What Is a Monte Carlo Simulation? How It Works, With Key Steps

Search Fintrev

Search our guides, reviews and glossary.

A Monte Carlo simulation is a computational method that estimates the probability of different outcomes by running a model thousands of times, each run using randomly drawn values for the inputs that are uncertain. Instead of one forecast, it produces a distribution of results: a range of possible outcomes with an estimate of how likely each part of that range is. It is also called a Monte Carlo analysis or a multiple probability simulation. Finance uses it to price options, measure portfolio risk, test retirement plans and estimate the odds that a project runs over budget. Engineering, physics, insurance and telecommunications use it for the same kind of question: what could happen, and how often, when key variables cannot be known in advance.

What Is a Monte Carlo Simulation?

Many forecasts contain at least one variable whose future value is unknown: next year’s stock return, the cost of a construction phase, the number of insurance claims, peak demand on a network. A deterministic model handles this by plugging in one number, usually an average or a best guess, and calculating one answer. A Monte Carlo simulation treats that variable as random and describes it with a probability distribution instead of a single value.

Every Monte Carlo model has the same four components:

1. A model. A set of equations or rules that turns inputs into an output, such as “portfolio value at the end of the year equals starting value times one plus the return, minus withdrawals.”
2. Uncertain inputs and their distributions. Each variable that cannot be known is described by a distribution: normal, lognormal, uniform, triangular, or a set of historical observations to resample from.
3. Random sampling. A random number generator draws one value for each uncertain input, and the model computes one outcome. That single pass is called a trial, iteration or path.
4. Aggregation. After thousands of trials, the outcomes are collected and summarized as a mean, median, percentiles, a histogram or the probability of crossing a threshold.

The result answers a different question than a point forecast does. A point forecast says “the portfolio will be worth $1.2 million.” A Monte Carlo analysis says something like “in 10% of trials the portfolio was worth less than $800,000, the median was $1.1 million, and in 10% of trials it exceeded $1.6 million.” (These figures are hypothetical, shown only to illustrate the format of the output.) The spread is the information. It describes risk, which a single number cannot.

How the Monte Carlo Simulation Assesses Risk

The case for the method starts with what happens when uncertainty is replaced by an average. Suppose a retiree plans withdrawals using an assumed average return of 6% a year. A spreadsheet that applies 6% every year shows a smooth path, and if the plan works at 6% it appears to work. Real returns vary, and their order matters once money is being withdrawn. A bad stretch early in retirement forces withdrawals from a shrunken portfolio, and later good years have less capital to compound. Two return sequences with the same average can leave very different ending balances. This is called sequence-of-returns risk.

A Monte Carlo simulation captures this because every trial generates its own sequence of returns. Counting how many trials end with money left gives an estimated probability of success for the plan under the stated assumptions. A plan that looks safe at the average can fail in a meaningful share of trials.

The same logic applies wherever outcomes depend on several uncertain variables at once:

• Projects. Each task’s duration and cost is uncertain. Adding up the averages gives a single total that hides the chance of a large overrun. Simulating each task’s range and summing them trial by trial estimates the probability of finishing over budget or late.
• Insurance. Claim frequency and claim size both vary. Simulating thousands of years of claims shows how large losses could become in bad years, which feeds into pricing and capital reserves.
• Derivatives and credit. Averaging payoffs across thousands of simulated price paths gives an estimate of an option’s value. Simulating a borrower’s asset values or cash flows estimates the probability of default.
• Networks. Telecommunications companies simulate demand to test whether capacity holds during peak events, such as Super Bowl Sunday, as well as on ordinary days.

History of the Monte Carlo Simulation

The idea of estimating a quantity through repeated random experiments is older than the name. In the 18th century, Georges-Louis Leclerc, Comte de Buffon, asked for the probability that a needle dropped on a floor of parallel lines crosses one of them. The answer involves π, so dropping many needles and counting crossings estimates π.

The modern method came out of the Manhattan Project. The mathematician Stanislaw Ulam reportedly got the idea while recovering from an illness and playing solitaire. He wondered whether it was easier to estimate the odds of winning by playing many hands and counting wins than by working out the combinatorics. He saw that the same approach could solve neutron diffusion problems too complex for direct calculation. Ulam shared the idea with John von Neumann, who programmed early simulations on ENIAC, one of the first electronic computers.

