011 min
Why precision matters
A precise instrument gives you a number you can repeat. That sounds like a guarantee of quality, and it is not. A team that reads a tight cluster of results often decides the result is right. The cluster only shows that the measurement process gives the same answer each time, including when the answer is wrong.
The decision this changes is whether to trust a number or to check it against something else. A performance engineer who sees five benchmark runs within 2% of each other can stop worrying about noise. The same engineer still has to ask whether the benchmark measures the thing users feel. A researcher who sees a survey score move by 0.1 points needs to know how much that score moves when nothing changes. Without that, a real change and a random change look the same.
The cost of confusing the two is a confident wrong decision. A very precise instrument with a fixed offset gives the same wrong answer every time, so nothing in the data warns you. Accuracy, the closeness of a result to the true value, needs a separate check against a reference.
The word also has a second job. In machine learning and search, "precision" names the share of flagged items that really belong. This article covers both meanings and says which one it uses in each section.
022 min
What precision actually measures
In measurement, precision is the spread of results when you measure the same thing again under stated conditions. The spread is a number, not a feeling. The usual choices are the standard deviation (how far results typically sit from their average), the variance (the standard deviation squared), and the coefficient of variation (the standard deviation divided by the mean, shown as a percentage so that different units can be compared).
The phrase "under stated conditions" matters. The international vocabulary of metrology defines three sets of conditions, and a precision figure means little until you know which one applies:
- Repeatability conditions: the same procedure, operator, instrument, operating conditions and location, over a short period of time.
- Reproducibility conditions: different locations, operators and measuring systems.
- Intermediate conditions: some of these change and some stay fixed, for example the same lab on different days.
Results spread more as you change more of the conditions. A scale that agrees with itself within a minute can disagree with the same model of scale in another building. Reporting "precision" without naming the conditions hides which of these cases was tested.
Precision is a property of the measuring process, not of one result. A single reading has no precision. You need several readings of the same thing to estimate it. This is the reason the margin of error on a survey depends on how many people answered: more answers give a better estimate of the spread.
Precision also has nothing to say about the true value. Take a thermometer that always reads 2 degrees too high. It can be extremely precise, because it repeats the same reading, and still be wrong by 2 degrees every time. That fixed offset is called systematic error. Accuracy needs both a small systematic error and a small spread.
031 min
Where the two meanings of precision come from
The measurement meaning comes from metrology, the science of measurement. The international vocabulary of metrology (JCGM 200:2012, published by the Joint Committee for Guides in Metrology) defines measurement precision as the closeness of agreement between measured values obtained by replicate measurements on the same or similar objects under specified conditions. It is usually expressed with standard deviation, variance or coefficient of variation.
The vocabulary adds a warning. It says people sometimes use measurement precision by mistake to mean measurement accuracy. The two are related, but they are different quantities.
The standard that sets out the full scheme is ISO 5725. Under ISO 5725-1, accuracy has two parts: trueness, the closeness of the average of many results to the true or accepted value, and precision, the repeatability or reproducibility of the results.
The classification meaning is newer and comes from information retrieval, the study of search systems. A standard textbook by Manning, Raghavan and Schütze names precision and recall as the two most frequent and basic measures of retrieval effectiveness. Machine learning libraries later adopted the same formula for any classifier that sorts items into positive and negative.
The two meanings share a word and a loose idea, "how much can I rely on this", but they are calculated differently. One divides spread by size. The other divides correct positives by flagged positives.
042 min
A worked example: two benchmark setups
This example is illustrative. A team wants to know how long a page takes to load. It tests two benchmark setups on the same page, five runs each, in milliseconds.
| Run 1 | Run 2 | Run 3 | Run 4 | Run 5 | Mean | Std. deviation | Coefficient of variation | |
|---|---|---|---|---|---|---|---|---|
| Setup A | 212 | 208 | 215 | 210 | 205 | 210 | 3.8 | 1.8% |
| Setup B | 180 | 240 | 205 | 260 | 165 | 210 | 39.8 | 19.0% |
Both setups report a mean of 210 ms. Setup A has a standard deviation of about 3.8 ms, which is 1.8% of the mean. Setup B has a standard deviation of about 39.8 ms, which is 19.0% of the mean. Setup A is about ten times more precise.
