Why Trend Selection Requires Judgment
Historical data can support several plausible trend estimates. A short period may reflect current conditions but contain little data; a long period may be more stable but include conditions that are no longer relevant. A regression may describe its historical observations well without being the best guide to the prospective period. An external indicator may add useful context, but it is not automatically a substitute for insurance experience.
Trend selection is therefore a reasoned choice among evidence, not a rule that says to use the newest estimate, the longest estimate, or the estimate from a preferred method. The goal is to select a rate that is supported by the data and appropriate to the costs expected during the prospective period. This process is aligned with the CAS Exam 5 (Basic Ratemaking and Estimated Claim Liabilities) syllabus.
What to Evaluate
Before comparing estimates, check that the underlying measures are comparable. For example, estimates should be based on consistently defined losses, exposure units, and time periods, with relevant adjustments made consistently. A difference caused by a change in measurement or adjustment is not necessarily evidence of a change in underlying trend. Earlier material on selecting historical and prospective trend periods and adjusting external indicators for insurance loss costs provides useful context for these checks.
| Consideration | Questions to ask | Typical implication |
|---|---|---|
| Length and volume | How many observations and years support the estimate? Is the result sensitive to one or two observations? | A very short or low-volume period may be unstable even if it is recent. |
| Stability | Do estimates from nearby periods or reasonable specifications tell a similar story? | Greater agreement supports a narrower selection range; disagreement calls for investigation and judgment. |
| Fit | Does the selected model describe the historical observations reasonably well for its intended use? | Poor fit weakens the case for relying heavily on that model’s estimate. |
| Relevance | Do the period and cost drivers represented in the data resemble those expected in the prospective period? | A less recent estimate may still be informative, but its relevance should be explained. |
These considerations answer different questions. A long series may improve the stability of an estimate while reducing its relevance to current conditions. A good historical fit does not establish that the same relationship will continue. Likewise, a recent estimate is not automatically persuasive if it is based on few observations. Weigh the considerations together rather than treating any one as decisive.
The estimates being compared may also share data. For example, a three-year estimate may be part of a five-year estimate, and both may use the same underlying experience. They are then not independent pieces of evidence. Counting each estimate as a separate “vote” can give too much apparent weight to a period simply because it was analyzed in several ways.
Annualized Rates and Prospective Factors
When a trend is modeled as a constant compound annual rate \(r\), the modeled level after \(t\) years is the starting level multiplied by \((1+r)^t\). This model assumes a constant annual percentage change over the interval; it does not assert that actual costs rise by exactly the same percentage in every year.
The annualized rate implied by a change from \(L(0)\) to \(L(n)\) over \(n\) years follows directly from the same assumption. Set \(L(n)=L(0)(1+r)^n\), divide by \(L(0)\), and take the \(n\)th root:
This derivation explains why annualized rates compound: each year’s percentage change applies to the level reached in the preceding year. Once a rate is selected, a factor for a prospective interval of \(t\) years is \((1+r)^t\), under the constant-rate assumption. The interval \(t\) must be expressed in years; for example, 18 months is 1.5 years.
A Practical Selection Process
A disciplined selection makes the judgment visible. First review the data and adjustments; then compare estimates on a common annual basis. Next consider which estimates are more stable, how well they describe the relevant experience, and how closely their periods and cost drivers match the prospective period. Finally, select and document a rate or range, including material uncertainty. The process does not require a formal numerical weighting scheme. If weights are used, they should represent a deliberate judgment and should not be mistaken for a universal statistical rule.
Check that the loss or cost measure, time basis, and applicable adjustments are consistent across the estimates being considered.
Record its period, data volume, method, fit, and sensitivity to the observations or assumptions used.
Consider whether the experience and cost influences represented by each estimate are appropriate for the period being priced.
Choose a rate or a supported range, explain why some evidence receives more weight, and identify what would make the selection change.
