Skip to main content
aifinhub
Backtesting & Validation Explainer

Monte Carlo Simulation

Monte Carlo simulation samples N paths from a probabilistic model (Brownian motion, GARCH, jump-diffusion, bootstrap-from-history), evaluates the quantity of interest on each path, and reports the empirical distribution of outcomes. Convergence rate is O(1/sqrt(N)) — quadrupling N halves the error.

By Orbyd Editorial · AI Fin Hub Team

On This Page

Definition

Monte Carlo simulation

Monte Carlo simulation samples N paths from a probabilistic model (Brownian motion, GARCH, jump-diffusion, bootstrap-from-history), evaluates the quantity of interest on each path, and reports the empirical distribution of outcomes. Convergence rate is O(1/sqrt(N)) — quadrupling N halves the error.

Why it matters

Most useful path-dependent quantities — drawdown distribution, time-to-ruin, options PnL with early-exercise — have no closed-form solution. Monte Carlo gets you a numerical answer in exchange for compute. The trade-off: the answer is only as good as the underlying model. A Gaussian Monte Carlo of crypto returns will undershoot tail risk badly.

How it works

Specify the stochastic process (parameters, distribution). Pick a seed for reproducibility. Generate N paths. Compute the path-dependent statistic on each. Report mean, median, quantiles, and the empirical distribution. Sensitivity-test by re-running with N/2 to verify convergence. Always report N, the seed, and the model assumption — without those the result isn't reproducible.

Example

Kelly-sized strategy, simulate 10,000 1-year paths

Paths

10,000

Median final equity

1.18x

5th percentile final equity

0.62x

95th percentile final equity

2.1x

Probability of >50% drawdown

8%

The mean return looks fine. The 5th percentile and the drawdown probability tell you the strategy has real ruin risk that a single backtest path would have hidden.

Key Takeaways

1

Always seed your RNG and report the seed.

2

Convergence is sqrt(N) — be honest about how many paths you need for the precision you're claiming.

3

Garbage model in, garbage distribution out: Monte Carlo doesn't compensate for a wrong stochastic process.

Try These Tools

Run the numbers next

FAQ

Questions people ask next

The short answers readers usually want after the first pass.

For median estimation, 10k is usually fine. For 99th-percentile tail estimation, 100k+. The standard error scales as 1/sqrt(N), so doubling precision needs 4x the paths. Always run a convergence check at N/2 vs N.

Sources & References

Related Content

Keep the topic connected

Planning estimates only — not financial, tax, or investment advice.