Probability Distributions Explained

A probability distribution is the full picture of how a random quantity behaves: it lists, for every possible value, how likely that value is. Where a single probability answers "how likely is this one outcome," a distribution answers "how are all the outcomes spread out." Recognizing which distribution your data follows is what tells you which summary statistics, charts and tests are appropriate.

Analysts do not need to derive distributions from scratch, but they do need to recognize the common families on sight and know what each one models. This page covers the discrete-versus-continuous split and the four distributions you will meet most often.

Discrete versus continuous

The first fork is whether the variable is discrete (countable) or continuous (measurable over a range).

Discrete variables take separate, countable values: number of orders, number of defects, heads in ten coin flips. Their distribution is a probability mass function (PMF) that assigns a probability to each exact value, and those probabilities sum to 1.

Continuous variables take any value in a range: height, temperature, response time. Their distribution is a probability density function (PDF), a smooth curve where probability is the area under the curve over an interval. For a continuous variable, the probability of any single exact value is essentially zero — you only ask about ranges, like "P(height between 160 and 170 cm)." The total area under any PDF is 1.

The binomial distribution

The binomial models the number of successes in a fixed number of independent yes-or-no trials, each with the same success probability p. Think "how many of 10 emails get opened if each has a 50% open rate."

The probability of exactly k successes in n trials is:

P(k) = C(n, k) * p^k * (1 - p)^(n - k)

where C(n, k) is the number of ways to choose k successes from n trials. Let us compute the chance of exactly 5 heads in 10 fair coin flips.

from math import comb

n, k, p = 10, 5, 0.5
prob = comb(n, k) * (p ** k) * ((1 - p) ** (n - k))
print(round(prob, 4))   # 0.2461

Check it by hand: C(10, 5) = 252, and 0.5^10 = 1/1024 ≈ 0.000977. So 252 × 0.000977 ≈ 0.246. Even though 5 is the single most likely count, it happens only about a quarter of the time — a useful reminder that "the average outcome" is far from guaranteed on any one run.

The Poisson distribution

The Poisson models the count of events in a fixed interval when events occur at a constant average rate, independently. Examples: support calls per hour, website errors per day, typos per page. It needs a single parameter, the mean rate λ (lambda).

P(k) = (lambda^k * e^(-lambda)) / k!

If a support line averages λ = 3 calls per hour, the probability of exactly 5 calls in an hour is:

from math import exp, factorial

lam, k = 3, 5
prob = (lam ** k) * exp(-lam) / factorial(k)
print(round(prob, 4))   # 0.1008

So about a 10% chance of exactly five calls in an hour. The Poisson is the standard model for rare-event counts and for staffing and capacity planning.

The uniform and normal distributions

The uniform distribution is the simplest: every outcome in a range is equally likely. A fair die is discrete uniform (each face 1/6). A random number between 0 and 1 is continuous uniform. It is the baseline "no preference" distribution.

The normal (Gaussian) distribution is the famous bell curve, defined by its mean and standard deviation. It is symmetric, and values cluster around the mean with predictable spread described by the 68-95-99.7 rule. It dominates statistics because, by the central limit theorem, the averages of many independent things tend toward it. The normal distribution has its own tutorial because so much rides on it.

How analysts use distributions

Choosing a distribution is really choosing a model for your process. Modeling conversions or click-throughs? That is binomial territory. Counting arrivals or defects per interval? Poisson. Working with measurements or, more importantly, with sample means? Normal. The distribution you assume determines the confidence intervals and hypothesis tests that are valid. Picking the wrong one — say, treating skewed count data as normal — produces confidence intervals that are quietly wrong.

Even when you never name a distribution explicitly, plotting a histogram and asking "what shape is this?" is a distribution question. A right-skewed histogram warns you away from mean-based summaries; a bell shape invites normal-based tools.

Common mistakes

Assuming everything is normal. Real data is often skewed or heavy-tailed. Income, wait times and counts are rarely normal. Check the histogram before reaching for normal-based methods.

Mixing up the PMF and PDF interpretation. For continuous variables, asking "what is the probability height equals exactly 170.0000 cm" has the answer zero. Only ranges have nonzero probability under a density curve.

Ignoring the binomial's assumptions. The binomial requires a fixed number of independent trials with constant probability. If trials influence each other or p drifts over time, the model breaks.

Confusing the parameter with the outcome. In the Poisson, λ is the average rate, not a count you observed. Feeding an observed count where the rate belongs gives nonsense.

In interviews

Distribution questions probe recognition and assumptions. Expect "Which distribution would you use to model X?" for scenarios like coin flips (binomial), arrivals per hour (Poisson), or heights (normal). A common follow-up is "What are its assumptions?" — name independence and constant probability or rate. You may be asked the difference between a PMF and a PDF, where the key point is exact-value probability versus area under a curve. Live binomial calculations for small n also appear, so keep the formula handy.

Where this fits in your learning path

Probability distributions build directly on probability basics and set up everything that follows. The normal distribution gets its own deep dive, and the central limit theorem explains why the normal shows up so relentlessly for sample means. Together they form the inference core of the data analytics learning path. See the data analyst roadmap for where distribution knowledge pays off in testing and forecasting.