Why Zero-Exposure Records Need Special Treatment
A frequency model relates a claim count to the time or amount of exposure over which claims could occur. A policy with positive exposure and no claims is informative: it records an opportunity for claims that happened not to produce one. A record with zero exposure and no claims is different. Under the usual frequency model, no claim could occur during zero exposure, so the record is certain regardless of the underlying claim rate. It therefore provides no information about that rate.
The important distinction is between a valid structural zero and a record that is wrong or incomplete. Zero exposure paired with zero claims can be valid. Zero exposure paired with one or more claims cannot arise under a model that defines claims as occurring during the recorded exposure. The latter requires investigation; it should not be repaired by blindly inserting a small positive exposure.
For a log-link frequency model, the mean is often written as \(\mu_i=e_i\exp(x_i^\mathsf{T}\beta)\), or equivalently with \(\log(e_i)\) as an offset when \(e_i>0\). At \(e_i=0\), the offset \(\log(e_i)\) is undefined. This is not a reason to substitute an arbitrary value: the likelihood can be understood directly from the model’s limiting mean, \(\mu_i=0\). A zero-exposure row is not an ordinary positive-exposure observation in the fitted model.
Likelihood at Zero Exposure
For \(e_i>0\), the Poisson probability of observing \(y_i\) claims is
To see what happens when exposure is zero, set the mean \(\mu_i=e_i\lambda_i\) to zero. The Poisson distribution with mean zero is concentrated entirely at zero: \(P(Y_i=0)=1\), while \(P(Y_i=y_i)=0\) for any \(y_i>0\). This degenerate distribution statement avoids any ambiguity about writing a zero raised to the zero power in the probability formula.
The same principle applies to a negative binomial frequency model parameterized by its mean and dispersion: as the mean approaches zero, its count distribution also places all probability on zero. A zero-exposure, zero-count row therefore contributes no frequency information under either model. This conclusion assumes that the count and exposure definitions are aligned and that the model has no separate process that generates counts independently of exposure.
The result does not mean that every zero-exposure record should be deleted from a data system. It means that the record contributes nothing to this particular frequency likelihood once its zero exposure and zero count are confirmed. Keeping it in an audit trail can be useful. The model-fitting data should distinguish such valid records from missing exposure, incorrectly assigned claims, or records outside the intended observation period.
Worked Example: Estimating a Rate With Valid Zero-Exposure Rows
A hypothetical portfolio has four independent policy records. Their exposure and claim counts are \((0,0)\), \((0,0)\), \((0.5,1)\), and \((1.5,2)\), where exposure is measured in policy-years. Assume a common Poisson claim rate \(\lambda\) per policy-year. Find the maximum-likelihood estimate and calculate the likelihood at that estimate.
Method and assumptions. Conditional on the exposures, the counts are independent with \(Y_i\sim\operatorname{Poisson}(e_i\lambda)\). The two zero-exposure records have zero claims, as required by the model. Their likelihood contributions are each 1.
For the positive-exposure records, the total claim count is \(1+2=3\), and total exposure is \(0.5+1.5=2\) policy-years. Ignoring factors that do not depend on \(\lambda\), the log-likelihood is
Differentiate and set the result to zero:
The fitted means for the two positive-exposure records are \(0.5(1.5)=0.75\) and \(1.5(1.5)=2.25\). Their probabilities of the observed counts are
The zero-exposure rows each contribute 1, so the full likelihood is \(1(1)(0.3542749)(0.2667918)\approx0.0945\). The estimate also checks directly as total claims divided by total positive exposure, \(3/2=1.5\). The method avoids treating the two certain zero counts as additional evidence for a lower rate.
Cancellation, Occurrence Dates, and Exposure Definitions
A policy cancellation can create an apparent contradiction if exposure, claim count, and claim timing are drawn from different records or defined over different periods. A claim reported after cancellation may still have occurred while coverage was in force. For frequency modeling, the relevant question is generally whether the claim’s occurrence belongs to the period represented by the exposure basis—not simply whether the policy was active on the reporting date.
For example, an earned-exposure extract might show zero because a policy is now marked canceled, while a claims extract includes a claim reported later for an occurrence during covered time. If the model counts claims by occurrence date, the exposure should represent that covered period. If the claim instead occurred outside the period represented by the exposure, it should not be assigned to that period’s count. The definitions must be made consistent before fitting.
Worked Example: A Claim Reported After Cancellation
A hypothetical policy is in force for one quarter of a rating year and is then canceled. A claim occurs during that covered quarter but is reported after cancellation. A claims extract assigns one claim to the record, while a preliminary policy extract incorrectly shows zero exposure because it uses the policy’s later canceled status. The intended exposure basis is policy-years, and the annual claim rate for this illustration is \(0.8\). Reconcile the record and calculate the probability of one claim over the covered period.
Method and assumptions. Counts are Poisson with mean equal to annual rate times covered policy-year exposure. The claim’s occurrence date is within the covered quarter, so the record should use \(e=0.25\), not zero.
