Why Sparse Claims Make Model Choice Difficult
When claims are sparse, many policies have no loss during the observation period and only a few policies contribute positive claims. This creates a practical tension. A frequency-severity model separates the claim count process from claim amounts and can use details recorded for individual claims. An aggregate Tweedie model works with each policy's total loss, retaining zero totals without requiring the claim amounts to be modeled separately. Neither structure automatically solves a lack of information: the right choice depends on what was recorded, what is to be predicted, and which parts of the data are too thin to estimate reliably.
Here, exposure is measured in policy-years unless stated otherwise. The observed pure premium is aggregate loss divided by earned exposure. It is an average loss cost, not necessarily the indicated rate after expenses, profit provisions, or other rating considerations.
The independence and identical-severity assumptions are modeling assumptions, not facts guaranteed by the data. If claim amounts vary systematically by coverage, cause, or another risk characteristic, the severity model needs a way to represent that variation. If the frequency and severity processes are dependent, the standard independent frequency-severity product may not describe the expected loss adequately. A Tweedie model also requires appropriate assumptions about how aggregate loss relates to predictors and exposure; it is not a remedy for every misspecification.
What Information Does Each Structure Use?
A frequency-severity fit can use the number of claims per policy and the amount of each observed claim. This is useful when claim characteristics matter to the model, or when separate frequency and severity patterns are needed for interpretation. Its cost is that the severity component is informed only by positive claims. If there are few claims, a severity estimate—or a severity comparison across several rating groups—can be highly sensitive to one observation.
An aggregate Tweedie fit uses policy-period totals, including zeros. It does not need a separate severity observation for each claim, so it can be a natural structure when the rating target is aggregate loss cost and only policy totals are available. But it still needs enough informative variation in the policy totals to estimate its parameters. A large number of zero policies does not, by itself, mean the positive-loss part of the distribution is well determined.
The two approaches can use the same underlying policy experience but differ in how they represent it. For a pure premium, a frequency-severity estimate is typically formed from an estimated expected count per unit exposure multiplied by an estimated mean claim severity. A Tweedie model estimates expected aggregate loss per unit exposure directly. The following justification shows why the product is the appropriate frequency-severity target under the stated assumptions.
Why Frequency Times Mean Severity Gives Pure Premium
Suppose that, conditional on exposure \(e\), \(N\) has a Poisson distribution with mean \(e\lambda\), where \(\lambda\) is the expected count per unit exposure. Assume the severities are independent of \(N\), independent of one another, and have common finite mean \(m_X=E[X]\). The aggregate loss is \(S=\sum_{j=1}^{N}X_j\), with \(S=0\) when \(N=0\).
Taking the expectation over the count gives the result:
Dividing by exposure, expected loss per unit exposure is \(\lambda m_X\). The factorization depends on the stated assumptions. In particular, if expected severity differs across claim-count conditions or the count and severity processes are dependent, the simple product may not represent the expected aggregate loss.
Worked Example: Sparse Claims With Claim-Level Detail
Worked Example: Sparse Claims With Claim-Level Detail
A hypothetical small book has 120 earned policy-years. It records three claims, with amounts of $600, $900, and $1,500. Each claim is linked to a policy, and claim cause is recorded. For a basic overall estimate, calculate observed frequency, observed mean severity, and observed pure premium. Then assess what the claim-level cause information can support.
Method and assumptions. Use the empirical claim count per policy-year and the empirical mean of the observed claim amounts. For interpretation, suppose counts follow a Poisson process with rate \(\lambda\) per policy-year and claim amounts are independent and identically distributed with finite mean \(m_X\). No interest or mortality assumption is involved. These are hypothetical observations, not a fitted model with demonstrated stability.
Step 1: Sum the observed claims.
Step 2: Calculate frequency. The observed frequency is three claims divided by 120 policy-years:
Step 3: Calculate mean severity.
Step 4: Calculate the frequency-severity pure premium.
Check. Total observed loss divided by exposure is \(3{,}000/120=25.00\) dollars per policy-year, matching the product. Claim cause is available at the claim level, which favors a frequency-severity structure if cause-specific severity or frequency effects are to be modeled. However, three claims cannot reliably support a detailed comparison across multiple causes: each cause group would contain at most three claims in total, and possibly fewer. The data support calculating an overall observed average; they do not, by themselves, establish stable cause-specific estimates.
