Why Territory Boundaries Matter
A territory factor can only be applied consistently if each risk can be assigned to a territory under a clear rule. The boundary design therefore matters before an actuary summarizes experience or selects factors. A boundary that is statistically attractive but cannot be applied reliably to individual policies is not a useful rating boundary.
Territories are geographic groupings used to distinguish risks by location within a rating plan. Their boundaries may follow administrative units, such as counties or municipalities; postal areas; or custom geographic lines. A territory system should cover the rating area without gaps or overlaps, and it should specify how to assign a risk whose address lies near a boundary. The appropriate location can depend on the product and rating rules. For example, a hypothetical property policy might use the insured building’s location, while a hypothetical personal auto policy might use the garaging address.
Territory design balances several considerations. More detailed boundaries can distinguish local geographic patterns, but they can also divide the available experience into smaller groups. Broader boundaries may make assignment and administration simpler and provide more data within each unit, but may combine places with different characteristics. No single boundary type is best in every setting.
Geographic Detail, Exposure Density, and Volume
Exposure density describes how much exposure is concentrated in a geographic area. For example, exposures per square mile can help identify densely populated urban areas and sparsely populated rural ones. Density is not the same as total exposure volume. A large rural territory can have low density but still contain many exposures; a small, sparsely populated territory can have both low density and low total volume.
This distinction matters when considering finer boundaries. A high-density area may support several smaller geographic units because many risks are located close together. In a low-density area, drawing similarly small units may result in territories with very little experience. But density alone does not establish that a boundary is appropriate: the boundary must also be definable, assignable, and supported by available location information.
Administrative boundaries often help because they are recognizable and can be maintained in business systems. They are not automatically appropriate, however. A county line may divide a continuous urban area, while a custom line may fit a proposed risk grouping but be difficult to communicate or maintain. When a boundary changes over time, the rating plan needs a reliable way to determine which version applies to a policy at a given time.
Worked Example: Comparing Candidate Geographic Groupings
A hypothetical insurer is reviewing five adjacent map areas for a property rating plan. The areas have the following land areas and in-force exposures. For an initial design discussion, the insurer is considering joining Areas C and D into one territory. Calculate exposure density for each original area and for the proposed combined unit, then describe what the comparison does and does not establish.
| Area | Land area (square miles) | Exposures |
|---|---|---|
| A | 10 | 900 |
| B | 20 | 300 |
| C | 40 | 200 |
| D | 25 | 500 |
| E | 100 | 100 |
Method: compare exposure density while keeping total volume visible. For an area with \(E\) exposures and land area \(A\), define exposure density as \(E/A\), measured here in exposures per square mile.
For Area A, the density is \(900/10=90\) exposures per square mile. For B it is \(300/20=15\). For C it is \(200/40=5\). For D it is \(500/25=20\). For E it is \(100/100=1\).
For the proposed C–D territory, the combined exposure count is \(200+500=700\), and the combined land area is \(40+25=65\) square miles. Its density is \(700/65=10.7692\), or about \(10.77\) exposures per square mile. The unrounded arithmetic is shown; reporting the density to two decimal places is sufficient here.
Area A has the greatest density, while Area E has the least. Combining C and D produces an intermediate density, but it also creates a territory with 700 exposures rather than 200 or 500 separately. This is a useful description of the candidate grouping, not evidence that the combined area has uniform risk. Before selecting the boundary, the insurer would also consider whether C and D are geographically connected, whether the boundary can be assigned consistently, and whether local differences should be preserved.
Common mistake avoided: Treating density as if it were total volume. Area E has the lowest density, but that fact alone does not imply that it has the fewest exposures; its count is 100, compared with 200 in Area C. A boundary decision should examine both measures.
Small Territories and Statistical Stability
A finer map can create smaller exposure groups. With fewer observations, measured experience may vary more from period to period simply because of random fluctuation. A simple count model illustrates this trade-off. It is a diagnostic illustration of sampling variation, not a territory-selection rule and not a substitute for an analysis of whether adjacent locations have different underlying risk.
