Expected maximum drawdown: why to plan for worse than your backtest

Contents

1. What maximum drawdown should you expect in live trading compared with your backtest?
2. How is maximum drawdown calculated, and why is the backtest number only one sample?
3. Why does the expected maximum drawdown grow the longer you trade?
4. How do you estimate max drawdown with Monte Carlo?
5. How many times the backtest drawdown should you plan for?
6. How large a drawdown should stop a strategy?
7. Frequently asked questions
8. References

What maximum drawdown should you expect in live trading compared with your backtest?

Your expected maximum drawdown in live trading is larger than the backtest figure: plan for about 1.5 to 2 times it, and check that against a Monte Carlo 95th percentile and the margin your account must keep. The backtest drawdown is one sample from one order of trades over one period; live trading adds time and a new order.

Think of the backtest as one roll of a set of dice. The trades are the dice; the order they landed in is the roll. A second roll of the same dice gives a different maximum drawdown, and the longer you keep rolling, the deeper the worst stretch becomes. The quickest way to see this with your own numbers is to paste a trade list into the free Monte Carlo simulator and compare the spread of simulated drawdowns with the single backtest value.

The claim this article defends is simple: the backtest maximum drawdown is close to the best case you should assume, not the typical case. The sections below show why, how to estimate the typical and the bad case, and how to turn those numbers into a planning size and a stop rule written before the first live trade.

How is maximum drawdown calculated, and why is the backtest number only one sample?

Maximum drawdown is the largest fall in account equity from a peak to the lowest point that follows it, before equity makes a new peak. It is measured in money or in percent of the account at the peak. The free drawdown calculator computes it from an equity series and also shows how long the recovery took. In words, the one formula is:

maximum drawdown = the largest value of (highest equity so far − current equity)

Here highest equity so far is the running peak, the best the account has ever been up to that bar; current equity is the account value on that bar; the difference is the drawdown at that moment; and the maximum drawdown is the largest such difference anywhere in the record. The drawdown period ends on the bar where equity rises above the old peak, which is the recovery point.

The backtest number is one sample because the maximum drawdown depends on the order of the trades, and the backtest shows only one order. The same wins and losses arranged differently produce a different deepest fall, even though the net profit is the same. Losses that landed months apart in the backtest can land next to each other live.

Why does the expected maximum drawdown grow the longer you trade?

The expected maximum drawdown rises with the length of the trading record, so a strategy traded for ten years should be expected to show a deeper worst stretch than the same strategy traded for three, with no change in its edge. The reason is the same as for losing streaks: the longer you flip a fair coin, the longer the longest run of tails you will have seen. A longer record simply gives bad orderings more chances to appear.

Magdon-Ismail and Atiya (2004) worked this out for a price path with a steady average gain and random noise around it. Their result: the expected maximum drawdown keeps rising as the record lengthens, quickly at first and then more slowly once the average gain starts to dominate the noise. A strategy with no edge shows a drawdown that grows roughly with the square root of time; a strategy with a positive edge grows more slowly, but it still grows. The expectancy per trade can stay exactly the same while the deepest fall gets deeper.

This matters because a backtest covers a fixed window and live trading does not. If the backtest spans six years and you intend to trade the rule for the next six, the combined record will be twelve years long, and the deepest fall in a twelve-year record is on average deeper than the deepest fall in either six-year half. Treat the backtest drawdown as a reading taken partway through, not as the end of the story.

How do you estimate max drawdown with Monte Carlo?

A Monte Carlo drawdown test reshuffles or resamples the backtest trades many times, computes the maximum drawdown of each new order, and reports the spread of those values; the 95th percentile of that spread is a more useful planning number than the single backtest figure. The method is named after the casino because it answers a question by generating many random outcomes rather than by solving an equation (Metropolis and Ulam, 1949). The 95th percentile is the value that 95 of every 100 simulated runs stay below. Two ways to generate the runs:

1. Reshuffle (permutation). Keep every trade, change only the order. Each run has the same net profit as the backtest and a different drawdown. This isolates the one thing the backtest fixed by accident: the sequence.
2. Resample with replacement (bootstrap). Draw trades at random from the list, allowing repeats, until a run is as long as the original (Efron, 1979). Some runs contain the big loser twice and the big winner never, so net profit varies too. This gives a wider and usually more honest spread when the trade count is small.

