Monte Carlo Simulation in Investing: What It Adds and Where It Breaks

Monte Carlo simulation runs a model many thousands of times, drawing a random value for each uncertain input on every run, and collects the outputs into a distribution. Instead of one answer you get a spread: where results cluster, how wide the range is, and how often the extreme outcomes show up under the assumptions you specified.

That last clause carries all the weight. A Monte Carlo output looks like a measurement of uncertainty. It is actually a very thorough description of the assumptions you fed it. Everything useful and everything dangerous about the method follows from that.

What the method actually does

The mechanics are less mysterious than the name suggests.

Take any model that turns inputs into an output. It might be a portfolio return path, a retirement corpus projection, a valuation, or a strategy’s equity curve. Some of its inputs are uncertain. Rather than picking one value for each, you describe each uncertain input as a distribution: a shape that says which values are more and less likely.

Then you run the model. On each run, the computer draws one value from each input distribution, computes the output, and stores it. Do that ten thousand times and you have ten thousand outputs. Sort them and you can read off the median, the spread, the proportion of runs that fell below some threshold, and the shape of the tails.

For a portfolio path, the structure is usually sequential rather than single shot. You simulate a return for period one, apply it, simulate period two from the new base, and so on to the horizon, which produces a full path rather than a single endpoint. Run many paths and you can ask path dependent questions: how often did a simulated portfolio fall more than a given amount from its peak, how long did the average underwater period last, how often did a withdrawal plan exhaust the corpus before the horizon.

That path dependence is the genuine contribution. Analytical formulas handle endpoints well. They handle questions about the journey badly, and most real investment questions are about the journey.

Where it earns its keep

Sequence risk. Two paths with identical average returns produce very different outcomes if one has its bad years early and the other has them late, particularly when money is being added or withdrawn. Simulation makes this visible in a way that a compound growth calculation cannot, which is the same blind spot described in what is CAGR.

Combining several uncertainties. When four or five inputs are each uncertain and they interact non linearly, working out the combined effect by hand is impractical. Simulation handles it mechanically.

Drawdown and underwater questions. How deep, how often, how long. Because each run produces a full path, you can compute a distribution of maximum drawdowns rather than a single realised one, which is a more honest representation of what a strategy could have delivered.

Threshold and goal questions. The proportion of runs that clear or miss a defined level is a natural output, and it is easier for a non technical reader to interpret than a standard deviation.

How to do it well

State every distribution and defend it. The most common failure is using a normal distribution for asset returns because it is convenient. Realised equity returns have fatter tails and negative skew: extreme moves happen more often than a normal distribution allows, and the bad ones are larger. A simulation built on normal returns will systematically understate how bad the worst case gets, and it will do so while producing thousands of runs that look like thoroughness.

Do not assume independence across periods without checking. Drawing each period’s return independently from the same distribution implies no momentum, no mean reversion, and no volatility clustering. Real return series show volatility clustering clearly: calm periods follow calm periods and turbulent ones cluster together. Independent draws will therefore understate the length of bad stretches even if they get the average right.

Handle correlation explicitly, and question whether it is stable. For a multi asset simulation you need the correlation structure, not just the individual distributions. Simulating each asset independently produces a portfolio that diversifies far better than any real one. But a fixed correlation matrix has its own problem, because correlations rise in crises exactly when diversification is needed. Fixed correlations therefore understate the tail too.

Consider bootstrapping instead of fitting a distribution. Rather than assuming a shape, you can resample actual historical returns, drawing periods at random from the observed record. Block bootstrapping, where you draw contiguous chunks rather than single periods, preserves some of the clustering and correlation structure. It avoids inventing a distribution, at the cost of being limited to the patterns present in the sample you drew from.

Calibrate on data that was knowable. If the historical sample used to set the distributions has been restated, cleaned, or drawn only from surviving companies, the simulation inherits that bias and multiplies it across every run. This is the same problem as survivorship bias in backtests and the reason point-in-time data matters.

Run enough paths, and check convergence. Results should stabilise as the number of runs increases. If the ninety fifth percentile still moves materially between ten thousand and fifty thousand runs, you are reading noise. Tail statistics need far more runs to stabilise than central ones.

Report the assumptions next to the chart, every time. A fan chart of simulated paths is one of the most persuasive images in finance and one of the easiest to misread. Publishing the input distributions, the correlation treatment and the calibration window alongside it is the minimum that lets a reader judge the output.

What it does not tell you

It cannot manufacture information. Simulation redistributes the assumptions you supplied across many runs. It adds no new knowledge about the world. If the expected return assumption is wrong, all ten thousand paths are wrong together, and the distribution is centred in the wrong place with impressive precision.

Precision is not accuracy. A simulation can report a percentile to two decimal places. That figure is precise about the model and says nothing about how well the model matches reality. This false precision is the single most common misuse, because the volume of computation feels like evidence.

The tails are the least reliable part, and the part people care about. Extreme outcomes are, by construction, driven by the shape of the input distributions in regions where you have the least historical data to calibrate from. The part of the answer that matters most for risk is the part with the weakest empirical foundation.

Regime change is invisible. A simulation calibrated on one market environment assumes that environment continues. Structural shifts in market composition, liquidity, interest rate regimes or the investor base are not in the sample and cannot appear in the output. This is the same reason backtest results do not repeat.

It does not model behaviour or constraints. Simulated paths are held with perfect discipline. Real portfolios face redemptions, committee decisions, mandate limits and human tolerance for pain. A path that recovers over several years is only survivable if someone stays invested through it.

Costs, taxes and implementation are often omitted. A simulation of gross returns describes an outcome nobody receives. Transaction costs, impact, and tax on realised gains all compound against the investor over the horizon, and a simulation that ignores them shifts the entire distribution favourably.

It can be tuned until it says what you want. Small, defensible looking changes to the input distributions move the output percentiles a lot. Because those choices are usually buried in an appendix or absent entirely, a Monte Carlo result is very hard for an outside reader to audit. That makes documentation of the inputs not a courtesy but the substance of the work.

The right way to use a Monte Carlo output is as a structured way to ask what if, run many times. It shows you the shape your assumptions imply. It never shows you the shape the future will take, and any presentation that blurs that distinction has made the tool more misleading than the single point forecast it replaced.

Related reading

• Portfolio and Backtest Metrics, Explained: the hub guide to the metrics and methods behind portfolio analysis.
• Scenario Analysis Explained: a small set of coherent futures rather than a simulated distribution.
• Sensitivity Analysis Explained: finding which single input actually drives the answer.
• Stress Testing a Portfolio: severe but conceivable shocks, without probabilities attached.
• Why Backtest Results Do Not Repeat: regime change, decay and the honest expectation.

This article is educational. Altys Labs is not a registered research analyst or investment adviser, and nothing here is investment advice or a recommendation to buy, sell, or hold any security.