Choosing Classification Granularity

Why Classification Granularity Matters

A rating plan can distinguish risks using broad groups or a large number of finely divided classes. More detail can separate risks with meaningfully different expected costs. But each additional class also creates work: the classification must be defined, assigned consistently, explained to users, and supported by data. If detail produces little useful differentiation, its complexity may not be justified.

The decision is not simply “more classes are better” or “fewer classes are safer.” The actuarial question is whether the extra distinction is sufficiently credible, interpretable, and usable to improve the rating plan. The analysis should compare class costs on a consistent basis and consider what changes when groups are combined.

Definition: Classification granularity is the degree of detail in a rating classification system. A finer system divides risks into more specific classes; a coarser system combines risks into broader classes. A cell is one resulting group of risks for which experience or an indicated factor is measured.

For this discussion, let $e_i$ denote the exposure in class $i$, and let $c_i$ denote its indicated loss cost per exposure. Loss costs must be on a comparable basis—for example, adjusted to the same period and coverage definition—before they are compared or combined. When classes are combined, their pooled loss cost is an exposure-weighted average, not generally the simple average of their individual costs.

What Pooling Does to Indicated Costs

Combining classes removes distinctions in the indicated costs of those classes. The pooled indication is appropriate as a summary for the combined group, but it will generally differ from each class’s separate indication. It therefore changes how indicated costs are allocated among the risks, even when the total indicated amount across the combined group is unchanged.

Key result: If classes have exposures $e_i$ and comparable indicated loss costs $c_i$, the combined loss cost is $$ c_{\text{pool}}=\frac{\sum_i e_i c_i}{\sum_i e_i}. $$ The total indicated loss amount is preserved by pooling: $(\sum_i e_i)c_{\text{pool}}=\sum_i e_i c_i$. The individual class indications are not preserved.

The weighting follows directly from the meaning of loss cost. Class $i$ represents an indicated loss amount of $e_i c_i$. Adding those amounts gives the total for the combined group. Dividing that total by the combined exposure gives the loss cost per exposure:

\frac{\text{combined indicated loss amount}}{\text{combined exposure}} =\frac{\sum_i e_i c_i}{\sum_i e_i}.

An unweighted average would give a small class the same influence as a large one, even though their contributions to the combined experience differ. The exposure weighting is therefore required when the target measure is loss per exposure.

Worked Example: Pooling Two Classes and Checking the Redistribution

Hypothetical data. A program has 400 exposures in Class A with an indicated loss cost of $180 per exposure, and 600 exposures in Class B with an indicated loss cost of $240 per exposure. The indications use the same coverage, period, and adjustment basis. Find the pooled cost and each class’s change in indicated loss amount if the classes are combined. Name the method and identify a common error.

Method. Use the exposure-weighted pooled loss cost. The common error is taking an unweighted average, or describing the direction of a class’s change backwards.

First calculate the separate indicated loss amounts:

\text{Class A}=400(180)=\$72{,}000,\qquad \text{Class B}=600(240)=\$144{,}000.

The combined indicated amount is $72,000 + $144,000 = $216,000, and combined exposure is 400 + 600 = 1,000. Thus:

c_{\text{pool}}=\frac{216{,}000}{1{,}000}=\$216\text{ per exposure}.

At the pooled cost, Class A’s indicated amount becomes $400(216)=\$86{,}400$, an increase of $86,400 − $72,000 = $14,400. Class B’s becomes $600(216)=\$129,600$, a decrease of $144,000 − $129,600 = $14,400. The changes net to zero: +$14,400 − $14,400 = $0. This check confirms that pooling preserves the combined indicated amount.

A simple average, $(180+240)/2=\$210$, is wrong because the class exposures differ. It would imply a combined indicated amount of $210,000 rather than $216,000.

Granularity Also Changes Factors and Premium Allocation

A class loss cost can be expressed as a relativity to a selected base cost. If the base loss cost is $c_0$, the corresponding cost ratio is $c_i/c_0$. The choice to combine classes changes the factors applied to the affected risks. The combined indication can still preserve total cost for the group, while shifting indicated amounts between its former classes.

