Estimating Claim Frequency Trends

Why Claim Counts Need an Exposure Denominator

A larger claim count does not necessarily mean that a portfolio has become more claim-prone. The count may have risen because the insurer covered more vehicles, homes, or other exposure units. To compare claim incidence across periods, the count must be related to the amount of exposure observed in each period.

For example, 120 claims from 10,000 vehicle-years represent a different frequency from 120 claims from 5,000 vehicle-years. The claim count is the same, but the first portfolio has twice as much exposure over which claims could occur. Frequency trending therefore begins by calculating a claim count per consistent exposure unit for each period.

Definition: For period \(t\), let \(N_t\) be the number of claims counted under a consistent claim-counting rule, and let \(E_t\) be the corresponding exposure in consistent exposure units. The observed claim frequency is \(\hat f_t=N_t/E_t\), measured as claims per exposure unit. The exposure period, coverage basis, and claim-counting definition must match.

An exposure unit could be a policy-year, vehicle-year, or another measure appropriate to the coverage. It is not interchangeable with premium: premium can change because of rate changes even when the amount of insured exposure does not. The count and denominator should also refer to compatible populations and periods. Dividing claims from one accident period by exposure from a different period can produce a misleading estimate.

Estimating Frequency and Its Trend

A standard model for count data assumes that claims arise independently at a constant rate per exposure unit within the group being analyzed. Under a Poisson count model, \(N_t\sim\operatorname{Poisson}(E_t\lambda_t)\), where \(\lambda_t\) is the underlying expected claim frequency in period \(t\). This model implies \(E[N_t]=E_t\lambda_t\), so dividing the observed count by exposure estimates the rate per unit.

\hat f_t=\frac{N_t}{E_t}

The formula is most directly interpretable when exposures are reasonably homogeneous within each period and the exposure measure adequately represents time at risk. If policyholders have materially different frequencies, the portfolio rate is an exposure-weighted average of their rates; it is not necessarily the rate for a fixed, unchanged risk. A changing portfolio composition can therefore affect the overall frequency even if each risk group’s own frequency is stable.

To summarize change between two observations, compare the frequencies, not the raw counts. If periods are separated by \(n\) years, the annualized compound trend \(g\) is the constant annual rate that would transform the first observed frequency into the second over those \(n\) years.

\hat f_{\text{later}}=\hat f_{\text{earlier}}(1+g)^n \qquad\Longrightarrow\qquad g=\left(\frac{\hat f_{\text{later}}}{\hat f_{\text{earlier}}}\right)^{1/n}-1

This is an endpoint annualization: it expresses the total change as a constant compound rate over the interval. It does not establish that the frequency actually changed at that constant rate in every intervening year. It is also not appropriate to apply the annual formula without adjusting for the elapsed time when the observations are more than one year apart.

Justification: By definition, an annual compound change of \(g\) multiplies a frequency by \(1+g\) each year. After \(n\) years the multiplier is \((1+g)^n\), so the later frequency equals the earlier frequency times that multiplier. Dividing by the earlier frequency and taking the \(n\)th root gives \(1+g=(\hat f_{\text{later}}/\hat f_{\text{earlier}})^{1/n}\), which yields the stated formula.

Worked Examples

Worked Example: Counts Rise as Exposure Volume Expands

A hypothetical insurer records 36 claims on 1,800 policy-years in Year 1 and 60 claims on 2,500 policy-years in Year 3. Calculate the frequency in each period, the total frequency change, and the annualized compound trend. The periods are two years apart.

Method and assumptions. Use the claim count divided by matching policy-year exposure. Assume both periods use the same coverage, claim-counting rule, and exposure definition, and that the portfolio is sufficiently comparable for an overall frequency comparison.

Year 1 frequency. Divide the claim count by exposure: \(36/1{,}800=0.0200\) claims per policy-year.

Year 3 frequency. \(60/2{,}500=0.0240\) claims per policy-year.

