Monte Carlo Simulation

Definition

Monte Carlo simulation is a computational technique that uses repeated random sampling from defined probability distributions to model the range of possible outcomes for a process with uncertain inputs. In financial modeling, it replaces single-point estimates with probability distributions — running the model thousands of times with randomly drawn input values to produce a distribution of outputs rather than a single number. [src1]

Key Properties

• Core mechanism: Define distributions → draw random samples → run model → repeat 1,000-10,000+ times → analyze output distribution [src1]
• Common distributions: Normal (returns), lognormal (prices), triangular (expert estimates), uniform (bounded unknowns) [src3]
• Key outputs: Mean, median, standard deviation, percentiles (P5/P95), probability of threshold [src2]
• Tools: Excel add-ins (@Risk, Crystal Ball), Python (numpy, scipy), R, MATLAB [src3]
• Applications: Portfolio VaR, option pricing, project NPV, credit risk, retirement planning [src2]

Constraints

• Distribution selection is critical: Using normal for prices (should be lognormal) or assuming symmetry for skewed variables produces misleading output. [src1]
• Correlation must be modeled: Treating correlated inputs as independent narrows the output distribution unrealistically. [src3]
• Underestimates tail risk: Simulations based on historical data assume the future resembles the past, systematically underestimating Black Swans. [src2]
• Computational requirements: 1,000+ iterations for stable means, 10,000+ for stable tail percentiles. [src4]
• False precision risk: "73.4% probability of profitability" looks precise but depends entirely on input assumptions. [src1]

Framework Selection Decision Tree

START — User needs to model financial uncertainty
├── How many uncertain inputs?
│   ├── 1-2 variables → Sensitivity Analysis
│   ├── 3-5 discrete scenarios → Scenario Analysis
│   └── 5+ continuous uncertain inputs → Monte Carlo (this unit)
├── What output format?
│   ├── Single-point estimate → Not Monte Carlo
│   ├── Probability distribution → Monte Carlo
│   ├── Input sensitivity ranking → Sensitivity Analysis first
│   └── Discrete comparison → Scenario Analysis
├── Are distributions defensible?
│   ├── YES (historical data, calibration) → Proceed
│   └── NO → Scenario Analysis instead
└── Audience comfortable with probabilistic outputs?
    ├── YES → Monte Carlo with confidence intervals
    └── NO → Scenario Analysis with 3-5 cases

Application Checklist

Step 1: Identify uncertain variables and assign distributions

• Inputs needed: Model inputs with uncertainty, historical data or expert judgment
• Output: Distribution type and parameters per variable
• Constraint: Do not default to normal — match distribution to variable characteristics [src1]

Step 2: Define correlations between inputs

• Inputs needed: Historical correlation data or logical assessment
• Output: Correlation matrix for uncertain inputs
• Constraint: Omitting correlations produces artificially narrow output distributions [src3]

Step 3: Run the simulation (1,000-10,000+ iterations)

• Inputs needed: Model with distribution-assigned inputs, correlation matrix, software
• Output: Raw output distribution
• Constraint: If 5K vs 10K iterations changes the mean by 1%+, more iterations needed [src4]

Step 4: Analyze and present results

• Inputs needed: Output distribution
• Output: Summary statistics (mean, P5-P95), probability of thresholds, histogram
• Constraint: Present as ranges and probabilities, not single numbers [src1]

Anti-Patterns

Wrong: Using normal distributions for all inputs

Assigning normal distributions to revenue allows negative values — which is impossible. [src1]

Correct: Matching distributions to variable characteristics

Use lognormal for prices/revenues, triangular for expert estimates, uniform for bounded unknowns. [src3]

Wrong: Running Monte Carlo without modeling correlations

Simulating interest rates, exchange rates, and commodities independently understates risk. [src3]

Correct: Building a correlation matrix

Define correlations using historical data or expert judgment. Even approximate correlations are better than independence. [src1]

Wrong: Presenting output as a single number

Reporting only the mean discards all information about risk, range, and tail outcomes. [src2]

Correct: Presenting the full distribution

Show histogram, percentiles (P5, P50, P95), and probability of key thresholds. [src1]

Common Misconceptions

Misconception: Monte Carlo is only for quantitative finance and options pricing.
Reality: Used in project finance, corporate budgeting, retirement planning, and any domain with multiple uncertain inputs. [src2]

Misconception: More iterations always mean better results.
Reality: Means converge at ~1,000 iterations; tails at ~10,000. Beyond that, improvement is marginal. Input quality is the binding constraint. [src4]

Misconception: Monte Carlo accounts for all risks, including Black Swans.
Reality: Simulations only reflect the distributions you define. Fat-tail events require explicit fat-tailed distributions. [src2]

Comparison with Similar Concepts

When This Matters

Fetch this when a user asks about Monte Carlo simulations, modeling financial uncertainty with probability distributions, calculating VaR, generating NPV distributions, or stress-testing with random sampling.