Binomial Counts as the Basis for Poisson Approximation

From Individual Opportunities to a Count

A count of claims, failures, or alerts often begins with a set of opportunities, each of which can produce either a defined event or no event. If there are many opportunities and each event is uncommon, it may be useful to consider a Poisson model for the total. The exact starting point, when the opportunities satisfy the right conditions, is a binomial count.

The key is not simply that the count is small or that the population is large. A binomial model describes a fixed number of independent opportunities that share the same probability of success. Stating what counts as one opportunity—and what counts as success—makes the model’s assumptions visible.

Definition: A random variable \(X\) has a binomial distribution with parameters \(n\) and \(p\), written \(X\sim\operatorname{Bin}(n,p)\), if it counts successes in \(n\) independent trials and each trial has the same success probability \(p\). Here, \(n\) is a fixed nonnegative integer and \(0\le p\le1\). The possible counts are \(0,1,\ldots,n\).

A “trial” is any clearly defined opportunity with two outcomes for the purpose of the count. For example, in a study of whether a component fails during a specified week, one component over that week can be one trial; failure is success for the count. The word success is a label for the event being counted, not a judgment that the outcome is desirable.

The assumptions should be checked in context:

Fixed number of trials: the count refers to a stated number \(n\) of opportunities.
Two relevant outcomes per trial: each opportunity either does or does not meet the success definition.
Independence: knowing the outcome of one trial does not change the probability of outcomes on the others.
Common probability: every trial has the same probability \(p\) of success.

If the number of opportunities is random, or if probabilities differ by trial, the count is not binomial in this form. Dependence can also invalidate the model. A plausible count alone does not establish these conditions; they come from the design or assumptions of the scenario.

The Exact Binomial Count

For \(X\sim\operatorname{Bin}(n,p)\), the probability of exactly \(k\) successes is

P(X=k)=\binom{n}{k}p^k(1-p)^{n-k},\qquad k=0,1,\ldots,n.

The factor \(p^k(1-p)^{n-k}\) is the probability of any one specified pattern with \(k\) successes and \(n-k\) failures. There are \(\binom{n}{k}\) different sets of \(k\) trials that could be the successes. These patterns are mutually exclusive, so their equal probabilities add to give the formula.

The binomial mean and variance are

E[X]=np,\qquad \operatorname{Var}(X)=np(1-p).

One justification uses indicator variables. Let \(I_j=1\) if trial \(j\) is a success and \(I_j=0\) otherwise. Then \(X=I_1+\cdots+I_n\), and each indicator has mean \(p\) and variance \(p(1-p)\). Linearity of expectation gives \(E[X]=np\). Independence makes the covariances between distinct indicators zero, so the variance of the sum is \(np(1-p)\). This also shows why independence matters to the variance formula.

Worked Example: Alerts Across Independent Sensor Readings

In a hypothetical monitoring system, 1,200 sensor readings are reviewed during a fixed interval. Assume each reading independently produces a false alert with probability \(0.0015\). Let \(X\) be the number of false alerts. Identify the model, calculate its mean and variance, and explain why a Poisson approximation might be considered.

Method. Define one trial as one reading and success as a false alert. There are \(n=1{,}200\) readings, the probability is common at \(p=0.0015\), and independence is an explicit assumption. Thus \(X\sim\operatorname{Bin}(1{,}200,0.0015)\).

The mean is

$$ E[X]=np=1{,}200(0.0015)=1.8. $$

The failure probability for one reading is \(1-p=0.9985\). Therefore,

\operatorname{Var}(X)=np(1-p) =1{,}200(0.0015)(0.9985) =1.8(0.9985) =1.7973.

There are many independent opportunities, while the probability of a false alert on any one reading is small. Those features motivate considering a Poisson count with parameter matching the binomial mean, \(1.8\). That is a modeling motivation, not a claim that the binomial count is exactly Poisson or that an approximation is automatically accurate.

Check. The variance, \(1.7973\), is slightly less than the mean, \(1.8\), as expected from the factor \(1-p=0.9985\). A common mistake is to use \(np\) for both quantities; \(np\) is the mean, while the binomial variance includes the additional factor \(1-p\).

Choosing What Counts as a Trial

A single situation can be counted at different levels. The total number of events across individual opportunities is not necessarily the same random variable as the number of groups containing at least one event. Each count needs its own trial definition and success probability.

Worked Example: Transactions Versus Blocks With an Alert

A hypothetical system processes 500 blocks, each containing 4 transactions. Assume all 2,000 transactions independently produce an alert with probability \(0.001\). Find the model and mean for (a) the total number of alerting transactions and (b) the number of blocks with at least one alert.

(a) Count alerting transactions. Each transaction is one trial, with success probability \(0.001\). Thus the total \(T\) has distribution \(\operatorname{Bin}(2{,}000,0.001)\), and

$$ E[T]=2{,}000(0.001)=2. $$

(b) Count blocks with one or more alerts. Now one block is the trial, and success means at least one of its four transactions alerts. A block has no alerts only if all four transactions avoid an alert. By independence, that probability is \((1-0.001)^4=0.999^4=0.996005996001\). Hence the probability of a successful block is

p_{\text{block}}=1-0.999^4 =1-0.996005996001 =0.003994003999.

The 500 block outcomes are independent because they are formed from disjoint sets of independent transactions, and each block has the same alert probability. If \(B\) is the number of blocks with at least one alert, then \(B\sim\operatorname{Bin}(500,0.003994003999)\), with

$$ E[B]=500(0.003994003999)=1.9970019995. $$

The two means are close, but the counts describe different events: \(T\) counts alerting transactions, while \(B\) counts affected blocks. A common mistake is to use \(0.001\) as the success probability for a block. The probability for a block is instead the probability of at least one alert among its four transactions.