The appropriate mean is
The probability of exactly one claim is
If the preliminary zero exposure were used without correction, a positive count would have probability zero under the stated model. The correct action is to reconcile the occurrence date and covered period, then use the exposure consistent with those definitions. The common mistake avoided here is treating the post-cancellation report date as proof that the claim occurred outside coverage.
Positive Claims With Zero Recorded Exposure
A row with \(e_i=0\) and \(Y_i>0\) is not a low-information observation. It is impossible under the stated exposure-based count model. For a Poisson model its likelihood contribution is zero, so its log-likelihood contribution is negative infinity. A fitting routine may reject the data, return an error, or behave unpredictably if the offset was altered to force the row into the model.
The correct response is to check the record’s exposure, count, and claim definition. Potential explanations include missing or incorrectly coded exposure, an inconsistent policy identifier, a claim allocated to the wrong period, a mismatch between occurrence and reporting dates, or a claim that is outside the modeled coverage. The resolution depends on the data and the intended model; it is not always appropriate to assign the claim to a different policy period automatically.
Worked Example: Why an Arbitrary Small Exposure Is Not a Fix
Suppose two hypothetical records are intended to share one Poisson rate. The first has one policy-year of exposure and one claim. The second is recorded with zero exposure and one claim. Determine what the model says, and show the effect of replacing the second exposure with an arbitrary \(0.01\) policy-years.
Method and assumptions. Under a common rate \(\lambda\), independent counts satisfy \(Y_i\sim\operatorname{Poisson}(e_i\lambda)\). For the second record as actually recorded, \(e_2=0\) and \(Y_2=1\), so \(P(Y_2=1)=0\). The joint likelihood is zero for every finite \(\lambda\); this data set cannot be fitted as stated under this model.
Now suppose someone replaces the zero exposure with \(0.01\), without evidence that this is the correct exposure. The resulting total count is \(1+1=2\), and the artificial total exposure is \(1+0.01=1.01\). The common-rate Poisson estimate would then be
Using only the valid first record would give \(1/1=1\) claim per policy-year, but that calculation is not a justification for silently dropping the second claim. The discrepancy signals that the record must be resolved. The \(0.01\) substitution produces a seemingly usable estimate by changing the data, not by repairing the model. The common mistake avoided is using a small exposure as a numerical workaround for a positive count with zero exposure.
A Practical Review Before Fitting
Before fitting a frequency model, inspect zero-exposure records alongside their claim counts and the definitions used to construct both fields. A compact review separates records that are harmless to the likelihood from records that need correction or an explicit modeling decision.
Check the unit of exposure and the observation period represented by each record. Verify that zero means genuinely no exposure, rather than missing data or a default code.
Confirm whether claims are assigned by occurrence date, report date, or another rule, and ensure that rule matches the exposure period.
A confirmed \((0,0)\) record has likelihood contribution 1. A \((0,y)\) record with \(y>0\) contradicts the exposure-based model and requires investigation.
Correct the underlying exposure or count when supported by the records, or exclude an unresolved invalid record from the fitted data under a documented rule. Do not invent a small exposure merely to make a logarithmic offset computable.
This distinction matters because frequency estimates depend on the relationship between counts and exposure. Valid zero-exposure, zero-count records do not lower the estimated rate; invalid records can create impossible likelihoods or distort an estimate if they are forced into the data. Careful reconciliation preserves the meaning of the fitted rate and makes the modeling assumptions auditable. This treatment of exposure-based claim-count likelihoods is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.
Check Your Understanding
For each problem, use the stated exposure and count definitions. Treat counts as conditionally Poisson unless the question asks for a data-quality conclusion.
- A policy has zero exposure and zero claims. Under an exposure-based Poisson model, what are its likelihood contribution and log-likelihood contribution? What changes if the count is one? (Answer: For zero claims, the likelihood contribution is 1 and the log-likelihood contribution is 0; for one claim, the likelihood contribution is 0 and the log-likelihood contribution is negative infinity.)
- Four records have \((e,Y)\) values \((0,0)\), \((0.25,1)\), \((0.75,0)\), and \((0,0)\). Assuming one common Poisson rate per policy-year, find the maximum-likelihood estimate. (Answer: \(1/(0.25+0.75)=1.0000\) claim per policy-year.)
- A hypothetical policy has exposure \(0.25\) policy-years and an annual rate of \(0.8\). Find the mean count and the probability of zero claims over that exposure. (Answer: Mean \(=0.25(0.8)=0.20\); \(P(Y=0)=e^{-0.20}\approx0.8187\).)
- A claim is reported after a policy is canceled, but its occurrence date is within the covered period. The policy extract shows zero exposure because it uses the later canceled status. What should be checked before fitting, and what is the governing exposure principle? (Answer: Reconcile the occurrence date, coverage dates, count allocation rule, and exposure extract; use exposure for the covered period represented by the model, not the later report or status date alone.)
- A portfolio has records \((0.5,0)\), \((1.5,3)\), and \((0,0)\), with a common Poisson rate. Find the estimate and state the effect of the final record. (Answer: \(\widehat{\lambda}=3/(0.5+1.5)=1.5000\) claims per policy-year; the zero-exposure, zero-count record contributes likelihood 1 and does not change the estimate.)