Generate custom courses on any topic — with hands-on practice, AI guidance, and visuals built in.
Already have an account?
A product team runs an A B test on a new checkout flow. Each visitor either converts or does not. After 20 visitors, the conversion rate might be 35%. After the next 20, it might drop to 25%. Nothing about the product changed in those minutes. The process was the same. What changed was the sample.
That tension shows up in ops metrics too. A warehouse might measure the on time shipping rate each day. A support team might track the fraction of tickets resolved within 24 hours. These are all built from the same kind of repeated random experiment.
One repeat looks like this. Take one order and check whether it shipped on time. Record yes or no. Then do the same for the next order, and the next. What is repeated is the same rule for classifying each trial. What is counted is the number of yes outcomes. The metric you report is the proportion yes out of total trials. The reason it matters is that decisions need a stable signal, and small samples produce a lot of swing even when the underlying process has not moved.
Flip a fair coin. Each flip has two possible outcomes, heads or tails. After flips, suppose you have seen heads. The empirical proportion of heads is . Early on, that proportion can be extreme. After 1 flip it is either or . After 2 flips it can be , , or .
As grows, two things tend to happen in practice. The proportion wanders less, and it spends more time near . You still get streaks, but each new flip is a smaller fraction of the total, so it has less ability to drag the overall proportion far away.
A common wrong intuition is that a fair coin should alternate heads and tails to keep the proportion near . Alternation is not required. You can get 8 heads in a row and the coin is still fair. Long run frequency is not a rule about short run patterns. It is about what the running proportion does when you repeat the same experiment many times.
Try running the experiment and watching the running proportion, then compare what you see when you group flips into smaller versus larger windows.
When the curve looks jumpy, it usually means the denominator is still small. When it looks calm, it usually means the denominator is large.
Golden rule
Treat probability as the long run fraction of times an event happens when you repeat the same random experiment under the same conditions.
That idea is a preview of a deeper result you will use later, the law of large numbers, which formalizes the pattern that long run averages stabilize.
In business, you often start with a yes no question, then ask how often yes happens.
An outcome is what happens on one trial. For a coin flip it is heads or tails. For an order it might be on time or late. An event is a set of outcomes you care about, like the event order is on time. Once you can repeat the trial, you can talk about a rate. The observed on time rate over a week is the number of on time orders divided by total orders that week.
The phrase base rate means the background probability before you condition on extra information. If last quarter the churn rate was 4% per month, that is a base rate you might start from. Uncertainty enters because you never observe the true long run probability directly. You observe samples, and samples vary.
Match the probability language to a few familiar business metrics so the experiment and the event are both unambiguous.