The Distributions Analysts Actually Use: Interview-Ready Mental Models
A grocery app can have the same average order volume on two evenings and still break down on only one of them. The difference is hidden in the distribution - whether orders arrive smoothly, in bursts, or with a fat tail of extreme demand.
- A probability distribution tells you not just the average outcome, but how likely each possible outcome is.
- Use binomial when there are fixed yes/no trials, such as 100 users converting or not converting.
- Use Poisson when counting rare or random events in a fixed time or space, such as tickets per hour.
- Use normal for symmetric continuous variation, especially averages, measurement errors and many aggregated metrics.
- Use exponential for waiting time until the next event, such as time until the next support call.
- Use lognormal when values are positive and right-skewed, such as order value, delivery time or income.
- The interview-safe rule: first identify the variable type, then the business process, then the risk in the tail.
The big idea is simple: analysts do not choose distributions because formulas look elegant. They choose them because the shape of uncertainty changes the decision - stock more, staff more, flag fraud, price risk, or ignore noise.
Core Explanation: The Distributions Analysts Actually Use
A random variable is a numerical outcome of an uncertain process - conversions tomorrow, claims this month, order value per customer, or waiting time between calls. A distribution is the probability pattern of that variable.
In business analytics, five distributions appear again and again because they match common operating realities.
Here is the fast way to choose.
Worked Example: Poisson for Staffing Risk
Suppose a support desk receives an average of 6 tickets per hour. If arrivals are random and roughly independent, model the hourly ticket count as X ~ Poisson(λ = 6).
The probability of exactly 10 tickets in an hour is:
P(X = 10) = e-6 × 610 / 10! ≈ 0.041, or about 4.1%.
So what? If one agent can comfortably handle 6 tickets per hour, the average says one agent is enough. The distribution says there is a meaningful chance of overload hours, so staffing should consider service-level risk, not just the mean.
The Shapes You Must Recognise
In interviews and projects, you rarely need to derive the formula. You need to recognise the shape and explain its business implication.
Definitions
- Probability distribution: A rule assigning probabilities to the possible values of a random variable.
- Random variable: A numerical outcome determined by chance in an uncertain process.
- Expected value: The long-run probability-weighted average value of a random variable.
- Variance: The expected squared deviation of a random variable from its mean.
- Standard deviation: The square root of variance, measured in the same units as the variable.
Case Study: Swiggy Instamart and the Cost of Average Thinking
Swiggy Instamart shows why quick-commerce decisions depend on distributions of demand, arrivals and delivery times - not just average orders.
Situation: Quick-commerce is operationally unforgiving. A dark store may look well planned on average, but demand can spike by hour, neighbourhood, weather, payday, festival, cricket match, or a competitor campaign. Some SKUs move smoothly; others are lumpy and unpredictable.
The move: The distribution lens changes the operating question. Instead of asking, “What is average demand for milk, chips or soft drinks?”, the planner asks, “What is the probability demand exceeds stock before replenishment?” Instead of asking, “What is average delivery time?”, the operations team asks, “What fraction of orders will breach the promised window?”
Outcome or lesson: The primary driver of execution is hyperlocal distribution-aware planning - matching inventory and capacity to the uncertainty pattern in each catchment. Supporting drivers include dark-store proximity, app-level demand signals, rider availability, substitutions and replenishment discipline. The strategic lesson is sharp: in operations-heavy businesses, the average may describe the business, but the tail often determines the customer experience.
How AI Changes The Distributions Analysts Actually Use
AI does not remove distributions; it makes them more important because modern models often output probabilities, quantiles and risk bands, not just single predictions.
- Probabilistic forecasting: Demand tools now forecast ranges such as P50, P75 and P90 demand. This helps planners decide whether to stock for the median day or the riskier high-demand tail.
- Anomaly and fraud detection: ML models learn the normal distribution of behaviour for a user, merchant or location, then flag transactions in the extreme tail for review.
- Simulation and scenario planning: Instead of one forecast, analysts simulate thousands of possible futures for cash flows, delivery loads, credit losses or campaign response.
Take a small CSV of business data, load it into ChatGPT Advanced Data Analysis, and ask: “Identify whether this variable looks normal, lognormal, Poisson or binomial; show a histogram, summary statistics, and one business decision affected by the tail.” Then verify the recommendation using your own business logic.
Interview Relevance
“If you were analysing customer arrivals, conversion rates and order values for an app, which probability distributions would you use and why?”
If you are unsure of the exact distribution, say: “I would first plot the histogram and compare the shape with the business process.” That sounds more analytical than forcing a formula.
Common Mistake
The costly mistake is using the normal distribution for everything. It fails when the variable is a count, bounded, rare-event driven, or heavily skewed. One-line fix: classify the variable first - yes/no, count, waiting time, symmetric value, or skewed positive value - and only then name the distribution.
What to Revise Next
Now that distributions are clear, revise how samples behave and how analysts convert sample evidence into business confidence.