Introduction
Operational loss frequency is a count over a defined exposure unit, such as the number of qualifying events in one business unit during one year. A frequency model affects not only the expected number of events, but also the probabilities of no events, several events, or an unusually event-heavy year. That distinction matters when translating operational risk into annual loss estimates.
The Poisson distribution is a useful baseline, but it imposes equality of the mean and variance. Heterogeneity in event rates or a mixture of periods with and without exposure can produce more variation than that baseline allows. This tutorial compares Poisson, negative binomial, and zero-inflated frequency models, and shows how their assumptions change annual count predictions. It is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.
Start With the Count and Its Exposure Unit
Frequency counts require a clear event definition and time window. A change in the reporting threshold, event-unit convention, or exposure base can change the observed counts even if the underlying risk has not changed. The earlier material on operational loss event units and exposure scaling is therefore important context: the models below assume that counts refer to comparable, consistently defined exposure periods.
For a count distribution, \(E[N]\) is its mean and \(\operatorname{Var}(N)\) is its variance. The ratio \(\operatorname{Var}(N)/E[N]\), when the mean is positive, is called the dispersion index. A Poisson distribution has dispersion index 1. A value materially above 1 is evidence of overdispersion relative to Poisson; a value below 1 is underdispersion. This comparison is a useful model-selection clue, not proof of a particular mechanism.
Poisson: The Homogeneous-Rate Baseline
In an annual loss application, \(\lambda\) is the expected event count per year at the specified exposure level. The usual interpretation assumes a stable event rate over the period, independent increments, and that events occur individually rather than in simultaneous clusters. For independent periods with the same exposure and rate, the counts are identically distributed. If exposure changes, the rate must be adjusted to the exposure unit rather than treating all periods as equivalent.
The equality of mean and variance is restrictive. If a portfolio combines units with different underlying rates, or if event rates fluctuate across years, the realized counts can vary more than a single fixed-rate Poisson model predicts.
Worked Example: Checking a Poisson Baseline Against Annual Counts
A hypothetical operations group records event counts over eight comparable business-unit-years: \(0,0,0,1,1,2,4,8\). Use the sample mean and sample variance to assess whether a Poisson model is a plausible dispersion baseline. State what the calculation can and cannot establish.
Method. Compute the sample mean \(\bar{N}\), then the unbiased sample variance \(s^2\), and compare them. Under a Poisson model, the theoretical mean and variance are equal, so a large gap is a warning that the simple model may not capture the observed variability.
The total count is \(0+0+0+1+1+2+4+8=16\). Thus $$ \bar{N}=\frac{16}{8}=2. $$ The sum of squared counts is \(0+0+0+1+1+4+16+64=86\). Using the computational formula for sample variance, $$ s^2=\frac{\sum N_j^2-n\bar{N}^2}{n-1} =\frac{86-8(2^2)}{7} =\frac{54}{7} =7.7143. $$ The sample variance is about \(7.7143/2=3.8571\) times the sample mean. This is substantial sample overdispersion relative to the Poisson equality.
Interpretation and check. The mean is exactly 2 and the sample variance is \(54/7\), so the arithmetic is internally consistent. The eight observations suggest that a fixed-rate Poisson model could understate year-to-year variation, especially the chance of high counts. They do not identify the cause: varying exposure, rate heterogeneity, clustering, or data inconsistency could contribute. A small sample also makes the variance estimate noisy. Common mistake avoided: dividing by \(n\) instead of \(n-1\) when calculating the unbiased sample variance, or treating a sample variance above the mean as conclusive proof of a particular model.
Negative Binomial: Allowing Rate Heterogeneity
This parameterization is equivalent to counting failures before \(k\) successes in a sequence of Bernoulli trials when the parameters permit integer \(k\). For frequency modeling, the mean-dispersion form is often more directly interpretable. The smaller \(k\) is, the larger the extra variance \(\mu^2/k\), holding the mean fixed.
A standard justification comes from treating the Poisson rate as heterogeneous across exposure units or periods. Suppose that, conditional on a latent rate \(\Lambda\), \(N\) is Poisson with rate \(\Lambda\), and \(\Lambda\) has a gamma distribution with shape \(k\) and scale \(\mu/k\). Then \(E[\Lambda]=\mu\) and \(\operatorname{Var}(\Lambda)=\mu^2/k\).
\(E[N]=E[E[N\mid\Lambda]]=E[\Lambda]=\mu\).
\(\operatorname{Var}(N)=E[\operatorname{Var}(N\mid\Lambda)]+\operatorname{Var}(E[N\mid\Lambda])\).
