Probability turns uncertainty into a number
Probability measures how likely an event is on a scale from 0 to 1, or equivalently from 0% to 100%. By the end, you will be able to calculate probability from real-world data, interpret claims such as “likely,” and spot estimates that promise more certainty than the evidence can support.
A probability of 0 means an event is impossible within the model being used. A probability of 1 means it is certain within that model. Values between them express degrees of uncertainty. A 0.7 probability and a 70% probability say the same thing because multiplying a decimal probability by 100 converts it to a percentage.
Probability describes a defined event. “Rain is likely” is incomplete until you know the place, time period, and meaning of rain. “At least 0.2 millimetres of rain at the airport between noon and 6 p.m.” is an event that can be checked.
The number is only as clear as the event. Suppose a delivery service says a parcel has a 90% chance of arriving “on time.” Does that mean before 5 p.m., before the end of the day, or within the promised three-day window? Changing the definition changes which past deliveries count as successes, so it can change the probability.
Probabilities also depend on information. Before a card is drawn from a well-shuffled standard deck, the probability of a heart is 13 out of 52, which simplifies to 1 out of 4. After someone tells you that the card is red, the relevant possibilities are the 26 red cards. Now 13 of those 26 are hearts, so the probability is 1 out of 2. The card did not change. Your information did.
For a fair six-sided die, three results are even, so .
This counting rule works only when the possible outcomes are equally likely. It fits a fair die or a thoroughly shuffled deck. It does not automatically fit train delays, loan defaults, medical outcomes, or football results. For those, observed data or a model must estimate how often the event occurs.
How do you calculate probability from real-world data?
To estimate probability from data, divide the number of observations matching the event by the number of relevant observations. Then report the sample, the event definition, and the time period, because the fraction has meaning only in relation to those choices.
A café records 240 weekday lunch orders during one month. Sixty include soup. If next month is expected to resemble the recorded month, the observed proportion gives an estimated 25% probability that a randomly selected weekday lunch order includes soup.
The arithmetic is visible:
For the café, .
The hat over signals an estimate. The calculation describes the sample exactly: 25% of the recorded orders included soup. Using that figure to predict a future order adds an assumption that conditions remain similar. A new menu, colder weather, a weekend crowd, or a price change could make the old sample a poor guide. The connection between price and purchasing is developed further in Price Elasticity of Demand and Supply.
Before dividing, check the denominator. If the question is about weekday lunch orders, breakfast orders and weekend orders do not belong in the total. If some orders have missing item records, hiding them in the denominator can distort the estimate. A precise division cannot repair a badly chosen dataset.
How likely is “likely”?
Words such as “likely,” “possible,” and “unlikely” do not have universal numerical meanings in ordinary speech. A careful source should attach a number, a range, or a published vocabulary scale, because different readers can map the same word to different probabilities.
Imagine that one analyst uses “likely” for any probability above 60%, while another reserves it for probabilities above 80%. They could agree on a 70% estimate and still choose different words. The disagreement is linguistic, not mathematical. Reading the number prevents the label from doing more work than it can support.
“A delay is likely tomorrow.” The event, evidence, location, time window, and numerical threshold are missing.
“Based on 50 comparable departures, 35 left more than 15 minutes late, an observed rate of 70%.” The count and definition can be inspected.
A percentage still needs interpretation. If 35 of 50 comparable departures were late, the observed rate is 70%. It does not follow that exactly seven of the next ten departures must be late. Probability describes a pattern across uncertain cases, not a fixed quota that reality must fill.
The two bars summarize the same dataset and must add to 100% because “more than 15 minutes late” and “not more than 15 minutes late” are complementary events. Real reports may include cancelled trips or missing records. Those categories need explicit treatment rather than silent removal.
A forecast is not a promise
A 70% forecast means that the forecasting method assigns probability 0.7 to a specified event, given its current information. It does not promise that the event will happen, and one occurrence cannot prove that the probability was right or wrong.
If rain follows a 70% forecast, that outcome was compatible with the forecast. If no rain follows, that outcome was also compatible, since the forecast assigned a 30% chance to no rain. To judge the forecasting method, collect many forecasts made under similar conditions and compare predicted probabilities with observed frequencies.
A method is well calibrated if events given about a 70% probability occur about 70% of the time over a large set of comparable forecasts. Calibration does not require every individual forecast to come true. It tests whether the stated confidence matches the long-run record.
Forecast quality has another dimension: discrimination. A useful weather model should assign higher rain probabilities on days that become rainy than on days that remain dry. A model that predicts the same base rate every day might be calibrated overall but offer little help about which particular day needs an umbrella.
This distinction matters in public decisions. Economic forecasts inform tax and spending choices, but they remain conditional estimates. The mechanisms behind those choices are covered in Fiscal Policy. A forecast should inform a decision without being mistaken for a guarantee.
What makes a dataset a fair basis for probability?
A dataset is a fair basis for estimating probability when its observations represent the cases named in the question, use consistent measurements, and avoid systematic selection. More rows reduce random noise, but they do not cure bias in who or what was recorded.
Suppose a school wants to estimate the probability that a student uses the bus. Asking only students waiting at the bus stop creates selection bias. A thousand answers gathered there can be less informative than a smaller sample drawn across the whole school. Sample size and sample quality solve different problems.
