Every equity curve has a worst stretch, and max drawdown is the number that describes it: the largest peak-to-trough decline in account equity, expressed as a percentage of the peak. Unlike a Sharpe ratio it needs no assumptions about a risk-free rate, no annualization and no distribution — just the curve you already have.
That simplicity is exactly why it is the metric to gate on. This article covers the formula, the difference between floating and realized drawdown, why the number belongs to one path rather than a distribution, why backtest drawdown systematically understates the live figure, and what a threshold worth enforcing looks like.
The definition and the formula
At each moment the equity is compared with the highest equity reached so far. The drawdown at time t is the decline from that running peak; the maximum drawdown is the largest of those declines over the whole test.
DD_t = (Peak_t - Equity_t) / Peak_t
MaxDD = max over t of DD_t
Both terms are percentages of the peak, which is the part people get wrong: a fall from 145 to 92 is not 53% and not 36.6% — it is 63 / 145 = 43.4% of the money you had at the top.
A worked number
Take a simple curve: start at 100, run up to 145, fall to 92, recover to 130.
| Point | Equity | Running peak | Drawdown from peak |
|---|---|---|---|
| Start | 100 | 100 | 0% |
| High | 145 | 145 | 0% |
| Trough | 92 | 145 | 43.4% |
| Partial recovery | 130 | 145 | 10.3% |
Max drawdown is 43.4%. Note the second number hiding in the same table: climbing from 92 back to 145 requires +57.6%, not 43.4%. The asymmetry between the fall and the climb is the entire reason this metric gets its own gate.
Mark-to-market versus closed-trade drawdown
Two conventions produce two different figures from the same trading, and the difference is usually in the trader's favour when the wrong one is used.
Closed-trade drawdown accounts only for realized results: it resets after each position is flat, so a trade that was 40% under water and then recovered contributes nothing. Mark-to-market drawdown marks the account at every moment, including while positions are open, and therefore contains floating losses.
Mark-to-market is the honest figure, because the account is never actually at the level the closed-trade version reports. The broker does not suspend the loss while you decide. A system whose closed-trade drawdown is 12% and whose floating drawdown is 35% has a 35% problem, and the margin call arrives based on the second number. Closed-trade accounting is legitimate only for reporting per-trade risk metrics such as MAE on a flat account, never as the account-level gate.
One path, not a distribution
Sharpe, Sortino and win rate are properties of a distribution of returns — they survive a reshuffle of the same trades. Max drawdown does not. It is path-dependent: reorder the same three trades and the metric changes, sometimes by tens of percentage points.
| Order of trades | Equity path | Final equity | Max drawdown |
|---|---|---|---|
| +100%, then −30%, then −30% | 100 → 200 → 140 → 98 | 98 | 51% |
| −30%, then +100%, then −30% | 100 → 70 → 140 → 98 | 98 | 30% |
Identical trades, identical final equity, 21 percentage points of difference in the reported risk. Three consequences follow.
- The metric is a single realization. Your backtest shows the drawdown of one sequence of trades on one stretch of history. It is the observed value, not the expected value, and not a worst case.
- Sample length decides the number more than the strategy does. Two years of data cannot contain a five-year drawdown; a longer test finds a larger one almost mechanically.
- A small edge does not shrink the tail proportionally. The path that produced the worst historical drawdown is not the path you will trade next, but it is also not irrelevant — it is a sample from the same generator.
Why survival depends on this metric
Return metrics describe how fast the account grows; drawdown describes whether it survives the growth. The arithmetic is not symmetric, and no amount of average return offsets it:
| Loss | Required gain to recover |
|---|---|
| 10% | +11.1% |
| 20% | +25% |
| 33% | +49.3% |
| 50% | +100% |
| 75% | +300% |
A 50% drawdown is not a setback; it is a doubling of the required edge. Past roughly 60–70% the account is in a region where recovery needs a run that most systems never produced in the first place, which is why ruin is a drawdown concept before it is a return concept. Leverage makes the same point mechanically: at 2× position sizing, a strategy's own 20% peak-to-trough move becomes a 40% account move, and the recovery requirement jumps from +25% to +67%.
This is also where drawdown differs from volatility. Volatility measures how much the equity wiggles, symmetrically, and a strategy with violent upside months looks risky to it. Drawdown measures only the sequence that hurts — a Sharpe ratio can be respectable on a curve that is one regime change away from a hole, because the volatility term only knows the sample it was fed.
