Cross-Validation is a technique for testing how well a model will perform on new data, by repeatedly training and testing it on different splits of the available data.
What it means in plain English
Judging a model on a single train/test split can be misleading, because results might depend on which examples happened to fall in the test set. Cross-validation reduces this luck: it splits the data into several parts, trains on some and tests on the rest, then rotates so every part serves as the test set once. Averaging the results gives a more reliable estimate of true performance.
It is a standard practice for trustworthy model evaluation.
A simple example
In 5-fold cross-validation, the data is split into five parts. The model is trained five times, each time holding out a different fifth for testing, and the five scores are averaged — giving a more dependable measure than a single test would.
Why it matters
Cross-validation guards against being fooled by a lucky or unlucky data split, and helps detect overfitting. It is one of the most important techniques for honestly estimating how a model will behave on data it has never seen.
Related terms
- Overfitting — cross-validation helps detect it.
- Dataset — what cross-validation splits and rotates.
- Benchmark — a related idea of measuring performance.
Frequently asked questions
What is cross-validation for?
It is a technique for evaluating how well a model will generalise to new data, by repeatedly training and testing on different splits of the data rather than a single split.
What is k-fold cross-validation?
The data is split into k parts; the model trains on k-1 parts and tests on the remaining one, repeating so each part serves as the test set once, then averaging the results for a more reliable estimate.
