Simulation Methods: Historical, Bootstrap, and Monte Carlo

Published: May 9, 2026

Article by Ryan O'Connell, CFA, FRM

A single expected return or point forecast tells you almost nothing about the range of outcomes your portfolio might experience. Simulation methods address this limitation by generating many possible scenarios and aggregating the results into probability distributions. This guide compares the three most common simulation approaches in finance — historical simulation, bootstrap simulation, and Monte Carlo simulation — covering when to use each method, their key assumptions, and the pitfalls that lead to unreliable results.

What Are Simulation Methods?

Simulation methods generate multiple scenarios to estimate probability distributions of future outcomes. Rather than producing a single forecast, they reveal the range of possibilities — including tail events that simple averages would hide.

Three major simulation approaches dominate finance applications:

• Historical simulation uses actual past outcomes directly as possible future scenarios
• Bootstrap simulation resamples historical observations with replacement to create new synthetic scenarios
• Monte Carlo simulation generates random draws from a specified probability distribution or model

Each method makes fundamentally different assumptions about how the future relates to the past. Choosing correctly matters: the wrong method can produce misleading risk estimates, especially in the tails.

Simulation vs Scenario Analysis vs Stress Testing

These three terms are often confused. Simulation generates many random scenarios to build probability distributions. Scenario analysis evaluates a small number of specific, predetermined scenarios (e.g., “what if inflation rises to 8%?”). Stress testing focuses specifically on extreme or adverse scenarios to evaluate portfolio resilience under crisis conditions. Simulation methods can be used for all three purposes, but the goals differ.

Historical Simulation

Historical simulation uses actual observed outcomes as possible future scenarios. The logic is simple: if you want to know what returns might occur tomorrow, look at what returns actually occurred on past days and assume any of them could repeat.

The method works by collecting a history of returns (or risk factor changes), ranking them from worst to best, and reading percentiles directly. For Value at Risk (VaR), the 5th percentile of historical returns becomes the 95% VaR estimate. No distributional assumptions are imposed — the data speaks for itself.

Key assumptions:

• The past is representative of the future — the distribution of returns is stable over time
• The historical sample includes relevant tail events
• Market structure and correlations have not fundamentally changed

Strengths:

• No distributional assumption (non-parametric)
• Captures fat tails if they are present in the sample
• Preserves historical cross-asset correlations automatically
• Simple to implement and explain

Weaknesses:

• Cannot extrapolate beyond observed history — if a 10% single-day crash never happened, it will not appear
• Equally weights all historical observations, including stale data from different market regimes
• Limited scenarios — with 500 days of history, you have exactly 500 scenarios
• Ignores time-series dependence like volatility clustering

For a detailed methodology, see our guide to historical simulation VaR.

Bootstrap Simulation

Bootstrap simulation extends historical simulation by resampling with replacement. Instead of using each historical return exactly once, the bootstrap randomly draws observations from the historical sample — potentially selecting some observations multiple times and others not at all — to create new synthetic scenarios.

The key insight is that resampling generates more scenario combinations than the original sample while preserving the empirical distribution of returns. With 500 historical observations, you can generate 10,000 bootstrap samples, each representing a plausible path.

Key assumptions:

• Historical observations are independent and identically distributed (i.i.d.)
• The past distribution is representative of the future
• Sampling with replacement adequately captures uncertainty

Strengths:

• Generates more scenarios than pure historical simulation
• Preserves the empirical distribution without parametric assumptions
• Enables standard error estimation for risk metrics (by resampling entire samples and computing VaR for each)
• Multi-period paths can produce cumulative outcomes not directly observed in history

Weaknesses:

• Individual-period returns are still bounded by historical observations
• Standard bootstrap assumes i.i.d. observations, which ignores volatility clustering and serial correlation
• Resampling individual assets independently destroys historical correlations — must resample entire cross-sections together

Block bootstrap and stationary bootstrap variants address the time-dependence weakness by resampling blocks of consecutive observations rather than individual returns, preserving short-term serial dependence.

Monte Carlo Simulation

Monte Carlo simulation generates random scenarios from a specified probability distribution or model rather than directly from historical data. The analyst defines the expected return, volatility, correlations, and distributional form, then draws random values to simulate many possible paths.

