Monte Carlo Simulation for Trading: A Practical Guide

What Is Monte Carlo Simulation?

Monte Carlo simulation is a computational technique that uses random sampling to estimate the probability of different outcomes. Named after the famous casino in Monaco, it works by running thousands of simulated scenarios based on statistical models and counting how many produce a particular result.

In trading, Monte Carlo simulation answers questions like:

• What is the probability that BTC will be above $80,000 in 3 days?
• Given historical volatility, what is the chance of a 10% crash within 1 week?
• What is the optimal position size given the probability distribution of outcomes?

These are exactly the questions that quantitative hedge funds ask every day — and now, with the right tools, retail traders can ask them too.

How It Works: Geometric Brownian Motion

The standard model for simulating stock/crypto prices is Geometric Brownian Motion (GBM):

S(T) = S(0) × exp((μ - 0.5σ²)T + σ√T × Z)

Where:

• S(0) = Current price
• μ = Annualized drift (expected return) estimated from historical data
• σ = Annualized volatility estimated from historical data
• T = Time horizon (in years)
• Z = Random number drawn from a standard normal distribution

By generating thousands of random Z values, you create thousands of possible future price paths. The percentage of paths where price ends above your target = the probability of that target being reached.

A Concrete Example

Say BTC is at $70,000. Historical 30-day volatility is 65% annualized. You want to know: what's the probability BTC reaches $73,500 (5% up) within 3 days?

Run 1,500 simulated paths. If 980 of them end above $73,500, the estimated probability is 980/1500 = 65.3%.

Why Standard MC Isn't Enough

Basic Monte Carlo has several well-known problems in practice:

Problem 1: Noise

With 1,000 simulations, the estimated probability might be 65% one bar and 58% the next, even with minimal price change. This is sampling noise from finite simulations.

Solution: Variance Reduction Techniques

• Antithetic Variates — For each random Z, also simulate -Z. These paired paths are negatively correlated, reducing variance by ~50% with zero bias.
• Control Variates — Use the closed-form Black-Scholes probability as a "control variable." Adjust the MC estimate using the known analytical solution, reducing variance by 50-90% when the distribution is near log-normal.
• Stratified Sampling — Instead of random sampling across the entire distribution, divide it into strata and sample from each uniformly. Ensures coverage of the full probability space.

Problem 2: Rare Events

If the true probability of a crash is 3%, you'd need 10,000+ simulations to get enough crash samples for a stable estimate. With 1,000 simulations, you might get 20-40 crash samples — far too few for reliable estimation.

Solution: Importance Sampling

Instead of sampling from the true distribution, shift the sampling distribution toward the rare event (crash) region. Then reweight each sample by the likelihood ratio to correct for the sampling bias. Result: 10-100x more crash samples from the same number of simulations.

Problem 3: Bar-to-Bar Jitter

MC recalculates from scratch every bar, causing probability estimates to jump around even when the market hasn't materially changed.

Solution: Particle Filter (Bayesian State Estimation)

A Particle Filter maintains a swarm of probability "particles" that update incrementally as new price data arrives. Unlike MC which starts fresh each bar, the Particle Filter learns continuously, producing smoother and more responsive probability estimates.

From Probabilities to Trading Decisions

Raw probability numbers aren't useful in isolation. A robust decision framework requires multiple independent confirmations:

The 4-Factor Decision Engine

1. MC Meaningful Edge — The winning side must have probability > 33% (1.5:1 break-even minimum)
2. Particle Filter Agreement — PF must independently agree with MC direction at > 54%
3. Bias Edge — The gap between P(BULL) and P(BEAR) must be > 5 percentage points
4. Model Convergence — MC and Black-Scholes analytical solution must agree within tolerance. Large divergence = fat tails or insufficient simulations

Only when 3-4 factors align does a high-confidence signal fire. This eliminates the majority of false signals and noise.

Position Sizing with Kelly Criterion

Once you have reliable probability estimates, optimal position sizing follows naturally. The Kelly Criterion calculates the percentage of capital to allocate:

f* = (p × b - q) / b

• p = Estimated win probability
• q = 1 - p (loss probability)
• b = Odds ratio
• f* = Fraction of capital to bet

Important: Full Kelly is mathematically optimal but aggressively volatile. Practical implementation uses Fractional Kelly (10-20% of full Kelly) to balance growth with drawdown management.

Risk Management: Tail Risk and VaR

Monte Carlo doesn't just estimate directional probability. The same simulation produces a full P&L distribution for risk management:

• Value at Risk (VaR 95%) — The loss level that will not be exceeded 95% of the time
• Expected Shortfall (ES 95%) — The average loss in the worst 5% of scenarios
• Crash Probability — The specific probability of a catastrophic move (e.g., BTC -10%)

When crash probability crosses a market-specific threshold, all bullish signals should be suppressed regardless of how strong the directional edge appears. No edge is worth taking when the tail risk is too high.

Calibration: Are the Probabilities Accurate?

The most important metric for any probability-based system is calibration. The Brier Score measures forecast accuracy:

BS = (1/N) × Σ(forecast - outcome)²

• < 0.10 = Excellent calibration — forecasts are highly accurate
• 0.10-0.20 = Good calibration — forecasts are usable for trading
• > 0.20 = Poor calibration — parameters need adjustment

Before risking real capital, observe the Brier Score across 50+ predictions. A rising Brier Score means the market regime has changed and the model needs recalibration.

Practical Application: Prediction Markets

With platforms like Polymarket and Kalshi offering real binary contracts, Monte Carlo probability estimates become directly tradable:

1. Polymarket shows "BTC > $80K by Friday" at 58¢
2. Your MC model estimates P(BULL) at 72% with Brier Score 0.11
3. The implied edge = +14 percentage points
4. Kelly sizing tells you to allocate 8.2% of capital
5. Action: Buy the BULL contract at 58¢, expect 72% win rate

This is arbitrage between the market's implied probability and your model's estimated probability.

Limitations to Understand

• GBM assumes log-normal returns — Real markets have fat tails, especially crypto during flash crashes
• Historical volatility lags — When regimes change, the lookback window may not reflect current conditions
• Not for events — Earnings, FOMC decisions, and regulatory announcements create discontinuities that no MC model can predict
• Particle Filter resolution — 200 particles provide reasonable but not precise probability tracking