Tutorials › CAS Exam 5 - Basic Ratemaking and Estimated Claim Liabilities › Sparse And Zero-Claim Experience

Home  /  Tutorials  /  CAS Exam 5 / Credibility Concepts and Data

CAS Exam 5 · Credibility Concepts and Data · Tutorial 518 of 989

Sparse And Zero-Claim Experience

Use exposure and a simple claim-count model to judge when zero claims are unsurprising, when they are unusual, and what limitations remain.

Beginner 11 min read

What You'll Learn

  • Calculate an observed claim frequency using the correct exposure base
  • Relate expected claim counts to the probability of observing zero claims
  • Compare a zero count with a hypothetical benchmark rate
  • Recognize when low exposure makes a zero observation weak evidence
  • Account for different risk levels within a segment when evaluating zero claims
  • Identify model assumptions and data checks that matter for sparse experience

Why Zero Claims Need Context

A segment with no observed claims may look safer than other segments. But a zero count can arise because the segment has a genuinely low claim rate, because it had little exposure, or simply because claims did not occur during the observation period. The count alone cannot distinguish these explanations.

The first step is to put the count beside the exposure that produced it. A zero count over a small amount of exposure is usually less informative than a zero count over a large amount. A simple probability model can make this comparison precise: given a proposed claim frequency and the segment's exposure, how likely would zero claims be?

Definition: For a segment with \(N\) observed claims and \(E\) units of exposure, the observed claim frequency is \(N/E\), provided \(E>0\). When \(N=0\), this calculated frequency is zero. It is a description of the observed period, not proof that the segment's underlying claim frequency is zero.

Exposure must use a consistent unit, such as earned vehicle-years or policy-years, and should correspond to the period and risk being measured. As emphasized in the earlier material on exposure as the unit of experience and on aligning experience periods, a count without its exposure base cannot be compared fairly with another count.

A Claim-Count Model for Zero Experience

For an initial diagnostic, suppose claims follow a Poisson model with a constant frequency rate \(r\) per exposure unit. Assume exposure is measured consistently, claims occur independently, the rate is stable over the period, and the count is complete and correctly recorded. Under this model, the claim count \(N\) has mean \(\lambda=rE\), where \(\lambda\) is the expected number of claims.

$$ N\sim\operatorname{Poisson}(\lambda),\qquad \lambda=rE $$

The probability of observing exactly zero claims is therefore:

$$ P(N=0)=e^{-\lambda}=e^{-rE} $$

This probability is a diagnostic conditional on the assumed rate and model. It is not the probability that the rate itself equals zero, nor does it tell us directly how much credibility to assign to the segment. Those are different questions.

Why the Zero-Claim Formula Holds

For a Poisson count, the probability of \(k\) claims is \(e^{-\lambda}\lambda^k/k!\). Setting \(k=0\) gives \(0!=1\) and \(\lambda^0=1\), leaving \(P(N=0)=e^{-\lambda}\). Since the expected count is rate multiplied by exposure, \(\lambda=rE\), so the zero probability is \(e^{-rE}\).

This derivation also explains the role of exposure. Holding the rate \(r\) fixed, increasing \(E\) increases \(\lambda\), and \(e^{-rE}\) decreases. Thus, a zero count becomes less likely under the same assumed rate as the exposure grows.

Worked Example: Is Zero Claims Unusual at This Exposure?

A hypothetical class of small commercial watercraft has 6.0 earned vessel-years of exposure and no reported claims. A hypothetical benchmark frequency is 0.30 claims per vessel-year. Under a Poisson model with a constant rate and complete claim data, assess how surprising the zero count is. Assume no discounting or mortality model is relevant; the calculation concerns claim counts only.

Step 1: Calculate the expected count under the benchmark. The rate is \(r=0.30\) claims per vessel-year and exposure is \(E=6.0\) vessel-years:

$$ \lambda=rE=0.30(6.0)=1.80\text{ claims} $$

Step 2: Calculate the probability of zero claims.

$$ P(N=0)=e^{-1.80}=0.1653 $$

The probability is 0.1653, or about 16.53%. That is not an implausibly rare outcome under the benchmark model. The observed frequency is \(0/6.0=0\) claims per vessel-year, but the calculation shows that this zero can occur by chance even if the benchmark rate applies.

