Accuracy
Last updated June 14, 2026
What is Accuracy in simple terms?
In simple terms, accuracy is a model's report card score: out of every prediction it made, how many did it get right? Simple and intuitive — but it can flatter a lazy model when one answer is rare.
What is Accuracy?
Accuracy is the most basic measure of a classification model's performance: the share of all its predictions that were correct — though on imbalanced data, where one outcome is rare, a high accuracy can be deeply misleading.
Accuracy is the first number most people reach for to judge a model that sorts things into categories: out of every prediction it made, what fraction did it get right? Score 92 correct out of 100 and the accuracy is 92%. It's intuitive, easy to explain, and a perfectly good headline when the categories are reasonably balanced — when there are roughly as many of each outcome to predict. For a model deciding whether photos show a cat or a dog, with plenty of both, accuracy tells you something real and useful at a glance.
The problem — and it's the single most important thing to understand about accuracy — is that it becomes dangerously misleading when one outcome is rare. Suppose you're building a model to catch a fraud that occurs in just 1 in 100 transactions. A model that does nothing at all, simply labeling *every* transaction "not fraud," will be right 99 times out of 100 and post a glittering 99% accuracy — while catching exactly zero cases of the fraud it was built to find. The accuracy looks excellent and the model is useless. This is called the accuracy paradox, and it's why a high accuracy figure should always prompt the question: how common is the thing being predicted? On imbalanced data, accuracy rewards a model for ignoring the rare case, which is usually the case that matters most.
That's not a reason to throw accuracy out — it's a reason to know when to trust it. For balanced problems it's a clear, sensible summary. For imbalanced ones you reach instead for metrics that don't let a lazy model hide: precision and recall, which separate false alarms from misses, and the F1 score, which combines them. A good habit is to never read accuracy alone; pair it with the confusion matrix, which shows the actual breakdown of right and wrong answers and instantly reveals whether a high accuracy is genuine skill or just the model coasting on the common answer.
Real-world example of Accuracy
Picture a weather app in a desert town where it rains only about 5 days a year. You build a model to predict "rain tomorrow?" and proudly report 95% accuracy. Impressive — until you notice the model has learned the laziest possible rule: always predict "no rain." In a place that's dry roughly 360 days out of 365, "no rain, every day" is correct about 99% of the time on the days it doesn't rain, and the model sails to a high accuracy while being completely worthless for the only prediction anyone cares about — the rare rainy day, which it never once calls. A resident relying on it would be caught in every single downpour. The accuracy number looked great precisely because the event it should predict almost never happens. That gap is the trap accuracy hides and other metrics expose.
Related terms
Frequently asked questions about Accuracy
What is the difference between accuracy and precision?
Accuracy measures *all* predictions: the share the model got right across every category. Precision measures only the predictions it flagged as positive: of those, how many were actually correct. Accuracy answers "how often is it right overall?"; precision answers "when it says yes, can I trust it?" They can diverge sharply — a model can have high accuracy yet poor precision (or vice versa), especially when one category is rare. Accuracy is the broad summary; precision zooms in on the quality of the positive flags specifically. **2. Mechanism — How does accuracy work?**
How does accuracy work?
You run the model on a set of examples whose true categories you already know, then count how many predictions matched the truth and divide by the total number of predictions. The result is a fraction from 0 to 1, usually shown as a percentage. Every prediction counts equally, whether it's the common outcome or the rare one — which is exactly the source of its weakness on imbalanced data, where the common outcome dominates the count and the rare, important one barely moves the number. **3. Application — What is accuracy used for?**
What is accuracy used for?
It's the standard quick summary of how a classification model is doing, and a fine default whenever the categories are roughly balanced — image labeling with plenty of each class, sentiment as positive or negative in even proportions, and the like. It's also the easiest metric to communicate to non-specialists. The key discipline is knowing its limit: for imbalanced problems like fraud, disease screening, or rare-defect detection, accuracy alone hides failure, and you should report precision, recall, and the F1 score alongside it.