Monte Carlo simulation is a method for generating many possible outcomes by repeatedly drawing from a specified probability model or resampling process. In finance it is used to study distributions of portfolio values, strategy paths, drawdowns, risk measures and other outcomes that depend on uncertain sequences.
Simulation starts with a model of uncertainty
A simulation does not create uncertainty from nothing. It needs rules describing how returns, shocks, volatility, correlations, trade outcomes or other variables are generated.
The most important research question is therefore not “How many paths were simulated?” but “What assumptions generated those paths?”
Path-dependent questions are a natural use case
Two sequences can contain the same set of trade outcomes but produce different drawdown paths depending on order. Simulation can reorder or generate outcomes to examine how sequence risk affects capital.
This makes Monte Carlo useful alongside Maximum Drawdown and Risk of Ruin, which ask different capital-risk questions.
Parametric simulation and resampling are different
A parametric simulation draws from an assumed distribution or fitted process. A bootstrap-style approach resamples historical observations or blocks of observations. Parametric methods can explore outcomes beyond the exact historical sequence, while resampling preserves more empirical features of the observed data.
Neither is automatically more realistic. A fitted distribution can miss tails; naive resampling can break time dependence or assume historical observations remain representative.
Dependence must be preserved when it matters
Randomly shuffling individual returns can destroy autocorrelation, volatility clustering or cross-asset dependence. If those features are material to the decision, the simulation design should preserve or model them.
GARCH Models can supply one conditional-volatility process, while block resampling is one possible way to retain local temporal dependence.
Correlation is not guaranteed to remain stable
Multi-asset simulations often rely on an estimated correlation or dependence structure. During stress, relationships can change. Using one historical correlation matrix for every simulated state can understate model uncertainty.
Correlation Risk and Scenario & Stress Testing address that wider problem.
Simulation count controls numerical noise, not model truth
More simulated paths can reduce Monte Carlo sampling error for a fixed model. They do not repair a misspecified return distribution, omitted transaction costs or unrealistic dependence assumptions.
A million precise simulations of the wrong model can still produce a misleadingly precise answer.
Risk estimates should state the requested statistic
Possible outputs include median ending capital, percentile outcomes, drawdown distribution, probability of crossing a loss threshold or an estimate of VaR/Expected Shortfall. Each answers a different question.
Value at Risk & Expected Shortfall explains the tail-risk measures themselves.
Monte Carlo differs from stress testing
Simulation samples outcomes under a probability or resampling model. Stress testing asks what happens under deliberately adverse scenarios, including scenarios whose probability may be difficult to estimate.
The two methods complement each other; Monte Carlo should not replace explicit stress scenarios simply because it produces more paths.
Validation should compare the model with later evidence
Distributional assumptions, volatility dynamics, correlations and costs should be checked against data not used to tune them. If simulated tail or drawdown behavior is much milder than later observations, the model needs review rather than a larger simulation count.
Use the distribution, not one attractive path
The point of Monte Carlo is to examine uncertainty across many possible outcomes. Selecting a visually appealing path or focusing only on the average defeats that purpose.
Monte Carlo simulation is a conditional risk experiment. It can make path uncertainty visible, but every result should be read as “under these assumptions,” not as a forecast of what the market will do.