Monte Carlo on trade order and where it stops helping
Posted: Sat Oct 31, 2026 8:21 am
Shuffle the order of backtest trades a thousand times and the drawdown stops being one number. Same trades, different sequences. A system with a historical max drawdown of 60 pips might show a 95th percentile drawdown of 140 pips across shuffles. The real history is one draw from that spread, usually a kind one.
So size the account against the shuffled 95th percentile, not the historical figure. That alone prevents a lot of blown accounts.
The catch: plain shuffling assumes trades are independent, and scalping trades are not. Losses cluster in particular conditions, like a trending session that keeps stopping out a mean-reversion setup. Shuffling breaks those streaks apart and tends to understate the damage. Block-shuffling whole days, or runs of trades, keeps the clustering and gives a harsher and more believable estimate.
And one thing Monte Carlo cannot do is repair a bad test. If costs, fills or data are wrong, the shuffled curves are just many versions of the same wrong answer.
Run both versions and compare. A big gap between plain and block results is itself information about how streaky the system is.
Do you shuffle single trades or whole days, and how different did the drawdown estimates come out?
So size the account against the shuffled 95th percentile, not the historical figure. That alone prevents a lot of blown accounts.
The catch: plain shuffling assumes trades are independent, and scalping trades are not. Losses cluster in particular conditions, like a trending session that keeps stopping out a mean-reversion setup. Shuffling breaks those streaks apart and tends to understate the damage. Block-shuffling whole days, or runs of trades, keeps the clustering and gives a harsher and more believable estimate.
And one thing Monte Carlo cannot do is repair a bad test. If costs, fills or data are wrong, the shuffled curves are just many versions of the same wrong answer.
Run both versions and compare. A big gap between plain and block results is itself information about how streaky the system is.
Do you shuffle single trades or whole days, and how different did the drawdown estimates come out?