What Does \(n\) Count?
In the Bühlmann credibility factor, \(n\) is not automatically the number of rows in a data file, policies on a schedule, or calendar periods observed. It represents the amount of comparable exposure contributing to the experience being rated. The meaning of one observation must be specified before \(n\), EPV, and \(K\) can be interpreted together.
For the standard Bühlmann model, let \(\Theta\) denote the unknown risk parameter for a risk, and let \(Y_1,\ldots,Y_n\) be observations on that risk. Conditional on \(\Theta\), the observations are assumed to be independent and identically distributed, with conditional mean \(\mu(\Theta)=E[Y_j\mid\Theta]\) and conditional variance \(\sigma^2(\Theta)=\operatorname{Var}(Y_j\mid\Theta)\). The model’s EPV is the expected conditional variance for one observation; its VHM is the variance across risks of the conditional mean for one observation.
A unit could be one policy-year, one month of exposure, or a specified block of exposure. A count observed over a unit and a frequency per unit are not the same random variable: the latter is the count divided by its exposure. Changing between them changes the variance scale. Likewise, grouping several elementary observations into one average changes EPV even though the risk’s underlying expected rate is unchanged.
The factor’s inputs are linked. If a block contains \(m\) independent, equal elementary units and the observation is their average, then the average retains the same conditional mean as one elementary unit, while its conditional variance is divided by \(m\). Its VHM therefore remains unchanged, its EPV is divided by \(m\), and its \(K\) is divided by \(m\). The number of such blocks for a fixed amount of exposure is also divided by \(m\). This is why a consistent change of unit does not change the resulting credibility weight.
Why a Consistent Change of Unit Preserves \(Z\)
The standard factor can be justified by comparing the two sources of variation in the experience mean. Conditional on \(\Theta\), the average of \(n\) independent equal-unit observations has variance \(\operatorname{EPV}/n\). The conditional mean \(\mu(\Theta)\) varies across risks with variance VHM. The credibility weight is the fraction of the total variance in the experience mean attributable to differences among risks:
Now group \(m\) elementary units into one block and use the average within each block. Conditional independence gives \(\operatorname{EPV}_{\text{block}}=\operatorname{EPV}_{\text{unit}}/m\); because the average has the same conditional mean as an elementary observation, \(\operatorname{VHM}_{\text{block}}=\operatorname{VHM}_{\text{unit}}\). Thus \(K_{\text{block}}=K_{\text{unit}}/m\). If the original exposure was \(n_{\text{unit}}\), the block count is \(n_{\text{block}}=n_{\text{unit}}/m\). Substitution verifies the invariance:
This argument requires the units within a block to be conditionally independent and to have the same conditional mean and variance. If the units are dependent, have materially different exposure conditions, or the statistic is not an average, the stated EPV scaling need not hold. For a total rather than an average, both the EPV and VHM must be rescaled to the total’s scale before calculating \(K\).
Worked Example: Monthly Counts Versus Annualized Frequency
A hypothetical portfolio is modeled with a risk parameter \(\Theta\), the expected number of claims per month. Given \(\Theta\), monthly claim counts are independent Poisson random variables with mean \(\Theta\). Across risks, \(\Theta\) has a Gamma distribution with shape \(9\) and rate \(3\), so \(E[\Theta]=9/3=3\) and \(\operatorname{Var}(\Theta)=9/3^2=1\). Compare the credibility weights based on eight annualized observations and on the corresponding monthly observations, using the same \(96\) months of exposure. No interest or mortality assumptions are relevant.
Method and monthly unit. For a single monthly count \(Y\), conditional on \(\Theta\), the Poisson variance is \(\Theta\). Therefore \(\operatorname{EPV}_{\text{month}}=E[\Theta]=3\). The conditional mean is \(\Theta\), so \(\operatorname{VHM}_{\text{month}}=\operatorname{Var}(\Theta)=1\). Hence \(K_{\text{month}}=3/1=3\). With \(n_{\text{month}}=96\),
Annualized observation. Let \(N\) be the total number of claims in a year. Conditional on \(\Theta\), \(N\) is Poisson with mean \(12\Theta\) and variance \(12\Theta\). Define the annualized monthly frequency as \(Y_{\text{year}}=N/12\). Its conditional mean is \(E[N/12\mid\Theta]=\Theta\), and its conditional variance is \(\operatorname{Var}(N/12\mid\Theta)=12\Theta/144=\Theta/12\). Therefore \(\operatorname{EPV}_{\text{year}}=E[\Theta]/12=3/12=0.25\). The conditional mean remains \(\Theta\), so \(\operatorname{VHM}_{\text{year}}=1\), giving \(K_{\text{year}}=0.25\). Eight years give
Check. Eight years contain \(8(12)=96\) months, and both calculations give \(Z=0.9697\) to four decimal places. The common mistake is to use \(n=8\) annual observations with the monthly \(K=3\), giving \(8/(8+3)=0.7273\). That calculation mixes annual \(n\) with monthly \(K\); it does not describe either consistently defined observation basis.
