All lessons

Module 3: Distributions · Lesson 3.5 · about 20 min

Waiting times and money: exponential and log-normal

Two distributions for data with a long right tail: the exponential for waiting times between events, and the log-normal for order values, incomes and prices. Why their mean sits above the median, and why logs help.

Key points, meanings and formulas on one page: choose "Save as PDF" in the print box.

Module 3 · about 20 minutes · school maths and logarithms. You’ll be able to use the exponential for waiting times, recognise log-normal money data, and explain why the median and percentiles describe both better than the mean.

Waiting times and money have a floor at zero and no ceiling, so a few huge values stretch the right tail and the mean ends up above the typical value.

In everyday terms: queue times at a bank: most people wait a few minutes, nobody waits less than zero, and now and then someone waits forty minutes. The long right tail drags the mean above what most people experience.

1. Data with a floor and no ceiling

Waiting times, session lengths, order values, incomes, house prices and claim sizes can’t go below zero but have no upper limit. They pile up near the low end with a long tail to the right. Two distributions describe most of them.

Try it: long right tails
median
mean
share above the mean
Waiting times and money are never negative and have no upper limit, so they pile up on the left with a long tail to the right. The mean lands above the median, often well above, and fewer than half the values are above the mean. Report the median (and a high percentile such as the 90th) for these, and expect a few huge orders or long waits.

2. The exponential: time between events

When events arrive as a Poisson process (Lesson 3.3) at rate λ per minute, the time between them is exponential:

f(t)=λe−λt,P(T>t)=e−λt
  • Mean 1/λ, SD also 1/λ.
  • Median ln⁡2/λ≈0.693/λ: about 69 % of the mean. Most gaps are shorter than average, a few are very long.
  • Memoryless: having waited 5 minutes doesn’t make the next call any closer. Useful for arrivals; often wrong for things that wear out or for service times.

3. The log-normal: money and sizes

A variable is log-normal when its logarithm is normal. That happens when a quantity is the product of many independent factors (an order value = items × price × quantity discounts…), just as normals come from sums.

If ln⁡X∼N(μ,σ2):

median=eμmean=eμ+σ2/2

The bigger σ (the spread of the logs), the further the mean runs above the median. In the simulation, push the spread up and watch the share of orders above the mean fall below 30 %.

4. What to do with skewed data

  • Report the median and high percentiles (P90, P95) alongside the mean.
  • Plot on a log scale. Order values from $5 to $5,000 are unreadable on a normal axis and clear on a log axis.
  • Analyse logs when testing or modelling money: differences in log values are percentage differences, which is how managers think about money anyway (“10 % higher”).
  • Keep the mean for totals. Revenue is still mean × count, so for budgets you still need the mean, outliers and all.
Exponential, rate λ
Time between independent events. Mean 1/λ; median 0.693/λ; P(wait > t) = e^(−λt).
Log-normal
A variable whose logarithm is normal: order values, incomes, prices, file sizes.
Median e^μ, mean e^(μ + σ²/2)
For a log-normal the mean always sits above the median, more so the bigger σ.
Physical meaning
Log-normal data come from multiplying many small effects (a bit more income, a bit more spending, a bit bigger basket), while normal data come from adding them. Taking logs turns multiplying into adding and the skew into a bell.

“Remove the big values and the data will be normal.” In skewed data the big values are part of the shape, not mistakes. Use medians and percentiles, or work on the log scale, instead of cutting the tail off.

Watch it happen, one step at a time:

The model. Order values are log-normal with median $50 and log-spread σ = 0.8 (so μ = ln 50 ≈ 3.912).
Mean. 50 × e^(0.8²/2) = 50 × e^0.32 ≈ $68.86, well above the $50 median.
Share above $100. On the log scale: z = (ln 100 − 3.912) ÷ 0.8 ≈ 0.866; area above ≈ 19 %.
95th percentile. e^(3.912 + 1.645 × 0.8) ≈ $186.
Report. "Typical order $50; average $69 because one order in twenty is over $186."
Maths and principle

Principle: sums of many independent pieces become normal; products of many independent pieces become log-normal, because the log turns a product into a sum.

Equations: exponential P(T>t)=e−λt, mean 1/λ, median ln⁡2/λ. Log-normal: median eμ, mean eμ+σ2/2, P(X≤x)=Φ(ln⁡x−μσ).

In code: Excel EXPON.DIST(t, λ, TRUE), LOGNORM.DIST(x, μ, σ, TRUE); Python scipy.stats.expon(scale=1/lam), scipy.stats.lognorm(s=sigma, scale=exp(mu)).

5. Worked example

Made-up numbers. Order values are log-normal with median $50 and log-spread σ = 0.8 (so μ = ln 50 ≈ 3.912).

  • Mean: 50 × e^(0.8²/2) = 50 × e^0.32 ≈ 50 × 1.377 ≈ $68.86.
  • Share above $100: ln 100 = 4.605; z = (4.605 − 3.912) ÷ 0.8 ≈ 0.866; P(Z > 0.866) ≈ 19 %.
  • 95th percentile: e^(3.912 + 1.645 × 0.8) = e^5.228 ≈ $186.

“Average order $69” hides that the typical order is $50 and one in twenty is over $186. A free-shipping threshold at $75 would be above the median: fine if the goal is to push typical customers to add an item.

6. Practice

  1. Website visits arrive at 2 per minute. What’s the chance of a gap longer than 1 minute?
  2. Exponential waits with mean 10 minutes: what are the median and the 90th percentile?
  3. Incomes in a region: median $55k, mean $72k. Which distribution would you try first?
Answers
  1. e−2 ≈ 13.5 %.
  2. Median 10 × 0.693 ≈ 6.9 min; P90: −ln⁡(0.1)×10 ≈ 23 min.
  3. Log-normal (positive, right-skewed, mean well above median).

Try it on your own data: the probability distribution explorer draws exponential and log-normal curves and gives their percentiles; the histogram builder shows whether your data has a long right tail.

New words: exponential distribution, memoryless, log-normal distribution, log scale, right tail.

  • Exponential: waiting times between independent events; P(wait > t) = e^(−λt), mean 1/λ.
  • Log-normal: logs are normal; typical for money and sizes.
  • In right-skewed data the mean sits above the median; report median and percentiles.
  • Logs turn multiplicative, skewed data into roughly symmetric data.

Check yourself

1Calls arrive on average every 4 minutes (Poisson arrivals). What is the median wait for the next call?
2Order values are log-normal. What happens if you take the log of each value?
3A customer has been on hold 5 minutes. With exponential waits, how does the chance of waiting another minute compare with a new caller's?

Sources

Your ticks stay in this browser. Save them as a file on the course home.

Updated 2026-11-22. Numbers in the examples are made up to show the idea unless a source is given. Simulations run in your browser; nothing you type is sent anywhere.