The frequency ratio is \(0.0240/0.0200=1.20\), so frequency increased by 20% over the two-year interval. Annualizing compounds over two years: \((1.20)^{1/2}-1\approx0.095445\), or approximately 9.54% per year.

The raw count increased from 36 to 60, a 66.7% increase, but exposure also increased from 1,800 to 2,500 policy-years. The count-only comparison exaggerates the change in claim incidence because it does not account for the larger exposure base. The answer is checked by \((1+0.095445)^2\approx1.20\), which recovers the observed frequency ratio. A common mistake is to report the 20% two-year increase as a 20% annual trend; the annualized rate is about 9.54%.

Worked Example: Annualizing Frequency Across Unequal Periods

A hypothetical specialty insurer observes 400 claims on 20,000 exposure units in 2021, 440 claims on 22,000 exposure units in 2022, and 600 claims on 24,000 exposure units in 2023. Calculate each annual frequency and the compound annual frequency trend from 2021 through 2023.

Method and assumptions. Calculate each period’s rate separately, then annualize the endpoint frequency ratio over the two years from 2021 to 2023. Assume a consistent exposure basis and claim definition. The endpoint annualization describes the average compound change, not the year-by-year path.

For 2021, \(\hat f_{2021}=400/20{,}000=0.0200\) claims per exposure unit. For 2022, \(\hat f_{2022}=440/22{,}000=0.0200\). For 2023, \(\hat f_{2023}=600/24{,}000=0.0250\).

The endpoint ratio is \(0.0250/0.0200=1.25\), a 25% increase over two years. Thus \(g=(1.25)^{1/2}-1\approx0.118034\), or approximately 11.8% per year.

The annual frequencies show that the increase did not occur evenly in the observations: the rate was unchanged from 2021 to 2022, then rose by 25% from 2022 to 2023. Nevertheless, the two-year endpoint annualization is correct because \((1.118034)^2\approx1.25\). A common mistake is to calculate the annual trend from the first and last claim counts alone. Those counts rise by 50%, but exposure rises by 20%; the frequency ratio is 1.25, not 1.50.

Diagnosing Portfolio Mix Effects

Exposure adjustment corrects for portfolio size, but it does not automatically correct for changes in the kinds of risks insured. If one group has a higher underlying frequency than another, shifting exposure toward the higher-frequency group can raise the total frequency even when neither group’s frequency changes.

For a portfolio divided into groups \(j\), the overall frequency is the sum of group claim counts divided by total exposure. Equivalently, it is a weighted average of group frequencies, with each group’s share of total exposure as its weight:

\hat f_{\text{overall},t} =\frac{\sum_j N_{j,t}}{\sum_j E_{j,t}} =\sum_j\left(\frac{E_{j,t}}{\sum_k E_{k,t}}\right)\hat f_{j,t}

The identity follows because \(N_{j,t}=E_{j,t}\hat f_{j,t}\). Substituting this into total claims and dividing by total exposure gives the exposure-weighted average. Consequently, a change in overall frequency can reflect a change in group frequencies, in exposure weights, or in both.

Worked Example: An Apparent Increase Caused by Changing Mix

A hypothetical insurer has two risk groups. Group A has a frequency of 0.010 claims per exposure unit, and Group B has a frequency of 0.040. In Period 1, the insurer has 900 exposure units in A and 100 in B. In Period 2, it has 400 in A and 600 in B. Assume the group frequencies are unchanged. Calculate period claim counts and overall frequencies, and determine whether the overall increase indicates a within-group trend.

Method and assumptions. Multiply each group’s exposure by its stated frequency to obtain its expected count; here the figures give whole-number counts. Use the same group-specific frequencies in both periods, so any change in the overall rate is attributable to the exposure mix.

In Period 1, Group A produces \(900(0.010)=9\) claims and Group B produces \(100(0.040)=4\). Total claims are \(9+4=13\), and total exposure is \(900+100=1{,}000\). The overall frequency is \(13/1{,}000=0.013\).