Worked Example: Reconciling Estimates With Different Periods
A hypothetical insurer is selecting an annual severity trend for a prospective period. The estimates, all calculated from consistently adjusted data, are: 3.6% from a ten-year regression, 5.0% from a five-year regression, 7.0% from a three-year annualized comparison, and 5.3% from an external cost indicator adjusted for insurance losses. The ten-year estimate fits its broad historical period reasonably well, but includes older conditions. The three-year estimate is recent but is based on relatively few claims and varies noticeably when one year is omitted. The five-year estimate has adequate volume and a reasonable fit. The adjusted indicator reflects current cost pressures, but is not based on this insurer’s claims.
Method. Compare the evidence by stability, fit, and relevance, then make a documented selection. The insurer gives the five-year estimate 70% of the judgmental weight and the adjusted indicator 30%. These weights express this hypothetical selection judgment; they are not a standard formula.
Step 1: Identify the evidence. The ten-year estimate of 3.6% is relatively broad but partly dated. The three-year estimate of 7.0% is current but unstable. The five-year estimate of 5.0% balances a relevant period with adequate volume. The 5.3% adjusted indicator corroborates the five-year result, with the limitation that it does not directly measure the insurer’s own claims.
Step 2: Apply the stated weights. The weighted judgment is
Step 3: Select and check. Rounding the selected annual trend to one decimal place gives 5.1%. This is close to the five-year estimate and the adjusted indicator, but below the unstable three-year result and above the older ten-year result. That placement is consistent with the stated reasons for giving greater weight to the five-year experience.
Common mistake avoided. Giving every estimate an equal vote would produce \((3.6\%+5.0\%+7.0\%+5.3\%)/4=5.225\%\), or 5.2% to one decimal place. This treats overlapping estimates as though they were independent and gives an unstable short-period estimate the same influence as the better-supported evidence. The equal average is not justified merely because four estimates were calculated.
Worked Example: Balancing Recent Volatility Against a Longer Record
A hypothetical specialty insurer compares annualized severity trends from the same consistent loss measure. The two-year estimate is 7.2%, the four-year estimate is 5.6%, and the eight-year estimate is 4.1%. The two-year estimate is based on low claim volume. The four-year estimate has a reasonable fit and includes most of the experience considered relevant to the upcoming pricing period. The eight-year estimate is more stable but includes several older years that are less representative of current repair costs.
Method. Select a rate using an explicit judgmental blend, placing 20% weight on the short estimate, 60% on the four-year estimate, and 20% on the eight-year estimate. The short estimate is included as evidence of recent experience, but its low volume limits its weight.
Step 1: Calculate each contribution.
Step 2: Add the contributions. The selected annual rate before rounding is \(1.44\%+3.36\%+0.82\%=5.62\%\). Rounded to one decimal place, the selection is 5.6% per year.
Step 3: Calculate a two-year factor. Under the constant compound-rate assumption, the factor is
Rounded to three decimal places, the factor is 1.115, equivalent to an increase of about 11.5% over two years. The factor is greater than \(1+2(0.056)=1.112\), because the second year’s trend applies to the increased first-year level.
Common mistake avoided. Choosing 7.2% solely because it is the newest estimate would ignore its low volume. Conversely, using 4.1% solely because it covers the longest period would overemphasize older experience. The selection explains how both observations inform, but do not dictate, the result.
Worked Example: Using an Adjusted Indicator as Supporting Evidence
A hypothetical insurer has an internal annual loss-cost trend estimate of 4.2%. A recent external repair-cost indicator, after adjustment to better reflect insurance loss costs, implies 6.0%. A broader internal estimate is 5.0%. The internal recent estimate has a good volume base and directly measures the insurer’s loss costs. The indicator is relevant to current repair prices but does not capture all influences on insured losses. The broader estimate is useful context but includes older experience.
