Common Errors In Frequency Model Assumptions

Why Frequency Model Assumptions Matter

A frequency model describes the number of claims or other events in a specified group or period. A familiar formula can produce a precise number even when its assumptions do not fit the scenario. The key diagnostic task is therefore to identify what is being counted, what creates an opportunity for a claim, and whether those opportunities behave as the model requires.

For a binomial count, the number of trials must be fixed, each trial can produce at most one counted event, and the trials must have a common claim probability and be independent. A Poisson count over exposure assumes a specified rate; a constant-rate Poisson process additionally requires that rate to remain stable over the period and that counts in disjoint intervals be independent. These are modeling assumptions, not details that can be repaired by choosing a different parameter after the calculation.

Definition: A binomial count \(N\sim\operatorname{Binomial}(n,p)\) counts successes in exactly \(n\) independent trials, each with the same success probability \(p\), with at most one success per trial. A homogeneous Poisson process with rate \(\lambda\) has independent increments and a constant event rate per unit of exposure. Its count over exposure \(t\) is Poisson with mean \(\lambda t\).

The setting determines what counts as a trial or as exposure. For example, a contract might be one trial if the outcome is simply whether that contract has any claim during a fixed term. It is not one binomial trial if the count includes multiple claims per contract. Similarly, the number of policy-years is an exposure measure, not necessarily a fixed number of one-claim trials.

Check the Binomial Conditions

A practical binomial check has four parts. First, can the number \(n\) of opportunities be stated before observing the outcomes? Second, can each opportunity contribute no more than one success? Third, is the probability \(p\) the same for every opportunity? Fourth, are the outcomes independent? Failure of any one condition means the ordinary binomial model is not justified as stated.

The equal-probability condition is distinct from independence. Independent risks can still have different claim probabilities, in which case the total is not generally binomial. Conversely, identical marginal probabilities do not establish independence: shared weather, common equipment, or a common operational decision can link outcomes.

Worked Example: A Fixed Set of One-Claim Opportunities

A hypothetical laboratory has five separate temperature monitors. During a specified month, each monitor either fails once or does not fail; a failed monitor is replaced immediately, so it cannot contribute a second failure to the monthly count. Assume each monitor has failure probability \(0.20\), and failures are independent. Find the probability of exactly two failures and check the model conditions.

Method and conditions. The five monitors are a fixed set of five trials. Each contributes zero or one failure, all have the same probability \(p=0.20\), and independence is given. Thus the monthly failure count has a binomial model with \(n=5\) and \(p=0.20\).

For exactly two failures, choose which two of the five monitors fail. The remaining three must not fail:

P(N=2)=\binom{5}{2}(0.20)^2(0.80)^3 =10(0.04)(0.512) =0.2048.

As an additional check, the expected count is \(np=5(0.20)=1\), which is between zero and the maximum possible count of five. The calculated probability, \(0.2048\), is also between zero and one. The model is appropriate here because the count is over a fixed number of independent, one-failure opportunities with a common probability.

Common mistake avoided. Treating “five monitors” as enough information by itself would be incomplete. If a monitor could fail several times and the count included every failure, the one-success-per-trial condition would fail even though the number of monitors remained fixed.

Why Independence Is More Than a Convenience

Independence affects how the variability of a total count is calculated. Let \(I_j\) equal 1 if trial \(j\) is a success and 0 otherwise. For a binomial count, \(N=I_1+\cdots+I_n\), and each indicator has mean \(p\) and variance \(p(1-p)\). Therefore,

E[N]=np,\qquad \operatorname{Var}(N) =\sum_{j=1}^{n}\operatorname{Var}(I_j) +2\sum_{j<k}\operatorname{Cov}(I_j,I_k) =np(1-p),

because independence makes every covariance term zero. If outcomes are dependent, those terms need not be zero, so the binomial variance formula generally fails. Positive dependence, for example, can make high-count and zero-count outcomes more likely than an independent model with the same marginal probabilities suggests.

