AI Glossary / AI and machine learning / Advanced
Monte Carlo integration
Definition
Monte Carlo integration estimates an integral or average by random sampling: draw points at random, evaluate the quantity at each, and let the sample average stand in for the exact answer.
What it means
Some averages are impossible to compute exactly: too many cases, too tangled a formula. Monte Carlo sidesteps the algebra by simulating it, trading exactness for an estimate that improves with more samples.
The method's guarantee is statistical: the estimate is unbiased, and its error shrinks with the square root of the sample count. Quadrupling the samples halves the error, which sets the price of precision.
That square-root law is the method's personality. A rough answer is nearly free; a precise one is expensive; and each extra decimal place costs a hundred times the samples.
Monte Carlo appears wherever uncertainty must be averaged over: risk simulations, option pricing, Bayesian computation, and the confidence intervals on a model's own evaluation scores.
For a business, the practical skill is reading a Monte Carlo answer correctly: it comes with an error bar, and decisions should use the bar, not just the point in the middle.
The name comes from the casino: chance is not the enemy of the calculation but the engine of it, and the mathematics of randomness supplies the guarantee.
How it works
Write the quantity of interest as an average over some random quantity: outcomes, scenarios, samples from a distribution.
Draw many independent samples of that random quantity, using a trustworthy generator and a recorded seed.
Evaluate the quantity on each sample and average the results; the average is the estimate.
Estimate the error from the spread of the evaluations, and report the interval alongside the point.
Add samples until the error bar is narrow enough for the decision, knowing each halving costs four times the samples.
Three practical examples
- An insurer simulating ten thousand random years of claims to estimate the average annual loss and its tail.
- A project team sampling random combinations of task delays to estimate the chance a launch slips past quarter end.
- A data scientist estimating a model's expected calibration error by resampling the test set two thousand times.
Think of it this way
Think of estimating the average depth of a lake: you cannot drain it, but a few hundred random soundings give an average you can trust, and the more soundings, the tighter the answer.
Worked example
the price of one more decimal
Suppose a risk team estimates average annual loss at 2.3 million from ten thousand simulated years, with an error bar of plus or minus 0.2. The figures are illustrative.
The board asks for the error bar down to 0.05: a fourfold tightening, which by the square-root law needs sixteen times the samples, one hundred sixty thousand runs.
The simulation takes a weekend instead of an hour, and the answer moves to 2.31 million, changing no decision the team was actually facing.
The exercise teaches the habit: ask how narrow the bar must be to change the decision, then buy exactly that much sampling and no more.
Illustrative case study
A fictional trading desk reported Monte Carlo estimates of portfolio risk from five hundred scenarios without error bars, and the numbers moved noticeably between runs.
Different analysts got different answers from the same model, and the disagreement was read as a bug until someone computed the sampling error, which explained the whole spread.
The invented lesson: a Monte Carlo number without its error bar is a rumor. Report the interval, or the estimate will be trusted exactly as much as it should not be.
Common misconceptions and mistakes
- Monte Carlo gives exact answers with enough samples. It gives arbitrarily tight answers, never exact ones; the error bar is permanent, only shrinkable.
- It is only for physics and finance. Any average over uncertainty qualifies, including model evaluation and project risk.
- More samples fix a bad model. They shrink sampling error only; a wrong scenario generator gives a precisely wrong answer.
Business use
Always ask for the error bar with a simulated estimate, and ask how many samples produced it.
Record the random seed so runs can be reproduced and disagreements diagnosed.
Buy precision to fit the decision: tighten the interval until the answer would not change the choice, then stop.
Remember the limit: Monte Carlo prices the model you gave it, and garbage scenarios produce confident garbage.
People also ask
What is Monte Carlo integration?
Estimating an average or integral by random sampling, letting the sample mean stand in for the exact value.
How fast does the error shrink?
With the square root of the sample count: four times the samples halves the error.
What is it used for in business?
Risk simulation, pricing, project uncertainty and putting confidence intervals on model metrics.
Does it need special software?
No: any scripting language with a random generator can run a sound Monte Carlo estimate.
Is Monte Carlo the same as simulation?
Simulation is the broader activity; Monte Carlo integration is using simulated samples specifically to estimate an average or integral.
Check your understanding
Why does each extra decimal place of Monte Carlo precision cost a hundred times the samples?
Answer: Because the error shrinks with the square root of the sample count, so tightening the estimate by a factor of ten requires squaring that factor in samples, which is ten squared, one hundred.
Related terms
Learn next
AI for Business Owners helps you read simulation-based numbers in risk and forecast reports: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]
More in AI and machine learning
- Modes of variation - Modes of variation describe systematic ways observations differ around an average pattern.
- Momentum - Momentum is a training technique that lets an optimizer keep moving in its recent direction: each update blends the new gradient with a running average of...
- Monolithic Machine Learning Pipeline - A monolithic machine learning pipeline is one built as a single welded unit: ingestion, cleaning, features, training and serving inside one program, so any...
- Moral graph - A moral graph is an undirected graph constructed from a directed acyclic graph by connecting every pair of nodes that share a child and then removing arrow...
- Moral outsourcing - Moral outsourcing is the shifting of responsibility for ethical judgment or harmful consequences onto an external person, organization or system.
- Multi-class - Multi-class describes a classification setting where each item belongs to exactly one of three or more categories: a ticket routed to one of ten teams, a...
Sources
[S13] Minka, Statistical Learning/Pattern Recognition Glossary; [S11] Murphy, Probabilistic Machine Learning: An Introduction. Full URLs are in the source ledger. Verified 11 October 2026.
Previous: Monte Carlo expectation maximization Next: Monte Carlo localization