In Period 2, Group A produces \(400(0.010)=4\) claims and Group B produces \(600(0.040)=24\). Total claims are \(4+24=28\), and total exposure is \(400+600=1{,}000\). The overall frequency is \(28/1{,}000=0.028\). The observed overall rate has increased by \((0.028/0.013)-1\approx1.1538\), or about 115.4%.

That large aggregate change is not a within-group frequency trend: Group A remains at 0.010 and Group B remains at 0.040. To check with a fixed mix, apply Period 1’s exposure weights of 90% A and 10% B to the unchanged group rates: \(0.90(0.010)+0.10(0.040)=0.013\). The standardized frequency remains 0.013, confirming no within-group change under this mix. The common mistake is to interpret the overall increase as a change in claim propensity without checking composition. Here exposure shifted toward Group B, which has the higher frequency.

Practical Checks and Common Traps

Before interpreting an estimated frequency trend, verify that the numerator and denominator are aligned and that the rate is expressed in the same units in each period. If one rate is claims per policy-year and another is claims per vehicle-year, the numerical ratio is not a meaningful trend until the measures are made comparable. Likewise, a partial-year exposure period must not be treated as a full year when calculating elapsed time.

The Poisson model provides a useful basis for estimating a count rate, but it does not make the estimate immune to random variation. A low count over limited exposure can produce a volatile observed frequency. The endpoint formula summarizes the observed change; it does not by itself establish that the change is statistically significant or will continue. When data are sparse or the portfolio changes materially, examine additional periods and relevant risk-group frequencies before selecting a trend.

Key takeaway: Calculate frequency as claims divided by matching exposure, and compare rates rather than counts. Annualize the ratio over the actual elapsed time. Then check whether the exposure mix changed, because an aggregate frequency trend can arise without any change in the frequencies within risk groups.

Check Your Understanding

For each problem, calculate using the stated exposure unit and distinguish a total-period change from an annualized trend where requested.

A hypothetical insurer reports 315 claims on 15,000 equipment-years. What is the observed frequency in claims per equipment-year? (Answer: \(315/15{,}000=0.0210\) claims per equipment-year.)
A portfolio has 450 claims on 30,000 exposure units in 2022 and 540 claims on 30,000 units in 2024. Find both frequencies and the annualized compound trend over the two-year interval. (Answer: 0.0150 and 0.0180; \((0.0180/0.0150)^{1/2}-1\approx9.54\%\) per year.)
Claims rise from 90 on 6,000 exposure units to 120 on 10,000 units one year later. Calculate the frequency change and explain why raw counts give a different impression. (Answer: frequency falls from 0.0150 to 0.0120, a 20.0% one-year decrease, even though counts rise by 33.3%.)
Groups A and B have unchanged frequencies of 0.020 and 0.050. Period 1 has 1,500 exposure units in A and 500 in B; Period 2 has 800 in A and 1,200 in B. Find both aggregate frequencies and the Period 2 frequency standardized to Period 1’s mix. (Answer: Period 1 \(=(30+25)/2{,}000=0.0275\); Period 2 \(=(16+60)/2{,}000=0.0380\); Period 2 standardized to Period 1 mix \(=0.75(0.020)+0.25(0.050)=0.0275\), so the aggregate change is due to mix.)
A hypothetical portfolio records 10 claims on 500 exposure units in Year 1, 90 on 4,500 in Year 2, and 120 on 4,000 in Year 3. Calculate the exposure-weighted frequency across all three years and the annualized endpoint trend from Year 1 to Year 3. (Answer: pooled frequency \(=(10+90+120)/(500+4{,}500+4{,}000)=220/9{,}000\approx0.02444\); endpoint frequencies are 0.0200 and 0.0300, giving \((0.0300/0.0200)^{1/2}-1\approx22.47\%\) per year.)

What You'll Learn