Working With Exposure Weights and Grouped Observations
In practice, an exposure unit need not correspond to one complete policy observed for one complete year. If exposure accumulates proportionally and the model supports that assumption, \(n\) can be measured in equivalent units. For example, 18 policy-months may represent \(18/12=1.5\) policy-years. This is not a claim that fractional records were observed; it is a statement about the amount of exposure represented.
When a record is an average over \(m\) independent units, calculate EPV for that average, not for the underlying total. The following example uses a Poisson-Gamma model, in which proportional exposure has a standard, explicit variance relationship.
Worked Example: Four-Month Blocks of Claim Counts
For a hypothetical risk class, let the monthly claim count conditional on \(\Theta\) be Poisson with mean \(\Theta\), independently from month to month. Assume \(\Theta\) has a Gamma distribution with shape \(16\) and rate \(4\), so its mean is \(16/4=4\) and its variance is \(16/4^2=1\). The experience consists of \(24\) months, recorded as six four-month blocks. Each observation is the average monthly count within its block, not the four-month total. Find \(K\) and \(Z\). No interest or mortality assumptions are relevant.
Method. For one monthly count, \(\operatorname{EPV}_{\text{month}}=E[\Theta]=4\) and \(\operatorname{VHM}_{\text{month}}=\operatorname{Var}(\Theta)=1\), so \(K_{\text{month}}=4\). A block observation is the total count over four months divided by four. Given \(\Theta\), the total has Poisson mean and variance \(4\Theta\). Dividing the total by four gives conditional mean \(\Theta\) and conditional variance \(4\Theta/16=\Theta/4\). Thus \(\operatorname{EPV}_{\text{block}}=E[\Theta]/4=1\), while \(\operatorname{VHM}_{\text{block}}=1\). It follows that \(K_{\text{block}}=1/1=1\).
There are \(n_{\text{block}}=24/4=6\) block observations. Therefore,
As a check, the same exposure represented by monthly observations gives \(Z_{\text{month}}=24/(24+4)=24/28=6/7=0.8571\). The common mistake is to use the four-month total’s EPV, which is \(4E[\Theta]=16\), as the EPV of the block average. That would give \(K=16\), even though the modeled observation is a total divided by four. The EPV must match the actual observation definition.
Totals, Averages, and the Scale of \(K\)
Multiplying an observation by a constant changes both variance components. If \(Y^*=cY\), then the conditional variance is multiplied by \(c^2\), and the variance of the conditional mean is also multiplied by \(c^2\). Consequently, their ratio \(K\) is unchanged when both components are transformed consistently. This is different from grouping independent observations and averaging them: averaging reduces EPV relative to VHM, so \(K\) changes along with the block size.
For example, if the observation is a four-unit total rather than its average, its conditional mean is four times the unit mean. Its VHM is therefore \(16\) times the unit VHM, and the total’s EPV is four times the unit EPV under conditional independence. The ratio for that total observation is \(4\operatorname{EPV}/(16\operatorname{VHM})=K_{\text{unit}}/4\). This \(K\) belongs with an observation counted in four-unit blocks. It should not be paired with a count of elementary units.
Worked Example: Converting a Frequency Model to Ten-Policy Blocks
A hypothetical portfolio has a risk parameter \(\Theta\), the expected claim count per policy-year. Conditional on \(\Theta\), policy-year claim counts are independent Poisson variables. Assume \(\Theta\) has a Gamma distribution with shape \(12.5\) and rate \(6.25\), giving \(E[\Theta]=12.5/6.25=2\) and \(\operatorname{Var}(\Theta)=12.5/6.25^2=0.32\). A risk has \(40\) policy-years of exposure, recorded as four ten-policy blocks, with the observation in each block defined as claim count per policy-year. Find the credibility weight. No interest or mortality assumptions are relevant.