Both methods assume that one trade does not depend on the trade before it. When losses cluster in a particular market regime, the real sequence has more clustering than a shuffle, and the simulated spread can be narrower than reality. That is a reason to read the 95th percentile as a floor for planning, not as a ceiling, and to combine it with the multiple in the next section. The steps for setting up such a test, and the mistakes that make it meaningless, are in Monte Carlo backtest: your real max drawdown is probably worse.

To show what such a distribution looks like, this article uses one real AlphaPrime (Valynox) run on NQ daily bars from 2019-01-02 to 2025-06-30, long and short, one contract, with commission and slippage charged on every trade. The rule is a stand-in: the Efficiency-Adaptive Moving Average rule it uses as its only entry and exit is not the subject of this article; it is simply a trade list to feed into a drawdown test. The run is the one candidate the search kept out of the 1,024 it tried, chosen for its return against its drawdown. Choosing the kept one after looking at the results flatters it: among a thousand variants, the one with the best return for its drawdown is partly the one whose trades happened to land in a kind order, so its realised drawdown should sit toward the favourable side of its own Monte Carlo spread. Bailey et al. (2014) describe this effect in detail. Nothing here says the rule earned this money or that a reader can expect it.

Two things in the box shape the Monte Carlo result. The run has 59 trades over six and a half years, and it was in the market 92.15% of the time, so single trades last for months and a few of them decide the drawdown. A distribution built from 59 trades is wide, because reordering a short list moves the big losers around a lot. And the realised maximum drawdown of $116,210 against a net profit of $193,250 is the one order history produced. The distribution itself is not part of the backtest: it comes from pasting the 59 trade results of this run into the free Monte Carlo simulator, which reshuffles the list many times and plots the maximum drawdown of each new order, so the figure below shows how many other orders of the same 59 trades would have been deeper or shallower than the one the backtest reported.

The realised drawdown that the distribution is built from is visible on the equity curve of the same run, as the deepest dip under the running peak. In AlphaPrime the drawdown panel of the Performance Report is drawn directly under the equity curve of each candidate, so the figure you plan around comes from the same run that produced the trade list, not from a separate calculation.

How many times the backtest drawdown should you plan for?

A practical planning size is 1.5 to 2 times the backtest maximum drawdown, checked against the Monte Carlo 95th percentile and against the margin the account must keep, with the multiple pushed toward 2 or beyond when the trade count is small or the candidate was picked from many. A multiple is a rule of thumb; the percentile is a measurement from your own trades; the margin check is what keeps the account tradable at the bottom of the drawdown. Use all three and plan for the largest.

Condition	Multiple toward 1.5	Multiple toward 2 or more
Trade count in the backtest	Several hundred trades	Under about 100 trades
How the candidate was chosen	One idea, tested once	Best of many variants
Monte Carlo 95th percentile	Below 1.5 times the backtest value	Above 1.5 times the backtest value
Losses cluster by regime	Losses spread across years	Losses concentrated in one stretch
Planned live horizon	Shorter than the backtest	Longer than the backtest

The margin check comes last and is the hardest limit. At the bottom of the planned drawdown, the account must still hold more than the maintenance margin for the open position, or the position gets reduced at the worst moment. How to read that ratio and size the account from it is covered in futures risk indicator: calculation and drawdown sizing. Exchanges and brokers change margin levels, so take the current figure from the exchange's contract page before you size.

When several strategies run in one account, do the planning on the combined equity curve, not on each strategy alone. Drawdowns that overlap in time add up, and the combined worst stretch is the one you actually sit through; portfolio backtest: the combined drawdown is the one you live through shows how to measure it.

