Learn AI / Compare /Accuracy vs precision vs recall

Accuracy vs precision vs recall

The vendor says 95% accurate. Is that good?

The short answer

On its own it means very little. If one case in a hundred is the one you care about, a system that finds none of them is still 99% accurate. Precision asks how many flagged cases were real; recall asks how many real cases were found. Ask for both.

Accuracy

Model Quality and Evaluation · Starter

In one sentence: The share of cases a system labels correctly out of all the cases it was tested on.
Why it matters: A single percentage can strongly influence whether leaders trust a system, buy a product, or change a workflow. If that percentage hides important mistakes, people may approve a tool that looks good on paper but fails where the consequences matter most.
The everyday version: Accuracy tells you what fraction of all answers were right.
The one thing to remember: Accuracy is the share of all tested cases labelled correctly. When one class is much more common, the score can look strong while performance on the rarer class remains poor.
The commonest misconception: High accuracy means the system works well in every important situation. Accuracy averages all cases together. If some cases are rare or more important, the average can stay high even while performance on those cases is poor.

Read the full page

Precision

Model Quality and Evaluation · Intermediate

In one sentence: The share of flagged results that are actually correct positives.
Why it matters: A system can look useful while still producing many false alarms. Precision helps you judge whether a positive result is trustworthy enough for your task, especially when acting on a wrong alert has costs.
The everyday version: When the system says ‘this one is a match’, precision tells you how often that positive signal is correct.
The one thing to remember: Precision is the share of positive flags that are correct. It does not measure how many real positives were missed, and it can change with the threshold or data being assessed.
The commonest misconception: High precision means the system finds most of the real positives. Not necessarily. A system can be very selective, flag only a few cases, and get most of those right. That gives high precision but may miss many true cases, which is low recall.

Read the full page

Recall

Model Quality and Evaluation · Intermediate

In one sentence: The share of real positive cases that a system successfully flags.
Why it matters: If a system misses too many real cases, a good-looking overall score can hide a serious failure. Recall helps people judge whether the system is safe enough for tasks where missed cases have real consequences.
The everyday version: Recall tells you how good a system is at catching the cases you do not want to miss.
The one thing to remember: Recall asks: among all the cases that were truly positive, what proportion did the system flag at the chosen threshold?
The commonest misconception: High recall means the system is generally accurate. Not necessarily. A system can catch most real positive cases and still make many false alarms. Recall looks at missed real cases, not overall correctness.

Read the full page

False positives and false negatives

Model Quality and Evaluation · Starter

In one sentence: False positives are wrong yes predictions, while false negatives are wrong no predictions.
Why it matters: A system can look good overall while still making the wrong kind of mistakes for your purpose. If you only hear that a tool is '90% accurate', you may miss that it wrongly flags many innocent cases, or misses many real ones. False positives and false negatives make those trade-offs visible. That helps non-specialists ask better questions: What harm comes from each kind of error? Who is affected? Are some groups hit more often? Is the threshold set for convenience, safety or fairness? Those questions matter in procurement, policy, programme design and day-to-day use.
The everyday version: There are four basic results when a system makes a yes-or-no call: correct yes, correct no, wrong yes and wrong no. The two wrong ones are false positives and false negatives, and they tell you what kind of mistakes the system makes.
The one thing to remember: Overall accuracy can conceal the mistake that matters. Count false positives and false negatives separately, then assess their consequences and how the decision threshold shifts the balance.
The commonest misconception: A high accuracy score means false positives and false negatives are not a problem. Not necessarily. If the positive cases are rare, a system can get high accuracy by mostly saying 'no'. It may still miss many real positives, or it may create too many false alarms. You need to inspect the four outcome counts, not just one overall score.

Read the full page