For setup A, the deviations from the mean are 2, −2, 5, 0 and −5. Their squares add up to 58. Divide by 4 (the number of runs minus one) to get a variance of 14.5, and take the square root to get 3.8.
Now change one thing. Suppose the team keeps setup B and takes 20 runs instead of 5, and the spread stays about 39.8 ms. The standard error of the mean, which is the standard deviation divided by the square root of the number of runs, falls from about 17.8 ms to about 8.9 ms. The instrument is no more precise. The average of its results is.
That is the first lesson: more runs narrow the uncertainty of the average, but they do not make one run less noisy. Only a better setup does that.
The second lesson is what the table cannot say. If the true load time for real users is 150 ms, both setups are wrong by the same 60 ms on average. Precision told the team which setup repeats. It did not tell them which setup is right.
051 min
A documented case: the Hubble Space Telescope mirror
The Hubble Space Telescope launched in 1990 with a primary mirror that had been made with great care and was the wrong shape.
Analysis of the flawed images showed that the primary mirror had been polished to the wrong shape. Although it was believed to be one of the most precisely figured optical mirrors ever made, smooth to about 10 nanometers, the outer perimeter was too flat by about 2200 nanometers.
So the mirror's surface was smooth to a tiny tolerance, and its overall curve was off by about 220 times that amount. The polishing process repeated well. It was just aimed at the wrong target.
The cause was in the test equipment. The Allen Commission, led by Lew Allen of the Jet Propulsion Laboratory, found that a reflective null corrector, a device used to test whether a non-spherical mirror has the right shape, had been assembled incorrectly. One lens was out of position by 1.3 mm. Every measurement made with that device carried the same offset, so the results agreed with each other and with the design, and all of them were wrong.
This is the practical meaning of the difference. An instrument can be precise and wrong, and the only way to find out is to compare it with something independent of it.

062 min
Precision in classification
In machine learning and search, precision answers a different question: of the items the system flagged as positive, how many really are? Google's Machine Learning Crash Course puts it as the proportion of all the model's positive classifications that are actually positive. The formula is true positives divided by true positives plus false positives.
A true positive is a flagged item that really is positive. A false positive is a flagged item that is actually negative, a false alarm. The counts come from a confusion matrix, the table of a classifier's four kinds of outcome.
The scikit-learn library defines precision the same way. Its best value is 1 and its worst is 0. When the model flags nothing as positive, the formula divides zero by zero, and scikit-learn returns 0 and raises a warning unless you tell it to do otherwise.
Take a spam filter, with illustrative numbers. Out of 1,000 emails, 120 are really spam.
| Threshold | Flagged as spam | Really spam among flagged | Precision | Recall |
|---|---|---|---|---|
| Lenient | 100 | 80 | 80 / 100 = 0.80 | 80 / 120 = 0.67 |
| Strict | 50 | 45 | 45 / 50 = 0.90 | 45 / 120 = 0.38 |
Making the filter stricter raised precision from 0.80 to 0.90. It also lowered recall, the share of real spam caught, from 0.67 to 0.38. The strict filter lets 75 spam emails through. Precision alone cannot show that cost, and this is why it is almost always reported next to recall, or combined with it in the F1 score.
Choose to favor precision when a false alarm is expensive. A fraud system that freezes real customers' cards, or a content filter that removes legitimate posts, needs most of its flags to be right. Choose to favor recall when a miss is expensive, as in screening for a serious illness.
Notice that classification precision is not about repeated measurement at all. A classifier can have a precision of 1.0 on one test set and a very different value on another, because it depends on how common the positive cases are in the data you test on.
071 min
False precision: more digits than the measurement supports
There is a third use of the word that shows up in dashboards and slide decks. A number written with many digits looks precise even when the measurement behind it is not.
