Why is it called a random forest?

Because it is a collection of trees, and randomness in row sampling and column selection is what makes the trees diverse enough for their votes to help.

When does a random forest struggle?

On very high-dimensional sparse data like raw text, on extrapolation beyond the training range, and when a fast, tiny model is required on a device.

How is it different from one decision tree?

One tree is transparent but unstable; the forest sacrifices that transparency for accuracy and stability by averaging hundreds of deliberately varied trees.

How many trees should a forest have?

Enough for the validation score to level off, typically a few hundred; past that point extra trees add compute cost without meaningful accuracy.

AI Glossary / AI and machine learning / Beginner

Random forest

Definition

A random forest is a prediction model made of many decision trees trained on different random slices of the data, whose answers are combined by voting or averaging into one stronger answer.

What it means

One decision tree is easy to read but twitchy: small changes in data can grow a very different tree, and a single tree tends to memorize noise. A random forest keeps the readability of tree logic and cancels the twitchiness with numbers.

The trick is deliberate diversity. Each tree sees a random sample of rows and, at each split, a random sample of columns, so trees make different mistakes. Errors that are random cancel when you average; patterns that are real survive.

The result is a model that works well out of the box on tables of business data, handles mixed column types, and tells you which columns it leaned on.

How it works

Draw many bootstrap samples of the training rows, each the same size as the original but with repeats allowed.

Grow a deep decision tree on each sample, letting it overfit its own slice rather than pruning it.

At every split in every tree, consider only a random subset of the columns, forcing variety between trees.

Combine the trees: majority vote for classification, average for regression.

Estimate performance on each row using only the trees that never saw it, the out-of-bag estimate, a built-in honesty check.

Three practical examples

Scoring which loan applicants are likely to default from application tables, where mixed numbers and categories suit trees.
Predicting next month's demand per product line from history, promotions and seasonality.
Ranking which machine sensors matter before investing in a deeper monitoring project, using the forest's importance scores.

Think of it this way

Think of asking a large jury where each juror studied a different random excerpt of the evidence. No juror saw everything, and each has quirks, but the vote of the whole room is steadier than any single opinion.

Worked example

the jury vote

Suppose a retailer wants to flag likely churners from fifty columns of account history. The figures are illustrative.

The forest grows five hundred trees, each on a different random sample of customers and random picks of columns at each split.

A customer flagged by four hundred and ten trees is scored high risk; one flagged by forty is not. The out-of-bag check says the forest catches eighty per cent of real churners with tolerable false alarms.

The general lesson: you trade a single readable tree for a dependable crowd, and you read the crowd through its votes and importance scores, not its branches.

Illustrative case study

A mid-sized insurer replaced a hand-tuned scoring sheet with a random forest for triaging claims. Accuracy rose modestly, but the larger win was stability: the old sheet swung whenever a few odd months landed, while the forest's vote barely moved. Auditors accepted the importance ranking as an explanation of which factors drove scores, though the team learned to say ranked influence, not cause. When a new fraud pattern appeared, retraining took an afternoon, because the method demanded little bespoke preparation of the data.

Common misconceptions and mistakes

More trees keep improving forever. Gains flatten after a few hundred trees; beyond that you are buying compute, not accuracy.
It cannot overfit. It resists overfitting but can still memorize noise, especially with very deep trees and tiny datasets.
Importance scores prove causation. They show what the model used, which may be proxies and correlations, not levers you can pull.
Forests need heavy data preparation. They cope with mixed types and missing-ish mess better than most methods, which is exactly why they are the usual first baseline.

Business use

Reach for it first on tabular business data before trying fancier methods; it is the standard baseline to beat.

Read the importance ranking with the team to sanity-check the model against domain knowledge.

Track out-of-bag or validation scores over time and retrain when data drifts, not on a calendar alone.

Do not promise explanations of individual predictions without extra tooling; forests vote, they do not narrate.

Keep one trained as a permanent baseline; any fancier replacement must beat it on held-out data before earning the swap.

Check your understanding

What is a random forest, and why does randomness make it better rather than worse?

Answer: A random forest is many decision trees, each trained on a random sample of rows and random subsets of columns, combined by vote or average. The randomness makes the trees err differently, and errors that differ cancel out when combined, so the crowd is more accurate and far more stable than any single tree grown on all the data.

Related terms

Learn next

AI for Business Owners helps you decide where dependable baseline models like this fit in your operation: https://aiforbusinessandlife.com/academy/ai-for-business-owners [C2]

Sources

[S2] Google, Machine Learning Glossary; [S24] scikit-learn documentation, Glossary of Common Terms and API Elements. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Random feature Next: Random-fuzzy variable