Random Variables and Probability Distributions

Introduction

Business decisions are made under uncertainty. A retailer does not know exactly how many customers will arrive tomorrow, a bank cannot predict with certainty which borrowers will default, and a service manager cannot know the precise time until the next customer request. Probability distributions provide a systematic language for representing these uncertain outcomes.

A random variable converts an uncertain outcome into a numerical value. A probability distribution describes the possible values of that variable and the probabilities associated with them. Together, they allow analysts to calculate expected performance, quantify risk, estimate service requirements, simulate scenarios, and select appropriate statistical models.

This chapter develops the principal discrete and continuous distributions used in business analytics: Bernoulli, binomial, Poisson, uniform, normal, and exponential distributions. It also explains expectation, variance, standard deviation, and selected business risk measures. The central analytical skill is not merely calculating probabilities, but choosing a distribution whose assumptions reasonably represent the business process.

Learning Objectives

By the end of this chapter, you should be able to:

1. Distinguish random variables from their possible outcomes.
2. Explain probability mass functions, probability density functions, and cumulative distribution functions.
3. Model binary outcomes using the Bernoulli distribution.
4. Calculate binomial and Poisson probabilities and interpret their assumptions.
5. Model continuous outcomes using uniform, normal, and exponential distributions.
6. Calculate expected values, variances, standard deviations, and selected risk measures.
7. Standardize normal variables using z-scores.
8. Select an appropriate distribution for a practical business problem and evaluate limitations.

1. Random Variables and Distributions

1.1 Outcomes, sample spaces, and random variables

An experiment is any process that produces an uncertain result. Its sample space, written as Ω, is the set of all possible outcomes. For example, a customer-service interaction may result in a resolution or no resolution; a daily store may receive 0, 1, 2, … customers; and a delivery time may be any positive number of hours.

A random variable, usually denoted by X, assigns a numerical value to each outcome. The assignment is governed by the process, not by the analyst’s preference. Random variables are classified as:

• Discrete: take countable values, such as the number of defects, claims, or purchases.
• Continuous: take values on an interval, such as revenue, waiting time, temperature, or delivery duration.

The distinction matters because discrete variables are described by probabilities assigned to individual values, whereas a continuous variable has probability distributed over intervals. For a continuous variable, the probability of any exact single value is zero; probabilities are assigned to ranges.

1.2 Probability mass, density, and cumulative functions

For a discrete random variable X, the probability mass function (PMF) is

It must satisfy p(x) ≥ 0 for every x and

where the sum is over all possible values of X.

For a continuous random variable, the probability density function (PDF) is f(x). A density is not itself a probability at a point. Probabilities are areas under the density:

The density must be nonnegative and have total area 1.

Both discrete and continuous variables have a cumulative distribution function (CDF):

For a discrete variable, F(x) = ∑_{t ≤ x} p(t). For a continuous variable, F(x) = ∫_−∞^x f(t)dt. The CDF is especially useful for calculating lower-tail probabilities, percentiles, and exceedance probabilities.

1.3 Expectation and variance

The expected value or mean of a discrete random variable is its probability-weighted average:

For a continuous random variable:

The expected value may not be an attainable outcome. For example, the expected number of daily complaints may be 4.7 even though the business cannot receive 4.7 complaints on a particular day.

Variance measures dispersion around the mean:

For a discrete variable, an equivalent computational formula is

where E(X²) = ∑ x²p(x). The standard deviation is

Variance is expressed in squared units, whereas standard deviation has the same units as X and is therefore usually easier to interpret.

For a linear transformation Y = a + bX,

and

provided a and b are constants. Adding a constant changes the mean but not the variance; multiplying by b changes the standard deviation by |b|.

2. Discrete Probability Distributions

2.1 Bernoulli distribution

A Bernoulli experiment has exactly two outcomes: success and failure. Let X = 1 indicate success and X = 0 indicate failure. If the probability of success is p, then the probability of failure is 1 − p, where 0 ≤ p ≤ 1. The PMF is

The Bernoulli mean and variance are

A Bernoulli model is suitable when there is one trial, one binary outcome, a clearly defined success, and a stable probability p. Examples include a loan being approved, a transaction being fraudulent, or a customer renewing a subscription.

Example. If the probability that a randomly selected online order is returned is 0.08, define X = 1 for a return and X = 0 otherwise. Then E(X) = 0.08, and Var(X) = 0.08 × 0.92 = 0.0736. The standard deviation is approximately 0.271. For one order, the expected return indicator is 0.08; across many independent orders, this probability becomes useful for forecasting the number of returns.

