Choosing the Experience Period Length

Why Experience-Period Length Matters

An experience period determines which years of a risk’s observed experience contribute to an experience rating. Choosing its length is a design decision with competing effects. A longer period generally includes more exposure and can reduce the influence of random fluctuations in any one year. But it also gives older experience more influence, even when the risk or its operating conditions have changed. A shorter period is more current, yet may be more sensitive to chance variation or one unusual loss year.

The choice is not simply “more data is better” or “the newest data is best.” The actuary should consider how much exposure each candidate period contains, whether older years remain relevant to the risk being rated, and how much unusual years influence the result. Those considerations inform the design; they do not, by themselves, specify a unique universally correct period length.

Definition: An experience period is the span of historical experience selected for use in determining an experience rating. Its length is the duration represented by that span. In comparing candidate lengths, an actuary should distinguish the amount of exposure included from the number of calendar years: two years with different exposure volumes do not necessarily carry equal statistical weight.

Exposure Volume and the Experience Measure

A useful comparison begins with a measure that accounts for exposure. Suppose each year $t$ has an observed loss amount $L_t$ and a corresponding expected loss amount $E_t$, measured on a consistent basis. The experience ratio for a selected set of years $W$ is the total observed loss divided by total expected loss:

R_W=\frac{\sum_{t\in W}L_t}{\sum_{t\in W}E_t}.

When the annual expected loss amounts are positive, this total ratio is also a weighted average of annual ratios $r_t=L_t/E_t$, with weights proportional to $E_t$. Thus, a year with twice the expected losses of another year has twice its weight in the combined ratio. This is why simply averaging annual ratios can be misleading when annual exposure volumes differ.

Derivation: Since $L_t=E_t r_t$, the total ratio is $R_W=\sum_{t\in W}E_t r_t/\sum_{t\in W}E_t$. The weights $w_t=E_t/\sum_{j\in W}E_j$ are nonnegative and sum to 1, so $R_W=\sum_{t\in W}w_t r_t$. This establishes that the combined result is an expected-loss-weighted average, not generally an arithmetic average of the annual ratios.

The ratio summarizes experience relative to its benchmark, but it does not establish that every historical year is equally predictive of the prospective period. Before comparing periods, the loss and expected-loss amounts must be on a consistent basis. The calculations below take those values as already prepared for comparison; they do not prescribe a method for adjusting historical amounts.

Worked Example: Comparing Two- and Four-Year Periods

A hypothetical contractor has the following annual experience ratios, with the same expected loss volume in each year: 2022, 1.20; 2023, 0.80; 2024, 2.40; and 2025, 1.00. Compare the latest two-year period with the four-year period. Assume the annual expected loss amount is $100,000 in each year. This is a deterministic calculation from stipulated data; no probability distribution, interest rate, or mortality assumption is used or applicable.

Method. Convert each annual ratio to observed losses by multiplying by the annual expected losses. For 2024, for example, observed losses are $2.40\times \$100{,}000=\$240{,}000$.

For the latest two years, 2024 and 2025, total observed losses are $\$240{,}000+\$100{,}000=\$340{,}000$, and total expected losses are $\$100{,}000+\$100{,}000=\$200{,}000$. The two-year ratio is therefore $\$340{,}000/\$200{,}000=1.700$.

For all four years, total observed losses are $\$120{,}000+\$80{,}000+\$240{,}000+\$100{,}000=\$540{,}000$. Total expected losses are $4\times \$100{,}000=\$400{,}000$, so the four-year ratio is $\$540{,}000/\$400{,}000=1.350$.

Check. With equal expected loss volumes, the results must equal the arithmetic averages of the included annual ratios: $(2.40+1.00)/2=1.700$, and $(1.20+0.80+2.40+1.00)/4=1.350$. The two-year period puts more weight on the high 2024 result. The four-year period adds volume and reduces that year’s relative influence, but also gives two older years more weight. The common mistake to avoid is reporting the two-year result as 1.000 by overlooking the 2.40 year.

How an Unusual Year Affects a Period

A loss year may be unusual because of random variation, a one-time event, or a change in the risk’s operations. The effect of including it depends not only on how extreme its ratio is, but also on its exposure weight and on which other years are included. A long period can dilute a single extreme year when other years have substantial volume; it does not guarantee that the year’s influence is immaterial.

