On the Distribution of Maximum Drawdown

1. MDD as a Random Variable

Ask a trader what their maximum drawdown is and they will give you a single number: "15%." Ask them what drawdown they should expect over the next 5 years and you will get a blank stare or a repetition of the same number. This reflects a fundamental misunderstanding: the observed MDD is a single realisation of a random variable, not a fixed property of the strategy. Run the same strategy over a different 5-year period — or even the same period with slightly different execution timing — and the MDD will be different.

The distribution of maximum drawdown is right-skewed and heavy-tailed. This means that the observed MDD from any finite sample systematically underestimates the expected MDD over longer horizons. A trader who calibrates their risk tolerance to an observed 15% drawdown is almost certainly underestimating the drawdown they will eventually experience.

2. Analytical Framework

Magdon-Ismail and Atiya (2004) derived the expected value and approximate distribution of MDD for a Brownian motion with drift. For a process with annualised Sharpe ratio S and volatility σ, the expected MDD over T years is:

where Q(·) is a function they derive from the distribution of the maximum of a Brownian bridge. For practical computation, a useful approximation (accurate to within 5% for S·√T > 0.5) is:

The critical insight is the √T scaling: expected MDD grows with the square root of the evaluation period. This means a strategy with a 10% expected MDD over 1 year has an expected MDD of approximately 14.1% over 2 years, 17.3% over 3 years, and 22.4% over 5 years — even if nothing about the strategy changes. Traders who compare Calmar ratios (annualised return / MDD) across strategies with different track record lengths without adjusting for this time scaling are making systematically biased comparisons.

3. Monte Carlo Estimation

The analytical formula assumes Gaussian returns, which understates MDD for strategies with fat tails, serial correlation, or time-varying volatility. A more robust approach is Monte Carlo simulation using the empirical return distribution:

import numpy as np

class MDDDistribution:
    """
    Monte Carlo estimation of maximum drawdown distribution.
    Supports both parametric and empirical resampling.
    """
    def __init__(self, returns, n_simulations=50000):
        self.returns = returns
        self.n_sim = n_simulations
        self.T = len(returns)
        self._mdd_samples = None

    def _compute_mdd(self, equity_curve):
        """Max drawdown from a cumulative return series."""
        running_max = np.maximum.accumulate(equity_curve)
        drawdowns = (running_max - equity_curve) / running_max
        return np.max(drawdowns)

    def simulate_empirical(self):
        """Bootstrap from empirical return distribution."""
        mdds = np.zeros(self.n_sim)
        for i in range(self.n_sim):
            # Resample with replacement (IID assumption)
            sim_returns = np.random.choice(self.returns,
                                           size=self.T, replace=True)
            equity = np.cumprod(1 + sim_returns)
            mdds[i] = self._compute_mdd(equity)
        self._mdd_samples = mdds
        return mdds

    def simulate_block_bootstrap(self, block_length=21):
        """Block bootstrap to preserve serial dependence."""
        mdds = np.zeros(self.n_sim)
        n_blocks = self.T // block_length + 1
        for i in range(self.n_sim):
            blocks = [self.returns[j:j+block_length]
                     for j in np.random.randint(0, self.T-block_length,
                                                 size=n_blocks)]
            sim_returns = np.concatenate(blocks)[:self.T]
            equity = np.cumprod(1 + sim_returns)
            mdds[i] = self._compute_mdd(equity)
        self._mdd_samples = mdds
        return mdds

    def confidence_interval(self, alpha=0.05):
        """Return (lower, median, upper) CI for MDD."""
        if self._mdd_samples is None:
            self.simulate_empirical()
        s = np.sort(self._mdd_samples)
        lower = s[int(alpha/2 * self.n_sim)]
        median = s[int(0.5 * self.n_sim)]
        upper = s[int((1-alpha/2) * self.n_sim)]
        return lower, median, upper

    def prob_exceeds(self, threshold):
        """P(MDD > threshold) from simulation."""
        if self._mdd_samples is None:
            self.simulate_empirical()
        return np.mean(self._mdd_samples > threshold)

4. Analytical vs. Monte Carlo Comparison

We compare the analytical formula to Monte Carlo estimates using four representative strategy profiles: a high-Sharpe low-vol strategy (S=1.5, σ=8%), a moderate strategy (S=0.8, σ=12%), a volatile strategy (S=0.5, σ=20%), and an aggressive strategy (S=0.3, σ=25%). For each, We compute the expected MDD and 95% CI over 3 years using both methods.

