Introduction: Credibility as a Stability Question
A segment’s observed claims can differ from its underlying average simply because claims are random. Limited-fluctuation credibility asks a focused question: is the experience sufficiently stable to fall within an acceptable distance of its target with a specified probability? This is a way to assess the precision of observed experience, not a guarantee that the experience is representative or that a resulting rate is adequate.
The assessment begins by defining the claim measure and its target. It then states how large a deviation is tolerable and how often the measure is required to fall within that range. The probability calculation depends on an explicit model for the observations. A conclusion is meaningful only relative to those choices and assumptions.
For a target of 250 and an allowable deviation of 10% of that target, for example, the acceptable interval is \(225\) to \(275\). A requirement of 80% means that, under the stated model and repeated sampling conditions, at least 80% of samples are expected to produce an estimate in that interval. It does not mean that an individual estimate has an 80% chance of being correct in every sense, nor does it assign an 80% weight to the segment’s experience.
Why More Observations Usually Improve Stability
The basic mechanism is averaging. Suppose \(X_1,\ldots,X_n\) are independent observations with the same mean \(\mu\) and finite variance \(\sigma^2\). Their sample average is \(\bar X=(X_1+\cdots+X_n)/n\). Independence makes the variance of the sum equal to the sum of the variances, so:
Therefore the standard deviation of the average is \(\sigma/\sqrt{n}\), not \(\sigma/n\). The expected value remains \(\mu\), while the typical random variation around it shrinks as the sample grows. This derivation requires independent observations with a common finite variance. Dependence or materially different risk distributions can change the result; simply counting more records does not automatically create the same stability.
This relationship explains the stability perspective: under an appropriate model, a larger volume of observations tends to make an average fluctuate less. To decide whether it fluctuates little enough, however, the analyst must compare its distribution with the stated tolerance and probability requirement.
Worked Example: Testing Stability of a Normal Average
A hypothetical group has 25 independent claim-cost observations. Assume each observation has distribution \(N(250,100^2)\), so the mean is 250 and the standard deviation is 100. Assess whether the average is within 10% of 250 with probability at least 80%. No interest rate or mortality assumption applies to this claim-cost model.
Step 1: Find the tolerance interval. Ten percent of the target is \(0.10(250)=25\). The acceptable interval for the average is therefore \(250-25=225\) through \(250+25=275\).
Step 2: Find the distribution of the average. The average of independent normal observations is normal. Its mean is 250 and its standard deviation is \(100/\sqrt{25}=100/5=20\). Thus \(\bar X\sim N(250,20^2)\).
Step 3: Standardize the interval. The lower endpoint is \((225-250)/20=-25/20=-1.25\); the upper endpoint is \((275-250)/20=25/20=1.25\). The required probability is \(\Phi(1.25)-\Phi(-1.25)=2\Phi(1.25)-1\), where \(\Phi\) is the standard normal cumulative distribution function.
Step 4: Evaluate and decide. Using \(\Phi(1.25)\approx0.8944\), the probability is \(2(0.8944)-1=0.7888\), approximately 0.7887 using unrounded values. This is less than 0.80, so the experience does not meet the stated standard. The common mistake is to compare the tolerance of 25 with the standard deviation of an individual observation, 100; the statistic being tested is the average, whose standard deviation is 20.
Check: The event is symmetric about 250, and the two tails together have probability about \(1-0.7887=0.2113\). Thus the probability inside the interval is about \(0.7887\), consistent with the calculation.
Stability for a Claim Count or Frequency
A count is discrete, so the probability of being within a tolerance is the sum of probabilities for the acceptable integer counts. A Poisson model is one possible assumption for a count: if \(N\) is Poisson with parameter \(\lambda\), then \(E[N]=\lambda\) and \(\operatorname{Var}(N)=\lambda\). For a sufficiently large \(\lambda\), a normal approximation with mean \(\lambda\) and standard deviation \(\sqrt{\lambda}\) can be used. A continuity correction places each integer count’s approximate normal interval from one-half below it to one-half above it.
Worked Example: Stability of a Hypothetical Claim Frequency
A hypothetical portfolio has 800 exposure units and an expected 400 claims. Assume its claim count \(N\sim\operatorname{Poisson}(400)\), and assess whether the observed frequency \(N/800\) is within 10% of the target frequency \(400/800=0.5\) with probability at least 95%. Use a normal approximation with a continuity correction. No interest rate or mortality assumption applies.
Step 1: Translate the frequency tolerance into counts. Ten percent of the target frequency is \(0.10(0.5)=0.05\). The acceptable frequency interval is \(0.45\) to \(0.55\). Multiplying its endpoints by 800 gives acceptable counts from \(360\) to \(440\), inclusive.
Step 2: Specify the count approximation. The Poisson mean and variance are both 400, so the approximating normal distribution has mean 400 and standard deviation \(\sqrt{400}=20\). The continuity-corrected count interval is \(359.5\) to \(440.5\).
Step 3: Standardize both endpoints. The lower standardized endpoint is \((359.5-400)/20=-40.5/20=-2.025\). The upper is \((440.5-400)/20=40.5/20=2.025\). Thus the approximate probability is \(\Phi(2.025)-\Phi(-2.025)=2\Phi(2.025)-1\).
Step 4: Evaluate and decide. Using \(\Phi(2.025)\approx0.9786\), the probability is approximately \(2(0.9786)-1=0.9571\). Since \(0.9571\ge0.95\), the stated 95% standard is met under the assumed model and approximation.