This distinction matters when assessing a proposal. An unchanged total for a set of classes does not mean every policy is unaffected. A factor change can increase the indication for one former class and decrease it for another. Nor does the arithmetic alone establish whether the distinction is useful: the data’s support, factor clarity, and implementation consequences also matter.

Worked Example: Combining Classes and Translating the Result to Premium

Hypothetical data. Three vehicle-use classes have comparable indicated loss costs: Class A, $100 per exposure on 1,000 exposures; Class B, $140 on 300 exposures; and Class C, $190 on 200 exposures. Suppose the indicated base rate is $500 per exposure and, solely for this illustration, the indicated rate is proportional to the loss-cost relativity with other provisions held fixed. Combine B and C. Find their pooled loss cost and factor relative to A, then calculate how the indicated premium allocation changes. Name the method and a common mistake.

Method. Pool B and C using their exposure weights, then compare the old and revised indicated amounts. A common mistake is to compare only the combined total and overlook the redistribution between B and C.

The separate indicated loss amounts are $300(140)=\$42{,}000$ for B and $200(190)=\$38{,}000$ for C. Their combined amount is $80,000 on $300+200=500$ exposures, so:

c_{BC}=\frac{42{,}000+38{,}000}{500}=\$160.

Relative to Class A’s $100 loss cost, the combined factor is $160/100=1.60$. Under the stated rate illustration, B’s former factor is 1.40 and C’s is 1.90; the combined factor of 1.60 lies between them.

The indicated premium for B before combining is $300(500)(1.40)=\$210{,}000$. At the combined factor, it is $300(500)(1.60)=\$240{,}000$, an increase of $30,000. For C, the former amount is $200(500)(1.90)=\$190{,}000$; the combined amount is $200(500)(1.60)=\$160{,}000$, a decrease of $30,000.

Before combining, B and C together have $210,000 + $190,000 = $400,000 of indicated premium. After combining, they have $500(500)(1.60)=\$400{,}000$. The total is preserved, while the allocation shifts by equal and opposite amounts. This premium calculation relies on the stated proportional-rate assumption; it is not a claim that loss cost alone determines every filed premium.

Cell Size, Stability, and Interpretability

Finer classifications divide a portfolio into smaller cells. Smaller cells can show genuine differences that broad groupings conceal, but their observed experience may also be more variable. A simple model illustrates this tension. Suppose claim counts in a class follow a Poisson distribution with mean equal to exposure times a constant claim frequency, and suppose each claim has a fixed severity $m$. If a cell has $N$ claims on $e$ exposures, its estimated loss cost is $mN/e$. Under this model, the relative standard error of the estimated frequency $N/e$, using the observed count as a plug-in estimate of its mean, is approximately $1/\sqrt{N}$. This illustration assumes independent Poisson counts and fixed severity; actual loss costs can be more variable when claim sizes vary.

Worked Example: Comparing Separate and Pooled Estimates

Hypothetical data. A home-protection program has two proposed usage groups. Group 1 has 600 exposures and 30 claims; Group 2 has 300 exposures and 9 claims. Assume independent Poisson claim counts with a common fixed claim severity of $4,000, and comparable observation periods and coverage. Calculate each group’s indicated loss cost, the pooled cost, and the plug-in relative standard errors for the frequency estimates. State what pooling does and does not show.

Method. Calculate frequency as claims divided by exposures, multiply by the assumed fixed severity, and use $1/\sqrt{N}$ for the approximate relative standard error under the stated Poisson model. The common mistake is to apply that approximation to total dollars while ignoring claim-severity variability; here severity is explicitly fixed.

For Group 1, frequency is $30/600=0.05$ claims per exposure, so indicated loss cost is $0.05(\$4{,}000)=\$200$. For Group 2, frequency is $9/300=0.03$, so indicated loss cost is $0.03(\$4{,}000)=\$120$.

Together, the groups have $30+9=39$ claims on $600+300=900$ exposures. Their pooled frequency is $39/900=0.043333\ldots$, and pooled loss cost is $0.043333\ldots(\$4{,}000)=\$173.33$ per exposure, rounded to the nearest cent.

