Separating Frequency Trend From Exposure Change

Why Claim Counts Need an Exposure Denominator

A portfolio can produce more claims simply because it insured more risk for more time. If claim counts rise between two periods, the count change alone does not show whether the underlying frequency changed. To assess frequency, relate claims to the amount of earned exposure in each period, using the same exposure definition and units throughout.

For example, a hypothetical insurer might write twice as many vehicle-years in one period as in another. If the risk mix and frequency per vehicle-year are unchanged, the expected number of claims doubles as well. That is portfolio growth, not a deterioration in underlying frequency.

Definition: Let \(N_j\) be the observed claim count and \(w_j\) the earned exposure in period \(j\), measured in a stated, consistent exposure unit. The observed claim frequency per exposure unit is \(f_j=N_j/w_j\), provided \(w_j>0\). An exposure unit might be a policy-year, vehicle-year, or another unit appropriate to the coverage being analyzed.

The denominator must represent the risk actually exposed during the period. A policy in force for half a year contributes half a policy-year if policy-years are the chosen basis. The exposure measure should be defined consistently across periods; changing from policy-years to vehicle-years, for instance, changes what the frequency measures.

Observed frequency is a rate calculated from data, not a guarantee of the underlying expected frequency. Counts fluctuate. For a standard claim-count model, one may assume \(N_j\sim\operatorname{Poisson}(w_j\lambda_j)\), where \(\lambda_j\) is the expected claim frequency per exposure unit in period \(j\). This model assumes a fixed rate within each period and independent Poisson claim counts conditional on the exposure and rate. It is not appropriate without qualification if there is material dependence, overdispersion, or unmodeled variation in risk.

Separating Exposure Change From Frequency Change

For two periods, compare the claim counts, exposures, and observed frequencies separately. The count ratio has an exact decomposition:

\frac{N_2}{N_1} = \frac{w_2}{w_1} \cdot \frac{f_2}{f_1}, \qquad \text{when }N_1>0,\ w_1>0,\ w_2>0.

This identity follows directly from \(N_j=w_j f_j\). Substituting \(N_2=w_2f_2\) and \(N_1=w_1f_1\) into the left-hand side gives the product of the exposure ratio and frequency ratio. Thus, the claim-count change can be described as an exposure-volume factor multiplied by a frequency factor. The identity describes observed values; it does not establish that an observed frequency difference is statistically reliable or caused by any particular event.

When periods have different lengths, compare rates per unit time or calculate an annualized trend. If the frequency changes by a factor \(R\) over \(k\) years, the compound annual frequency trend is \(R^{1/k}-1\), assuming a constant proportional change per year over that interval. This annualization is a summary of the endpoints, not proof that the frequency changed smoothly every year.

Worked Example: More Claims Because of Growth, Not Frequency

A hypothetical book of personal watercraft coverage has 20,000 earned vehicle-years and 240 claims in Period 1. In Period 2 it has 25,000 earned vehicle-years and 330 claims. Calculate the observed frequency in each period, separate the count change into its exposure and frequency factors, and find the Period 2 count expected if frequency had stayed at the Period 1 level.

Method and assumptions. Use claims per vehicle-year. For interpreting expected counts, assume a Poisson count with mean \(w_j\lambda_j\), where \(w_j\) is vehicle-years and \(\lambda_j\) is the period's expected frequency. The model assumes a constant rate within each period and independent Poisson claims. No interest or mortality assumption is relevant.

Period 1 frequency. Divide claims by exposure:

f_1=\frac{240}{20{,}000}=0.012 \quad\text{claims per vehicle-year.}

Period 2 frequency.

f_2=\frac{330}{25{,}000}=0.0132 \quad\text{claims per vehicle-year.}

Frequency factor. The observed frequency ratio is \(0.0132/0.012=1.10\), so frequency is 10% higher. The exposure factor is \(25{,}000/20{,}000=1.25\), or a 25% increase. Their product is \(1.25\times1.10=1.375\), which agrees with the count ratio \(330/240=1.375\).

Count at unchanged frequency. If Period 2 had the Period 1 frequency, its expected claim count would be:

$$ 25{,}000(0.012)=300. $$

The observed count is 330, which is 30 above that unchanged-frequency benchmark. The decomposition is consistent: the 25% exposure increase alone would take the count from 240 to 300; the additional frequency factor of 1.10 takes 300 to 330. Check: \(240\times1.25\times1.10=330\). The common mistake is to call the full 37.5% count increase a frequency increase; the exposure-adjusted frequency increase is 10%.