Check. The expected count is greater than one, but it is not large; a zero outcome remains possible. The result is consistent with the direction expected from \(e^{-\lambda}\). The method avoids treating the observed zero frequency as a known underlying rate.

Worked Example: How Much Exposure Makes Zero Claims Less Likely?

A hypothetical equipment-protection segment has no claims in its first 10 equipment-years. A proposed benchmark frequency is 0.20 claims per equipment-year. Assume a Poisson count with a constant rate, independent claims, and complete data. Calculate the probability of zero claims at 10 equipment-years, then find the exposure at which that probability falls to 5%. No interest or mortality assumption is needed.

Step 1: Evaluate the observed exposure. The expected count under the benchmark is:

$$ \lambda=0.20(10)=2.00 $$

Therefore:

$$ P(N=0)=e^{-2.00}=0.1353 $$

Zero claims have probability 0.1353 under the benchmark, so the observed zero is not especially unusual at this exposure.

Step 2: Solve for the exposure associated with a 5% zero probability. Let \(E^*\) denote that exposure. Set the zero probability equal to 0.05 and solve:

$$ e^{-0.20E^*}=0.05 $$

Taking natural logarithms gives \(-0.20E^*=\ln(0.05)\), so:

$$ E^*=\frac{-\ln(0.05)}{0.20} =\frac{2.9957}{0.20} =14.9785\text{ equipment-years} $$

At about 14.98 equipment-years, the benchmark model assigns a 5% probability to zero claims. At the actual 10 equipment-years, zero claims are more likely than that. This threshold is a probability calculation under the stated model; it does not establish a rate or determine a credibility weight.

Check. Substitution gives \(e^{-0.20(14.9785)}=e^{-2.9957}\), approximately 0.05. The method avoids comparing counts without accounting for their different exposure amounts.

Different Risk Levels Require an Exposure-Aware Comparison

A segment may combine risks with different expected frequencies. In that case, the expected claim count should be calculated separately for each known risk group and then added. Applying an unweighted average of group rates to the entire exposure can give the wrong benchmark, especially when the groups have different amounts of exposure.

$$ \lambda_{\text{total}}=\sum_j r_jE_j,\qquad P(N_{\text{total}}=0)=e^{-\lambda_{\text{total}}} $$

This calculation assumes independent Poisson counts within groups with fixed rates. The sum of independent Poisson counts is Poisson, with mean equal to the sum of the group means. If those assumptions are not reasonable, the resulting probability may not describe the actual chance of a zero count.

Worked Example: Combining Two Risk Groups Correctly

A hypothetical service-contract segment has no observed claims. It contains 3.0 contract-years in a higher-risk equipment group, with a benchmark frequency of 0.50 claims per contract-year, and 9.0 contract-years in a lower-risk group, with a benchmark frequency of 0.10. Assume independent Poisson counts within the two groups, constant group rates, and complete data. Find the probability of zero claims under the benchmark. No discounting or mortality model is used.

Step 1: Calculate each group's expected count. For the higher-risk group:

$$ \lambda_1=0.50(3.0)=1.50 $$

For the lower-risk group:

$$ \lambda_2=0.10(9.0)=0.90 $$

Step 2: Add the expected counts.

$$ \lambda_{\text{total}}=1.50+0.90=2.40 $$

Step 3: Calculate the zero probability.

$$ P(N_{\text{total}}=0)=e^{-2.40}=0.0907 $$

Thus, the zero count has probability 0.0907 under this exposure mix and benchmark. The corresponding exposure-weighted benchmark rate is \(2.40/12.0=0.20\) claims per contract-year, which gives the same mean when multiplied by the total 12.0 contract-years.

