Selecting And Applying A Frequency Trend

Why Selecting a Frequency Trend Requires Judgment

A frequency trend measures how claim counts per unit of exposure change over time. It can be used to project an observed frequency to a future period, but the calculation is only as meaningful as the data behind it. Historical changes may reflect a real change in claim incidence, or they may result from changes in exposure measurement, portfolio mix, claim reporting, or operations. The earlier material on diagnosing operational shifts in frequency data helps identify such comparability issues; the next task is to select a trend that is supportable and apply it over the right span of time.

A trend is not automatically the latest observed percentage change, nor is it necessarily the average of every available estimate. The actuary should determine which observations are comparable, how much weight to give each source, whether a structural change is present, and whether the selected rate is appropriate for the projection period. The result is a reasoned assumption, not a mechanical answer produced by one formula.

Definition: A constant annual multiplicative frequency trend \(g\) is the annual effective rate of change in expected claims per unit of exposure. If frequency at a reference date is \(f_0\), frequency \(t\) years later is modeled as \(f_t=f_0(1+g)^t\). The model assumes the same frequency definition and exposure basis over time, comparable portfolio composition or an appropriate adjustment for mix, and a constant multiplicative trend over the period. It is not appropriate to apply mechanically across a material unadjusted operational or data-definition change.

Selecting a Supportable Trend

Candidate estimates may come from successive-period comparisons, a regression fitted to multiple periods, or a recent stable experience window. These approaches can produce different answers because they use different time spans and respond differently to random fluctuation and structural changes. Selection requires reviewing the evidence rather than treating any one estimate as inherently authoritative.

A useful review asks whether claim counts and exposures are defined consistently, whether the periods have comparable coverage and mix, whether operational changes plausibly altered observed counts, and whether the estimates tell a coherent story. A sudden change associated with a reporting or intake process should not be interpreted as a change in underlying claim incidence without support. Conversely, a real change in risk conditions should not be discarded merely because it makes the recent estimate differ from older experience.

When two comparable frequencies \(f_1\) and \(f_2\) are separated by \(n\) years, the constant annual trend implied by the endpoints is:

g=\left(\frac{f_2}{f_1}\right)^{1/n}-1

This is an annualized rate, not the total percentage change over the whole interval. It summarizes the endpoints under the constant multiplicative trend model; it does not establish that the trend was actually constant in every intervening year.

Justification: Under the stated model, the later frequency equals the earlier frequency multiplied by one annual factor for each year: \(f_2=f_1(1+g)^n\). Dividing by the positive earlier frequency gives \(f_2/f_1=(1+g)^n\). Taking the positive \(n\)th root yields \(1+g=(f_2/f_1)^{1/n}\), so subtracting one gives the annualized trend formula. The derivation depends on comparable frequencies and a constant multiplicative rate; if the observations are not comparable, the calculation is still arithmetic but its interpretation as a trend is not justified.

The trend factor over a projection period of \(t\) years is \((1+g)^t\). The time interval should be measured between the dates represented by the experience frequency and the projection frequency. If annual data are represented by exposure-weighted average dates, use the time between those average dates, not automatically the number of calendar years named in the data.

Establish comparability.
Check that claim counts and exposures use consistent definitions and identify known shifts in reporting, operations, or portfolio mix.

Review candidate evidence.
Compare estimates from relevant periods and methods. Give less weight to periods affected by unexplained volatility or known breaks.

Select and document a rate.
State the selected annual trend and the evidence supporting it. Do not imply more precision than the data justify.

Apply the correct duration.
Measure years from the experience reference date to the projection reference date, compound the annual factor, and then apply it to frequency or expected counts as required.

Worked Applications

Worked Example: Annualizing Comparable Endpoint Frequencies

A hypothetical specialty vehicle portfolio has 12,500 exposure units in each of two experience periods whose average dates are two years apart. The first period has 1,000 claims and the second has 1,092 claims. Assume the counts and exposures have been normalized to the same portfolio basis and no operational break is identified. Find the endpoint-implied annual trend, select a rate rounded to one decimal place in percentage terms, and project the first frequency forward two years using that selected rate.

Method and comparability check. Both periods have the same stated exposure basis, so calculate their frequencies as claims divided by exposure:

f_1=\frac{1{,}000}{12{,}500}=0.08000,\qquad f_2=\frac{1{,}092}{12{,}500}=0.08736

There are two years between the average dates, so the endpoint ratio is \(0.08736/0.08000=1.092\). The implied annual trend is:

g=(1.092)^{1/2}-1 =\sqrt{1.092}-1 \approx 1.04498804-1 =0.04498804

Thus the implied trend is approximately \(4.498804\%\) per year. Rounded to one decimal place in percentage terms, select \(4.5\%\). The two-year factor from the selected rate is:

$$ (1.045)^2=1.092025 $$

The projected frequency is \(0.08000(1.092025)=0.087362\), or \(0.0874\) to four decimal places. This is close to the observed later frequency \(0.08736\), as expected from rounding the implied annual rate. The calculation checks: the selected two-year factor is close to the observed ratio of \(1.092\).

Common mistake avoided. Do not confuse a two-year total increase of \(9.2\%\) with an annual trend of \(9.2\%\), and do not miscalculate the square root. The annualized rate is about \(4.4988\%\), which rounds to \(4.5\%\), not \(4.4\%\).