Sharpe, drawdown and Calmar
The metrics answer different questions, and none of them substitutes for the others.
| Metric | Question it answers | Blind spot |
|---|---|---|
| Sharpe | How much return per unit of volatility? | Says nothing about the worst stretch or about trial count |
| Sortino | Same, counting only downside moves | Still distributional, still sample-bounded |
| Max drawdown | What was the worst hole, and how long to climb out? | One path; no information about frequency |
| Calmar (CAGR / MaxDD) | How much growth per unit of worst loss? | Inherits every property of both inputs |
The same-Sharpe comparison is the useful one. Two strategies can both post a Sharpe near 1.1 on the same years: one grinds through an 8–12% drawdown, the other spends most of its time flat and then loses 45% in a stretch the sample never contained. The Sharpe cannot see the second hole, because the event that digs it is not in the returns it was computed from. That is the honest reading of "same Sharpe, different tail": the ratio is bounded by its sample, while the tail is a property of the process. Calmar makes the difference visible in a single number — 12% growth over an 18% drawdown is 0.67; the same 12% growth over a 45% drawdown is 0.27.
Why the backtest figure is a lower bound
The maximum drawdown in a backtest is the largest value observed in the sample. As a statement about the future it is biased low, for a reason that has nothing to do with the quality of your work:
- Order statistics grow with the sample. The expected maximum of more draws is larger. Trade the same rules for ten years instead of two and the worst stretch found will almost certainly be worse, even if nothing about the strategy changed.
- Regimes are missing. A test covering a calm bull period has no mechanism for producing the drawdown that a rate shock, a liquidity gap or a correlated de-risking event would create. Absence of evidence inside the sample is not evidence about the process.
- The live account runs forward into unseen data. The gap between the backtest number and the first serious live drawdown is measured, not hypothetical — it is the same gap discussed in backtest versus live trading and in out-of-sample testing.
The practical translation: treat the backtest's max drawdown as the floor of the plausible range, not its centre. A number that grows each time you extend the test is telling you it has not converged yet — and a search that kept the best curve among many compounds the problem, since the retained variant is precisely the one with the luckiest path: how many backtests is too many.
What a sane gate looks like
A gate is a statement with two halves: a level and the evidence behind it. In practice it reads like "max drawdown must stay at or below 25% of capital, on at least five years of data that include a bear regime, with commission and slippage inside the run." Each clause does work:
- The level is set by capital, not by the strategy — it is the most you can lose before the position sizing, the funding plan and your own behaviour break.
- The data requirement stops a two-year bull sample from certifying a level it cannot possibly have measured.
- The cost clause prevents the most common inflation, since a system that trades often can show a shallow backtest drawdown only if friction is absent — the items listed in backtest costs.
- The account-level clause requires the gate to be applied to equity including open positions, at the position size you will actually run.
Common errors
- Strategy drawdown instead of account drawdown. Reporting the system's trades while deposits, withdrawals or idle cash flatten the curve. Size the whole account or size nothing.
- Ignoring leverage and position sizing. The strategy's percentage and the account's percentage are different numbers whenever sizing is not 100% of one instrument.
- One lucky year. A single short sample can produce an immaculate 6% drawdown that has no relationship to the strategy's behaviour outside that window.
- Counting only closed trades. Floating losses are still losses; see the mark-to-market section above.
- Judging it alone. A tiny drawdown achieved by a strategy that earns nothing is not risk control. The gate only has meaning next to a return metric and a trial count.
How a judging report uses it
In a report the drawdown is read against the gate, not against intuition, and never in isolation. As an interface example — figures the interface is designed to show, not results from a live account — a run reporting max drawdown 18% against a 25% gate, with the Sharpe 0.94 versus a required 1.47 at 11 effective trials produces a split verdict: the risk gate passes, the return bar fails. Neither number overturns the other; each constrains what the other can claim. That pairing is the whole design: drawdown answers "can the account survive this," the threshold answers "is the edge likely to be real," and only both together justify running the rules forward.
How to check this yourself
- Compute drawdown on mark-to-market equity, at the position size you intend to trade.
- Confirm the figure includes commission and slippage, and that fills do not happen on the signal bar itself.
- Check what the sample contains. If it has no losing regime, the number measures the window, not the strategy.
- Extend the test and see whether the max drawdown grows. Growth means it has not converged.
- Compare the number with your capital limit first, then with the Sharpe and its threshold.
- Require the variant to survive out-of-sample data before treating the drawdown as a property of the rules.
To see the metric move in real time, open the Replay simulator, run a saved graph across a stretch of history with a deep decline, and read the drawdown next to the gate rather than next to the return.