The name is generally credited to their Los Alamos colleague Nicholas Metropolis. It refers to the Monte Carlo casino in Monaco, which Ulam’s uncle was said to frequent, and to the role of chance in games such as roulette and dice. Metropolis and Ulam published a paper titled “The Monte Carlo Method” in 1949.

The method spread as computing power grew. In finance, Phelim Boyle’s 1977 paper applied Monte Carlo simulation to option pricing. Later work extended it to options that can be exercised early, including the least-squares method published by Francis Longstaff and Eduardo Schwartz in 2001.

How Monte Carlo Simulations Work

The Core Loop

The process inside any Monte Carlo model is a loop:

1. Draw a random value for each uncertain input from its assigned distribution.
2. Run the model with those values and record the output.
3. Repeat thousands or millions of times.
4. Analyze the collection of recorded outputs.

The intuition comes from the law of large numbers. As the number of independent trials grows, the average of the results converges toward the true expected value of the model’s output, and the histogram of results converges toward the true distribution implied by the model’s assumptions.

How Many Trials Are Enough?

The precision of a Monte Carlo estimate improves with the square root of the number of trials. The standard error of an estimated mean is roughly:

Standard error ≈ σ ÷ √N

where σ is the standard deviation of the simulated outcomes and N is the number of trials. Quadrupling the number of trials only halves the sampling error, so going from 1,000 to 4,000 trials gives a meaningful improvement, while going from 100,000 to 400,000 gives a much smaller absolute gain. Estimates about the tails, such as the worst 1% of outcomes, need more trials than estimates of the average, because only a small fraction of trials land in the tail.

This sampling error is separate from model error. Ten million trials of a model with a wrong volatility assumption produce a very precise estimate of the wrong distribution.

Random Numbers and Seeds

Computers generate pseudorandom numbers: deterministic sequences that pass statistical tests for randomness. A starting value called a seed determines the sequence. Fixing the seed makes a simulation reproducible, which matters when results need to be audited or compared across model versions. Analysts also use variance reduction techniques, such as antithetic variates (pairing each random draw with its mirror image), to get more precision from the same number of trials.

A Simple Illustration: Estimating π

A classic teaching example shows the mechanics without any finance. Draw a square with sides of length 1 and a quarter circle of radius 1 inside it. The quarter circle covers π/4 of the square’s area. Now generate random points with x and y coordinates between 0 and 1, and count how many fall inside the quarter circle (where x² + y² ≤ 1). The fraction inside approaches π/4 as the number of points grows, so four times that fraction estimates π. With a few hundred points the estimate is rough. With millions it is close. Nothing about π is random, but random sampling still recovers it, which is the essence of the Monte Carlo method.

The 4 Steps in a Monte Carlo Simulation of an Asset Price

A standard way to simulate future prices for a stock or other asset can be built in Microsoft Excel or any similar spreadsheet. It models an asset’s price movement as two components:

• Drift: the constant directional tendency of the price.
• A random input: a shock that represents market volatility.

Historical price data supply the building blocks: the average periodic return, standard deviation, variance and drift. This setup corresponds to a model called geometric Brownian motion, the same model behind the Black-Scholes option pricing formula.

Step 1: Calculate Periodic Daily Returns

From the asset’s historical closing prices, compute a series of daily returns using the natural logarithm:

Periodic daily return = ln(Day’s price ÷ Previous day’s price)

This log return differs from the ordinary percentage change, (Day’s price − Previous day’s price) ÷ Previous day’s price. Log returns add across time. The sum of daily log returns equals the log return over the whole period, which makes them convenient for building multi-day paths. In Excel, the formula is =LN(B3/B2) copied down the column of prices.

Step 2: Calculate Average Return, Standard Deviation, Variance and Drift

Apply three functions to the full series of daily returns:

• AVERAGE gives the average daily return.
• STDEV.P gives the standard deviation (σ).
• VAR.P gives the variance (σ²).

