Selecting Frequency and Severity Trend Assumptions

Why Separate Frequency and Severity Trends?

A change in total losses can arise because claims are becoming more or less frequent, because the average claim is changing in size, or because both are happening. Separating these effects makes the assumptions easier to interpret and test. It also helps avoid treating a change in exposure volume as a change in underlying claim frequency.

For ratemaking, the components must be measured on consistent bases: claim counts and exposure should correspond to the same experience periods, while severity should use losses and claim counts from those same periods. Losses should be developed consistently where necessary, and changes in coverage, claim handling, or exposure mix should be considered before interpreting the observed movements as trends.

Definition: For a period with $N$ claims, total losses $L$, and exposure $E$, the observed claim frequency is $N/E$, the observed average severity is $L/N$ when $N>0$, and the pure premium is $L/E$. An annual frequency trend $T_f$ is the assumed proportional change in frequency per year; an annual severity trend $T_s$ is the assumed proportional change in average severity per year. A positive trend indicates an increase, and a negative trend indicates a decrease.

The exposure base must match the coverage being analyzed. It might be earned vehicles, insured locations, or another appropriate unit. A frequency measured per policy is not directly comparable with one measured per vehicle-year if the average number of vehicles per policy is changing. Likewise, severity comparisons should use the same treatment of development, loss adjustment expenses, limits, deductibles, and claim definitions.

Annualizing and Combining Trends

If a measure changes from $M_0$ to $M_n$ over $n$ years and is assumed to change at a constant compound annual rate $T$, then $M_n=M_0(1+T)^n$. Solving gives the annualized trend. The compound form is appropriate because each year's percentage change applies to the value after the prior year's change.

T=\left(\frac{M_n}{M_0}\right)^{1/n}-1

This formula can be applied separately to frequency and severity, provided both measurements are positive and comparable. It does not establish that a constant trend is appropriate: it summarizes the endpoints under that assumption. Intermediate observations, changes in data quality, and known operational or coverage changes should also be reviewed before selecting an assumption.

Key result: If frequency is multiplied by $1+T_f$ per year and severity by $1+T_s$ per year, then pure premium is multiplied by $(1+T_f)(1+T_s)$ per year. The annual overall loss trend is therefore $T_L=T_f+T_s+T_fT_s$, not simply $T_f+T_s$.

The reason follows directly from the definitions. Since $L/E=(N/E)(L/N)$, pure premium equals frequency times severity whenever $N>0$. After one year, those factors become $f(1+T_f)$ and $s(1+T_s)$. Their product is $fs(1+T_f)(1+T_s)$, so the proportional change in pure premium is $(1+T_f)(1+T_s)-1$. The cross-product $T_fT_s$ accounts for the fact that both components change in the same year.

\frac{\text{future pure premium}}{\text{current pure premium}} = \frac{f(1+T_f)s(1+T_s)}{fs} =(1+T_f)(1+T_s)

For $n$ years under constant selected component trends, the cumulative overall loss factor is $[(1+T_f)(1+T_s)]^n$. This is not the same as adding the two annual trend rates and applying that sum as a compound rate. For small rates, the cross-product may be small, but it is still part of the exact calculation.

Worked Examples

Worked Example: Annualizing Separate Endpoint Changes

A hypothetical property program has the following experience. In year 1, it records 800 claims over 10,000 earned exposure units, with average severity of $1,200. Three years later, it records 1,000 claims over 10,000 earned exposure units, with average severity of $1,500. Assume the periods are comparable, and estimate constant annualized frequency and severity trends from the endpoints. Then calculate the overall annual loss trend.

Method. Calculate frequency per exposure unit in each period, annualize the endpoint ratios over three years, and combine the component trends multiplicatively.