The separate plug-in relative standard errors are $1/\sqrt{30}=0.1826$, or 18.26%, and $1/\sqrt{9}=0.3333$, or 33.33%. For the pooled count they are $1/\sqrt{39}=0.1601$, or 16.01%. These calculations show that pooling gives a larger claim count for the pooled frequency estimate under the model, but also replaces two different indicated frequencies with one common frequency. They do not establish whether the observed difference between groups is real or whether the classes should be combined; that decision also requires judgment about risk differentiation and practical use.

A Practical Decision Process

Granularity selection is a balance among several considerations. A useful distinction should be supported by a credible difference in expected cost, have a clear interpretation, and be feasible to apply consistently. A distinction that is statistically measurable but difficult to define or explain may be a poor operational choice. Conversely, combining groups solely because a factor table is easier to administer can erase a meaningful risk difference.

Define the candidate distinction.
Specify what makes one class different from another and confirm that the classification can be assigned consistently.

Compare indications on a common basis.
Use consistent exposure definitions, coverage, and adjustments so that observed differences are interpretable.

Assess the value of the added detail.
Ask whether the cost difference is practically material and sufficiently supported to justify separate treatment, rather than relying on the number of displayed decimal places.

Quantify the effect of combining.
Calculate the exposure-weighted pooled indication and identify how indicated amounts would shift among the original groups.

Check usability and explainability.
Consider data collection, classification rules, administration, and whether users can understand the resulting factors.

The process is not a mechanical rule that chooses the class count with the highest apparent fit. The preferred granularity is a defensible compromise: enough detail to represent material differences, but not so much that factors become unstable, opaque, or burdensome. A decision to combine should be documented with both the cost impact and the practical reason for the chosen grouping.

Common Mistakes and Exam Traps

Using an unweighted average. Pooling loss costs requires exposure weights when loss cost is measured per exposure.
Getting the direction of change backwards. A class below the pooled cost increases when assigned the pooled cost; a class above it decreases.
Assuming pooling preserves every class amount. It preserves the combined indicated amount under the calculation, not the separate indications or their allocation.
Treating more detail as automatically better. Additional classes may distinguish risks, but they also reduce cell size and add operational complexity.
Over-interpreting a simple uncertainty calculation. A Poisson count model with fixed severity is an illustration, not a complete model of loss-cost variability.

Choosing classification granularity is therefore both a quantitative and a practical exercise. Exposure-weighted pooling provides a transparent way to calculate what combining classes does to indicated costs. The actuarial judgment lies in deciding whether the distinctions being removed are valuable enough to retain, given the support in the data and the requirements of a usable rating plan. This treatment is aligned with the CAS Exam 5 (Basic Ratemaking and Estimated Claim Liabilities) syllabus.

Check Your Understanding

For each hypothetical situation, calculate the requested quantity and check whether pooling preserves the combined indicated amount.

A class with 250 exposures has a loss cost of $200, and another with 750 exposures has a loss cost of $160. Find the pooled cost and the change in indicated amount for each class. (Answer: pooled cost $170; first class decreases by $7,500 and second increases by $7,500.)
Two classes have loss costs of $90 and $120 on 200 and 300 exposures, respectively. Use the first class as the base and combine the second class with a third class having loss cost $150 on 300 exposures. Find the pooled cost for the combined group and its relativity to the base. (Answer: pooled cost $135; relativity 1.500.)
A proposed segment has 400 exposures and 20 claims; a second has 200 exposures and 8 claims. Assume independent Poisson counts and fixed claim severity of $3,000. Find both indicated loss costs, the pooled loss cost, and the pooled plug-in relative standard error. (Answer: $150 and $120; pooled cost $140; pooled relative standard error $1/\sqrt{28}=0.1890$, or 18.90%.)
Two classes have exposures of 500 and 250, and indicated costs of $180 and $240. If combined, calculate the pooled cost and the changes in indicated amounts for both classes. (Answer: pooled cost $200; first class increases by $10,000 and second decreases by $10,000.)
A rating plan has two classes with comparable indicated costs of $110 and $170 and exposures of 800 and 200. The base cost is $110. If the classes are combined, find the pooled cost and its factor relative to the base; explain whether the combined indication preserves the separate indications. (Answer: pooled cost $122; factor 1.109; it preserves the combined indicated amount but not each class’s separate indicated amount.)

What You'll Learn