Why Experience Rating Has More Than One Objective
Experience rating uses a risk’s own past experience, alongside a broader benchmark, to inform its prospective cost. Its purpose is not simply to make each risk’s indication move as much as possible when its observed losses change. Historical experience contains information about future losses, but it also contains random variation. A sound approach must balance how much it responds to risk-specific evidence against how much stability, equity, and practical usability it preserves.
These objectives can conflict. Giving individual experience more weight can make estimates more responsive to genuine differences between risks, but it can also make indicated costs more volatile when the experience is sparse or unusually favorable or unfavorable. Using a broad benchmark can stabilize estimates, but may not reflect persistent differences among risks. There is no universally best balance without specifying the purpose, data, and constraints of the rating application.
Core Objectives and Their Trade-Offs
Predictive accuracy means using information that improves estimates of future cost, rather than merely reproducing past outcomes. A high historical loss does not necessarily imply a high future expected loss: it may reflect a persistent risk characteristic, chance variation, or both. Credibility methods address this distinction by blending experience with a complement. Earlier material on Bayesian and Bühlmann credibility develops ways to quantify that blend.
Risk differentiation means that risks with materially different expected future costs can receive different indications when the data support that distinction. If a method gives every risk the same indication regardless of credible experience, it may miss useful information. Conversely, a method that treats every observed difference as a true difference can overreact to noise.
Stability concerns the extent to which indications change from one evaluation to another. Stability can help make costs more predictable and reduce the effect of random fluctuations. But stability is not an end in itself: excessive smoothing can delay recognition of a genuine change in expected cost.
Equity asks whether the costs assigned to risks reasonably reflect their expected costs, subject to the information available and applicable constraints. It is not the same as making every risk’s premium equal, nor does it mean that each risk must be charged exactly its realized losses. Realized losses vary by chance, while a prospective indication is intended to estimate future cost.
Aggregate adequacy asks whether the total indicated amount for a portfolio is consistent with the prospective cost level the rating is intended to cover. Individual indications may be reasonable yet collectively produce a total that is too high or too low. Aggregate adequacy is therefore a portfolio-level check, not a substitute for evaluating individual risk estimates.
Finally, administrative feasibility matters. A method may be statistically attractive but depend on data that are unavailable, unreliable, costly to maintain, or difficult to explain. An implementable method needs consistent data, repeatable calculations, and a clear way to communicate its results. Practical constraints can therefore limit the feasible choices even when a more complicated approach appears theoretically appealing.
| Objective | Question to ask | Potential tension |
|---|---|---|
| Predictive accuracy | Does the indication use information that helps estimate future cost? | More detail may require more data and more complex calculations. |
| Risk differentiation | Does supported evidence produce relevant differences among risks? | Too much response can turn random variation into large indications. |
| Stability | Are changes over time proportionate to credible changes in expected cost? | Smoothing can also delay recognition of genuine changes. |
| Equity | Are indicated costs reasonably related to expected costs? | Available experience may be sparse or uneven across risks. |
| Aggregate adequacy | Are total indications consistent with the intended portfolio cost level? | Improving one risk’s indication does not by itself ensure a sound total. |
| Administrative feasibility | Can the required information and calculations be applied reliably? | A simpler method may sacrifice some predictive detail. |
Why Credibility Weighting Balances Responsiveness and Stability
A common form of a credibility indication is \(P=ZY+(1-Z)m\), where \(Y\) is a risk’s experience estimate, \(m\) is a complement, and \(Z\) is the credibility assigned to the experience, with \(0\leq Z\leq1\). The complement is a benchmark appropriate to the rating context. When \(Z\) is near 1, the indication responds strongly to the risk’s own experience; when \(Z\) is near 0, it stays close to the complement. This form is useful for explaining the objective trade-off, but it does not imply that the same \(Z\) is appropriate for every risk or application.
A simple model shows why a credibility weight can balance responsiveness against random variation. Suppose the underlying expected cost \(\theta\) varies around a benchmark \(m\), with variance \(\tau^2\), and the observed experience estimate is \(Y=\theta+\varepsilon\), where the random error \(\varepsilon\) has mean zero and variance \(\sigma^2\). Assume \(\theta\) and \(\varepsilon\) are independent. Consider estimates \(P=ZY+(1-Z)m\). These assumptions are illustrative; real experience may not satisfy them, and the resulting calculation is not automatically a complete rating design.