Since \(\operatorname{Var}(N\mid\Lambda)=\Lambda\) and \(E[N\mid\Lambda]=\Lambda\), the variance is \(E[\Lambda]+\operatorname{Var}(\Lambda)=\mu+\mu^2/k\).
The extra term is the between-rate contribution. A gamma mixture is a coherent explanation for overdispersion when the portfolio contains latent rate variation. It is not the only explanation, and selecting a negative binomial distribution solely because a sample variance exceeds its mean would be too strong an inference.
Worked Example: Comparing Poisson and Negative Binomial Tail Counts
For a hypothetical portfolio, suppose the expected annual event count is 2. Compare the predicted probability of at least 5 events under a Poisson model with \(\lambda=2\) and a negative binomial model with \(\mu=2\) and \(k=2\).
Method. For each model, calculate the probability of counts 0 through 4 and subtract their sum from 1. This directly compares annual high-count predictions at the same mean.
For Poisson, $$ P(N\ge 5)=1-e^{-2}\left(1+2+\frac{2^2}{2!}+\frac{2^3}{3!}+\frac{2^4}{4!}\right). $$ The terms in parentheses sum to \(1+2+2+1.3333+0.6667=7\). Therefore \(P(N\ge5)=1-7e^{-2}=1-0.9473=0.0527\), to four decimal places.
For the negative binomial, the success-probability form has \(p=k/(k+\mu)=2/4=0.5\). Its probability at \(n\) is \(\binom{n+k-1}{n}p^k(1-p)^n\). For \(k=2\), the probabilities at 0 through 4 are respectively $$ 0.2500,\quad 0.2500,\quad 0.1875,\quad 0.1250,\quad 0.078125. $$ Their sum is \(0.890625\), so \(P(N\ge5)=1-0.890625=0.109375\), or \(0.1094\) to four decimal places.
Check and interpretation. Both distributions have mean 2, but the negative binomial variance is \(2+2^2/2=4\), while Poisson variance is 2. Accordingly, its probability of at least five events is about 10.94%, compared with 5.27% under Poisson. The tail difference follows from the larger dispersion, not from a larger expected count. Common mistake avoided: comparing distributions with different means, or using the negative binomial parameter \(p\) as though it were the mean.
Zero-Inflated Poisson: A Separate Chance of No Events
The structural zero is a modeling representation of a distinct state in which events cannot occur during the period, such as a hypothetical unit being inactive. The Poisson component still permits a zero by chance. A ZIP model is appropriate only when that separate zero-generating state is substantively plausible; a large observed zero frequency alone does not establish it.
To verify the variance, let \(B\) equal 1 when the Poisson state is active and 0 in the structural-zero state, with \(P(B=1)=1-\pi\), and let \(Z\sim\operatorname{Poisson}(\lambda)\) independently. Then \(N=BZ\). Its mean is \((1-\pi)\lambda\), and \(E[N^2]=(1-\pi)(\lambda+\lambda^2)\). Subtracting the square of the mean gives $$ \operatorname{Var}(N)=(1-\pi)\lambda+\pi(1-\pi)\lambda^2. $$ When \(\pi>0\), this variance exceeds the mean by \(\pi(1-\pi)\lambda^2\). Thus ZIP can represent both an excess of zeros and additional dispersion, though its positive-count distribution has a particular form that may or may not suit the application.
Worked Example: Separating Structural Zeros From Poisson Zeros
Suppose a hypothetical set of units has probability \(\pi=0.40\) of being inactive for a year. In an active year, its annual event count is Poisson with \(\lambda=2.5\). Calculate the overall mean, variance, probability of zero events, and probability of at least 3 events. Compare the last two probabilities with a Poisson model having the same overall mean.
Method. First use the ZIP mixture formulas. For the tail, multiply the active-state Poisson tail by the probability of being active. Then calculate the corresponding probabilities for a Poisson distribution with mean equal to the ZIP mean.
The active probability is \(1-\pi=0.60\). The overall mean and variance are $$ E[N]=0.60(2.5)=1.5, $$ $$ \operatorname{Var}(N)=1.5+(0.40)(0.60)(2.5^2)=1.5+1.5=3.0. $$ The ZIP zero probability is \(0.40+0.60e^{-2.5}\). Since \(e^{-2.5}=0.0821\) to four decimals, this is \(0.40+0.0493=0.4493\).
