Size the Distribution

Your backtest hands you a single number for max drawdown, and it's the most reassuring lie in the whole report.

That number is one path. One ordering of your trades, the exact sequence the market happened to serve up in your sample. Shuffle the same trades into a different order and the drawdown moves, sometimes a lot. The market has no obligation to deal them in the order your backtest got.

What the single number hides

Take a strategy with a clean equity curve and a worst historical drawdown of 8%. Feels survivable. You size the book so an 8% drawdown is comfortable and you move on.

Now resample. Take the exact trade results, same win rate, same average win and loss, and reshuffle the sequence a few thousand times. Each shuffle is a plausible alternate history the strategy could just as easily have lived. Plot the drawdown from each one and you get a distribution. The median might sit near your 8%. The 95th percentile path will sit deeper, often meaningfully deeper. That deeper number is the one you actually have to survive, because there was nothing special about the ordering you happened to backtest.

The historical max drawdown is a floor. Live trading gets one more draw from the same distribution, and it has never once promised to hand you the good sequence.

Depth and duration

Depth is only half of it. A drawdown also has a duration: how long you sit underwater before the curve makes a new high. I measure drawdowns on both axes, depth and time, because the one that breaks a trader is rarely the deep-and-fast one. It's the shallow-and-endless one, the 6% that takes nine months to recover, the stretch long enough that you start overriding the system out of sheer boredom and doubt.

On my own book, XAQP, the worst historical drawdown sits at -5.83% (as of July 2026). I don't size to that number. I size to the tail of the resampled distribution, and I stress the duration alongside the depth, because the number that matters is the one I haven't lived yet.

How to actually do it

You don't need anything exotic. Two passes.

Resample the trade sequence with replacement. Draw trades at random, with repeats allowed, build a few thousand synthetic equity curves the same length as your real one, and read the drawdown distribution off them. With-replacement matters: it lets clusters form that your single history didn't happen to show, and clusters are where the real pain lives.

Then jitter the parameters. Nudge each input a little and re-run. If a small nudge collapses the curve, the drawdown distribution you just built was measured on a knife-edge, and the real risk is worse than any resample suggested.

Size the strategy so the 95th percentile path, on both depth and duration, is one you can hold through without touching it. If that sizing makes the returns look boring, the returns were never as good as the single-path backtest flattered them to be.

Why this changes the sizing

The real shift is in the sizing itself. The position size falls out of the distribution, and the single-path anecdote never had the authority to set it.

A strategy sized to its median drawdown is sized to a coin flip. A strategy sized to its 95th percentile drawdown is sized to survive the sequence it will eventually be handed. I'd rather run at two-thirds the size and know it survives the tail than run at full size on the comfort of one lucky ordering. The first one is still in the book in three years. The second gets cut at the worst possible moment, right when the drawdown that was always in the distribution finally shows up live.

Work in progress, always. But the sizing question has one true answer, and it lives in the distribution.

Personal commentary, not advice. Past performance is not indicative of future results. Capital at risk.