Since \(Y=\theta+\varepsilon\), the error is \(P-\theta=(1-Z)(m-\theta)+Z\varepsilon\).
The two terms are independent and centered, so the mean squared error is \((1-Z)^2\tau^2+Z^2\sigma^2\).
Differentiating with respect to \(Z\) gives \(-2(1-Z)\tau^2+2Z\sigma^2=0\). Therefore \(Z=\tau^2/(\tau^2+\sigma^2)\).
The result makes the trade-off explicit under the stated assumptions. Greater variation in underlying expected costs, \(\tau^2\), supports more weight on individual experience; greater noise in that experience, \(\sigma^2\), supports less. The formula is justified for this particular model and mean-squared-error objective. It should not be treated as a universal rule when the assumptions or rating objective differ.
Worked Example: Balancing an Experience Signal Against a Benchmark
A hypothetical coverage has a benchmark expected loss cost of $1,000 per exposure unit. For illustration, assume the risk’s expected cost \(\theta\) has a normal prior distribution with mean $1,000 and variance 60,000 dollars squared. Assume its observed experience estimate \(Y\), conditional on \(\theta\), is normally distributed with mean \(\theta\) and variance 90,000 dollars squared. The observation is $1,300. Use the model to calculate the posterior expected cost and describe the objective trade-off. The normal model is an approximation; it may be unsuitable if assigning material probability to negative loss costs is a concern.
Method. Under this normal prior and normal observation model, the posterior mean is a credibility-weighted average. The observation variance represents uncertainty in the experience estimate, while the prior variance represents variation in underlying expected costs.
The credibility assigned to experience is
Thus the benchmark receives weight \(1-Z=0.6\). The posterior expected cost is
The experience raises the indication by $120 above the benchmark, but not by the full $300 difference between the observed estimate and benchmark. The method uses the signal while limiting the influence of its assumed noise. Check: the weights sum to 1, and the result lies between $1,000 and $1,300, as a weighted average should.
Common mistake. Reversing the weights gives \(0.6(1{,}300)+0.4(1{,}000)=\$1{,}180\). That incorrectly assigns the larger weight to experience. The variance formula assigns more weight to experience when the prior variance is large relative to observation variance, not when observation variance is large.
Evaluating Individual and Portfolio Results
An indication should be reviewed at more than one level. For an individual risk, ask whether its own experience is credible enough to justify a distinct result and whether the movement from the benchmark is proportionate to the evidence. For the portfolio, consider whether the exposure-weighted total is reasonable. These checks answer different questions: one can be satisfied while the other is not.
The following example treats indicated amounts as expected loss costs only. It does not include expenses, profit provisions, or other premium components. The exposure weighting is essential: a per-unit cost for a risk with many exposure units contributes more to the portfolio total than the same cost for a risk with few units.
Worked Example: Differentiating Two Risks and Checking the Total
A hypothetical insurer considers two risks in one coverage segment. The complement for each is $900 per exposure unit, and the same credibility weight \(Z=0.30\) applies to both. Risk A has an experience estimate of $1,500 and 20 exposure units; Risk B has an estimate of $600 and 30 exposure units. Calculate both indicated costs and compare the exposure-weighted total with the complement-based total.
Method. Apply the same credibility blend to each risk, then multiply each indicated cost by its exposure count before summing.
For Risk A,
For Risk B,
The indications differ by \(1{,}080-810=\$270\) per exposure unit, reflecting the different experience estimates while retaining the same complement and credibility weight. The total indicated amount is
There are \(20+30=50\) exposure units. At the complement alone, the total would be \(50(900)=\$45{,}000\). The indicated total is $900, or \(900/45{,}000=2\%\), higher. This is an aggregate comparison, not proof that either individual indication is accurate.
Check: the exposure-weighted average indicated cost is \(45{,}900/50=\$918\) per unit, which is $18 above the $900 complement. Common mistake. Taking the unweighted average \((1{,}080+810)/2=\$945\) treats the two risks as if they had equal exposure. The correct average is $918 because Risk B has more exposure units.