Common mistake avoided. Do not treat the existence of claim-level fields as proof that a detailed severity model is estimable. The fields make such a model possible in principle, but the number and distribution of claims determine how much information is available.
Zero-Loss Policies Are Data, Not Missing Observations
A policy with no claim contributes no positive severity observation to a frequency-severity model, but it does contribute to the claim-count information. It also contributes a zero aggregate response to a Tweedie model. Omitting such policies changes the target population and can distort the estimated loss cost. In particular, calculating claim frequency using only policies that had a claim does not estimate frequency per policy-year.
Under a Poisson count model with rate \(\lambda\) per policy-year, a policy with exposure \(e\) has expected count \(e\lambda\). Its probability of no claims is:
The compound Poisson-gamma form underlying the Tweedie model for \(1<p<2\) also has a probability mass at zero. Thus, zero totals are compatible with both approaches. The difference is not that one model can include zeros and the other cannot; it is how the data are represented and which components are estimated separately.
Worked Example: Correctly Using Exposure and Zero Policies
A hypothetical program covers 80 policies for one-half of a policy-year each, giving 40 earned policy-years. Two distinct policies each have one claim; all other policies have no claims. The two claim amounts are $900 and $1,500. Estimate observed frequency, mean severity, and pure premium. Then, using the observed frequency as a Poisson rate, calculate the modeled probability of zero claims for one policy with one-half-year exposure and the expected number of zero-claim policies among the 80.
Method and assumptions. Use a Poisson claim count with rate per policy-year and independent claim amounts with a common finite mean for the frequency-severity summaries. Use the fitted rate only as a plug-in value for the zero-probability calculation. No interest or mortality assumption is involved.
Step 1: Calculate earned exposure and aggregate loss.
Step 2: Calculate observed frequency.
Step 3: Calculate mean severity and pure premium.
Step 4: Calculate the zero probability for half-year exposure. The expected count for this exposure is \(0.5(0.05)=0.025\), so:
Step 5: Calculate the expected zero-claim policy count.
Check. The observed number of zero-claim policies is \(80-2=78\), close to the model-based expected count of 78.02. The observed zero proportion is \(78/80=0.9750\), also close to 0.9753. These comparisons are descriptive, not proof that the Poisson model is correct.
Common mistake avoided. Dividing two claims by 80 policies gives 0.025 claims per policy, but the policies were exposed for only half a year. The correct rate per policy-year is \(2/40=0.05\), twice that incorrect result.
Assessing Stability Before Choosing a Structure
Sparse data create different stability concerns in the two approaches. A frequency-severity analysis has only as many direct severity observations as there are claims. Consequently, a single large claim can materially alter the estimated mean severity, and splitting claims into several groups makes each group thinner still. Frequency estimates can also be unstable when counts are few, particularly for narrow rating segments.
An aggregate Tweedie model uses all policy totals, including zeros, but the number of policies is not the same as the number of informative positive outcomes. If nearly all totals are zero, estimates that depend on the distribution of positive losses can remain uncertain. Very few positive totals can also make estimates sensitive to a small number of large losses. The aggregate structure can avoid a separate severity fit; it cannot create observations that were not collected.
A useful decision sequence is to separate the data and target questions from the estimation question:
Determine whether the data contain policy-period totals, claim records, or both, and verify that exposure is measured consistently.
If the target is aggregate pure premium and only policy totals are available, an aggregate model is a natural candidate. If claim-level behavior or claim attributes matter, a frequency-severity structure may use more of the available detail.
Check total claims and claims within any proposed severity groups; also inspect the number of positive policy totals for an aggregate fit.
Assess whether estimates change materially when individual large claims or small groups are examined, and whether the assumed frequency, severity, or aggregate structure is reasonable.
With sparse experience, a simpler structure may be more defensible than a richly segmented fit. Pooling categories or using information beyond the small segment may be necessary, but those steps require appropriate justification and are not accomplished merely by changing from one model family to another. When both policy totals and claim records exist, fitting both structures can be a useful sensitivity comparison, provided each is evaluated on a comparable target and validation basis.
