A backtest shows one path: the exact sequence of trades history happened to produce. Monte Carlo simulation asks what the neighboring paths look like, by reshuffling and resampling those trades thousands of times. The answer usually changes how much you trust the drawdown number.
Two strategies can share identical trade statistics and produce very different experiences, because outcomes arrive in different orders. A backtest's maximum drawdown is the drawdown of one particular ordering. If the same trades had arrived in a less friendly sequence, the drawdown could have been far deeper, and sequence is substantially luck. Trusting a single curve means trusting that luck.
Take the backtest's individual trade results. Shuffle their order thousands of times, or resample them with replacement, and rebuild the equity curve each time. The output is a distribution instead of a number: across ten thousand alternative orderings, the median maximum drawdown, the 95th percentile drawdown, the spread of final returns, and the probability of hitting a fixed loss threshold along the way. The original backtest turns out to be one draw from this distribution, and frequently a lucky one.
Three readings matter. Tail drawdown: if the 95th percentile drawdown is double the backtest's headline number, size positions for the tail, not the headline, or the first bad sequence will force quitting at the worst moment. Dispersion: if final results vary enormously across orderings, the strategy's outcome is sequence-dominated and the sample of trades is too thin to trust, which connects directly to sample size. Ruin probability: the fraction of paths that breach your real risk limit is a more honest risk statement than any single curve.
Reshuffling trades assumes the trades themselves are the truth. If the backtest is overfitted, every simulated path inherits the same fiction, so Monte Carlo cannot rescue a dishonest backtest; it can only stress an honest one. Simple reshuffling also breaks the natural clustering of markets, where losses arrive in streaks because volatility clusters, so block-based resampling that preserves streaks is the more conservative variant. And nothing in the method predicts a regime the history never contained.
Monte Carlo belongs after the strategy has passed the honesty checks: costs included, parameters stable, out-of-sample verdict positive. At that point it converts one equity curve into a risk distribution, and the sizing decision should be made against that distribution. Sonar Sciences's performance framework treats drawdown expectations this way: the question is never what the curve did, but what the curve family implies.
Monte Carlo simulation reshuffles or resamples a backtest's trades thousands of times to build the distribution of outcomes the same trade population could have produced in different orders. It turns single numbers like maximum drawdown into percentile ranges, exposing how much of the original result was sequence luck.
If the 95th percentile maximum drawdown is 30 percent, then in 5 percent of simulated orderings the drawdown was worse than 30 percent. Sizing against a high percentile rather than the backtest's single observed drawdown protects against ordinary bad luck in sequencing, which no strategy avoids forever.
No. It stresses the sequence of trades, not their validity. An overfitted strategy produces fictional trades, and every reshuffle of fiction is still fiction. Overfitting is caught by parameter-stability checks, out-of-sample testing, and walk-forward analysis; Monte Carlo then quantifies risk for strategies that survived those.
Bring one strategy you already trust. The Studio validates it against four years of real data, costs included, for free.
Start building