Stability Is a Choice, Not a Synonym for Accuracy
A more responsive method can produce larger changes in consecutive indications when experience shifts. That may be appropriate when the experience is informative, but it may also pass random fluctuation through to the indication. A more stable method dampens those changes; it can also understate or delay a real change in expected cost. Neither stability nor responsiveness alone establishes predictive accuracy. Their value depends on the quality of the experience signal and on the intended use of the indication.
Worked Example: Comparing Responsiveness Across Two Periods
A hypothetical risk has a benchmark of $1,000 per exposure unit. Its experience estimates in two successive evaluations are $1,400 and $600. Compare the indications using \(Z=0.25\) and \(Z=0.60\), and calculate the change between periods for each choice.
Method. Apply each fixed weight to both observations. A higher \(Z\) should produce a larger response to the same change in experience.
With \(Z=0.25\), the first indication is \(0.25(1{,}400)+0.75(1{,}000)=\$1{,}100\). The second is \(0.25(600)+0.75(1{,}000)=\$900\). The change is \(900-1{,}100=-\$200\).
With \(Z=0.60\), the first indication is \(0.60(1{,}400)+0.40(1{,}000)=\$1{,}240\). The second is \(0.60(600)+0.40(1{,}000)=\$760\). The change is \(760-1{,}240=-\$480\).
The higher credibility choice produces a larger movement, $480 rather than $200. In both cases, the first indication is above the benchmark and the second is below it. Check: the experience estimates differ by $800, so the change in indication is \(Z(600-1{,}400)\): \(-\$200\) at \(Z=0.25\) and \(-\$480\) at \(Z=0.60\). Common mistake. Applying the benchmark weight to experience reverses the intended responsiveness; the experience receives \(Z\), while the complement receives \(1-Z\).
Common Exam Traps and Practical Judgment
- Equating observed losses with expected future cost. Past experience is evidence, not a guarantee. Random variation and changes in circumstances can make direct pass-through inappropriate.
- Calling the largest experience weight the most accurate. A larger weight increases responsiveness, but can increase the effect of noise. Accuracy depends on the quality of the data and the predictive model.
- Confusing individual equity with aggregate adequacy. An exposure-weighted portfolio total does not establish that each risk’s cost is appropriately differentiated.
- Ignoring the unit of measurement. Compare per-unit amounts with per-unit amounts, and multiply by exposure before calculating a portfolio total.
- Assuming the mathematically best method is automatically implementable. Data availability, consistency, explanation, and operational burden are part of the decision.
For exam-style evaluation, state the objective being assessed, identify the evidence relevant to it, and explain the trade-off. For example, “the higher credibility weight improves responsiveness to the observed experience” is more precise than “the higher weight is better.” A complete assessment also asks whether that responsiveness is justified by predictive information, whether the resulting changes are acceptable, and whether the method can be applied consistently.
Check Your Understanding
For each problem, distinguish the calculation from the objective it helps assess.
- A hypothetical risk has a benchmark of $700 per exposure unit, an experience estimate of $1,100, and \(Z=0.25\). Calculate its indicated cost. (Answer: $800 per exposure unit.)
- Two hypothetical risks have the same $900 benchmark and \(Z=0.20\). Their experience estimates are $1,400 and $650. Calculate both indicated costs and their difference. (Answer: $1,000 and $850 per exposure unit; difference $150.)
- Under the normal credibility model, the prior variance is 40,000 and the observation-error variance is 60,000 dollars squared. The benchmark is $800 and experience is $1,300. Calculate \(Z\) and the indicated cost. (Answer: \(Z=0.40\); $1,000 per exposure unit.)
- A benchmark is $1,200. Experience changes from $1,500 to $900. Compare the change in indications at \(Z=0.20\) and \(Z=0.70\). (Answer: \(-\$120\) at \(Z=0.20\); \(-\$420\) at \(Z=0.70\).)
- A portfolio has 10 exposure units indicated at $1,050 per unit and 30 units indicated at $950 per unit. Calculate the exposure-weighted average and compare it with a $1,000 benchmark. (Answer: $975 per unit, which is $25, or 2.5%, below the benchmark.)