2.2 Binomial distribution

The binomial distribution counts successes in n Bernoulli trials. It requires:

1. A fixed number n of trials.
2. Two outcomes per trial.
3. The same success probability p for each trial.
4. Independence between trials, or an approximation to independence.

If X is the number of successes, then

where C(n,x) = n!/[x!(n−x)!] is the number of ways to select x successes from n trials. The mean and variance are

and the standard deviation is √[np(1−p).

Worked example. A retailer estimates that 6% of shipped packages arrive late. Assume the late status of 20 packages is independent and the same probability applies to each package. Let X be the number of late packages. Then X ∼ Binomial(n = 20, p = 0.06).

The probability of exactly two late packages is

Since C(20,2) = 190, this is approximately 0.221. The expected number is 20 × 0.06 = 1.2, and the variance is 20 × 0.06 × 0.94 = 1.128. Thus, the standard deviation is approximately 1.062 packages. The expected value is a planning average, not a promise that every group of 20 packages will contain exactly 1.2 late shipments.

The binomial model can also calculate cumulative probabilities such as P(X ≤ 2), obtained by summing the probabilities for x = 0, 1, and 2. In software, cumulative functions should be used carefully, with the event definition and parameter values documented.

2.3 Poisson distribution

The Poisson distribution models the number of events occurring in a fixed interval of time, space, length, or volume when events occur independently at a stable average rate. If λ is the expected number of events in the interval, then

where e is approximately 2.71828 and λ > 0. Its distinctive property is

The rate must match the interval. If a help desk receives an average of 12 calls per hour, then the expected number in 30 minutes is λ = 6, assuming a stable rate. If the average is 12 calls per hour, the expected number in an eight-hour shift is λ = 96.

Example. A warehouse receives an average of 3 urgent requests per day. Assuming a Poisson process, the probability of exactly 5 requests on a day is

The probability of no urgent requests is P(X = 0) = e⁻³ ≈ 0.0498. The expected number and variance are both 3, and the standard deviation is √3 ≈ 1.732 requests per day.

The Poisson model is often used for arrivals, claims, defects, incidents, and failures. It is inappropriate when events strongly influence one another, when the rate changes substantially during the interval, or when the process has a fixed upper limit that is important. If a population is finite and sampling is without replacement, a hypergeometric model may be more suitable.

2.4 Relationship among Bernoulli, binomial, and Poisson models

A Bernoulli variable records one binary trial. A binomial variable counts successes over a fixed number of Bernoulli trials. A Poisson variable counts events over an interval with a rate parameter. When n is large, p is small, and λ = np, a binomial distribution may sometimes be approximated by a Poisson distribution. This approximation should be evaluated rather than assumed; a common practical guideline is that p is small and n is sufficiently large, but the acceptable error depends on the decision context.

3. Continuous Probability Distributions

3.1 Uniform distribution

A continuous variable X has a uniform distribution on the interval [a,b] when every equal-length subinterval has equal probability. Its density is

and zero elsewhere. Its mean and variance are

The interval width must use consistent units. If a delivery is equally likely to occur at any time between 2:00 p.m. and 2:30 p.m., then the waiting time in minutes is uniform on [0,30]. The probability of waiting less than 10 minutes is 10/30 = 1/3, because the favorable interval is one-third of the total interval.

Uniformity is a strong assumption. It may be a reasonable approximation when only minimum and maximum values are known and no value within the interval is more plausible, but real business processes often have clustering, peaks, or operational constraints.

3.2 Normal distribution

The normal distribution is a continuous, bell-shaped, symmetric distribution defined by mean μ and standard deviation σ, where σ > 0. Its density is

The parameters determine location and spread. The mean, median, and mode are equal. Approximately 68%, 95%, and 99.7% of observations lie within one, two, and three standard deviations of the mean, respectively, when the population is exactly normal. These percentages are useful guidelines, not substitutes for exact probability calculations.

A major advantage of the normal distribution is standardization. If X ∼ Normal(μ,σ²), then

has a standard normal distribution with mean 0 and standard deviation 1. Therefore,

A z-score measures how many standard deviations an observation lies above or below the mean.