Using Exposure in Trend Analysis

When estimating frequency trend across multiple periods, exposure belongs in the analysis rather than being treated as if every period had equal volume. One standard model is a Poisson log-linear model:

Model: For period \(j\), assume \(N_j\sim\operatorname{Poisson}(w_j\lambda_j)\), with \(\log(\lambda_j)=\alpha+\beta t_j\). Here \(t_j\) is time measured in years from a stated reference period, \(\alpha\) is the log expected frequency at \(t_j=0\), and \(\beta\) is the annual log-frequency trend. The log exposure \(\log(w_j)\) is an offset, so the model has \(\log(E[N_j])=\log(w_j)+\alpha+\beta t_j\). The frequency trend per year is \(e^\beta-1\).

The offset matters because the expected count is exposure times frequency. If exposure doubles but the expected frequency stays constant, the expected count doubles. The model therefore compares frequency after accounting for each period's exposure rather than interpreting raw count differences as trend. The Poisson model assumes that each period's count has variance equal to its mean, conditional on its exposure and modeled rate. If the data show more variability than that assumption permits, uncertainty estimates from a basic Poisson model may be too small.

A direct ratio of observed frequencies can summarize two periods when both counts and exposures are positive. A regression model is useful when several periods are available and one wants to estimate a trend using all of them. It can also include a period with zero claims, for which the direct log of observed frequency is undefined. The regression model still requires appropriate assumptions and a sensible assessment of fit; exposure adjustment does not remove random variation or correct a poorly specified model.

Worked Example: Annualizing a Frequency Change Across Two Years

A hypothetical specialty equipment portfolio records 1,000 claims over 100,000 earned equipment-years in 2022 and 1,452 claims over 120,000 earned equipment-years in 2024. Calculate the two observed frequencies, the cumulative frequency change, the annualized trend, and the 2024 count expected at the 2022 frequency.

Method and assumptions. Use equipment-years as the consistent exposure unit. Assume a Poisson count in each period with mean equal to exposure multiplied by the period frequency. Treat the periods as two years apart and use a constant compound annual change only for annualizing the endpoint ratio. No interest or mortality assumption is relevant.

Observed frequencies.

f_{2022}=\frac{1{,}000}{100{,}000}=0.0100, \qquad f_{2024}=\frac{1{,}452}{120{,}000}=0.0121.

The cumulative frequency factor is \(0.0121/0.0100=1.21\), so the frequency increased 21% over two years. The annualized trend is:

1.21^{1/2}-1=1.10-1=0.10=10\%\text{ per year.}

Count decomposition. Exposure increased by \(120{,}000/100{,}000=1.20\), while frequency increased by 1.21. The total count factor is \(1.20\times1.21=1.452\), matching \(1{,}452/1{,}000=1.452\).

Unchanged-frequency benchmark. At the 2022 frequency, expected 2024 claims would be \(120{,}000(0.0100)=1{,}200\). Applying the observed frequency factor gives \(1{,}200(1.21)=1{,}452\). Check: the decomposed count matches the observed 2024 count. The trap is to compare 1,452 directly with 1,000 and report a 45.2% frequency increase; that is the count increase, combining 20% more exposure and 21% higher frequency.

Partial-Year Exposure and Pooled Rates

Exposure adjustments are especially important when policies are earned for different portions of a year. A count for a period should be divided by the exposure earned in that period, not by the number of policies at a single date. Consider a hypothetical portfolio with 400 policy-years and 20 claims in one period, followed by 300 policy-years and 18 claims in the next. The raw count falls by 10%, but frequency rises from \(20/400=0.05\) to \(18/300=0.06\) claims per policy-year, an increase of \(0.06/0.05-1=20\%\). A reduction in exposure can conceal a higher rate when only counts are reviewed.

When combining periods with different exposure volumes, the pooled frequency is total claims divided by total exposure. It is an exposure-weighted average of the period frequencies, not generally their simple arithmetic average:

f_{\text{pooled}} = \frac{\sum_j N_j}{\sum_j w_j} = \frac{\sum_j w_j f_j}{\sum_j w_j}.

The equality follows because \(N_j=w_jf_j\) for each period. Each period's rate receives weight proportional to its exposure. This pooled rate answers a different question from the average of the period-specific rates, which gives every period equal weight regardless of its volume.