Method. Give the recent internal estimate 60% weight, the adjusted indicator 25%, and the broader estimate 15%. This is a stated judgmental reconciliation, not a claim that these weights are statistically optimal.
Step 1: Calculate the weighted rate.
Step 2: Select the annual rate. Rounded to one decimal place, the selection is 4.8% per year. It is above the internal estimate because the adjusted indicator and broader experience support some upward allowance, but below the indicator’s 6.0% because the indicator is not a complete measure of the insurer’s losses.
Step 3: Apply the rate over 18 months. Eighteen months is 1.5 years, so the trend factor is
The factor rounded to three decimals is 1.073. As a check, it represents an increase of approximately 7.3% over 18 months, which is greater than one year’s 4.8% increase and less than two full years of trend, \(1.048^2=1.098304\).
Common mistake avoided. Using the unadjusted indicator as the selected trend would treat a related cost index as if it measured the insurer’s full loss-cost change. The indicator is supporting evidence; its relevance and limitations must be considered before it affects the selection.
Common Selection Traps
- Choosing the newest estimate automatically. Recency helps only when the estimate is sufficiently stable and relevant; a short period with few observations can be highly variable.
- Choosing the longest period automatically. More years can improve stability, but older observations may represent different cost conditions.
- Treating fit as a forecast guarantee. A model can describe historical observations well without establishing that its rate will persist in the prospective period.
- Averaging estimates without considering overlap. Estimates that reuse the same experience are not independent votes. Any averaging or weighting needs an actuarial rationale.
- Applying a simple percentage instead of compounding. For a fractional or multi-year interval, use \((1+r)^t\) under the constant-rate model, with \(t\) measured in years.
A useful selection record states the candidate estimates and periods, identifies the strongest and weakest evidence, explains why the chosen rate fits the prospective period, and notes important uncertainty. If reasonable alternative choices would materially affect the result, show that sensitivity rather than implying that the selected rate is uniquely determined by the data.
Check Your Understanding
For each problem, state the selection logic as well as the requested calculation.
- A hypothetical insurer has annualized estimates of 4.0% over eight years, 5.2% over four years, and 8.0% over two years. The two-year estimate is based on low volume; the four-year estimate has a reasonable fit and represents the period judged most relevant. If the selected rate is 5.2%, calculate the factor for 18 months, rounding to three decimals. (Answer: \(1.052^{1.5}\approx1.079\).)
- Two candidate annual trends are 4.6% from a six-year series and 6.2% from a two-year series. The shorter estimate has less data, while recent costs are considered more relevant. If a judgmental selection gives 75% weight to the six-year estimate and 25% to the two-year estimate, calculate the selected rate to one decimal place. (Answer: \(0.75(4.6\%)+0.25(6.2\%)=5.0\%\).)
- A hypothetical analysis reports estimates of 3.8% from a long period, 5.4% from a recent internal period, and 5.8% from an adjusted external indicator. The recent internal estimate has the strongest volume base; the external indicator is relevant but incomplete. Explain why simply averaging all three may be questionable, and identify which evidence should receive the greatest weight. (Answer: The estimates may overlap or have different relevance and quality, so an unweighted average is not automatically justified; the recent internal estimate should receive the greatest weight given the stated evidence.)
- A selected annual trend is 4.5%. Calculate the compound trend factor over three years, rounded to three decimals, and compare it with the simple addition of three annual percentage points. (Answer: \(1.045^3=1.141166\), so the factor is 1.141; the simple-addition factor 1.135 understates the compound factor.)
- A four-year estimate is 5.1%, a two-year estimate is 6.8%, and an adjusted current-cost indicator is 5.6%. The two-year result is unstable; the four-year data are stronger, and the indicator supports some additional upward pressure. Give one defensible qualitative selection approach. (Answer: Give the four-year estimate the greatest weight, use the adjusted indicator as supporting evidence, and limit the influence of the unstable two-year estimate; the precise selected rate depends on the documented judgment.)