An extreme result should not be removed solely because it is inconvenient or increases the indicated experience. Whether a loss is eligible for a particular treatment depends on the plan’s terms and the applicable data and loss adjustments. For choosing period length, the practical question is how sensitive each candidate period is to the year, and whether the year is representative of the risk being rated.

Worked Example: Unequal Exposure and a High-Loss Year

A hypothetical specialty equipment service firm has the following annual data. The ratio in each year is observed losses divided by expected losses. Compare the latest three years with all five years, and quantify the latest three-year result if the high-ratio year is excluded solely as a sensitivity check. Treat all amounts as already comparable. The data are deterministic; no probability distribution, interest rate, or mortality assumption is used or applicable.

Year, oldest to newest	Expected loss units	Annual ratio	Observed loss units
Year 1	40	0.90	36
Year 2	50	1.00	50
Year 3	60	2.80	168
Year 4	70	1.10	77
Year 5	80	1.00	80

Step 1: Latest three years. Their expected loss volume is $60+70+80=210$ units. Their observed loss total is $168+77+80=325$ units. The ratio is $325/210=1.547619$, or 1.5476 to four decimal places.

Step 2: All five years. Expected loss volume is $40+50+60+70+80=300$ units. Observed losses total $36+50+168+77+80=411$ units. The five-year ratio is $411/300=1.3700$.

Step 3: Sensitivity check. Excluding Year 3 from the latest three-year calculation leaves observed losses of $77+80=157$ and expected losses of $70+80=150$. The resulting ratio is $157/150=1.0467$. Thus, the high-ratio year materially affects the three-year result; the full five-year result is lower because it includes more expected loss volume at ratios below 2.80.

Check and interpretation. The latest three-year result, 1.5476, is higher than the five-year result, 1.3700, and the sensitivity result, 1.0467, confirms the influence of Year 3. The sensitivity calculation is diagnostic, not authorization to omit the year. A common error is to average the three annual ratios equally: $(2.80+1.10+1.00)/3=1.6333$. That is wrong here because expected loss volumes differ; the correct exposure-weighted result is 1.5476.

Volume, Statistical Stability, and Relevance

A simple model clarifies why additional years can improve stability. Let $R_t$ be an annual experience measure with a common mean $\theta$, common variance $\sigma^2$, and independence from year to year. For equal-volume years, the average over $n$ years is $\bar R_n=(R_1+\cdots+R_n)/n$. Its expected value remains $\theta$, while its variance is $\sigma^2/n$. This result depends on the stated assumptions: dependence between years, changing underlying risk, or materially different annual volumes can change the calculation.

\operatorname{Var}(\bar R_n) =\operatorname{Var}\left(\frac{1}{n}\sum_{t=1}^{n}R_t\right) =\frac{1}{n^2}\sum_{t=1}^{n}\operatorname{Var}(R_t) =\frac{\sigma^2}{n}.

The middle equality uses independence, which makes the covariance terms zero. The resulting standard error is $\sigma/\sqrt{n}$: increasing the number of equally informative years reduces random variation, but with diminishing gains. By contrast, if the underlying risk has changed, an older year may have a different mean from the current risk. Adding that year can then reduce random noise while introducing a less relevant signal.

Worked Example: Comparing Stability Under an Explicit Model

A hypothetical service portfolio has five annual, comparable experience ratios, oldest to newest: 1.40, 0.90, 1.10, 1.00, and 1.20. Compare the latest two-year and five-year averages. For an uncertainty comparison, assume the annual ratios are independent and normally distributed with common mean $\theta$ and variance $0.04$: $R_t\sim N(\theta,0.04)$. The normal model is an explicit approximation for this exercise, not a claim that all annual ratios follow a normal distribution. No interest rate or mortality assumption is relevant.

Step 1: Calculate each window’s estimate. The latest two-year average is $(1.00+1.20)/2=1.10$. The five-year average is $(1.40+0.90+1.10+1.00+1.20)/5=5.60/5=1.12$.

Step 2: Calculate standard errors under the model. The annual standard deviation is $\sqrt{0.04}=0.20$. For two years, the standard error is $0.20/\sqrt{2}=\sqrt{0.02}=0.1414$. For five years, it is $0.20/\sqrt{5}=\sqrt{0.008}=0.0894$, to four decimal places.

