Common Errors In Frequency Exposure Adjustments

Why Exposure Adjustments Go Wrong

A frequency statement is meaningful only when its exposure base is clear. “A rate of 0.12” could mean 0.12 claims per policy-year, per vehicle-quarter, or per some other unit. Multiplying a rate by a number of policies without checking the time observed can produce a count that is wrong by a substantial factor. A different error arises with binomial models: changing the number of trials while leaving the probability per trial unchanged may silently change the risk being modeled.

The essential check is to ask what the random variable counts. A count of claim events is not necessarily the same as a count of policies that experience at least one claim. Likewise, a frequency rate per exposure unit is not a probability for an entire policy period. These distinctions are aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Definition: Exposure is the amount of time or risk opportunity observed, measured in a stated unit such as policy-years or vehicle-months. A claim count is the number of claim events observed. A frequency rate is a claim count relative to exposure, such as claims per policy-year. A claim probability is the probability of an event over a specified interval; it is not automatically equal to a frequency rate.

Start With the Exposure Base

For a homogeneous Poisson frequency model, suppose claims occur at rate \(\lambda\) per exposure unit. If the total exposure is \(E\), then the aggregate claim count \(N\) satisfies \(N\sim\operatorname{Poisson}(\lambda E)\). The exposure units must match: a rate per policy-year must be multiplied by policy-years, not simply by the number of policies if policies were observed for different lengths of time.

E=\sum_{j=1}^{m} e_j,\qquad E[N]=\lambda E,\qquad P(N=0)=e^{-\lambda E}.

Here \(e_j\) is the exposure for policy \(j\), expressed in the same unit used for \(\lambda\). For example, 75 policies each observed for four months contribute \(75(4/12)=25\) policy-years. This conversion assumes that the exposure is appropriately represented by time at risk and that the rate is constant over the periods being combined. If risk changes by season, coverage, or another relevant feature, a single homogeneous rate may not be appropriate.

Key distinction: Under the Poisson model, \(\lambda E\) is the expected number of claim events. The probability that a particular policy has at least one claim over exposure \(e\) is \(1-e^{-\lambda e}\), not \(\lambda e\) in general. The two are approximately equal only when \(\lambda e\) is small.

Worked Example: Converting Policy-Months to Policy-Years

A hypothetical insurer models claims at a Poisson rate of 0.12 claims per policy-year. It observes 75 policies for four months each. Find the expected total number of claims and the probability of no claims in the observed portfolio. Give the probability to four decimal places.

Method. Convert the observation period to policy-years, then apply the Poisson mean and zero-count formulas. This avoids treating 75 policies as 75 full years of exposure.

Each policy contributes \(4/12=1/3\) policy-year. Thus the total exposure is \(E=75(1/3)=25\) policy-years.

The Poisson mean is \(\lambda E=0.12(25)=3\) claims. Therefore \(N\sim\operatorname{Poisson}(3)\), and the expected count is 3.

The probability of no claims is \(P(N=0)=e^{-3}=0.049787\ldots\), which is 0.0498 to four decimal places. The probability of at least one claim is \(1-0.0498=0.9502\), using the unrounded zero probability before rounding.

Check. The rate is per policy-year and the portfolio has 25 policy-years, so multiplying gives claims: \(0.12\) claims per policy-year times 25 policy-years equals 3 expected claims. Omitting the four-month adjustment would incorrectly give \(0.12(75)=9\).

Why Poisson Exposure Scaling Works

The exposure scaling formula follows from the assumed rate per unit of exposure. For a short exposure increment \(h\), the Poisson model assigns an expected count of \(\lambda h\). If exposure is divided into independent increments \(e_1,\ldots,e_m\), their claim counts have means \(\lambda e_1,\ldots,\lambda e_m\). The sum of independent Poisson counts is Poisson, with mean equal to the sum of their means.

N_j\sim\operatorname{Poisson}(\lambda e_j) \quad\Longrightarrow\quad N=\sum_{j=1}^{m}N_j\sim \operatorname{Poisson}\left(\sum_{j=1}^{m}\lambda e_j\right) =\operatorname{Poisson}(\lambda E).

The zero-count probability then follows directly from the Poisson probability mass function: \(P(N=0)=e^{-\lambda E}(\lambda E)^0/0!=e^{-\lambda E}\). This derivation depends on the constant-rate Poisson model and independent increments. Exposure multiplication is not a license to combine unlike risks under one rate without justification.

Binomial Trials: Change the Period Carefully

A binomial model counts successes across a fixed number \(n\) of independent trials, each with the same success probability \(p\). In frequency applications, a trial might be a policy-month with a claim or no claim, or a policy-period with at least one claim. The interpretation must be stated: one policy can have a claim in more than one policy-month, so the number of claim events and the number of affected policies can differ.

N\sim\operatorname{Binomial}(n,p),\qquad E[N]=np,\qquad P(N\geq 1)=1-(1-p)^n.

The formula \(np\) is the expected number of successful trials. If several trials belong to each policy, it is not automatically the expected number of distinct policies with one or more successes. For \(r\) independent periods per policy, each with success probability \(p\), a policy has probability \(1-(1-p)^r\) of at least one success. Across \(m\) policies, the expected number of affected policies is \(m[1-(1-p)^r]\).

Worked Example: Claim Events Versus Policies With Claims

In a hypothetical portfolio, 300 devices are each observed for four months. In each device-month, a claim occurs with probability 0.02, independently across devices and months. Find (a) the expected number of claim events and (b) the expected number of devices with at least one claim during the four months.

Method. Count device-months for the event total, then calculate the four-month probability for the distinct-device total. This avoids equating successful monthly trials with affected devices.