Then calculate drift:

Drift = Average daily return − (Variance ÷ 2)

The subtraction of half the variance comes from geometric Brownian motion. In that model, an asset with expected simple return μ has log returns that drift at μ − σ²/2. Volatility pulls the typical compounded outcome below the arithmetic average, an effect often called volatility drag. Because the Step 1 series is already in log returns, some practitioners skip the subtraction. The recipe as commonly taught keeps it.

Drift can also be set to zero. That choice reflects a particular theoretical view, for example, that past average returns are not a reliable guide to future direction. Over short horizons the difference is small, because daily drift is tiny relative to daily volatility.

What Is Annual Drift?

Drift is measured per period, so a daily drift needs scaling for a longer horizon. Under the usual assumption of independent returns and about 252 trading days a year (a convention, not a fixed number), annual drift ≈ daily drift × 252. Annual variance ≈ daily variance × 252, so annual volatility ≈ daily standard deviation × √252. An annual drift of 0.07, for instance, means the model’s median path grows by a factor of e^0.07 (about 7.25%) per year before random shocks. The compound annual growth rate of a simulated path will scatter around that figure from trial to trial.

Step 3: Generate the Random Input and the Next Day’s Price

Each day’s random shock is drawn from a normal distribution scaled by the asset’s volatility:

Random value = σ × NORMSINV(RAND())

RAND() returns a uniform random number between 0 and 1. NORMSINV (or NORM.S.INV in newer versions of Excel) converts it into a draw from the standard normal distribution (mean 0, standard deviation 1). Multiplying by σ scales the shock to the asset’s historical daily volatility.

The next day’s price is then:

Next day’s price = Today’s price × e^(Drift + Random value)

Step 4: Repeat to Build Paths and Many Trials

In Excel, e raised to a power x is EXP(x). So the price formula becomes =Previous_price*EXP(Drift + Random_value). Copy it down for as many rows as the horizon has trading days, and each row represents one day. One column of rows is one simulated price path.

Repeat the column across hundreds or thousands of columns (or use a data table to recalculate) and you have thousands of independent paths. Collecting the ending prices, or any statistic along each path, gives a distribution from which you can estimate the probability that the price follows a given trajectory or ends above or below a given level.

Example of a Monte Carlo Simulation

The following numbers are hypothetical and do not describe any real security.

Building One Price Path

Assume a stock trades at $100. From its history, the average daily log return is 0.0005 and the daily standard deviation is 0.02. That makes the variance 0.02² = 0.0004.

Drift = 0.0005 − (0.0004 ÷ 2) = 0.0005 − 0.0002 = 0.0003

Day 1. The random number generator returns a standard normal draw of −1.2. Random value = 0.02 × (−1.2) = −0.024 Exponent = 0.0003 + (−0.024) = −0.0237 Price = $100 × e^(−0.0237) ≈ $100 × 0.9766 = $97.66

Day 2. The next draw is 0.5. Random value = 0.02 × 0.5 = 0.010 Exponent = 0.0003 + 0.010 = 0.0103 Price = $97.66 × e^(0.0103) ≈ $97.66 × 1.0104 = $98.67

Continuing for 252 days produces one possible year for the stock. Its ending price is a single sample, no more meaningful on its own than one roll of a die.

Running Thousands of Trials

Repeat the year 10,000 times with fresh random draws and you get 10,000 ending prices. Sort them, and the 500th lowest value marks the 5th percentile, the 5,000th marks the median, and the 9,500th marks the 95th percentile. Counting how many trials end below, say, $80 and dividing by 10,000 estimates the probability of a loss of at least 20% under this model.

Because the model compounds log returns, ending prices are lognormally distributed. The distribution is skewed: losses are capped at 100%, while gains have no ceiling. The mean ending price sits above the median, pulled up by a small number of very strong paths.

A Retirement Planning Example

Retirement software applies the same structure to a household’s finances. A hypothetical setup might look like this:

• Starting portfolio: $1,000,000
• Annual withdrawal: $40,000 in the first year, increased each year by an assumed inflation rate
• Horizon: 30 years
• Annual return: drawn each year from a distribution with an assumed mean and standard deviation for the chosen asset mix

Each trial generates 30 years of returns, applies withdrawals and records whether the portfolio was depleted before year 30. If, hypothetically, 8,700 of 10,000 trials end with money remaining, the reported probability of success is 87%. The failing 13% of trials typically show poor returns early in the horizon or prolonged high inflation in the model. Inflation assumptions often draw on history such as the U.S. inflation rate by year, and that choice can shift the result noticeably.