Worked Example: Correcting a Misleading Count Comparison

A hypothetical small commercial property program has 8,000 earned location-years and 96 claims in 2023. In 2024 it has 9,600 earned location-years and 132 claims. An analyst says the claim count rose 37.5%, so frequency rose 37.5%. Calculate the correct frequency change and explain the error.

Method and assumptions. Calculate claims per location-year in each period. For expected-count interpretation, assume Poisson counts with means \(w_j\lambda_j\) and a constant frequency within each period. This is a hypothetical example; no interest or mortality assumption applies.

The observed frequencies are:

f_{2023}=\frac{96}{8{,}000}=0.012, \qquad f_{2024}=\frac{132}{9{,}600}=0.01375.

The frequency factor is \(0.01375/0.012=1.1458333\), so the observed frequency increase is approximately 14.58%. The exposure factor is \(9{,}600/8{,}000=1.20\), a 20% increase. Their product is \(1.20\times1.1458333=1.375\), which equals the count factor \(132/96=1.375\).

At the 2023 frequency, the 2024 exposure would produce \(9{,}600(0.012)=115.2\) expected claims. The higher observed frequency factor gives \(115.2(1.1458333)=132\). The analyst's 37.5% figure is the count change, not the rate change. Check: \(96\times1.20\times1.1458333=132\), subject to displayed rounding. A frequency estimate need not be a whole number of claims because it is an expected count benchmark.

Interpretation and Common Traps

Separating frequency from exposure is an essential first step, but it does not establish why frequency changed. The comparison is most informative when exposure units, coverage definitions, claim-count rules, and the broad risk population are comparable across periods. If the composition of the portfolio changes, the aggregate frequency can change even when the frequency within each underlying group does not. That is a distinct issue from exposure volume and requires an explicit treatment of portfolio mix.

An observed frequency increase is also subject to random fluctuation. A difference between \(N_1/w_1\) and \(N_2/w_2\) is a descriptive comparison; by itself it does not show that the underlying rates differ beyond expected sampling variation. A modeled trend should be interpreted with its assumptions and uncertainty in view.

Using counts as rates. A larger count may reflect more exposure rather than more claims per exposure unit.
Using inconsistent exposure units. Frequencies calculated on different denominators are not directly comparable.
Ignoring partial-year earning. Use earned exposure for the period, not just a count of policies written or in force on a selected date.
Giving every period equal weight when pooling. The pooled rate is total claims divided by total exposure.
Confusing description with inference. A calculated rate change is not, by itself, evidence of a statistically meaningful or causal change.

Key takeaway: First calculate frequency as claim count divided by consistent earned exposure. Then separate the exposure factor from the frequency factor. In a count model, include exposure in the expected count—commonly as a log-exposure offset—rather than treating claim counts from unequal exposure periods as directly comparable.

Check Your Understanding

For each hypothetical scenario, keep claim count, earned exposure, and frequency distinct.

A book has 12,000 earned policy-years and 144 claims in one period, then 15,000 earned policy-years and 195 claims in the next. What are the two observed frequencies and the frequency percentage change? (Answer: 0.012 and 0.013 claims per policy-year; frequency increased 8.33%.)
A portfolio has 50,000 earned machine-years and 400 claims in 2021, then 70,000 earned machine-years and 616 claims in 2024. What is the cumulative frequency change and the compound annual trend over the three-year interval? (Answer: 10% cumulative; approximately 3.23% per year.)
A hypothetical program's exposure rises from 8,000 to 9,600 location-years while its frequency rises from 0.012 to 0.01375 claims per location-year. If the first-period count was 96, what is the second-period count implied by the two factors? (Answer: 132 claims; the count factor is \(1.20\times1.1458333=1.375\).)
In one period, 100 earned units produce 3 claims; in another, 900 earned units produce 45 claims. What is the pooled frequency, and why is it not the unweighted average of the two period rates? (Answer: 0.048 claims per unit; the pooled rate weights each period by exposure, whereas the unweighted average gives each period equal weight.)
Two periods have exposures of 25,000 and 30,000 vehicle-years and counts of 100 and 180. Calculate the count increase and the frequency increase, and state the exposure factor. (Answer: counts increased 80%; frequency rose from 0.004 to 0.006, a 50% increase; the exposure factor is 1.20.)

What You'll Learn