Worked Example: Shared Weather Creates Dependent Claims

Suppose three hypothetical roadside sensors are exposed to the same storm. With probability \(0.20\), the storm damages all three sensors; otherwise none is damaged. Let \(N\) be the number of damaged sensors. Compare the actual count with the binomial model that would be incorrectly obtained by treating the sensors as independent, each with its correct marginal damage probability.

Step 1: Find each marginal probability. Each sensor is damaged exactly when the shared storm event occurs, so each has probability \(p=0.20\). The mean count is therefore \(E[N]=3(0.20)=0.60\).

Step 2: Describe the actual count. The actual distribution puts probability \(0.80\) on \(N=0\) and probability \(0.20\) on \(N=3\). Thus,

P(N\ge 2)=0.20,\qquad E[N]=0(0.80)+3(0.20)=0.60,

and

\operatorname{Var}(N) =E[N^2]-(E[N])^2 =9(0.20)-(0.60)^2 =1.80-0.36 =1.44.

Step 3: Calculate what the incorrect independence assumption would imply. If the three outcomes were independent with \(p=0.20\), the count would be \(\operatorname{Binomial}(3,0.20)\). The probability of at least two damages would be

\binom{3}{2}(0.20)^2(0.80)+(0.20)^3 =3(0.04)(0.80)+0.008 =0.104.

Its variance would be \(3(0.20)(0.80)=0.48\), not the actual \(1.44\). The independent model understates the probability of at least two damages, \(0.104\) rather than \(0.20\), and understates the variance. The mean agrees because the three marginal probabilities agree; matching the mean does not establish that a binomial model is valid.

Common mistake avoided. The incorrect calculation uses the correct marginal probability but assumes away the shared cause. The model check must consider how outcomes are related, not just whether each risk has the same individual probability.

Check the Poisson Rate and Exposure

A Poisson count model requires care about what “rate” means. For a homogeneous Poisson process with constant rate \(\lambda\) per unit of exposure, the count in exposure \(t\) has mean \(\lambda t\). This is appropriate when the rate is stable over the interval being modeled and the process has independent increments. The rate’s units must match the exposure’s units.

A constant-rate assumption can be inappropriate when claim intensity changes by season, time of day, location, or another systematic factor. If the rate varies deterministically over time, a nonhomogeneous Poisson process uses a time-varying rate \(\lambda(s)\); the expected count over an interval from \(a\) to \(b\) is the integrated rate \(\int_a^b\lambda(s)\,ds\). Counts in disjoint intervals can still be independent, but their means need not be proportional to interval length.

An important qualification is that changing rates do not automatically invalidate a Poisson model for the total count. If the rate is deterministic and the process has independent increments, the total over the full period is Poisson with mean equal to the integrated rate. A constant-rate model using that same total mean can give the correct distribution for the full-period total, while still giving incorrect probabilities for counts in particular subperiods.

Worked Example: A Rate That Changes Between Two Periods

A hypothetical delivery portfolio is observed over two equal six-month periods. The expected claim count is 1 in the first period and 3 in the second, with independent Poisson counts in the two periods. Compare the probability of no claims in the first period under the stated changing-rate model with the probability from an incorrectly assumed constant rate across the year. Also check the full-year total-count implication.

Step 1: Use the specified first-period mean. The first-period count is Poisson with mean 1. A Poisson count with mean \(\mu\) has zero-count probability \(e^{-\mu}\), so

P(N_1=0)=e^{-1}=0.367879\ldots\approx 0.3679.

Step 2: Calculate the incorrect constant-rate result for the subperiod. The total expected count is \(1+3=4\) over the year. A constant-rate assumption would spread this mean evenly over the two equal periods, assigning mean \(4/2=2\) to the first half-year. It would therefore give

P(N_1=0)\text{ under the constant-rate assumption} =e^{-2}=0.135335\ldots\approx 0.1353.