A Project Cost Example

A project has three phases, each with a cost range estimated by the team (all hypothetical):

• Phase A: $200,000 to $300,000, most likely $240,000
• Phase B: $400,000 to $700,000, most likely $500,000
• Phase C: $100,000 to $200,000, most likely $130,000

Adding the most likely values gives $870,000. A Monte Carlo model assigns each phase a triangular distribution with those three points, draws a cost for each phase in every trial and sums them. Because each range stretches further above its most likely value than below, the simulated total will tend to exceed $870,000 more often than not. The simulation quantifies how often, and how large a contingency budget would cover, say, 90% of trials.

Monte Carlo Simulation Results Explained

The raw output of a simulation is a long list of outcomes. Several summaries turn it into something usable.

The distribution and its center. A histogram of outcomes shows the shape: symmetric or skewed, narrow or wide, with or without long tails. The mean is the probability-weighted average outcome. The median is the outcome that half the trials beat. For skewed results such as compounded investment returns, the median is usually the better description of a typical result.

Percentile bands. Many tools display results as a fan chart: the median path surrounded by bands such as the 10th to 90th percentile. The width of the bands is the model’s estimate of uncertainty at each point in time, and it widens with the horizon.

Standard deviations. When outcomes are normally distributed, about 68% fall within one standard deviation of the expected value, about 95% within two and about 99.7% within three. These rules do not apply directly to lognormal price outcomes or other skewed distributions, where percentiles from the simulation itself are more reliable.

Probabilities of specific events. Often the most useful number is a count: the share of trials in which the portfolio runs out, the project exceeds budget, or losses exceed a threshold. Value at risk (VaR) is one example. A one-day 99% VaR of $1 million means that in 99% of simulated days, losses did not exceed $1 million. Monte Carlo is one of the standard ways to calculate it.

How Is the Monte Carlo Simulation Used in Finance?

The main applications in finance include:

• Option pricing. The simulation generates many paths of the underlying asset under risk-neutral assumptions, in which the asset drifts at the risk-free rate. It calculates the option’s payoff on each path, averages the payoffs and discounts the average to present value at the risk-free rate. For a plain European option, closed-form formulas such as Black-Scholes are faster. Monte Carlo earns its place with path-dependent contracts, such as Asian options (whose payoff depends on an average price) and barrier options, and with products tied to several underlying assets.
• Portfolio valuation and comparison. Alternative portfolios can be simulated side by side to compare their range of outcomes, the probability of losses and the depth of drawdowns. Multi-asset simulations need correlations between assets, typically introduced with a technique such as Cholesky decomposition. Results can then be summarized with measures like the Sharpe ratio across trials.
• Fixed income. In interest rate models the short rate, the rate on very short-term borrowing, is the random variable. Simulated rate paths value bonds with embedded options and mortgage-backed securities, whose cash flows change as rates move. That path dependence makes a single yield to maturity less informative.
• Risk management. Banks and asset managers use Monte Carlo VaR and expected shortfall to estimate potential losses across large books of positions, including non-linear instruments that simpler methods handle poorly.
• Credit analysis. Portfolio credit models simulate correlated defaults across many borrowers.
• Retirement and financial planning. Planners compare withdrawal rates, asset allocations, retirement dates and Social Security claiming choices by the likelihood that a client runs out of money.
• Corporate finance. In capital budgeting, uncertain revenue, cost and growth inputs are simulated to produce a distribution of net present value instead of a single estimate, showing the probability that a project destroys value.

Monte Carlo in Trading

Traders use the method in two ways. The first is forward-looking: simulating price paths to size positions or set stop levels. The second tests a strategy’s track record. A backtest produces one sequence of trades. Reshuffling or resampling those trades thousands of times shows the range of drawdowns and ending equity the same trades could have produced in a different order. A strategy with a 15% backtest drawdown might show much deeper drawdowns in a meaningful share of sequences, which bears on how much capital and leverage it can carry. The resampling assumes trades are independent, an assumption that fails for strategies whose results cluster in particular market regimes.