False precision occurs when numerical data are presented in a manner that implies better precision than is justified. The Wikipedia article gives a plain example. A speedometer reading of 60 mph converts to 96.56064 km/h, but if the speedometer is accurate to 1 mph, the honest figure is 97 km/h.
In product work, the same thing happens with a conversion rate reported as 3.2471% from 400 visitors. At that sample size the rate could plausibly move by a full percentage point or more if the same test were run again, so the digits after the first one are noise. The reader sees the digits and assumes the result is exact.
A simple rule is to report no more digits than the spread of your results supports. If repeated runs vary by about 4 ms, write 210 ms, not 210.37 ms. If a survey has a margin of error of 3 points, write 52%, not 52.4%.
The same logic applies to estimates. A roadmap that says a feature will take 37.5 days implies a precision that nobody who plans software has. "About six weeks" is more honest and more useful.
081 min
How precision shows up in product, design, and engineering work
Benchmarks and performance tests. Run the same test several times and report the spread, not only the mean. If the spread is larger than the change you are looking for, you cannot see the change. The remedy is more runs, a quieter machine, or a better test.
[A/B testing](/concept/ab-testing). A test result is a measurement with a spread. A difference between two variants that is smaller than the run-to-run spread of the metric cannot be told apart from noise. This is also why teams check how much a metric moves in an A/A test, where both groups see the same version.
User research scores. A usability score that moves from 72 to 74 across two small rounds of testing usually sits inside the spread you would see by repeating the same round. A PM who reads that as improvement is reading noise.
Classifiers and search. For a spam filter, a recommender or a search page, precision tells you how often a flagged or shown item is right. A support team reviewing flagged tickets by hand cares because every false positive costs a person's time.
Estimation. Planning numbers carry a spread too. A team that says "three weeks" and finishes in two to five is not imprecise by accident. The honest report is the range.
091 min
When to report precision, and when to report something else
Report precision when the question is "can I trust this number to repeat?" It is the right check when you compare two versions of something, when you need to know whether a change is bigger than the noise, or when you choose between instruments.
Before you report it, the condition must hold that you have several independent measurements of the same thing. If you measured once, you have no precision to report. If the measurements are not independent, for example the same cached result returned five times, the spread will look too small.
Report accuracy instead, or as well, when the question is "is this number right?" That needs a reference value that does not come from the same instrument.
Report classification precision when false alarms are the main cost. Report recall when misses are. If both matter, report both, or report the F1 score and say which error you care about more.
If the honest answer is "it depends", it depends on three things: which conditions were held fixed, how many measurements you have, and what the cost of each kind of error is.
101 min
Where precision goes wrong
Treating a tight result as a correct one. The Hubble case above is the full version. A fixed offset shows up as perfect agreement.
Not naming the conditions. "Precise to 1%" can mean the same operator on the same machine within one minute, or a different lab on a different continent. These are very different claims.
Too few measurements. A spread estimated from three runs is itself very uncertain. Two close runs in a row can happen by luck.
Averaging away the problem. Averaging many readings reduces random spread, and it does nothing about a systematic offset. Hubble's mirror could have been measured a thousand times with the same faulty device and the answer would have stayed wrong.
False precision. Reporting digits that the spread does not support.
Reading classification precision on its own. A classifier that flags a single very obvious case, and nothing else, has a precision of 1.0 and is nearly useless. Precision without recall rewards a system for flagging less.
Mixing up the two meanings. A measurement team and a machine learning team can both say "we need higher precision" and mean different things. Ask which quantity before agreeing.
111 min
Precision vs. accuracy, recall and validity
The nearest neighbor is accuracy. The deciding fact is that precision compares results with each other, while accuracy compares results with the true value. You can check precision with no reference at all. You cannot check accuracy without one.
Precision is also not validity. A measure is valid when it captures the thing you care about. A very precise count of page views is not a valid measure of whether users understood the page.