Step 3: Compare precision and recency. The five-year standard error is $0.0894/0.1414\approx0.6325$ of the two-year standard error, a reduction of about $1-0.6325=0.3675$, or 36.75%. The five-year estimate is only 0.02 higher in this data set, but it includes three older years. If the underlying risk has remained stable, the additional volume supports greater precision under the model. If the older years reflect materially different operations, the model’s common-mean assumption may fail and the longer estimate may be less relevant.

Check. The averages use the specified years, and the variance calculation uses the stipulated annual variance of 0.04. The common trap is to claim that five years always gives a better estimate: the variance result follows from independence and a stable common mean, not merely from counting calendar years.

A Practical Comparison of Candidate Lengths

A disciplined comparison makes the trade-offs explicit rather than relying on a rule of thumb. For each reasonable candidate length, calculate the exposure-weighted experience measure, record the total exposure included, and note the years that drive the result. Then assess whether the added older experience is still relevant to the risk being rated. A change in operations or coverage can weaken the case for using historical observations, even when those observations increase volume.

Measure the included volume.
Record the exposure or expected loss volume in each candidate period. Calendar-year count alone does not describe the statistical weight of the data.

Calculate a comparable result.
Use the same loss and exposure basis for each candidate, and weight annual experience by its relevant volume.

Inspect sensitivity.
Identify unusually high or low years and calculate how much they change each candidate result. Treat this as a diagnostic unless plan rules specify otherwise.

Assess relevance.
Consider whether older observations describe the risk that will be rated. A longer window is not automatically preferable if the risk has changed.

State the trade-off.
Explain the chosen period in terms of volume, stability, recency, and sensitivity to unusual experience. Do not present a universal optimal length without specifying the plan and assumptions.

Common Mistakes

Giving every calendar year equal weight. An arithmetic average of annual ratios is appropriate only when their relevant denominators are equal. Otherwise use the ratio of totals, or the equivalent denominator-weighted average.
Assuming more years always improve relevance. More observations can reduce random variation under appropriate assumptions, but older experience can be less representative after a risk change.
Removing a high-loss year just because it is high. A sensitivity calculation shows influence; it does not establish that the year should be excluded from the rating.
Confusing a high ratio with high volume. A very high annual ratio can have limited influence if its expected loss volume is small; a less extreme year can have substantial influence if its volume is large.
Claiming a precision result without assumptions. The variance reduction $\sigma^2/n$ requires independent annual observations with a common mean and variance and equal weighting.

Key takeaway: A shorter experience period emphasizes recency but may be more volatile; a longer period adds exposure and can stabilize results but may give less relevant years too much influence. Compare periods using consistent, exposure-weighted measures, and test the influence of unusual years before making a reasoned design choice.

Check Your Understanding

For each problem, compare the indicated experience periods using the stated data and assumptions.

A hypothetical risk has three years of expected loss volumes 50, 100, and 150 units, with annual ratios 0.80, 1.20, and 1.40, respectively. Find the three-year ratio of total observed losses to total expected losses. (Answer: 1.2333.)
Four equal-volume years have ratios, oldest to newest, 1.20, 0.80, 2.40, and 1.00. Find the latest two-year average and the four-year average. (Answer: two-year 1.700; four-year 1.350.)
A hypothetical account’s latest three years have expected loss volumes 40, 60, and 100 units and ratios 1.50, 0.90, and 1.20. Calculate the weighted three-year ratio, then recalculate it with the first of those years omitted as a sensitivity check. (Answer: three-year 1.170; excluding the first year, 1.0875.)
Annual experience measures are independent, have a common mean, and have known variance 0.09. Compare the standard errors of an equally weighted two-year average and a six-year average. (Answer: two-year 0.2121; six-year 0.1225.)
A risk’s operations changed substantially three years ago. A five-year period has more exposure than a two-year period, but includes three pre-change years. State the central trade-off when selecting between them. (Answer: the five-year period provides more volume and may reduce random variation if experience is stable, while the two-year period better reflects post-change operations; the choice depends on the relevance of the older experience.)

What You'll Learn