Common mistake and correction. A tempting but incorrect calculation is to take the unweighted average of the two group rates, \((0.50+0.10)/2=0.30\), and multiply by all 12.0 contract-years. That gives an incorrect expected count of \(0.30(12.0)=3.60\) and zero probability \(e^{-3.60}=0.0273\). The rates do not apply to equal exposure amounts. Use each group's rate with its own exposure, then add the expected counts.

Check. The correct total expected count, 2.40, equals the exposure-weighted average rate, 0.20, multiplied by 12.0. The calculation avoids letting the smaller high-risk group receive the same weight as the larger low-risk group.

What a Sparse or Zero Count Can—and Cannot—Say

A zero count can be consistent with a low underlying rate, but it can also be consistent with an ordinary rate when exposure is limited. The Poisson probability helps answer a narrow question: if a specified rate and the model assumptions were true, how likely would zero claims be at this exposure? It does not identify the true rate from one observation.

Sparse data also create a boundary problem for the raw frequency. When \(N=0\), \(N/E=0\), regardless of whether the exposure is one policy-year or one thousand. The same calculated frequency therefore conceals very different amounts of information. Always report the count and exposure together, and treat a zero estimate cautiously when the exposure is small.

The calculation is only as useful as its inputs. Before interpreting a zero, verify that the exposure is positive and correctly earned, the observation period is aligned with the claims being counted, and claims are sufficiently complete and mature for the selected measure. Confirm that the segment definition is consistent and that the benchmark rate applies to the risks and coverage in the segment. These checks matter because a missing, delayed, or misclassified claim can produce an apparent zero that is not a true zero count.

The Poisson model also has limitations. A constant rate and independent claim occurrences may be poor descriptions when risks vary substantially, claim events cluster, or conditions change over the period. For a mixture of rates, the group-specific calculation above is more appropriate than treating the whole segment as homogeneous. Even then, an incorrect grouping or benchmark can make the result misleading. When assumptions are doubtful, report the zero count and exposure, explain the limitations, and avoid presenting the model probability as a fact about the underlying risk.

Key takeaway: A zero observed frequency is a fact about the experience period; it is not automatically a reliable estimate of future frequency. Interpret it alongside exposure, a relevant benchmark, data completeness, and the assumptions used to model claim counts. A probability calculation can show whether zero is surprising under a specified model, but it does not by itself determine a rate indication or credibility weight.

Check Your Understanding

For each hypothetical case, assume the stated Poisson model, constant rates, and complete data unless the question says otherwise. Probabilities are rounded to four decimal places.

  1. A segment has 4.5 earned policy-years, zero claims, and a benchmark frequency of 0.40 claims per policy-year. Find the expected count and the probability of zero claims. (Answer: Expected count \(1.80\); zero probability \(e^{-1.80}=0.1653\).)
  2. A new coverage group has zero claims in 2.0 exposure units. Under a benchmark frequency of 0.15 claims per unit, calculate the probability of zero claims. Is zero unusual under this model? (Answer: \(e^{-0.30}=0.7408\); no, zero is quite likely.)
  3. A hypothetical segment has zero claims in 12 exposure units. At a benchmark rate of 0.18 claims per unit, find the expected count and zero probability. (Answer: Expected count \(2.16\); zero probability \(e^{-2.16}=0.1153\).)
  4. A hypothetical portfolio has 3 exposure units in a group with rate 0.50 and 9 exposure units in a group with rate 0.10. If there are no claims, calculate the correct expected count and zero probability. (Answer: Expected count \(0.50(3)+0.10(9)=2.40\); zero probability \(e^{-2.40}=0.0907\).)
  5. A segment has zero reported claims in 20 earned vehicle-years. A proposed benchmark is 0.25 claims per vehicle-year. Calculate the model probability of zero claims and state one important check before interpreting it. (Answer: Expected count \(5.00\); zero probability \(e^{-5.00}=0.0067\); for example, verify claim completeness and maturity before concluding that the zero is unusual experience.)

Previous

Next

Back to Tutorials
Clinical Biostats