Monte Carlo Reshuffling: Seeing the Drawdowns That History Skipped
Posted: Tue Oct 27, 2026 3:56 am
A backtest shows exactly one ordering of trades, yet that sequence is only one of many that the same edge could have produced. Reshuffling the trade list thousands of times creates alternative equity paths with identical average results but very different drawdowns.
That matters because maximum drawdown depends heavily on streaks. A run of losers that happened to fall in a quiet part of the sample may arrive at the worst moment in live trading. The reshuffled distribution gives a range of plausible drawdowns rather than a single flattering number.
Variants exist. Simple shuffling assumes trades are independent, which can understate risk if losses cluster during volatile sessions. Resampling with replacement, or shuffling in blocks to preserve streak structure, can be more cautious. Randomly skipping a percentage of trades also tests sensitivity to missed fills.
Results are most useful for sizing decisions. If a drawdown at a high percentile exceeds what an account or a trader's temperament can absorb, the position size is too large for the edge, regardless of how the original curve looked. Monte Carlo cannot create information missing from the sample, so a small or biased trade list produces equally unreliable simulations.
That matters because maximum drawdown depends heavily on streaks. A run of losers that happened to fall in a quiet part of the sample may arrive at the worst moment in live trading. The reshuffled distribution gives a range of plausible drawdowns rather than a single flattering number.
Variants exist. Simple shuffling assumes trades are independent, which can understate risk if losses cluster during volatile sessions. Resampling with replacement, or shuffling in blocks to preserve streak structure, can be more cautious. Randomly skipping a percentage of trades also tests sensitivity to missed fills.
Results are most useful for sizing decisions. If a drawdown at a high percentile exceeds what an account or a trader's temperament can absorb, the position size is too large for the edge, regardless of how the original curve looked. Monte Carlo cannot create information missing from the sample, so a small or biased trade list produces equally unreliable simulations.