How is a decision forest different from a random forest?

Random forest is the best-known kind of decision forest; the terms are often used loosely, with random forest naming the version that also randomizes the fields considered at each split.

How many trees should a forest have?

Commonly a few hundred. Add trees until held-out accuracy stops improving, then stop, because more trees only add computation.

Can a forest predict numbers as well as categories?

Yes. The trees' numeric predictions are averaged instead of voted.

AI Glossary / AI and machine learning / Beginner

Decision forest

Definition

A decision forest is a model made of many decision trees whose answers are combined, usually by voting or averaging, so the group predicts better than any single tree.

What it means

One decision tree is easy to read but fragile: small changes in the training data can grow a quite different tree, and its answers swing with it.

A forest grows many trees, each on a different random slice of the data and with a random choice of which input fields each split may consider. The trees end up making different mistakes.

Combining their answers cancels most of those mistakes out. Where one tree overreacts to a quirk in its slice, the majority usually does not, so the forest's answer is steadier and more accurate.

The price is readability. A single tree can be drawn on a page and explained to anyone; a forest of a hundred trees cannot. You trade a simple story for better numbers.

For a business, forests are often the first model worth trying on spreadsheet-style data: they handle messy columns, need little tuning, and perform well enough to be the baseline other models must beat.

How it works

Take the training data and draw many random samples from it, usually with replacement, so each sample overlaps the others but is not identical.

Grow a decision tree on each sample, but at each split only let the tree consider a random subset of the input fields, which forces variety between trees.

Do not prune these trees heavily; in a forest the individual trees are allowed to be deep and somewhat overfit, because the averaging step is what controls the overfitting.

To predict, pass the new case through every tree and combine: majority vote for categories, average for numbers.

Measure the forest's accuracy on data it never trained on, and tune the number of trees and the size of the random field subsets if more accuracy is needed.

Three practical examples

Credit screening: hundreds of small trees each vote on whether an applicant is likely to repay, and the majority vote sets the risk grade.
Retail demand forecasting: a forest averages its trees' predictions of next month's unit sales per store.
Equipment maintenance: trees trained on sensor history vote on whether a pump is heading towards failure this quarter.

Think of it this way

Think of asking a panel of doctors for a second, third and hundredth opinion instead of trusting one. Each doctor has seen a slightly different set of patients, so each has a blind spot, but the consensus of the panel is rarely fooled by any single blind spot.

The limit: the panel's answer is better, but no one can tell you a simple story about why. A forest inherits that opacity.

Worked example

scoring loan applicants with a forest

Suppose a lender trains one decision tree on past loan outcomes and finds it performs well on old data but erratically on new applicants. The example figures are illustrative.

The team grows 200 trees instead, each on a random 70 per cent of the loan book, with each split choosing from a random half of the fields.

On held-out applicants the single tree gets 81 per cent of outcomes right; the forest gets 87 per cent, and its accuracy varies far less from month to month.

The lender keeps the forest for scoring but keeps a single shallow tree as a display aid, so staff can still show customers roughly what drives decisions.

Illustrative case study

A fictional online retailer relied on one decision tree to flag orders likely to be returned. After a website redesign changed customer behavior, the tree's accuracy collapsed without warning.

Switching to a forest did not just raise accuracy; it made accuracy stable, because no single tree's habits could dominate the answer.

The invented lesson: if a single simple model worries you because it might be lucky, a forest is the cheap way to buy consistency.

Common misconceptions and mistakes

A forest is just a complicated tree. It is many independent trees combined; no single tree inside it resembles the forest's overall rule.
More trees always help. Accuracy rises quickly then flattens; past that point extra trees add cost, not quality.
Forests remove the need to check data quality. They tolerate messy data better than most models, but systematic errors in the data still poison every tree.

Business use

Reach for a forest first on table-shaped business data: it is a strong baseline that is hard to beat without more complex machinery.

Pair it with a variable-importance report so stakeholders can see which fields drive decisions, even though individual trees are not the story.

When you must explain individual decisions, as in regulated lending, decide whether the accuracy gain justifies the loss of a simple explanation, and document that choice.

Check your understanding

Why does a decision forest usually beat a single decision tree on new data?

Answer: Each tree is trained on a different random view of the data, so their errors differ, and combining their answers cancels much of that individual error out.

Related terms

Learn next

AI for Business Owners shows how to choose a sensible first model and what to demand from its accuracy reports: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Sources

[S2] Google, Machine Learning Glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Decision boundary Next: Decision lists