Worked example. Suppose weekly customer spending is approximately normal with mean $80 and standard deviation $15. The probability that a customer spends more than $110 is found by

Thus P(X > 110) = P(Z > 2). Using the standard normal table, this is approximately 0.0228, or 2.28%. The normal model suggests that high-spending customers above $110 are relatively uncommon.

To calculate an interval probability, subtract CDF values:

where Φ is the standard normal CDF. Normal modeling should be checked using histograms, box plots, Q–Q plots, subject-matter knowledge, and, where appropriate, formal diagnostics. A normal model can produce negative values, so it is unsuitable for inherently nonnegative variables when the probability of negative predictions is material.

3.3 Exponential distribution

The exponential distribution models a nonnegative waiting time until the next event in a Poisson process with constant rate λ. Its density is

and zero for x < 0. The CDF and survival probability are

where λ is measured in events per unit of time and x uses the corresponding time unit. The mean and variance are

The exponential distribution has the memoryless property:

This means that, under the model, the elapsed waiting time does not alter the distribution of the remaining wait. It does not mean that all real systems are memoryless.

Example. A support center receives, on average, 6 calls per hour. If arrivals follow a Poisson process, the waiting time until the next call is exponential with λ = 6 per hour. The probability of waiting more than 10 minutes is obtained by converting 10 minutes to 1/6 hour:

The expected waiting time is 1/6 hour, or 10 minutes. The model is not appropriate if arrival rates vary substantially by time of day or if calls arrive in clusters.

4. Business Risk Measures and Interpretation

Probability distributions support more than point forecasts. They describe uncertainty around forecasts and allow decision-makers to compare downside exposure.

4.1 Expected loss and variability

If a loss variable L has possible losses lᵢ and probabilities pᵢ, expected loss is

Expected loss is useful for budgeting and pricing, but it does not describe how severe an individual loss may be. Variance and standard deviation measure dispersion, while percentiles identify thresholds exceeded with specified probabilities.

For a continuous loss variable, the α-quantile qα satisfies

For example, the 95th percentile is a value not exceeded in approximately 95% of modeled cases. The corresponding exceedance probability is approximately 5%.

Value at Risk (VaR) at level α is commonly defined as a loss quantile:

For a normal loss model with mean μ and standard deviation σ,

where zα is the α quantile of the standard normal distribution. This formula applies only when the loss distribution is appropriately modeled as normal and losses are defined consistently. VaR does not describe the average loss beyond the threshold.

Expected Shortfall (ES) at level α is the conditional average loss in the tail beyond VaR. For continuous L,

Expected Shortfall often provides more information about extreme severity, although its estimate may be sensitive to limited tail data and model assumptions.

4.2 Portfolio and operational interpretation

If independent random costs are added, expected total cost equals the sum of expected costs. Variances also add under independence. If costs are correlated, covariance matters:

Ignoring positive dependence can materially understate risk. In operations, distributions can support staffing decisions, safety-stock calculations, service-level targets, and contingency planning. In finance and marketing, they support credit-risk estimates, campaign response modeling, and customer-value analysis.

5. Selecting a Suitable Distribution

Distribution selection should begin with the measurement process rather than with a favorite formula.

A sound selection process includes the following steps:

1. Define the unit of observation and the variable.
2. Identify whether it is discrete or continuous.
3. Examine support: can values be negative, bounded, or only nonnegative?
4. Consider whether trials are independent and whether rates are stable.
5. Estimate parameters from reliable historical or experimental data.
6. Compare model implications with observed data using plots and diagnostics.
7. Test sensitivity to plausible alternative models.
8. Document assumptions, estimation methods, and limitations.

A distribution is a model, not a fact. Even a mathematically elegant model can create poor decisions if its assumptions are hidden or violated.

6. Illustrative Case: Contact-Center Capacity Planning

Background

A fictional online retailer operates a customer-support center. Historical data suggest that incoming calls average 18 per hour during a particular weekday period. Management must decide whether the existing team can maintain a service objective during a 30-minute planning interval.

Problem

The manager wants to estimate the number of calls in 30 minutes and the probability that more than 12 calls arrive. The manager also wants to understand the waiting time until the next call.

Analysis and approach

Assume, as an initial model, that calls occur independently at a stable rate. Since the question concerns a count in a fixed interval, use a Poisson distribution. The interval is half an hour, so

Let X be the number of calls in 30 minutes. Then X ∼ Poisson(9). Its expected value and variance are both 9, and its standard deviation is 3 calls.