How large a drawdown should stop a strategy?

Write the stop level before trading starts, set it above the planned drawdown rather than at the backtest figure, and treat a drawdown beyond it as a signal to pause and revalidate, not to re-tune on the fly. A stop set at the backtest maximum drawdown will be hit by a healthy strategy fairly often, because the previous sections show that the backtest value is near the favourable end of what to expect. A stop set at the planned level plus a small buffer is hit only when the strategy is doing something the Monte Carlo spread said was unlikely, which is exactly the signal you want.

1. Set the stop level before going live. Take the planning drawdown from the previous section and add a buffer, for example one average losing trade, so that a single trade cannot trigger the stop by itself. Write the number down with the date.
2. Track live drawdown against it each week. Compute the running peak and the current drawdown from live equity, in the same units as the plan. A weekly check is often enough for a daily-bar strategy; an intraday strategy needs a daily check.
3. Pause when the level is crossed. Stop taking new signals and flatten the position according to the exit rule. Do not change a parameter while the position is open.
4. Revalidate on new data before any restart. Run the unchanged rules on the bars that arrived since the backtest ended and compare the result against the thresholds you set before. A strategy that passes the same validation on the new data can restart; one that fails goes back to research as a new candidate. The walk-forward and threshold steps in strategy validation are built for this kind of recheck.

The reason to pause rather than re-tune is the same reason to freeze rules after an out-of-sample test: changing a setting because the live curve looked bad turns the live data into tuning data, and the next drawdown figure means less than the last one. A pause costs you some time out of the market. A re-tune costs you the ability to tell whether the strategy ever worked.

The next step takes ten minutes: export the trade list from your current backtest, paste it into the free Monte Carlo simulator linked in the first section, read the 95th percentile of maximum drawdown, and compare it with 1.5 and 2 times the backtest figure. Whichever is largest is the number to put next to your margin requirement before the first live trade.

Frequently asked questions

How is maximum drawdown calculated?

Track the highest equity the account has reached so far. On each bar, subtract the current equity from that running peak; the result is the drawdown at that moment. The maximum drawdown is the largest of these values over the whole record, measured in money or as a percent of the peak. It ends only when equity makes a new high.

How many times the backtest maximum drawdown should I plan for?

A common planning range is 1.5 to 2 times the backtest figure. Move toward 2 or more when the backtest has fewer than about 100 trades, when the strategy was chosen as the best of many variants, or when a Monte Carlo reshuffle of the trades puts the 95th percentile above the 1.5 level. Then check that the account still holds more than the required margin at the bottom of the planned drawdown.

Does a Monte Carlo test predict the real maximum drawdown?

No. It shows the range of maximum drawdowns the same trades could have produced in different orders, which is wider than the single backtest value. The 95th percentile of that range is a planning number, not a forecast. Because the method assumes trades are independent of each other, it can understate drawdowns for strategies whose losses cluster in one market regime, so use it together with a multiple rather than instead of one.

When should a drawdown stop a strategy?

When live drawdown crosses a level you wrote down before trading started, set above the planned drawdown rather than at the backtest figure. Crossing it means the strategy is doing something its own trade distribution said was unlikely. Pause new signals, exit by the strategy's own rule, and rerun the unchanged rules on the data that arrived since the backtest before deciding whether to restart.

References

• Magdon-Ismail, M., & Atiya, A. F. (2004). Maximum Drawdown. Risk, 17(10), 99–102.
• Metropolis, N., & Ulam, S. (1949). The Monte Carlo Method. Journal of the American Statistical Association, 44(247), 335–341.
• Efron, B. (1979). Bootstrap Methods: Another Look at the Jackknife. The Annals of Statistics, 7(1), 1–26.
• Bailey, D. H., Borwein, J., López de Prado, M., & Zhu, Q. J. (2014). Pseudo-Mathematics and Financial Charlatanism: The Effects of Backtest Overfitting on Out-of-Sample Performance. Notices of the American Mathematical Society, 61(5), 458–471.