Method. For one policy-year, the conditional count variance is \(\Theta\), so \(\operatorname{EPV}_{\text{unit}}=E[\Theta]=2\). The conditional mean varies across risks with variance \(\operatorname{VHM}_{\text{unit}}=0.32\). Hence \(K_{\text{unit}}=2/0.32=6.25\). The unit-basis weight for \(40\) policy-years is
A block observation is the total count across ten policy-years divided by ten. Given \(\Theta\), the total is Poisson with mean and variance \(10\Theta\). Its per-policy-year average has conditional mean \(\Theta\) and conditional variance \(10\Theta/100=\Theta/10\). Thus \(\operatorname{EPV}_{\text{block}}=2/10=0.2\), \(\operatorname{VHM}_{\text{block}}=0.32\), and \(K_{\text{block}}=0.2/0.32=0.625\). With four blocks,
Check and trap. The two weights agree because \(40\) policy-years equal four ten-policy blocks. Using four blocks but retaining \(K_{\text{unit}}=6.25\) would give \(4/(4+6.25)=0.3902\). That is incorrect because it combines a block count with a one-policy-year \(K\).
A Consistent Setup Before Calculating
Before applying the factor, write down exactly what one observation represents and whether it is a total, a rate, or an average. Then express \(n\), EPV, and VHM on that same basis. A short audit can prevent most unit errors:
State the exposure and the quantity recorded, such as monthly count, annual count, or annualized frequency.
Use the chosen observation’s conditional distribution to calculate EPV; use the conditional mean across risks to calculate VHM.
Set \(K=\operatorname{EPV}/\operatorname{VHM}\), then express \(n\) in the same observation units.
When the underlying assumptions permit conversion, recompute EPV and \(K\) rather than carrying them over unchanged, and verify that equivalent exposure gives the same \(Z\).
The formula is not a unit-conversion shortcut that makes every choice of \(n\) interchangeable. Its assumptions matter: equal conditional means and variances across units, conditional independence, and a compatible definition of exposure. In particular, if observations are correlated within a block, averaging does not generally divide EPV by the block size. If risks have different exposure quality or the variance is not proportional to exposure, a simple count of exposure units may not be adequate.
Check Your Understanding
For each problem, first identify the observation unit before applying the Bühlmann factor. Round weights to four decimal places where needed.
- A model has EPV \(120\) and VHM \(30\) per policy-year. A risk has five policy-years of experience. Find \(K\) and \(Z\). (Answer: \(K=4\), \(Z=5/9=0.5556\))
- Monthly observations have EPV \(8\) and VHM \(2\). Experience is six four-month blocks, each recorded as the average of the four monthly observations. Find block EPV, block \(K\), and \(Z\); verify using the monthly basis. (Answer: block EPV \(=2\), block \(K=1\), \(Z=6/7=0.8571\); monthly check \(=24/(24+4)=0.8571\))
- Conditional on a risk parameter \(\Theta\), policy-year claim counts are Poisson with mean \(\Theta\). Across risks, \(\Theta\) is Gamma with shape \(6\) and rate \(4\). For \(20\) policy-years, calculate EPV, VHM, \(K\), and \(Z\). (Answer: EPV \(=1.5\), VHM \(=0.375\), \(K=4\), \(Z=20/24=0.8333\))
- One elementary unit has EPV \(500\) and VHM \(100\). An observation is redefined as the average over five independent units, and a risk has \(15\) elementary units of exposure. Calculate the block EPV, block \(K\), and credibility weight. (Answer: \(n_{\text{block}}=3\), block EPV \(=100\), block \(K=1\), \(Z=3/4=0.7500\))
- A risk has \(12\) policy-years of experience. Per policy-year, EPV is \(16\) and VHM is \(4\). The experience is regrouped into three four-policy-year blocks, each observation being the average per policy-year. A colleague uses \(n=3\) and \(K=4\). Identify the error and calculate the correct weight. (Answer: the colleague mixed block \(n\) with elementary-unit \(K\); block EPV \(=4\), block VHM \(=4\), block \(K=1\), and \(Z=3/(3+1)=0.7500\))