What Professions Use the Monte Carlo Simulation?

Financial analysts, quants, risk managers, planners, actuaries and project managers rely on it for the uses described above. It also appears in nearly every other field that must measure risk:

• Engineers use it for reliability analysis, estimating how likely a component or system is to fail under variable loads.
• Physicists use it to model particle transport, its original purpose, and in statistical mechanics.
• Meteorologists run ensembles of forecasts with perturbed starting conditions to estimate the probabilities of weather outcomes.
• Supply chain and operations analysts simulate demand, lead times and inventory levels.

What Factors Are Evaluated in a Monte Carlo Simulation?

The inputs depend on the model, but an investment simulation usually involves:

• Historical price or return data for each asset, the raw material for every other parameter.
• Average return and drift, the expected directional movement per period.
• Standard deviation and variance, the size of the random shocks.
• Correlations between assets, which determine how much diversification the simulated portfolio actually gets.
• The distribution assumed for returns: normal, lognormal, fat-tailed (such as Student’s t) or resampled historical returns.
• Time step and horizon: daily, monthly or annual steps over the planned period.
• Number of trials, which governs sampling precision.
• Cash flows in planning models: contributions, withdrawals, fees, taxes and inflation.

How Does a Monte Carlo Model Take Past Trends Into Account?

A standard parametric model uses the past only through its parameters. History is condensed into a mean, a standard deviation and perhaps correlations, and each simulated period draws a fresh, independent shock from that distribution. The model does not remember that the previous simulated month was a crash, and it does not reproduce momentum, mean reversion or volatility clustering unless those features are built in.

There are ways to bring more of history in:

• Historical bootstrapping draws actual past returns at random instead of from a fitted distribution, which keeps the real shape of the data, including fat tails.
• Block bootstrapping resamples consecutive stretches of history (for example, 12 months at a time) to keep some short-term patterns.
• Regime-switching and stochastic volatility models let parameters change over time, so calm and turbulent periods alternate.

Each approach still assumes the future will resemble the past in some way. The choice is about which features of the past to keep.

Advantages and Disadvantages of a Monte Carlo Simulation

Advantages

• It shows a range, not a point. Decision makers see downside and upside along with their estimated likelihood.
• It handles complexity. Many uncertain inputs, non-linear payoffs, path dependence and cash flows can be modeled together where no closed-form formula exists.
• It captures sequence effects of withdrawals, contributions and rebalancing.
• It supports sensitivity analysis. Re-running with different assumptions shows which inputs drive results most.
• It is flexible. The same loop works for options, budgets and networks, and a basic version fits in a spreadsheet.

Disadvantages

• Garbage in, garbage out. Results are only as good as the distributions and parameters chosen. A small change in assumed return or volatility can move a retirement plan’s success rate substantially.
• Normal distributions understate extreme events. Historical market returns have fatter tails than a normal distribution implies, so a model built on normality assigns too little probability to crashes.
• Correlations are unstable. Assets that move independently in calm markets often fall together in a crisis, which an average correlation does not reflect.
• Stationarity is assumed. Most models treat the statistical properties of returns as fixed over the horizon.
• False precision. A result such as “83.4% probability of success” looks exact, but the decimals reflect sampling, not knowledge of the future.
• Computational cost. Large portfolios, nested simulations and tail estimates can require heavy computing resources.

Limitations of a Monte Carlo Simulation

The disadvantages above share a root: a Monte Carlo simulation explores the consequences of its assumptions thoroughly, but it does not test whether they are right. If the return distribution has no room for a 40% one-year decline, no number of trials will produce one. If inflation is fixed at an assumed rate, the model says nothing about a decade of high inflation. A basic price model also ignores macroeconomic trends, company leadership, market sentiment and business cycles. It effectively assumes that price changes follow statistically stable patterns drawn from history, even though real markets do not always behave that way.