Worked Example: Selecting Among Estimates After an Operational Change

A hypothetical insurer is evaluating frequency for a new projection. Three candidate annual trends are available: 4.2% from an adjusted multi-year regression, 4.8% from a recent comparison of two stable periods, and 7.2% from the latest year. Review shows that the latest year coincided with a new claim-intake process that changed how quickly claims were recorded; the count definition is not consistent across that break. The two earlier sources use consistent definitions and comparable portfolio bases. Select a trend from the evidence and project a frequency of 0.120 over 18 months.

Selection. The 7.2% estimate is not directly comparable because the intake change may affect recorded counts. It should not be treated as evidence of a real increase without further validation. The 4.2% and 4.8% estimates are based on comparable data and are reasonably close. For this hypothetical case, select 4.5% as a rounded central assumption, documenting that it reflects the two comparable sources rather than the disrupted latest year.

Eighteen months is \(18/12=1.5\) years. Under the constant annual multiplicative trend assumption, the factor is:

(1.045)^{1.5}=1.045\sqrt{1.045}\approx 1.0682538

Apply the factor to the starting frequency:

f_{\text{projected}}=0.120(1.0682538) \approx 0.1281905

The projected frequency is \(0.1282\) to four decimal places. As a reasonableness check, the increase is about \(6.8\%\) over one and a half years, greater than a single year's \(4.5\%\) increase but less than a two-year increase of about \(9.2\%\).

Common mistake avoided. Using the latest 7.2% merely because it is most recent would allow an unexplained recording change to drive the projection. Another error would be applying the annual factor only once for an 18-month interval; the correct duration is 1.5 years.

Worked Example: Applying One Trend to Several Projection Periods

A hypothetical portfolio has a frequency of 0.080 claims per exposure unit at the average date of its 2024 experience. After reviewing comparable historical evidence, an actuary selects a constant annual trend of 4.0%. Project expected claim counts for 12,000 exposure units at the average date of 2026 and 8,000 exposure units at the average date of 2027. Treat each calendar-year midpoint as the corresponding average date.

Method. The 2026 midpoint is two years after the 2024 midpoint. The 2027 midpoint is three years after it. Trend the base frequency separately to each target date:

f_{2026}=0.080(1.04)^2 =0.080(1.0816) =0.086528

f_{2027}=0.080(1.04)^3 =0.080(1.124864) =0.08998912

Expected claims equal projected frequency multiplied by the relevant exposure:

E[N_{2026}]=0.086528(12{,}000)=1{,}038.336

E[N_{2027}]=0.08998912(8{,}000)=719.91296

The combined expected count is \(1{,}038.336+719.91296=1{,}758.248\), or 1,758.25 claims to two decimal places. As a check, total exposure is 20,000 units, and the combined expected frequency is \(1{,}758.248/20{,}000=0.0879124\), between the projected 2026 and 2027 frequencies. This is consistent with the mix of exposures across the two years.

Common mistake avoided. Do not apply the same two-year factor to both groups. Their average dates differ, so their durations from the 2024 reference date are two and three years, respectively.

Common Selection and Application Traps

Annualizing incorrectly. A total change over several years is not the annual rate. Use the appropriate root of the cumulative ratio under the constant multiplicative model.
Using incomparable data. A change in claim recording, exposure definition, or portfolio basis can create an apparent trend that is not a change in underlying frequency.
Projecting from the wrong date. Use the interval between the frequency's experience reference date and the projection reference date. For grouped annual data, midpoints or exposure-weighted average dates may be more appropriate than period labels.
Applying a factor rather than a rate as if they were interchangeable. An annual trend of 4.5% corresponds to a factor of 1.045 per year. The factor over \(t\) years is \((1.045)^t\), not \(1+4.5t\%\) except as a rough approximation for small changes.
Reporting false precision. Calculations may carry several digits internally, but the selected assumption should reflect the quality and stability of the evidence.

A selected frequency trend matters because it converts experience at one point in time into an estimate at another. A defensible selection keeps data-quality judgments visible, while correct compounding ensures that the projection reflects the actual time interval. The resulting projected frequency can then support expected-count calculations, provided the exposure basis is also appropriate.

Key takeaway: Select a trend from comparable, relevant evidence; document why disrupted or weak estimates receive less weight; and apply the selected annual factor for the elapsed time between the experience and projection dates. The constant-trend formula is useful only when its comparability and stability assumptions are reasonable.

Check Your Understanding

For each problem, identify the relevant period and assumptions before calculating.

A hypothetical portfolio's comparable frequencies are 0.075 and 0.084 at average dates two years apart. Find the implied annual trend to two decimal places in percentage terms. (Answer: \(\sqrt{0.084/0.075}-1=5.83\%\) per year.)
A selected annual trend is 4.0%. A frequency of 0.110 at the experience reference date is projected three years forward. Find the projected frequency to four decimal places. (Answer: \(0.110(1.04)^3=0.1237\).)
A frequency of 0.090 is at the base date, the selected trend is 5.0%, and the target date is two years later. With 5,000 exposure units at the target date, calculate expected claims. (Answer: \(0.090(1.05)^2(5{,}000)=496.125\) expected claims.)
Two comparable, stable sources indicate annual trends of 3.0% and 3.2%. A third estimate is 7.5% from a year in which claim intake procedures changed and recorded counts are not comparable. Which evidence should carry less weight, and why? (Answer: The 7.5% estimate should carry less weight because the procedural break undermines comparability; the two stable estimates are the relevant evidence for selecting a trend.)
A base frequency of 0.065 is measured at the midpoint of one year. The target is the midpoint two calendar years later, so the elapsed time is two years; the selected annual trend is 3.6%. Find the projected frequency to four decimal places. (Answer: \(0.065(1.036)^2=0.0698\).)

What You'll Learn