Under those assumptions, the standard deviation of the count is \(\sqrt{\lambda}\). The coefficient of variation, defined as standard deviation divided by mean, is therefore \(1/\sqrt{\lambda}\). Thus, a smaller expected claim count corresponds to greater relative random variation under this model.
The justification follows directly from the Poisson mean and variance: both equal \(\lambda\), so the standard deviation is \(\sqrt{\lambda}\); dividing that by the mean gives \(1/\sqrt{\lambda}\). If exposures share a common claim intensity \(q\), then for \(E\) exposures the expected count is \(\lambda=Eq\). This makes clear why dividing a fixed body of exposures into smaller territories can increase the relative fluctuation of counts in each territory.
Worked Example: Quantifying the Cost of a Fine Split
A hypothetical insurer is considering splitting a geographic area into two territories. One side has 250 exposures and the other has 650. For this illustration only, assume a common claim-count intensity of \(0.08\) claims per exposure and the Poisson model stated above. Calculate each side’s expected claim count and coefficient of variation, then calculate the corresponding quantities if the areas are combined.
Method: compare Poisson relative variability at each proposed scale. For 250 exposures, the expected count is \(\lambda_1=250(0.08)=20\). Its standard deviation is \(\sqrt{20}=4.4721\), so its coefficient of variation is \(4.4721/20=0.2236\), or \(22.36\%\).
For 650 exposures, the expected count is \(\lambda_2=650(0.08)=52\). Its standard deviation is \(\sqrt{52}=7.2111\), and its coefficient of variation is \(7.2111/52=0.1387\), or \(13.87\%\).
If the two areas are combined, they contain \(250+650=900\) exposures. The expected count is \(900(0.08)=72\), the standard deviation is \(\sqrt{72}=8.4853\), and the coefficient of variation is \(8.4853/72=0.1179\), or \(11.79\%\). Percentages are rounded to two decimal places.
The combined unit has lower relative count variability under the assumed model. The calculation does not show that the two areas should be combined: it assumes a common claim-count intensity, which would not capture a genuine difference between the areas. It only quantifies one cost of using smaller groups.
Common mistake avoided: Concluding that the combined territory is automatically the better rating choice because its coefficient of variation is lower. The model’s common-intensity assumption is not proof that the underlying risk is the same on both sides. Boundary detail and the ability to distinguish real geographic differences must be considered together.
Data Availability and Operational Assignment
A proposed boundary is only as useful as the information available to place risks on the correct side of it. Location records may use different geographic reference systems, addresses may be incomplete, and some source fields may identify only a broad area. A boundary that requires more precision than the records provide can create uncertain assignments, inconsistent treatment, or a territory that cannot be applied reliably.
Data availability also affects how practical a territory system is to maintain. A recognizable administrative boundary may be easier to implement than a custom polygon, even if the custom polygon looks more attractive on a map. For a proposed split, ask whether the location data can distinguish the two sides, whether assignment can be repeated using the same rule, and whether the boundary can be kept consistent as addresses and map references change.
Worked Example: Testing Whether a Proposed Split Is Assignable
A hypothetical insurer has 1,000 property exposures in a county that is currently rated as one territory. It is considering dividing the county into East and West territories. All 1,000 records contain a county identifier. Of these records, 420 have a verified address east of the proposed line, 340 have a verified address west of the line, 180 have only a postal-area code that crosses the line, and 60 have no usable street-level location. Determine what proportion can be placed on a side using the available location detail and recommend an initial design choice.
Method: quantify direct assignability before adopting a finer boundary. The records that can be assigned directly are the 420 East records plus the 340 West records, for \(420+340=760\) exposures. The fraction with a verified side is \(760/1{,}000=0.76\), or \(76\%\). The remaining \(180+60=240\) records, or \(24\%\), cannot be placed reliably on one side using the stated location detail.
The county boundary remains assignable for all 1,000 records because every record has a county identifier. On these facts, retaining the county as one territory is the more operationally reliable initial choice. A split might become feasible if location information improves or if a defensible, consistently applicable assignment rule becomes available for the records near or missing the proposed boundary.
