Term · MDD

Maximum Drawdown

Published May 9, 2026

Definition

Maximum drawdown is the largest cumulative loss from a previous peak to a subsequent trough during a measurement period, measured in percent of the peak. It captures the worst path-dependent loss an investor could have experienced if they bought at the wrong time.

Drawdown is a path-dependent statistic — its value depends not just on the distribution of returns but on the order in which they're realized. Two return series with identical mean and variance can have very different maximum drawdowns: a sequence of alternating +5% and -5% has small drawdowns, while the same +5% and -5% values clustered together (five -5% in a row, then five +5%) has a 22.6% drawdown. This path-dependence is precisely why drawdown is the single most diagnostic statistic for evaluating returns-generation realism.

For a real US equity portfolio over 1928–2025, the maximum drawdown was approximately -86% (1929–1932). The S&P 500 had a -55% drawdown in 2008–2009 and a -34% drawdown in early 2020. A regime-switching synthetic returns generator should produce a corpus whose 30-year-horizon drawdown distribution has these tail events represented; a random-walk IID-normal generator produces drawdowns with maximum values around -45% to -55% over 30 years, dramatically understating the worst-case experience.

Drawdown is also the metric that retirement-planning users care about most viscerally. A '95% probability of success' from a Monte Carlo simulator means little; '5% chance of a 60%+ loss in any 5-year window' is concrete. Engines that report only success probability and not drawdown distribution are hiding the path-shape of the failure cases.

Formula

Maximum Drawdown

MDD = max_t [(peak_t − value_t) / peak_t]

peak_t: = the running peak value through time t
value_t: = the portfolio value at time t

Example

Running peak 110, value drops to 88. Drawdown = (110 - 88) / 110 = 20%. The maximum across all t is the MDD.

Why this matters for synthetic data

Maximum-drawdown distributions are the second-strongest diagnostic (after kurtosis) for returns-generation realism. The empirical distribution of 30-year-horizon maximum drawdowns from a 60/40 US portfolio has a median around -32%, a 90th percentile around -55%, and a tail extending to -75%. A synthetic corpus whose 30-year drawdown distribution has a median below -30% and a 90th percentile above -55% is reproducing the right tail behavior; a corpus tightly clustered around -25% is using a generator that smooths tail events.

Common pitfalls

Computing max drawdown from monthly data and reporting it alongside other engines that use daily data — the daily drawdown is always larger because intra-month troughs are visible.
Confusing 'recovery time' with 'drawdown duration' — recovery is from trough back to prior peak; drawdown duration is peak-to-trough.
Reporting maximum drawdown as a single number without the conditional distribution — '15% drawdown' over 5 years is different from '15% drawdown' over 30 years.
Using drawdown to compare strategies of different volatility — a low-vol strategy will always have smaller drawdowns; the meaningful comparison is risk-adjusted (e.g. Calmar ratio = annualized return / max drawdown).

Examples

Drawdown computation

Portfolio values over 8 months: 100, 110, 105, 95, 88, 92, 98, 108. Running peak: 100, 110, 110, 110, 110, 110, 110, 110. Drawdowns from peak: 0%, 0%, -4.5%, -13.6%, -20%, -16.4%, -10.9%, -1.8%. Maximum drawdown: 20% (peak at month 2 = 110, trough at month 5 = 88).

Frequently asked questions

Is max drawdown a better risk metric than standard deviation?+

For investor-utility purposes, often yes — investors experience drawdowns viscerally and standard deviation only abstractly. For statistical-modeling purposes, standard deviation is more tractable (closed-form, time-additive, parametric) while drawdown is path-dependent and harder to model. Most institutional platforms report both: standard deviation for risk-budgeting and Sharpe-ratio computation; max drawdown for client communication and tail-risk reporting.

How does drawdown interact with the Calmar ratio?+

Calmar = annualized return / max drawdown over the same period. It's a return-per-unit-of-tail-risk measure, useful for comparing strategies whose risk profile is concentrated in the tail rather than spread across the distribution. Hedge funds report Calmar more than Sharpe for this reason.

Why do random-walk simulators understate drawdown?+

Two reasons. First, random walks have no volatility clustering — bad months don't follow bad months — so the path-clustering that creates large drawdowns is structurally absent. Second, IID-normal returns have thin tails, so the magnitude of any single bad month is bounded. The combined effect is that random-walk 30-year drawdown distributions have medians around -22% and 90th percentiles around -38%, both substantially smaller than the empirical historical distribution.