Year 1 frequency is $800/10{,}000=0.08$ claims per exposure unit. Year 4 frequency is $1{,}000/10{,}000=0.10$. The frequency ratio is $0.10/0.08=1.25$, so the annual frequency trend is:

T_f=1.25^{1/3}-1=1.077217\ldots-1=0.077217\ldots

Thus the annualized frequency trend is 7.722% (rounded to three decimal places). The severity ratio is $1{,}500/1{,}200=1.25$, so the annual severity trend is also $1.25^{1/3}-1=7.722\%$.

The overall annual loss trend is:

T_L=(1.077217\ldots)^2-1 =1.160397\ldots-1 =0.160397\ldots

The selected endpoint-based annual loss trend is 16.04%, rounded to two decimal places. As a check, pure premium was $0.08(1{,}200)=\$96$ in year 1 and $0.10(1{,}500)=\$150$ in year 4. The cumulative loss ratio is $150/96=1.5625$; its three-year annualized rate is $1.5625^{1/3}-1=16.04\%$, consistent with the component calculation.

Common mistake avoided: Dividing each endpoint change by three gives an arithmetic average change, not a compound annual trend. Also, the overall trend is not twice 7.722%; the multiplicative combination includes the cross-product.

Worked Example: Interpreting Log-Linear Component Experience

A hypothetical equipment coverage has four consecutive years of experience. Frequency per exposure unit is 0.08000, 0.08400, 0.08820, and 0.09261. Average severities are $1,000.00, $1,030.00, $1,060.90, and $1,092.73. Assume comparable exposure, claim definitions, and loss development. Use a log-linear trend interpretation to estimate the annual component trends and overall trend.

Method. A log-linear model for a positive measure $M_t$ is $\ln(M_t)=\alpha+\beta t$, where $t$ is year index. If the log values lie on a straight line, the annual multiplicative trend is $e^\beta-1$. Here each successive frequency is 1.05 times the previous one, and each successive severity is 1.03 times the previous one.

For frequency, the endpoint ratio over three years is $0.09261/0.08000=1.157625=1.05^3$. Therefore the log-linear slope is $\beta_f=\ln(1.05)=0.048790\ldots$, and $e^{\beta_f}-1=1.05-1=5.00\%$.

For severity, the endpoint ratio is $1{,}092.727/1{,}000=1.092727=1.03^3$; the displayed final severity is rounded to cents. The corresponding slope is $\beta_s=\ln(1.03)=0.029559\ldots$, giving an annual severity trend of 3.00%.

Combining the selected component rates:

$$ T_L=(1.05)(1.03)-1=1.0815-1=0.0815 $$

The overall annual trend is 8.15%. The check is direct: the modeled pure-premium factor after one year is $1.05\times1.03=1.0815$, so it rises by 8.15%, not 8.00%.

Common mistake avoided: Treating a log-linear slope $\beta$ itself as the percentage trend. The slope is on the log scale; the percentage trend is $e^\beta-1$.

Selecting Assumptions From Observed Experience

An observed trend is evidence for an assumption, not a mechanical instruction to use a particular rate. Selection requires checking that the data measure the intended changes. Consider the following questions before adopting a frequency or severity trend:

Is the denominator consistent? Use the same exposure definition and coverage basis across periods. A change in exposure mix can change the aggregate frequency even if the risk within each group is unchanged.
Are counts and losses aligned? Frequency and severity should refer to the same claim population and experience periods. A change in claim-counting rules can create an apparent frequency movement without a change in underlying risk.
Are loss amounts comparable? Apply a consistent development basis and account for changes in limits, deductibles, coverage, and loss adjustment expense treatment. A nominal severity increase does not by itself identify its cause.
Is the pattern stable? Review annual observations, not only the endpoints. A single unusual year or a structural change can make an endpoint annualization misleading.
Does the selected rate have a defensible interpretation? Distinguish observed experience from judgment about which portions are expected to continue. Document adjustments and the period used to estimate the trend.

When several comparable annual observations are available, a log-linear fit can summarize a persistent multiplicative pattern. A straight-line fit to the original dollar amounts instead describes an additive dollar change, which generally implies a changing percentage rate. Neither model is automatically right; the choice should match the intended assumption and the observed pattern.