In classification, the pair that gets confused is precision and recall. Precision is the share of flagged items that are right. Recall is the share of right items that were flagged. They share the true positive count in the top and differ in the bottom.
| Term | Compares | Needs a reference value? |
|---|---|---|
| Not in the library yetPrecision (measurement) | ||
| ResearchAccuracy | ||
| Not in the library yetPrecision (classification) | ||
| BusinessRecall |
121 min
Where the definitions disagree
Not everyone uses "accuracy" and "precision" the way ISO does, and that causes real confusion.
In the older and still common usage, accuracy means only closeness to the true value, which ISO 5725 calls trueness. Under this usage, a result can be accurate, precise, both or neither, and the two properties are independent.
In the ISO 5725 usage, accuracy needs both a small systematic error and a small spread. So high accuracy requires high precision. Under this usage a measurement process cannot be accurate and imprecise at the same time.
Both usages appear in textbooks and in software documentation.
For a practitioner, the safe move is to avoid the dispute. Say "spread" or "repeatability" when you mean the first, and "systematic error" or "offset from the reference" when you mean the second. Then no reader has to guess which definition you hold.
132 min
Precision FAQ
What is precision?
Precision is how closely repeated measurements of the same thing agree with each other. A set of readings that are all close together is precise, whether or not they are near the true value. In classification, precision is the share of positive predictions that are correct.
What is the difference between precision and accuracy?
Precision compares results with each other, and accuracy compares results with the true value. A set of readings can be precise and consistently wrong. Under the ISO 5725 definition, high accuracy requires high precision, and it also requires a small systematic error.
How do you measure precision?
Measure the same thing several times under stated conditions and calculate the standard deviation, variance or coefficient of variation of the results. A smaller value means higher precision. The condition set must be stated, such as the same operator and instrument or different labs.
What is the formula for precision in machine learning?
Precision equals true positives divided by true positives plus false positives. For 100 flagged emails where 80 are really spam, precision is 80 divided by 100, which is 0.80. The value runs from 0 to 1, and higher is better.
What is the difference between precision and recall?
Precision is the share of flagged items that are correct, and recall is the share of correct items that were flagged. Precision punishes false alarms. Recall punishes misses. Raising a decision threshold usually raises precision and lowers recall.
Can a result be precise but not accurate?
Yes. A scale with a fixed offset reads the same wrong value every time. The Hubble Space Telescope mirror was smooth to about 10 nanometers and still had the wrong overall shape, because the test device used to shape it was assembled with one lens 1.3 mm out of position.
What is false precision?
False precision is reporting a number with more digits than the measurement supports. A conversion rate of 3.2471% from 400 visitors implies a certainty the data does not have. A good rule is to report no more digits than the spread of your results justifies.
Does taking more measurements improve precision?
It improves the precision of the average, not of a single measurement. The standard error of the mean shrinks with the square root of the number of runs. A single run stays as noisy as before, and more runs do nothing about a systematic offset.
?8 questions
Questions people ask
What is precision?
What is the difference between precision and accuracy?
How do you measure precision?
What is the formula for precision in machine learning?
What is the difference between precision and recall?
Can a result be precise but not accurate?
What is false precision?
Does taking more measurements improve precision?
§7 sources
Sources
Show all 7 sourcesHide sources
Joint Committee for Guides in Metrology (JCGM), "International vocabulary of metrology: basic and general concepts and associated terms (VIM)", JCGM 200:2012, entries 2.15 (measurement precision), 2.20 and 2.24 (repeatability and reproducibility conditions).
Google for Developers, "Classification: Accuracy, recall, precision, and related metrics", Machine Learning Crash Course.
scikit-learn documentation, "precision_score".
Manning, Raghavan and Schütze, "Evaluation of unranked retrieval sets", in Introduction to Information Retrieval, Cambridge University Press, 2008, Stanford NLP Group web edition.
Wikipedia, "Hubble Space Telescope", section on the primary mirror error and the Allen Commission.
Wikipedia, "Accuracy and precision", section on the ISO 5725 definition.
Wikipedia, "False precision".





