Why Partial Credibility Is Useful for Claim Counts
A small group’s observed claim frequency can move substantially when only a few claims are added or removed. Yet ignoring that experience entirely may also discard relevant information. Partial credibility provides a rule for combining the observed frequency with a complement, such as an expected or manual frequency. For claim counts, the square-root rule connects the credibility weight to the observed claim volume and a specified full-credibility threshold.
The calculation has two distinct parts. First, use the observed number of claims and the full-credibility count to obtain a credibility factor. Second, use that factor to blend the observed frequency with the complement. The threshold is a count of claims, not a number of policies or an exposure amount unless the problem explicitly provides a conversion between those measures.
Here, frequency is measured as claims per unit of exposure, so \(E\), \(\hat{\lambda}\), and \(\lambda_0\) must use compatible exposure units. For example, if \(E\) is in policy-years, both frequencies must be claims per policy-year. The weight on the complement is \(1-Z\), ensuring that the two weights sum to one.
How the Square-Root Rule Connects to Claim Counts
The usual count-based rule is grounded in a Poisson model. Suppose claim counts for homogeneous risks over fixed exposure \(E\) follow \(N\sim\operatorname{Poisson}(\lambda E)\), where \(\lambda\) is the underlying claim frequency. Under this model, the count has mean and variance \(\lambda E\). The full-credibility standard supplies a threshold \(n_0\) consistent with the chosen precision rule; this tutorial takes that threshold as given.
The square-root relationship follows from how the relative fluctuation of a Poisson count changes with its size. A Poisson count with mean \(m\) has standard deviation \(\sqrt{m}\), so its standard deviation relative to its mean is \(1/\sqrt{m}\). At the full standard, that relative scale is \(1/\sqrt{n_0}\). At an observed count \(n\), it is approximately \(1/\sqrt{n}\). The ratio of the full-standard relative scale to the observed-count relative scale is therefore:
This explains why the partial-credibility factor grows with the square root of the observed count relative to the threshold. It is a limited-fluctuation credibility rule, not a claim that the observed count alone determines the exact probability distribution of a future rate. In particular, the blend with \(\lambda_0\) is the specified credibility procedure; the square-root calculation does not, by itself, derive a unique complement or a full statistical model for it.
Confirm that \(n\) is the observed claim count for the same claim definition and group to which \(n_0\) applies.
For experience below the full standard, compute \(Z=\sqrt{n/n_0}\).
Divide the count by its exposure: \(\hat{\lambda}=n/E\).
Compute \(Z\hat{\lambda}+(1-Z)\lambda_0\), retaining consistent units throughout.
Worked Calculations
Worked Example: Blending a Small Claim Count
A hypothetical specialty equipment group has 36 claims over 1,200 policy-years. Its stated full-credibility threshold is 100 claims, and the complement frequency is 0.025 claims per policy-year. Calculate the partial factor and credibility-weighted frequency. Assume claim counts follow a Poisson model with parameter \(\lambda E\); at the observed frequency, the fitted count parameter is \(0.030(1{,}200)=36\). No interest or mortality assumption is relevant to this frequency calculation.
Method. Apply the count-based square-root rule, then blend the observed frequency with the complement.
Step 1: Calculate the factor. The ratio of observed claims to the full standard is \(36/100=0.36\). Therefore, \(Z=\sqrt{0.36}=0.60\).
Step 2: Calculate the observed frequency. The observed frequency is \(36/1{,}200=0.030\) claims per policy-year.
Step 3: Blend. The complement weight is \(1-0.60=0.40\). Thus,
The credibility-weighted frequency is 0.028 claims per policy-year. As a check, the observed frequency is above the complement, so a blend with positive weight on the observed frequency should be above 0.025 and below 0.030. The result satisfies both conditions. A common error is to use 0.36 as the credibility factor; the square root is required.
Worked Example: A Different Full Standard
A hypothetical mobile-device portfolio has 64 claims during 3,200 device-years of coverage. Its full-credibility count threshold is 256 claims, and the expected frequency is 0.015 claims per device-year. Under a Poisson count model with parameter \(\lambda E\), the fitted parameter at the observed frequency is \(0.020(3{,}200)=64\). Calculate the factor and weighted frequency. Interest and mortality assumptions do not apply.
Method. Use the supplied count threshold and blend rates expressed per device-year.
Step 1: Calculate the factor. The count ratio is \(64/256=0.25\), so \(Z=\sqrt{0.25}=0.50\). The complement receives weight \(1-0.50=0.50\).
Step 2: Calculate the observed frequency. The observed frequency is \(64/3{,}200=0.020\) claims per device-year.
Step 3: Calculate the weighted frequency.