The method is named after the Monte Carlo casino — a reference to its reliance on randomness. It was formalized in the 1940s for nuclear weapons research and is now the dominant simulation approach in finance for portfolio risk, derivatives pricing, and retirement planning.

Key assumptions:

• The specified model and distribution are correct — if you assume normality but returns have fat tails, results will understate tail risk
• Estimated parameters (mean, volatility, correlations) are accurate
• The correlation structure you model reflects true dependence

Strengths:

• Can model scenarios that never occurred in history — stress scenarios, regime changes, or hypothetical events
• Can explicitly model correlations, copulas, and complex dependence structures when included in the model
• Handles path-dependent instruments (options, structured products) via full revaluation
• Scales to thousands or millions of scenarios with modern computing

Weaknesses:

• “Garbage in, garbage out” — results are only as good as the model specification
• Distribution misspecification is the primary source of model risk
• Requires parameter estimation, which introduces estimation error
• Tail convergence requires many simulations (10,000+ for stable VaR estimates)

Historical Simulation vs Bootstrap vs Monte Carlo

The three methods differ along several critical dimensions. The right choice depends on your data, your model confidence, and your risk measurement goals.

How to Choose a Simulation Method

Start with three questions:

1. Do you trust a parametric model? If yes, Monte Carlo gives flexibility. If not, use historical or bootstrap.
2. Do you need to model scenarios that never occurred? If yes, you must use Monte Carlo. Historical methods cannot extrapolate.
3. Is time dependence critical? If volatility clustering matters, use filtered historical simulation (which applies GARCH to historical data) or Monte Carlo with a GARCH component.

Example: VaR Estimation Using All Three Methods

Consider a simple portfolio allocated 60% to U.S. large-cap stocks (SPY, the S&P 500 ETF) and 40% to aggregate bonds (AGG, the Bloomberg U.S. Aggregate Bond ETF), evaluated using 250 trading days of history.

For comprehensive VaR methodology, see our Value at Risk guide and detailed method articles on historical VaR and Monte Carlo VaR.

Common Mistakes

Simulation methods are powerful but easily misapplied. These are the most frequent errors:

1. Using stale data: A historical sample dominated by a calm market regime will understate risk when volatility spikes. Equally weighting observations from 2017 (low volatility) and 2008 (crisis) may not reflect current conditions.

2. Distribution misspecification: Assuming normality when returns have fat tails leads Monte Carlo to underestimate extreme losses. Financial returns often exhibit excess kurtosis and negative skewness.

3. Ignoring correlation changes: Correlations between assets spike during market stress. A model calibrated to calm-period correlations will understate portfolio risk precisely when diversification benefits disappear.

4. Resampling assets independently: In bootstrap simulation, drawing stock returns and bond returns independently destroys their historical correlation structure. Always resample entire cross-sections (all assets on the same day) together.

5. Insufficient scenarios for tail estimation: Estimating the 99% VaR requires enough simulations that the 1st percentile is stable. With 1,000 simulations, only 10 observations define the 1% tail — too few for reliability. Use 10,000+ for VaR and 100,000+ for Expected Shortfall.

6. Confusing precision with accuracy: Monte Carlo can produce estimates to many decimal places, but that precision is meaningless if the underlying model is wrong. A stable, precise estimate of the wrong number is still wrong.

7. Mixing regimes without adjustment: Combining historical data from different volatility regimes (e.g., pre- and post-crisis) without adjustment can produce estimates that reflect neither regime accurately.

Limitations of Simulation Methods

Regime changes: The past may not predict the future. Structural changes in markets (new regulations, changed central bank policy, technological disruption) can invalidate historical relationships.

Model risk: Monte Carlo is only as good as its assumptions. The “correct” distribution is unknowable, and small changes in distributional assumptions can produce large changes in tail estimates.

Time horizon limitations: One-day VaR estimates do not scale cleanly to longer horizons. The square-root-of-time rule assumes i.i.d. returns, which is empirically false. Multi-period path simulation is more appropriate for longer horizons.

Nonlinear instruments: Options and structured products require full revaluation at each scenario — delta-normal approximations can severely understate risk when positions are far from the money or when volatility shifts dramatically.