The difference occurs because the true first-period mean is 1, not 2. The constant annual average rate does not describe the timing of claims correctly.

Step 3: Check the full-year total. The sum of independent Poisson counts with means 1 and 3 is Poisson with mean \(1+3=4\). A constant-rate Poisson model with annual mean 4 gives the same full-year total-count distribution. In particular, both models give annual probability of zero claims \(e^{-4}=0.018315\ldots\), or \(0.0183\) to four decimal places. That agreement for the annual total does not make the constant-rate assumption valid for the first half-year.

Common mistake avoided. Using the annual average rate for every subperiod confuses a correct total exposure mean with a correct time pattern. State the interval of the requested count before assigning its Poisson mean.

A Practical Assumption Audit

Before applying a frequency formula, write down the count’s scope and inspect the assumptions that generate it. The following questions help expose errors without being distracted by a familiar formula.

Define the count.
Specify the event being counted, the population or exposure, and the time interval. “Claims” may mean contracts with at least one claim or all claim occurrences; these are different counts.

Check for fixed one-event trials.
For a binomial count, identify a fixed \(n\), a maximum of one success per trial, a common probability, and independent outcomes. Do not infer these conditions merely from a finite number of policies.

Check common causes.
Ask whether shared weather, equipment, behavior, or another factor can affect multiple risks. Equal individual probabilities do not rule out dependence.

Check rate, units, and interval.
For a Poisson model, match rate units to exposure units and assess whether a constant rate is plausible over the interval. If the rate changes, distinguish the total-period mean from the means for subperiods.

These checks matter because a calculation can be arithmetically flawless and still answer the wrong probability question. A binomial model can misstate clustering when risks are dependent. A stable-rate Poisson model can misstate when claims occur, even if its full-period mean is correct. State the assumptions explicitly, and make clear which conclusions depend on them.

Key takeaway: A fixed number of independent, equal-probability, one-claim opportunities supports a binomial model. A constant-rate Poisson model requires a stable rate over the specified exposure interval. Check dependence and changing rates before trusting a frequency calculation.

Check Your Understanding

For each scenario, identify the assumption issue and give the appropriate conclusion about the stated model.

A repair plan covers 14 machines. The count is the number of machines with at least one breakdown this quarter; each machine has probability \(0.10\) of a breakdown, independently. Is a binomial model supported, and what are its parameters? (Answer: Yes; \(N\sim\operatorname{Binomial}(14,0.10)\), since the 14 fixed machines are independent one-success trials with a common probability.)
A portfolio contains 20 policies, but each policy can generate several claims and the count includes all claim occurrences. Is the count automatically binomial with \(n=20\)? (Answer: No; a policy can contribute more than one counted event, so the one-success-per-trial condition fails.)
Four independent accounts have claim probabilities \(0.05, 0.05, 0.10,\) and \(0.10\), with each account contributing at most one claim. Is their total count binomial? (Answer: No; the probabilities are not all equal, although the accounts are independent and each is a one-claim opportunity.)
Two buildings each have marginal fire-claim probability \(0.08\), but one shared electrical fault can affect both. Is a binomial model justified from the marginal probabilities alone? (Answer: No; dependence from the shared fault must be assessed, and equal marginal probabilities do not establish independence.)
A homogeneous Poisson model has a constant rate of 6 claims per year. What mean does it assign to a three-month period, and what assumption is essential to this scaling? (Answer: Mean \(6(3/12)=1.5\) claims; the constant annual rate must apply throughout the year.)
Independent Poisson counts have means 2 in the first half of a year and 5 in the second. What is the full-year total distribution, and does a constant-rate model with the same total mean reproduce the first-half count distribution? (Answer: The total is Poisson with mean 7; no, a constant-rate model assigns first-half mean 3.5 rather than 2.)

What You'll Learn