For the ZIP probability of at least 3 events, only the active state can produce a positive count: $$ P(N\ge3)=0.60\left[1-e^{-2.5}\left(1+2.5+\frac{2.5^2}{2}\right)\right]. $$ The terms in parentheses sum to \(1+2.5+3.125=6.625\). Thus the active-state probability of at least 3 is \(1-6.625e^{-2.5}=1-0.5438=0.4562\), and the unconditional probability is \(0.60(0.4562)=0.2737\).
A same-mean Poisson model has \(\lambda=1.5\). Its zero probability is \(e^{-1.5}=0.2231\). Its probability of at least 3 is $$ 1-e^{-1.5}\left(1+1.5+\frac{1.5^2}{2}\right) =1-(0.2231)(3.625) =0.1912 $$ to four decimal places. The ZIP model therefore predicts substantially more zero years (0.4493 versus 0.2231) and more years with at least 3 events (0.2737 versus 0.1912).
Check and interpretation. The ZIP variance is 3.0, twice its mean of 1.5, as expected from the positive structural-zero contribution. The active probability is included in the tail calculation; omitting the factor 0.60 would report a conditional rather than an unconditional probability. Common mistake avoided: treating the active-state probability as the unconditional tail probability, or comparing the ZIP model with a Poisson model using the active-state rate instead of the same overall mean.
Choosing Between the Models
| Model | Core assumption | Mean and variance | What it can represent |
|---|---|---|---|
| Poisson | One stable rate for comparable exposure periods | \(\lambda,\lambda\) | Baseline frequency with equal mean and variance |
| Negative binomial | Rate variation, represented by a gamma mixture | \(\mu,\mu+\mu^2/k\) | Overdispersion and a heavier count tail than Poisson at the same mean |
| Zero-inflated Poisson | A structural-zero state plus an active Poisson state | \((1-\pi)\lambda,(1-\pi)\lambda+\pi(1-\pi)\lambda^2\) | Excess zeros together with overdispersion |
Begin with the event definition, exposure unit, and operational explanation for the data pattern. Then use the mean-variance relationship and the observed zero frequency as evidence about whether a simple Poisson model is adequate. A variance larger than the mean may support considering a negative binomial or ZIP model, but those alternatives encode different mechanisms. The negative binomial attributes extra variation to rate heterogeneity; ZIP explicitly introduces a structural-zero state.
Frequency choices affect annual event-count predictions and, consequently, the distribution of aggregate operational losses when frequency is later combined with severity. A model with the same mean can imply very different probabilities of multiple events. These differences are especially consequential when estimates are used to assess a tail-sensitive risk measure or capital requirement. Frequency alone does not determine aggregate loss: the severity distribution and the dependence assumptions also matter.
Common Mistakes
- Assuming the mean identifies the distribution. A Poisson and an overdispersed model can share the same mean but assign different probabilities to zero and large counts.
- Calling every zero a structural zero. A Poisson process naturally produces zero counts. ZIP should be used only when a separate inactive or event-impossible state is credible.
- Mixing conditional and unconditional quantities. In a ZIP model, an active-state tail probability must be multiplied by \(1-\pi\) to obtain the portfolio-wide probability.
- Treating sample overdispersion as conclusive. A sample variance above the mean is a warning, not an explanation. Exposure differences, changing definitions, dependence, and limited data can all affect the comparison.
Check Your Understanding
For each problem, identify the relevant model quantity and retain the stated exposure unit.
- A portfolio has a Poisson annual count with \(\lambda=1.2\). Calculate the probability of zero events and the probability of at least 2 events. (Answer: \(P(N=0)=0.3010\); \(P(N\ge2)=0.3374\).)
- A negative binomial annual count has mean \(\mu=4\) and dispersion parameter \(k=4\). Calculate its variance and dispersion index. (Answer: variance \(=8\); dispersion index \(=2\).)
- A ZIP annual count has structural-zero probability \(\pi=0.25\) and active-state Poisson rate \(\lambda=4\). Calculate the unconditional mean, variance, and probability of zero events. (Answer: mean \(=3\); variance \(=6\); \(P(N=0)=0.2637\).)
- A negative binomial count has \(k=3\) and mean \(\mu=3\), using the failures-before-\(k\)-successes parameterization. Calculate \(P(N\ge2)\). (Answer: \(0.6875\).)
- Two hypothetical portfolios have the same mean count of 2.5 per year. One has a variance near 2.5; the other has a variance near 7.5 and many zero years. State the main frequency-model clue and one reason that the clue alone is insufficient to select ZIP. (Answer: The second portfolio is overdispersed relative to Poisson, and excess zeros may make ZIP worth considering; however, rate heterogeneity, exposure differences, or data-definition issues could also create the pattern, so a structural-zero mechanism needs operational support.)