The probability of exactly 12 calls is

The probability of more than 12 calls is

Using a Poisson CDF gives approximately 0.208, or 20.8%. This indicates a meaningful chance of a relatively high arrival volume, not a guarantee that staffing will be inadequate. Staffing capacity also depends on handling time, agent availability, skill requirements, and abandonment behavior.

For the waiting time T until the next call, use an exponential distribution with rate 18 per hour. The probability that no call arrives during the next 5 minutes is

Thus, the probability of at least one call in five minutes is approximately 0.7769.

Outcome and learning points

The model gives the manager a transparent first approximation: nine calls per half hour on average, standard deviation three, and approximately one-fifth probability of exceeding 12 calls. However, the manager should validate whether rates are stable within the period. If calls cluster after email campaigns, or if the rate changes minute by minute, a nonhomogeneous Poisson model, an empirical distribution, or simulation may be more realistic. The analysis also illustrates that the Poisson count and exponential waiting-time models are linked through the same constant-rate arrival assumption.

7. Technology, Ethics, and Governance

Statistical software and spreadsheet functions make distribution calculations rapid, but they do not select a defensible model automatically. Analysts should record the distribution, parameters, data period, exclusions, and validation results. Reproducible code or documented spreadsheet formulas are preferable to unexplained outputs.

In AI-enabled analytics, probability distributions may be embedded in demand forecasts, fraud scores, credit decisions, and workforce systems. Governance concerns include:

• Data quality: biased, incomplete, or outdated data can distort estimated probabilities.
• Fairness: probabilities used for individual decisions may reproduce historical discrimination.
• Calibration: predicted probabilities should correspond to observed frequencies across relevant groups.
• Transparency: stakeholders should understand whether a probability represents a forecast, a long-run frequency, or a model-based estimate.
• Privacy and security: customer-level data used to estimate distributions must be appropriately protected.
• Tail risk: rare events may be underrepresented in historical samples, causing dangerous underestimation of extreme losses.

A responsible analyst distinguishes uncertainty in the business process from uncertainty in the estimated parameters. A fitted probability of 0.08 is not exact; it is an estimate subject to sampling error, changing conditions, and model risk.

8. Limitations and Best Practices

The distributions in this chapter are powerful foundations, but they simplify reality. Independence may fail because customers influence one another, arrivals may vary by time, and measurements may be skewed or heavy-tailed. Parameter estimates may also be unstable when data are scarce.

Best practices include:

1. Plot the data before fitting a distribution.
2. Check units and interval conversions carefully.
3. Separate model assumptions from observed facts.
4. Use out-of-sample validation where forecasting is involved.
5. Report intervals or ranges, not only expected values.
6. Examine sensitivity to alternative distributions and parameters.
7. Avoid interpreting a high-probability event as a certainty.
8. Recalibrate models when processes, policies, or markets change.
9. Communicate assumptions in language decision-makers can understand.
10. Use simulation when several uncertain processes interact and a closed-form distribution is inadequate.

Chapter Summary

A random variable represents an uncertain numerical outcome. Discrete variables use probability mass functions, while continuous variables use density functions and interval probabilities. The cumulative distribution function gives the probability that a variable is at or below a specified value.

The Bernoulli distribution models one binary trial; the binomial distribution counts successes in a fixed number of Bernoulli trials; and the Poisson distribution models event counts over an interval with a stable rate. The uniform distribution assigns equal density across a bounded interval. The normal distribution models continuous, approximately symmetric measurements and supports standardization through z-scores. The exponential distribution models waiting times associated with a constant-rate Poisson process and has the memoryless property.

Expected value describes long-run average performance, while variance and standard deviation describe dispersion. Quantiles, Value at Risk, and Expected Shortfall help represent downside exposure. Distribution choice must follow the measurement process, support, dependence structure, and operational assumptions. Statistical software can calculate probabilities, but professional analysis requires validation, documentation, ethical awareness, and sensitivity analysis.

Key Terms

Random variable: A numerical function assigning values to uncertain outcomes.

Discrete random variable: A random variable with countable possible values.

Continuous random variable: A random variable that can take values across an interval.

Probability mass function: A function assigning probabilities to values of a discrete variable.

Probability density function: A function whose area over an interval gives the probability for a continuous variable.

Cumulative distribution function: The probability that a random variable is less than or equal to a specified value.

Expected value: The probability-weighted long-run average of a random variable.

Variance: The expected squared deviation from the mean.

Standard deviation: The square root of variance, expressed in the original units.