Two practical points follow. First, a Monte Carlo probability is conditional on the model. Two planning tools with different return and inflation assumptions can give the same household different success rates, and both can be computed correctly. Second, the inputs deserve the most scrutiny. A useful check is to re-run the simulation with more conservative assumptions, such as lower returns, higher volatility or fatter tails, and see how much the conclusions change.

The method does not tell you which outcome will occur, and real results can fall outside every simulated path. It also does not replace judgment about things that are hard to quantify, such as a company’s competitive position, policy changes or behavioral responses. For that reason, some analysts pair simulations with explicit stress scenarios.

Monte Carlo Simulation vs. Related Methods and Terms

Monte Carlo Simulation vs. Historical Backtesting

A backtest replays one actual historical sequence of returns. It shows what did happen, once. A Monte Carlo simulation generates many sequences that could have happened under the model’s assumptions. A backtest is limited to history’s single path, while a simulation is limited by its model. Bootstrapping sits in between, building new sequences out of real historical returns.

Monte Carlo Simulation vs. Scenario Analysis and Stress Testing

Scenario analysis evaluates a small number of hand-picked cases, such as a base case, a recession and a rate shock. Stress testing applies specific extreme shocks. Neither assigns probabilities across a full range of outcomes. Monte Carlo does assign probabilities, but those probabilities come from the model. Many risk frameworks use both: simulation for the overall distribution and stress tests for named events the model may underweight.

Monte Carlo Simulation vs. Sensitivity Analysis

Sensitivity analysis changes one input at a time while holding the others fixed, to see how much the output responds. Monte Carlo varies all uncertain inputs at once, including their interactions. Sensitivity analysis shows which input matters most, and Monte Carlo shows how the uncertainties combine.

Monte Carlo Method vs. Monte Carlo Algorithm

In computer science, a Monte Carlo algorithm is a randomized algorithm that runs in bounded time but may return an incorrect result with some small probability. It contrasts with a Las Vegas algorithm, which always returns a correct result but takes a random amount of time. A primality test that answers “composite” or “probably prime” is the standard example, and more runs reduce the chance of error. Both ideas use randomness, but the algorithm seeks a correct yes-or-no answer rather than an estimated distribution.

Markov Chain Monte Carlo

Markov chain Monte Carlo (MCMC) is a family of methods for sampling from distributions that are hard to sample directly, widely used in Bayesian statistics. The Metropolis algorithm, published in 1953 and named for Nicholas Metropolis, is one of its foundations. In finance, MCMC appears mainly in estimating model parameters, not in simulating portfolios.

The Monte Carlo Fallacy

The Monte Carlo fallacy, also called the gambler’s fallacy, is the belief that a random event becomes more likely because it has not happened recently. For example, a gambler may think red is “due” after a long run of black. The name comes from a 1913 night at the Monte Carlo casino, when black came up 26 times in a row and gamblers lost heavily betting on red. A correctly built simulation assumes the opposite: each independent draw is unaffected by the draws before it.

How to Run a Monte Carlo Simulation

A basic simulation follows the same sequence whatever the tool:

1. Define the question. For example: “What is the probability this portfolio lasts 30 years?” or “What contingency covers 90% of cost outcomes?”
2. Build the deterministic model first. Check that it gives the right answer when every input is fixed.
3. Assign distributions to the uncertain inputs. Base them on historical data, expert ranges or both, and record the reasoning.
4. Set correlations between inputs that tend to move together.
5. Choose the number of trials and the time step. Increase trials until key results stop shifting meaningfully between runs.
6. Run and summarize with the median, percentiles and the probabilities of the events in question.
7. Test the assumptions by re-running with alternative inputs.

Spreadsheets handle small models. Dedicated add-ins and languages such as Python or R handle larger ones, with libraries for random number generation and vectorized calculation that make millions of trials practical.

FAQs

The Bottom Line

Investors, planners, risk managers and project teams can use Monte Carlo results to see downside risk and the odds of specific events, such as a portfolio running out of money. Every probability depends on the model's assumptions, so the inputs deserve the most scrutiny. Re-running with lower returns, higher volatility or fatter tails shows how much the conclusions change, and stress scenarios can cover events the model may underweight.