Table 1: Expected MDD and 95% CI over 3 years. Monte Carlo uses the block bootstrap with Student-t innovations (ν=4) and block length 21 days. The analytical formula underestimates E[MDD] by 8–17% relative to MC because it assumes Gaussian returns.

The analytical formula consistently underestimates MDD compared to the Monte Carlo approach, by 8% for the high-Sharpe strategy to 17% for the aggressive strategy. The gap widens for lower-Sharpe strategies because fat-tailed returns have a disproportionate impact when the drift is small relative to the volatility — the extreme negative returns dominate the drawdown experience.

The confidence intervals are strikingly wide. For the moderate strategy (the most common profile among institutional systematic traders), the 95% CI spans [8.7%, 27.3%]. A trader who observes a 15% MDD over 3 years and concludes their strategy's "true" MDD is 15% is ignoring the fact that the same strategy could have produced a 27% MDD with 2.5% probability — a drawdown severe enough to trigger most institutional risk limits.

5. The Time-Scaling Trap in Detail

To illustrate the time-scaling problem, We compute expected MDD at multiple horizons for the moderate strategy (S=0.8, σ=12%):

Table 2: Expected MDD growth with horizon for a moderate strategy. The 95% upper bound at 10 years (42.8%) is nearly 3× the 1-year estimate (18.1%).

At the 20-year horizon, even this moderate strategy has a 95% upper bound MDD of 54%. This is not a deficiency of the strategy — it is a mathematical property of stochastic processes. Any strategy that will eventually be traded for decades should plan for drawdowns substantially larger than anything observed in a 3-5 year backtest.

6. Fat Tail Adjustment Factor

The ratio of Monte Carlo E[MDD] to analytical E[MDD] is a measure of how much fat tails amplify the drawdown experience relative to the Gaussian baseline. We call this the fat tail adjustment factor (FTAF) and compute it for varying degrees of freedom ν in a Student-t return distribution:

Table 3: Fat tail adjustment factor for different tail heaviness. *Kurtosis is undefined for ν ≤ 4 but finite moments exist for ν > 2. FTAF is computed for Sharpe = 0.8 over 3 years.

For equity futures (ν ≈ 4), the analytical formula understates expected MDD by approximately 22%. For commodity futures (ν ≈ 3), the understatement is 38%. This is the price of analytical tractability: the Gaussian assumption makes the mathematics clean but produces materially wrong risk estimates for the instruments that most systematic traders actually trade. The practical recommendation: multiply the analytical E[MDD] by the appropriate FTAF from Table 3 to get a more realistic estimate, or use the Monte Carlo approach directly.

7. Implications for Risk Budgeting

These results have direct implications for setting risk limits. A fund manager who sets a 20% maximum drawdown limit based on a 3-year backtest showing 12% MDD is implicitly assuming that the backtest MDD is close to the expected MDD. our analysis shows that for a typical strategy profile, the expected MDD over the next 3 years is approximately 30% higher than the observed 3-year MDD, and the 95% upper bound is approximately 2× the observed value. A 20% limit on a strategy with observed 12% MDD provides a buffer of only 8 percentage points — which our analysis suggests is insufficient at the 95% confidence level.

We recommend setting drawdown limits at 1.5–2.0× the expected MDD (computed analytically or via Monte Carlo) rather than at a multiple of the observed MDD. For a strategy with E[MDD] = 16%, a limit of 24–32% provides adequate room for normal drawdown variation while still flagging genuinely anomalous behaviour.

8. Drawdown Duration Distribution

A related quantity is the maximum drawdown duration — the longest period between equity highs. This metric is often more psychologically relevant than depth: many traders can tolerate a 20% drawdown that recovers in a month but cannot endure a flat equity curve that lasts for a year. The distribution of maximum drawdown duration is also right-skewed and grows with √T. For the moderate strategy, the expected maximum duration is approximately 85 trading days over 3 years, with a 95% upper bound of 210 days — nearly 10 months without a new high. Strategy developers should report expected drawdown duration alongside depth to provide a complete picture of the drawdown experience.

9. Conclusion

Maximum drawdown is not a constant — it is a random variable with a right-skewed, heavy-tailed distribution that grows with √T. The observed MDD from a backtest is a single draw from this distribution and systematically underestimates the expected future MDD. For a typical institutional systematic strategy, the 95% CI for the true expected MDD is approximately [0.6×, 1.8×] the observed value, with fat tails widening this interval further. Traders and allocators should use Monte Carlo simulation (or the analytical formula with a fat-tail adjustment) rather than the observed MDD as the basis for risk budgeting, drawdown limits, and cross-strategy comparison.