Worked Example: Selecting Component Rates From Comparable Annual Estimates

A hypothetical portfolio analyst reviews three comparable annual intervals. The estimated frequency trends are 1%, 3%, and 2%; the corresponding severity trends are 4%, 6%, and 5%. After confirming that the intervals use consistent exposure and loss bases, the analyst decides to give each interval equal weight as a simple summary. Select component trends and calculate the overall annual trend.

Method. Use the arithmetic mean of each component's annual estimates as the explicitly chosen summary, then combine those selected annual assumptions multiplicatively. Equal weighting is a selection choice here, not a universal rule.

The selected frequency trend is $(1\%+3\%+2\%)/3=6\%/3=2\%$. The selected severity trend is $(4\%+6\%+5\%)/3=15\%/3=5\%$. Therefore:

$$ T_L=(1.02)(1.05)-1=1.071-1=0.071 $$

The selected overall annual trend is 7.10%. As a check on the averaging step, the three interval-specific combined rates are $(1.01)(1.04)-1=5.04\%$, $(1.03)(1.06)-1=9.18\%$, and $(1.02)(1.05)-1=7.10\%$. Their arithmetic mean is $21.32\%/3=7.1067\%$, which differs slightly from 7.10%. The difference occurs because averaging component trends and then multiplying is not identical to averaging the interval-specific products.

Common mistake avoided: Adding the selected 2% and 5% to report 7.00%. The exact combined result is 7.10%. Another mistake is to treat equal weighting as mandatory without first considering whether the intervals are actually comparable.

Common Traps and Practical Interpretation

Frequency and severity are useful precisely because they help explain the total loss movement. They should not be treated as independent of data quality or coverage design. For example, increasing a deductible can reduce both the number of claims reported and the average paid amount. If those changes are mistakenly read as underlying frequency and severity trends, the resulting assumptions may not describe a consistent coverage basis.

A few exam-style checks help keep the calculation sound:

State the experience measure and its denominator before calculating a trend.
Count the number of elapsed annual intervals, not the number of observed data points. Four annual observations from year 1 through year 4 span three years.
Use compound annualization for proportional changes across multiple years.
Combine frequency and severity factors before subtracting one to obtain the exact overall rate.
Check the result against the pure-premium identity: frequency times severity.

The distinctions among frequency, severity, and their combined loss trend are aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus. In a ratemaking analysis, the numerical result is only as useful as the consistency of the experience data and the reasoning behind the selected assumptions.

Check Your Understanding

For each problem, identify the appropriate component calculations before combining or interpreting the trend.

A hypothetical liability book has frequency of 0.060 claims per exposure unit and average severity of $500 in the first year. Two years later, these are 0.072 and $605. Assuming constant compound trends, calculate the annual frequency trend, annual severity trend, and overall annual loss trend. (Answer: frequency 9.5445%; severity 10.00%; overall 20.50%.)
A selected annual frequency trend is 3%, while selected annual severity trend is -2%. Calculate the annual overall loss trend and state whether pure premium rises or falls. (Answer: 0.94% increase.)
A hypothetical program records 900 claims over 15,000 exposure units in one year and 1,080 over 18,000 in the next. Average severity rises from $1,500 to $1,650. Calculate the annual frequency trend, severity trend, and overall loss trend. (Answer: frequency 0%; severity 10%; overall 10%.)
For a prospective assumption, an analyst selects frequency trend of 1.5% and severity trend of 4.5%. Calculate the exact annual overall loss trend and the amount by which it exceeds the sum of the component rates. (Answer: overall 6.0675%; it exceeds 6.00% by 0.0675 percentage points.)
Four consecutive annual frequency observations cover years 1 through 4. An analyst annualizes the endpoint ratio by taking its cube root. Explain why three, not four, is the correct exponent denominator, and give the general annualization formula. (Answer: there are three elapsed year-to-year intervals; $T=(M_4/M_1)^{1/3}-1$.)

What You'll Learn