it would not resolve an unknown or risk-specific change.

Practical Interpretation and Exam Traps

Before applying a credibility formula across periods, ask whether the rates and the collective mean refer to the same basis. A careful analysis distinguishes among:

Common time trend: adjust observations to a common period if the trend is specified and applicable.
Exposure-mix change: compare rates using a consistent mix, or retain the class-level detail needed to separate mix from rate effects.
Change in the risk itself: adjust only when the effect is adequately understood; otherwise qualify the estimate and consider whether the model’s risk definition remains appropriate.

A frequent exam trap is to calculate $Z_i$ correctly and assume that this validates the estimate. Credibility measures the weight assigned to an experience average under a model; it does not certify that observations from different periods are comparable. Another trap is to apply trend to the aggregate rate after combining periods when different periods require different adjustments. In general, adjust each period on the stated basis first, then apply its exposure weight.

Key takeaway: Bühlmann-Straub credibility averages experience around a stable conditional risk mean. If period means change, first determine whether a known trend or mix adjustment can restore a common basis. If not, interpret the estimate as a summary of historical experience, not automatically as the current or future expected rate.

Check Your Understanding

For each problem, identify the relevant basis before applying a credibility calculation.

A risk has rates of $70 and $100 at exposures 3 and 1. Add $10 to the first rate to express both on the second-period basis. If $K=4$ and the collective mean on that basis is $75, calculate the adjusted credibility estimate. (Answer: Adjusted mean $85.00; $Z=0.5$; estimate $80.00.)
Two stable classes have rates of $120 and $200. Historical exposure proportions are 75% and 25%; future proportions are 25% and 75%. Calculate the two aggregate rates and state whether the difference proves that either class rate changed. (Answer: Historical $140.00; future $180.00; no, the difference is entirely explained by mix.)
A risk has rates of $90 and $130 at exposures 2 and 3. Add $20 to the first rate to put it on the second-period basis. With $K=5$ and a same-basis collective mean of $118, calculate the credibility estimate. (Answer: Adjusted mean $122.00; $Z=0.5$; estimate $120.00.)
A documented change adds $50 to the historical rate on the current basis. Historical experience is $100, $W=4$, $K=4$, and the current-basis collective mean is $140. Find the adjusted credibility estimate and the estimate if the change is ignored. (Answer: Adjusted estimate $145.00; unadjusted estimate $120.00.)
Two periods have different aggregate rates, but neither a trend adjustment nor class-level exposure data are available. Can a unique mix-adjusted current rate be calculated from those facts alone? State the interpretation limit. (Answer: No; the cause of the difference cannot be separated from the available information, so the historical credibility estimate is not established as a current-basis mean.)

What You'll Learn