The weighted frequency is 0.0175 claims per device-year. As a check, equal credibility weights put the result halfway between 0.015 and 0.020; their midpoint is 0.0175. The main trap here is treating the threshold as an exposure amount: the factor uses 64 claims and 256 claims, while the exposure is used to calculate the observed frequency.
Worked Example: Combining Claim Counts Across Exposure Groups
A hypothetical repair plan covers two groups with different durations. Group A has 24 claims over 900 policy-years; Group B has 12 claims over 600 policy-years. Assume the groups are homogeneous for this calculation, and their independent counts follow Poisson models with parameters \(\lambda(900)\) and \(\lambda(600)\). The stated full-credibility threshold is 100 claims, and the complement frequency is 0.020 claims per policy-year. Find the combined partial factor and weighted frequency. No interest or mortality assumption is involved.
Method. Combine compatible claim counts and exposures before calculating the frequency. For independent Poisson counts with the same underlying frequency, the combined count has mean \(\lambda(900+600)\).
Step 1: Combine the data. The total count is \(24+12=36\) claims. Total exposure is \(900+600=1{,}500\) policy-years.
Step 2: Calculate the factor. The count ratio is \(36/100=0.36\), so \(Z=\sqrt{0.36}=0.60\). The complement weight is 0.40.
Step 3: Calculate and blend the frequency. The observed frequency is \(36/1{,}500=0.024\) claims per policy-year. Then
The weighted frequency is 0.0224 claims per policy-year. The arithmetic checks: it lies between the complement 0.020 and the observed frequency 0.024, and it is closer to the observed frequency because \(Z=0.60\). Pooling the counts but dividing by only one group’s exposure would produce an inconsistent frequency.
Common Mistakes and Conditions for Use
The calculation is simple, but it relies on carefully matching the count, threshold, and frequency. In particular, \(n_0\) must be a full-credibility threshold stated in claims for the applicable standard. An exposure threshold cannot be substituted for it without a valid conversion. If a problem gives unequal policy durations, use total exposure in calculating the frequency; do not treat each policy or each year as an identical unit unless the stated exposure measure supports that interpretation.
A frequent error is using the count ratio rather than its square root. In Worked Example 1, this incorrect approach would set \(Z=36/100=0.36\). It would produce
That is not the square-root rule. The correct factor is \(\sqrt{36/100}=0.60\), giving \(0.028\) claims per policy-year. Another error is to blend counts directly when the desired estimate is a frequency. The two quantities being averaged must have the same units: here, both components of the blend are claims per policy-year.
The Poisson basis assumes a stable underlying frequency and a suitable homogeneous risk group over the exposure being analyzed. Material changes in coverage, claim definition, or risk composition can make a single pooled frequency misleading. The credibility factor also depends on the full-credibility standard supplied: different precision criteria can lead to different values of \(n_0\), and therefore a different \(Z\).
Check Your Understanding
For each hypothetical situation, calculate the square-root factor and the credibility-weighted frequency. Treat the stated thresholds as applicable to the described claim counts.
- A data-security product has 25 claims over 1,000 policy-years. The full standard is 100 claims, and the complement is 0.040 claims per policy-year. What are \(Z\) and the weighted frequency? (Answer: \(Z=0.500\); weighted frequency \(=0.0325\) claims per policy-year.)
- A hypothetical travel benefit has 81 claims over 2,700 covered-trip-years. The full standard is 225 claims and the complement is 0.020 claims per covered-trip-year. Calculate the factor and blended frequency. (Answer: \(Z=0.600\); weighted frequency \(=0.0260\) claims per covered-trip-year.)
- A fleet has 28 claims over 1,200 vehicle-years in one segment and 20 claims over 800 vehicle-years in another. Assume a common frequency; the full standard is 192 claims and the complement is 0.030 claims per vehicle-year. Combine the data and calculate the blended frequency. (Answer: \(n=48\), \(E=2{,}000\) vehicle-years, \(Z=0.500\); weighted frequency \(=0.0270\) claims per vehicle-year.)
- A hypothetical equipment warranty has 75 claims over 2,500 contract-years. The full standard is 300 claims, and the complement is 0.040 claims per contract-year. Calculate the observed frequency and its credibility-weighted value. (Answer: observed frequency \(=0.030\), \(Z=0.500\); weighted frequency \(=0.0350\) claims per contract-year.)
- A specialty home system group has 144 claims over 6,000 policy-years. The threshold is 400 claims, and the complement frequency is 0.018 claims per policy-year. Calculate the factor and blended frequency, then check whether the result lies between the two input frequencies. (Answer: \(Z=0.600\); weighted frequency \(=0.0216\) claims per policy-year, between 0.018 and 0.024.)