Why Maximum Drawdown Is a Noisy Statistic
Maximum drawdown feels like a hard, objective fact about a strategy's worst moment. It's actually an extreme-value statistic computed from one path out of many the same strategy could have produced, and it swings enormously between equally valid histories.
Prerequisites: Max Drawdown, The Standard Error of the Sharpe Ratio
Two strategies with identical annualized return of 12% and identical volatility of 15% are backtested over the same five-year window on the same underlying signal, differing only in the random order the market happened to deliver good and bad days. Strategy X's backtest shows a maximum drawdown of 18%. Strategy Y — mathematically the same edge, same risk, just a different draw of the same return distribution — shows a maximum drawdown of 34%. A committee reviewing tear sheets side by side would very likely fund X and shelve Y, treating 18% versus 34% as a real difference in risk management. It is not; it's a difference in which five years of noise landed on top of an identical process.
Why drawdown is so much noisier than it looks
A mean return and a Sharpe ratio are averages — the central limit theorem works in your favor, and more days make the estimate more stable. Maximum drawdown is an extreme-value statistic — the worst point along one particular path — and extreme values don't average out the same way; they depend heavily on the specific sequencing of good and bad periods, exactly the part of the data a backtest cannot resample or repeat. Run the identical strategy through 1,000 different orderings of the same daily returns (a block bootstrap preserving autocorrelation) and you get 1,000 different maximum drawdowns from one single edge and risk profile.
Worked example: bootstrapping the drawdown distribution
Take a strategy's 1,260 daily returns (5 years), block-bootstrap them 2,000 times in 20-day blocks (long enough to preserve short-term autocorrelation), and recompute maximum drawdown on each resampled path. Suppose the resulting distribution has:
| Percentile | Max drawdown |
|---|---|
| 10th | 14% |
| 25th | 19% |
| 50th (median) | 24% |
| 75th | 31% |
| 90th | 41% |
The single historical backtest happened to land at 18% — near the 15th percentile of what the strategy could plausibly have produced. Reporting "18% max drawdown" rather than "median 24%, with a real chance of 40%+" understates the tail risk a live allocator will experience, since the live path is just one more draw from the same distribution.
Worked example: two backtests, one strategy
Strategies X and Y are re-examined by concatenating both five-year windows into ten years and block-bootstrapping that combined series. The resulting distribution has a median of 27% and a 90th percentile of 46% — both individual backtests (18% and 34%) fall well within it, at roughly the 12th and 61st percentiles. Neither was special; both were ordinary draws from the same risk, and the 16-point gap that looked decisive in committee is smaller than the routine spread the bootstrap shows for one unchanged strategy.
The skew visible when you switch the parent distribution above to a heavy-tailed shape is the same mechanism: extreme-value statistics inherit skew far more than averages do, which is why drawdown swings so much more than Sharpe across otherwise-identical resamples.
Maximum drawdown from a single backtest is one draw from a wide distribution of drawdowns the same strategy could have produced. Report a bootstrapped range — median and a tail percentile — rather than the single historical number.
The classic confusion: comparing two realized drawdowns as if the smaller one proves better risk control. Without a bootstrap range around each number, you cannot tell a real difference in risk management from two ordinary draws off the same distribution — as the pooled example shows, the "worse" strategy is sometimes just the one that drew the harder five years.
What this means in practice
Never report a bare maximum-drawdown number without a resampled distribution around it. Block-bootstrap the return series, recompute drawdown on each resample, and report a median and a tail percentile alongside the historical figure. Use that width, not the single number, to set drawdown-based risk limits — see Drawdown Control.
Related concepts
Practice in interviews
Further reading
- Magdon-Ismail & Atiya, Maximum Drawdown
- Bailey & López de Prado, Drawdown-Based Stop-Outs are Ruinous