When Probabilities Differ

If each opportunity is independent but its success probability varies, the total is a sum of independent Bernoulli indicators, but it is generally not a binomial count. For independent trials with probabilities \(p_1,\ldots,p_n\), the mean is \(\sum_j p_j\) and the variance is \(\sum_j p_j(1-p_j)\). These follow by adding the individual means and variances as in the indicator argument. The common-probability condition is what permits the binomial formula with one \(p\).

Worked Example: Two Groups With Different Event Probabilities

A hypothetical review covers 400 contracts in one class, each with an independent event probability of \(0.002\), and 200 contracts in another class, each with an independent event probability of \(0.006\). Let \(X\) count all events. Determine the mean and variance, and decide whether \(X\) is binomial.

Method. The trials are independent, but the probabilities differ between classes. The sum is therefore not binomial under the stated model. Add the group contributions to calculate its mean:

$$ E[X]=400(0.002)+200(0.006)=0.8+1.2=2. $$

For the variance, add each group’s number of trials times its Bernoulli variance:

\begin{aligned} \operatorname{Var}(X) &=400(0.002)(1-0.002)+200(0.006)(1-0.006)\\ &=400(0.002)(0.998)+200(0.006)(0.994)\\ &=0.8(0.998)+1.2(0.994)\\ &=0.7984+1.1928\\ &=1.9912. \end{aligned}

A tempting shortcut is to pool the 600 contracts and assign each the average probability \(2/600=1/300\). That would produce a binomial variance of

600\left(\frac{1}{300}\right)\left(1-\frac{1}{300}\right) =2\left(\frac{299}{300}\right) =1.993333\ldots,

which is not the actual variance, \(1.9912\). More importantly, matching the average probability does not make the individual probabilities equal or make the total count binomial. The small numerical difference does not remove the modeling distinction.

Check. The mean is \(2\), obtained from \(0.8\) expected events in the first class and \(1.2\) in the second. The variance contributions are \(0.7984\) and \(1.1928\), for a total of \(1.9912\). The count is not binomial because the trial probabilities are not common.

Why Rare Opportunities Suggest a Poisson Model

The binomial model gives an exact description when its assumptions hold. Its connection to Poisson modeling is a useful conceptual one: when there are many opportunities and each has a low probability of success, the total may consist of a relatively small number of events spread across those opportunities. A Poisson model can provide a simpler count description, with its rate chosen to match the binomial expected count \(np\).

This reasoning explains why a Poisson approximation is considered, but it does not itself establish the approximation’s accuracy. A large \(n\) by itself is not enough: if \(p\) is not small, the setting does not have the same rare-opportunity character. Nor should a heterogeneous or dependent collection of opportunities silently be labeled binomial. The model and its assumptions must be identified before choosing a calculation.

In practical work, the distinction is useful even when an approximation is ultimately used. The binomial model spells out the underlying opportunities; the Poisson model describes a count through a rate. Keeping that distinction clear prevents confusing an exact model with an approximation and prevents using a count’s average as if it determined the full distribution.

Key takeaway: A binomial count requires a fixed number of independent trials with one common success probability. Many rare opportunities can motivate a Poisson approximation, but the motivation does not make the binomial count exactly Poisson or settle whether the approximation is accurate. This treatment is aligned with the SOA VEE Mathematical Statistics syllabus.

Check Your Understanding

For each problem, identify the relevant count and use the assumptions stated. Answers are provided for checking.

A hypothetical instrument makes 900 independent calibration cycles. Each cycle has probability \(0.002\) of producing a flagged reading. Let \(X\) count flagged readings. State its binomial model, mean, and variance. (Answer: \(X\sim\operatorname{Bin}(900,0.002)\), \(E[X]=1.8\), \(\operatorname{Var}(X)=900(0.002)(0.998)=1.7964\).)
A system runs 1,500 independent jobs, each with probability \(0.0004\) of an error. Find the binomial parameters, mean, and variance, and say whether many rare opportunities motivate considering a Poisson approximation. (Answer: \(X\sim\operatorname{Bin}(1{,}500,0.0004)\), mean \(0.6\), variance \(1{,}500(0.0004)(0.9996)=0.59976\); the many low-probability independent opportunities motivate consideration, but do not guarantee accuracy.)
A hypothetical portfolio has 300 accounts, each with two independent transactions. Every transaction independently generates a notice with probability \(0.003\). Let \(X\) count notices across transactions. Give the model and its mean and variance. (Answer: There are 600 transaction trials, so \(X\sim\operatorname{Bin}(600,0.003)\), mean \(1.8\), variance \(600(0.003)(0.997)=1.7946\).)
A study counts events among 1,000 independent units with probability \(0.001\) and 500 independent units with probability \(0.004\). Find the mean and variance of the total and decide whether it is binomial. (Answer: Mean \(=1{,}000(0.001)+500(0.004)=3\); variance \(=1{,}000(0.001)(0.999)+500(0.004)(0.996)=0.999+1.992=2.991\); it is not binomial because the probabilities differ.)
Fifty independent blocks each contain 20 independent trials. Each trial has probability \(0.01\) of an event. Let \(Y\) count blocks with at least one event. Give the model and exact mean in terms of the block probability. (Answer: The block probability is \(1-0.99^{20}\); \(Y\sim\operatorname{Bin}(50,1-0.99^{20})\), with mean \(50(1-0.99^{20})\).)

What You'll Learn