Check: The interval is centered at the mean count of 400, and each continuity-corrected endpoint is 40.5 counts from that mean. Dividing \(40.5\) by the standard deviation 20 gives \(2.025\) on both sides. Omitting the continuity correction would instead use endpoints 360 and 440, changing the probability calculation.
Compare Stability at Different Volumes
A stability result depends jointly on the model’s variability, the number of observations, the tolerance, and the required probability. The following example compares two sample sizes while keeping the target and tolerance fixed. It illustrates why experience can be more stable at higher volume without suggesting that volume alone resolves every credibility question.
Worked Example: Comparing Two Sample Sizes
Suppose individual claim measures are independent and normally distributed as \(N(500,120^2)\). Compare whether an average based on 64 observations and one based on 49 observations fall within 5% of the target 500 with probability at least 95%. Assume the same distribution applies to every observation. No interest rate or mortality assumption applies.
Step 1: Set the tolerance. Five percent of 500 is \(0.05(500)=25\), so the acceptable interval for either average is \(475\) to \(525\).
Step 2: Evaluate 64 observations. The average has standard deviation \(120/\sqrt{64}=120/8=15\). The standardized endpoints are \((475-500)/15=-25/15=-1.6667\) and \(+1.6667\). The probability is \(2\Phi(1.6667)-1\approx2(0.9522)-1=0.9044\). This is below 0.95, so 64 observations do not meet the standard.
Step 3: Evaluate 49 observations. The average has standard deviation \(120/\sqrt{49}=120/7\approx17.1429\). The standardized endpoints are \(-25/(120/7)=-175/120\approx-1.4583\) and \(+1.4583\). The probability is \(2\Phi(1.4583)-1\approx2(0.9276)-1=0.8552\). This also falls below 0.95.
Decision and check. Neither sample size meets the stated standard, though the 64-observation average is more stable: its probability inside the fixed interval is about 0.9044 rather than 0.8552. The result agrees with the standard deviation comparison, \(15<17.1429\). The common mistake is to infer that a larger sample necessarily meets a standard; it improves stability here, but the probability still needs to be checked against the required 95%.
How to Interpret the Result
A stability assessment is conditional. It answers whether the statistic is likely to lie within a specified interval if the assumed model and inputs are appropriate. It does not establish that the target is correct, that the group is homogeneous, that claims have been measured consistently, or that a premium is adequate. Those are separate questions. For example, an average may be highly stable around a target that is no longer relevant to the risks being insured.
The tolerance also needs a clear interpretation. A fixed-dollar tolerance treats the same absolute deviation as acceptable regardless of the target. A percentage tolerance scales with the target; for a positive target \(\theta\), a tolerance of \(r\) proportion corresponds to \(a=r\theta\). If the target is zero or close to zero, a percentage tolerance may be unhelpful, so an absolute tolerance or another meaningful scale may be needed.
The distributional assumptions matter. The normal examples used independent observations from the same normal distribution, making the distribution of their average exactly normal. The count example used a normal approximation to a Poisson distribution and a continuity correction. When a model is a poor description of the claim process, the resulting probability can be misleading even if its arithmetic is correct.
Specify the claim measure or estimator being tested and its target value.
Give the acceptable deviation, how it is measured, and the required probability.
Identify the distribution, its parameters, and relevant assumptions such as independence or equal distributions.
Find the probability within the tolerance interval and compare it with the required probability.
Report whether the stated standard is met under the model; do not treat that conclusion as proof of data quality or rate adequacy.
Check Your Understanding
For each problem, use the stated model and decide whether the stability standard is met. Give probabilities to four decimal places where needed.
- A hypothetical set of 36 independent observations follows \(N(80,12^2)\). Assess whether their average is within 5% of 80 with probability at least 95%. (Answer: standard deviation of the average \(=2\); probability \(=2\Phi(2)-1\approx0.9545\), so it meets the standard.)
- A hypothetical claim count has distribution \(\operatorname{Poisson}(225)\). Using a normal approximation with a continuity correction, assess whether the count is within 10% of 225 with probability at least 90%. (Answer: acceptable integer counts 203 through 247; corrected limits 202.5 and 247.5; standardized limits \(\pm1.5\); probability \(\approx0.8664\), so it does not meet the standard.)
- Independent observations follow \(N(400,80^2)\), and there are 16 observations. Assess whether their average is within 5% of 400 with probability at least 95%. (Answer: average standard deviation \(=80/4=20\); tolerance \(=20\); probability \(=2\Phi(1)-1\approx0.6827\), so it does not meet the standard.)
- A hypothetical portfolio has 2,500 exposure units and count \(N\sim\operatorname{Poisson}(625)\), giving a target frequency of 0.25. Use a continuity-corrected normal approximation to test a tolerance of 8% of the target frequency against a 95% standard. (Answer: frequency tolerance \(=0.02\), acceptable counts 575 through 675, corrected standardized limits \(\pm2.02\), probability \(\approx0.9566\), so it meets the standard.)
- Independent observations follow \(N(120,30^2)\), with 25 observations. Is the average within 5% of 120 with probability at least 90%? (Answer: average standard deviation \(=30/5=6\); tolerance \(=6\); standardized limits \(\pm1\); probability \(\approx0.6827\), so it does not meet the standard.)