There are \(300(4)=1{,}200\) device-month trials. The total number of claim events is binomial with \(n=1{,}200\) and \(p=0.02\), so its expected value is \(1{,}200(0.02)=24\) claims.

For one device, the probability of no claim in any of its four months is \((1-0.02)^4=0.98^4\). Since \(0.98^2=0.9604\), \(0.98^4=0.9604^2=0.92236816\). Thus the probability of at least one claim for a device is \(1-0.92236816=0.07763184\).

The expected number of devices with at least one claim is \(300(0.07763184)=23.289552\), or 23.29 devices to the nearest hundredth.

Check. The expected 24 claims count successful device-months; the expected 23.29 counts devices with one or more such successes. The latter is smaller because a device with claims in multiple months contributes multiple events but only one affected device.

Changing the Observation Period in a Binomial Model

A frequent adjustment error is to split an annual probability evenly across shorter periods without checking the model. Suppose the annual probability of at least one claim is \(p_{\text{annual}}\), and assume two equal half-year periods have independent, identical claim opportunities. If \(q\) is the half-year probability, then no claim in the year requires no claim in either half-year. Therefore \(1-p_{\text{annual}}=(1-q)^2\), which determines \(q\). This is a model-based conversion, not a universal consequence of knowing the annual probability alone.

q=1-\sqrt{1-p_{\text{annual}}},\qquad p_{\text{annual}}=1-(1-q)^2.

Worked Example: Converting an Annual Claim Probability

A hypothetical product has an annual probability of at least one claim of 0.06 per contract. Assume the two half-year periods have independent, identical claim opportunities. For 500 contracts, find the probability of at least one claim in a half-year and the expected number of contracts with a claim in that half-year.

Method. Use the complement relationship for two independent half-years. This avoids simply dividing the annual probability by two.

Let \(q\) be the half-year probability. The annual probability of no claim is \(1-0.06=0.94\). Under the stated assumptions, \( (1-q)^2=0.94\), so \(1-q=\sqrt{0.94}=0.96953597\ldots\). Therefore \(q=1-\sqrt{0.94}=0.03046403\ldots\).

For 500 contracts, the expected number with a claim in one half-year is \(500(0.03046403\ldots)=15.232014\ldots\), or 15.23 contracts to the nearest hundredth.

Check. Recombine the half-years: \(1-(1-q)^2=1-0.94=0.06\), as required. Simply setting \(q=0.03\) would instead give annual probability \(1-0.97^2=0.0591\), not 0.06. The annual probability alone does not establish this conversion unless the interval assumptions are suitable.

Common Errors and Exam Checks

Before calculating, write down the count being modeled, the exposure unit, and the probability interval. Then check whether changing the observation period changes the number of trials, the probability per trial, or both. Several tempting calculations fail this simple audit.

Label the frequency or probability.
Write “claims per policy-year” or “probability per policy-month,” rather than recording an unlabeled decimal.

Convert to matching exposure units.
For a rate per policy-year, convert months observed into fractions of a year before multiplying.

Define the outcome being counted.
Decide whether the target is claim events, successful period-level trials, or distinct policies with claims.

Reconcile periods and probabilities.
If a binomial interval changes, adjust the success probability consistently with the stated risk model; verify that recombining intervals gives the original probability.

Common mistake—scaling a rate without its exposure base. If a rate is 0.12 per policy-year for 75 policies observed four months each, the incorrect calculation \(75(0.12)=9\) treats every policy as a full year. The correction is \(75(4/12)(0.12)=3\) expected claims.

Common mistake—confusing claims with affected policies. In the device example, \(1{,}200(0.02)=24\) is the expected number of claim events. It is not the expected number of distinct devices with claims; that is \(300[1-(0.98)^4]=23.289552\).

Common mistake—changing \(n\) but not \(p\). If one annual Bernoulli trial is replaced by two half-year trials, using the original annual probability for each half-year would change the annual probability to \(1-(1-p_{\text{annual}})^2\). Under the independent equal-period assumption, use \(q=1-\sqrt{1-p_{\text{annual}}}\) instead.

Key takeaway: Exposure scaling is valid only when the rate and exposure use matching units and the model supports combining those exposures. For binomial counts, always state what one trial represents; changing the period can change both the trial count and the trial probability.

Check Your Understanding

For each problem, identify whether the requested quantity is a claim-event count or a count of policies with claims.

A hypothetical portfolio has a Poisson rate of 0.09 claims per machine-year. Forty machines are observed for six months each. Find the expected claim count and the probability of no claims, to four decimal places. (Answer: 1.8 expected claims; \(e^{-1.8}=0.1653\).)
A Poisson rate is 7 claims per 100 policy-years. A group of 250 policies is observed for eight months each. Find the expected number of claims. (Answer: \(7(250)(8/12)/100=11.6667\) expected claims.)
Each of 150 accounts has a 0.01 probability of a claim in each month, independently, and is observed for two months. Find the expected number of claim events and the expected number of accounts with at least one claim. (Answer: 3 expected claim events; \(150[1-(0.99)^2]=2.985\) expected affected accounts.)
The annual probability of at least one claim for a contract is 0.19. Assume independent, identical half-year periods. For 300 contracts, find the half-year probability and expected number of contracts with a claim in that half-year. (Answer: \(1-\sqrt{0.81}=0.10\); 30 expected contracts.)
Eighty instruments are each observed for three quarters. The probability of a claim in each instrument-quarter is 0.04, independently. Find the expected number of claim events and the expected number of instruments with at least one claim. (Answer: \(80(3)(0.04)=9.6\) expected events; \(80[1-(0.96)^3]=9.22112\), or 9.22 expected instruments.)

What You'll Learn