Bernoulli distribution: A distribution for one binary trial.

Binomial distribution: A distribution for the number of successes in fixed independent Bernoulli trials.

Poisson distribution: A distribution for event counts in a fixed interval at a stable average rate.

Uniform distribution: A distribution with constant density over a specified interval.

Normal distribution: A symmetric continuous distribution defined by its mean and standard deviation.

Exponential distribution: A distribution for waiting time until an event in a constant-rate Poisson process.

Parameter: A numerical quantity defining a probability distribution.

z-score: The number of standard deviations an observation is from the mean.

Value at Risk: A specified quantile of a modeled loss distribution.

Expected Shortfall: The average loss beyond a specified tail quantile.

Memoryless property: The exponential-distribution property that elapsed waiting time does not change the remaining-wait distribution.

Review Questions

Short Answer Questions

1. What is the difference between a discrete and a continuous random variable?
2. State the assumptions required for a binomial distribution.
3. What does the parameter λ represent in a Poisson distribution?
4. Why is the probability at one exact point zero for a continuous random variable?
5. Interpret a z-score of −1.5.

Long Answer Questions

1. Explain the relationship among Bernoulli, binomial, and Poisson distributions, including an example of each.
2. Derive or explain the mean and variance of a binomial random variable and interpret them in a business context.
3. Compare the normal and exponential distributions in terms of shape, support, parameters, and applications.
4. Explain why expected loss alone is insufficient for managing operational or financial risk.
5. Discuss a systematic procedure for selecting and validating a probability distribution.

Analytical / Application Questions

1. A subscription business expects 4% of 50 randomly selected customers to cancel in a month. Specify an appropriate model, calculate the expected number of cancellations and variance, and identify assumptions that should be examined.
2. A delivery center receives an average of 10 urgent requests per day. Using a Poisson model, calculate the probability of no urgent requests and explain how the corresponding waiting-time model could be used.
3. Weekly demand is approximately normal with mean 1,000 units and standard deviation 120 units. Calculate the z-score for demand of 1,240 units and explain how a manager could use the result in inventory planning.

Multiple Choice Questions with Answers

Question 1

Which distribution models one trial with exactly two possible outcomes?

A. Bernoulli

B. Poisson

C. Uniform

D. Exponential

Correct Answer: A. Bernoulli

Question 2

If X follows a binomial distribution with n = 20 and p = 0.10, what is E(X)?

A. 0.10

B. 2

C. 10

D. 20

Correct Answer: B. 2

Question 3

For a Poisson random variable with λ = 6, what is its variance?

A. 2.45

B. 3

C. 6

D. 36

Correct Answer: C. 6

Question 4

A continuous variable is uniformly distributed from 10 to 30. What is its mean?

A. 10

B. 15

C. 20

D. 30

Correct Answer: C. 20

Question 5

A normal variable has mean 80 and standard deviation 15. What is the z-score for x = 110?

A. −2

B. 0.5

C. 2

D. 6.33

Correct Answer: C. 2

Question 6

If arrivals follow a Poisson process with rate 12 per hour, which distribution models the waiting time until the next arrival?

A. Binomial with p = 12

B. Exponential with λ = 12 per hour

C. Uniform from 0 to 12 hours

D. Normal with mean 12

Correct Answer: B. Exponential with λ = 12 per hour

Question 7

For a Bernoulli variable with p = 0.30, what is the variance?

A. 0.09

B. 0.21

C. 0.30

D. 0.70

Correct Answer: B. 0.21

Question 8

Which condition is essential for the standard binomial model?

A. The number of trials is random

B. Each trial has the same success probability and trials are independent

C. The variable must be continuous

D. The event rate must equal the variance

Correct Answer: B. Each trial has the same success probability and trials are independent

Question 9

If X has an exponential distribution with λ = 4 per hour, what is the probability that the wait exceeds 0.5 hour?

A. e⁻² ≈ 0.1353

B. e⁻⁰·⁵ ≈ 0.6065

C. 1 − e⁻² ≈ 0.8647

D. 0.5⁴ = 0.0625

Correct Answer: A. e⁻² ≈ 0.1353

Question 10

What does Value at Risk at the 95% level represent under the loss-quantile definition?

A. The average loss in every period

B. The loss value not exceeded in approximately 95% of modeled cases

C. The probability of a 95% gain

D. The variance of the loss distribution

Correct Answer: B. The loss value not exceeded in approximately 95% of modeled cases