This recommendation is about assignability, not about whether East and West have different underlying risk. The stated data do not establish that. The insurer would need suitable experience and analysis before making a claim about geographic differences.
Common mistake avoided: Counting only the records that are easy to place and treating them as if they represented all 1,000 exposures. Here, 76% are directly assignable, but the other 24% remain part of the population the rating plan must handle.
A Practical Framework for Defining Territories
Territory design is best treated as a sequence of decisions rather than a search for the smallest possible map units. The following process separates geographic, data, and statistical considerations so that a candidate boundary is not accepted merely because it appears precise.
Specify what location is rated for the product and how the risk is assigned when a record is incomplete, ambiguous, or near a boundary.
Ensure territories are mutually exclusive, cover the intended area, and have boundaries that can be described and maintained. Consider whether administrative units offer practical benefits.
Use density to understand geographic concentration, and separately examine total exposures in each candidate territory. Neither measure alone establishes that a territory is suitable.
Determine whether available location information supports the proposed detail and whether the boundary can be applied consistently over the relevant period.
Retain detail when it is meaningful and supportable; consider broader units when smaller territories would be difficult to assign or would leave too little experience for stable assessment.
The framework does not impose a universal minimum number of exposures or a fixed preferred map scale. Those choices depend on the product, the geographic setting, the data, and the purpose of the rating plan. A useful design makes its compromises explicit: what geographic distinctions it preserves, what information supports those distinctions, and what limitations remain.
Common Boundary-Design Traps
- Equating a small map unit with a better territory. More detail can create smaller experience groups and may exceed the precision of available location data.
- Using density as a proxy for volume. Exposures per square mile and total exposures answer different questions; review both.
- Assuming administrative boundaries are risk boundaries. They may be practical and stable, but their convenience does not prove that risks are uniform within them or different across them.
- Ignoring ambiguous assignments. A territory system must handle all eligible risks, not only those with complete location records.
- Interpreting random variation as geographic difference. Small groups can have volatile observed counts even when a common-risk model is appropriate. Conversely, larger groups can conceal real differences if unlike areas are combined.
Check Your Understanding
For each hypothetical problem, identify the relevant boundary-design consideration and check your arithmetic before comparing candidate territory structures.
- A proposed territory contains 360 exposures across 24 square miles. Calculate its exposure density. If an adjacent territory has 500 exposures across 40 square miles, which territory has the greater density, and which has the greater exposure volume? (Answer: The first territory has \(360/24=15\) exposures per square mile; the second has \(500/40=12.5\). The first has greater density, but the second has greater volume: 500 versus 360 exposures.)
- Under a Poisson claim-count model, a candidate territory has 400 exposures and a common assumed intensity of 0.05 claims per exposure. Calculate the expected count and coefficient of variation. (Answer: \(\lambda=400(0.05)=20\), and \(\operatorname{CV}=1/\sqrt{20}=0.2236\), or \(22.36\%\).)
- A proposed boundary leaves 900 exposures in a county. Street-level information assigns 630 to the north side and 180 to the south side; the remaining records have only a postal code that crosses the boundary. What proportion is directly assignable, and how many exposures are unresolved? (Answer: \(630+180=810\) are assignable, so \(810/900=90\%\) are directly assignable; 90 exposures are unresolved.)
- Two adjoining areas each have 300 exposures. Under a common Poisson claim-count intensity of 0.04 per exposure, compare the coefficient of variation for either separate area with that for the combined area. (Answer: Each separate area has \(\lambda=12\) and \(\operatorname{CV}=1/\sqrt{12}=28.87\%\); together they have \(\lambda=24\) and \(\operatorname{CV}=1/\sqrt{24}=20.41\%\). The comparison assumes a common intensity.)
- A hypothetical rating plan proposes a custom boundary that separates a dense city from a rural area, but 30% of the risks have only location codes that cross the proposed line. The current county boundary can be assigned for every risk. State one design trade-off and a defensible initial response. (Answer: The custom boundary may preserve a geographic distinction but is not directly assignable for 30% of risks; retaining the assignable county boundary initially is defensible while better location information or a consistent assignment method is assessed.)