State whose probability you want, such as all enrolled students rather than students visible in one place.
Decide what counts, such as using the bus for the trip to school on the survey day.
Give members of the target group a known or defensible chance of selection.
Ask if nonresponses or failed measurements are concentrated among a particular kind of case.
Measurement can also create bias. A fitness app estimates running distance from sensor readings and algorithms. If its method consistently undercounts routes among tall buildings, the resulting pace data can be precise to two decimal places and still be systematically wrong. Precision in display is not evidence of accuracy.
Historical data can reproduce historical patterns. A hiring model trained on earlier decisions may learn who was selected rather than who was capable. Building and testing such systems requires probability, data checks, and an account of error; those ideas connect directly to Foundations for AI Developers.
That sentence is not an argument for tiny samples. With a sound sampling method, more independent observations usually make an estimated proportion more stable. It is a warning to examine the route by which data entered the table before admiring the number of rows.
Conditional probability changes the denominator
Conditional probability is the probability of an event after restricting attention to cases where another condition holds. It answers a narrower question by changing the relevant group, which often changes the denominator and sometimes changes the answer sharply.
Consider 200 parcel deliveries. Of these, 50 travel during severe weather, and 20 of those are late. Across all 200 deliveries, suppose 30 are late. The overall late probability in the sample is 30 out of 200, or 15%. Among severe-weather deliveries, it is 20 out of 50, or 40%.
Using counts, .
The vertical bar means “given.” The calculation asks about late delivery given severe weather, so only severe-weather cases belong in the denominator. Using all 200 deliveries would answer a different question.
| Observed deliveries | Late | Not late | Total |
|---|---|---|---|
| Severe weather | 20 | 30 | 50 |
| Other conditions | 10 | 140 | 150 |
| Total | 30 | 170 | 200 |
Conditional probability is easy to reverse by mistake. In this table, 20 of the 50 severe-weather deliveries are late, so . But 20 of the 30 late deliveries faced severe weather, so , about 66.7%. The two questions use different denominators.
This reversal error appears in medical testing, fraud alerts, and court reporting. “The probability of a positive test given illness” is sensitivity. “The probability of illness given a positive test” also depends on how common the illness is in the tested group. The order of the condition and event cannot be swapped.
How much uncertainty should an estimate show?
An estimate from a sample should usually show that another reasonable sample could produce a different result. Sample size, variability, measurement quality, and sampling method determine how much uncertainty belongs around the reported proportion.
Take two fictional but fully specified samples. In the first, 7 of 10 customers choose option A. In the second, 700 of 1,000 do. Both observed proportions are 70%, yet the second is more stable under repeated random sampling. One changed response moves the first estimate by 10 percentage points and the second by only 0.1 percentage points.
The estimate is 70%. One observation changing category produces 60% or 80%, so individual cases have large influence.
The estimate is also 70%. One observation changing category produces 69.9% or 70.1%, so individual cases have much less influence.
A confidence interval is one standard way to express sampling uncertainty. Its exact calculation depends on the method and assumptions. It does not create certainty, and it does not include every possible source of error. A narrow interval around a biased estimate is still misleading.
Decimal places can counterfeit confidence. A model output of 63.847% is not automatically more trustworthy than “about 64%.” Report only the precision supported by the data, measurements, and model.
Dependence also matters. One thousand sensor readings taken one second apart may not supply one thousand independent pieces of evidence if the same local condition affects them all. Repeated measurements can make a file large without adding the amount of new information that its row count suggests.
The mathematical tools for samples, distributions, and uncertainty sit within the wider study of mathematics. The useful habit is simple: pair every probability estimate with a question about where its uncertainty came from and what the calculation leaves out.
Good probability claims make better decisions
A good probability claim names the event, population, time period, evidence, and uncertainty. A good decision then combines that probability with the consequences of acting, waiting, or being wrong. Likelihood alone does not determine the best choice.
Suppose an outdoor event has a 30% chance of rain. Cancelling may waste a clear day, while continuing without cover may damage equipment. The decision depends on costs, alternatives, and risk tolerance. A low-probability event can deserve action when its consequences are severe. A high-probability event can require little action when its consequences are minor.
A shop can pay £80 for temporary flood barriers. A forecast assigns a 10% flood probability, and an unprotected flood would cause £2,000 of damage. In a simplified calculation, the expected flood loss is . Since £80 is less than £200, buying the barriers has the lower expected cost, if those numbers and assumptions are sound.
Expected value multiplies each possible outcome by its probability and adds the results. It is useful for repeated or comparable decisions, but it does not erase practical constraints. A person may be unable to absorb a rare large loss even when the long-run average looks acceptable. Insurance exists partly because the timing and size of losses matter, not only their average.
The takeaway: Translate words such as “likely” into numbers, inspect the event and denominator, ask how the data were selected, separate a forecast from a promise, and weigh probability against consequences before acting.
Probability does not remove uncertainty. It disciplines how uncertainty is described. Once the event, evidence, denominator, and assumptions are visible, disagreement becomes easier to locate. People can argue about the data or the decision without hiding behind a vague word.