Worked Example: A Large Claim and an Unstable Severity Average
Worked Example: A Large Claim and an Unstable Severity Average
A hypothetical portfolio has 100 policy-years and five claims with amounts of $400, $500, $600, $700, and $2,800. Calculate its observed frequency, mean severity, and pure premium. Then suppose the largest claim had instead been $1,000, with exposure and claim count unchanged. Recalculate the quantities and explain the implication for choosing a model.
Method and assumptions. Use empirical frequency and mean claim amount under the usual independent Poisson-frequency and identically distributed severity interpretation. Treat the alternative amount as a hypothetical sensitivity scenario, not a revised observation. No interest or mortality assumption is involved.
Step 1: Original aggregate loss and frequency.
Step 2: Original mean severity and pure premium.
Step 3: Recalculate under the alternative largest claim. The revised aggregate is \(400+500+600+700+1{,}000=3{,}200\) dollars. Frequency remains \(5/100=0.05\), while mean severity becomes \(3{,}200/5=640\) dollars. The revised pure premium is:
Check. Direct division gives \(5{,}000/100=50.00\) in the original scenario and \(3{,}200/100=32.00\) in the alternative, matching the products. Changing just one claim amount changes the observed pure premium by \(50.00-32.00=18.00\) dollars per policy-year.
This sensitivity signals that the severity estimate is based on little information and is influenced by the largest observation. A frequency-severity model preserves that claim-level detail, but five claims are not a strong basis for estimating a complex severity distribution or many severity relativities. An aggregate Tweedie model can estimate aggregate loss cost without separately fitting a severity distribution, but the same large loss still affects a policy total. Neither structure makes the experience insensitive to the observation.
Common mistake avoided. Do not conclude that an aggregate model removes the influence of a large claim. It changes the modeled response and parameterization; it does not erase the loss from the data.
Common Exam Traps and Practical Checks
- Dropping zero-loss policies. Zeros contribute to the exposure and frequency information in a frequency-severity model, and are valid aggregate responses in a Tweedie model.
- Using policy count instead of earned exposure. If policy terms differ, claim counts must be related to exposure on the correct time basis.
- Assuming many policies imply many claims. A large data set can still contain very few positive claims, leaving severity estimation weak.
- Choosing a model from the label alone. A Tweedie specification is not automatically more stable, and a frequency-severity specification is not automatically more informative; the available observations and target determine what each can use.
- Confusing model fit with prediction reliability. A model can fit the observed aggregate pattern while estimates for thin groups remain uncertain. Stability, assumptions, and validation still matter.
Check Your Understanding
Work each problem using the stated exposure and assumptions. Monetary amounts are hypothetical.
- A hypothetical book has 200 policy-years, four claims, and total claim payments of $6,400. Find observed frequency, mean severity, and pure premium. (Answer: 0.020 claims per policy-year; $1,600 per claim; $32.00 per policy-year.)
- Sixty policies are each covered for one-quarter of a year. One claim occurs, and the estimated Poisson rate is based on the full earned exposure. What is the observed frequency per policy-year, and what is the modeled probability of zero claims for one of these policies? (Answer: \(1/(60\times0.25)=0.0667\) claims per policy-year; \(\exp[-0.25(0.0667)]\approx0.9835\).)
- A portfolio has 50 policy-years, three claims of $300, $900, and $1,200, and 47 zero-claim policies. Calculate the frequency-severity pure premium and verify it using aggregate loss divided by exposure. (Answer: frequency \(3/50=0.06\); mean severity \(2,400/3=\)$800; pure premium $48.00 per policy-year, also \(2,400/50=$48.00\).)
- A hypothetical segment has 600 policy-years but only two claims. Claim cause is recorded, and the intended analysis would estimate a separate severity mean for each of four causes. What is the main model-choice and stability concern? (Answer: Frequency-severity can use claim cause, but two claims cannot support four cause-specific severity means; an aggregate Tweedie fit avoids separate severity estimates but does not eliminate uncertainty from the very few positive losses.)
- Two hypothetical data sets have equal exposure and aggregate loss. One contains claim-level records and the other only policy-period totals. The proposed target is overall aggregate pure premium, with no need to model claim attributes. Which structure can be fit to each data set, and what should not be inferred from the data-format difference? (Answer: An aggregate Tweedie model can use either set if policy totals are available or can be formed; frequency-severity requires claim counts and amounts. The format alone does not establish that either model is more stable or accurate.)