Monte Carlo Simulation: Stress-Testing Your Own Results
Your trade history is one sequence out of many that could have occurred. Monte Carlo simulation explores the others, and shows how bad things might get.
The method is simple. Take your list of past trade results in R. Randomly reorder them, or randomly draw them with replacement, to build many new sequences of the same length. For each sequence, compute the equity curve, maximum drawdown, and ending value. Repeat thousands of times.
The output is a distribution. You might find that your actual maximum drawdown was 8 percent, but in 5 percent of simulated sequences it exceeded 18 percent. That tells you the drawdown you experienced was on the lucky side, and that you should be prepared for worse.
It also shows the probability of ruin under your sizing, and the range of outcomes over the next hundred trades, which is wider than most people expect.
Spreadsheets can perform simple simulations with random functions, and free tools exist for the purpose. Programmers can do it in a few lines of code.
Remember the limits. The simulation assumes future trades resemble the past ones and that outcomes are independent. It cannot capture regime changes or behavior changes.
Still, it improves risk planning, by supplying realistic worst-case figures for sizing decisions.
Practical step: if you can, run a simple simulation of your trade list, and find the 95th percentile of maximum drawdown.
Monte Carlo Simulation: Stress-Testing Your Own Results
Re: Monte Carlo Simulation: Stress-Testing Your Own Results
Even a basic spreadsheet version is useful. Shuffle the R column a few hundred times with a random sort and record the worst drawdown each time. It takes maybe twenty minutes to set up. In Google Sheets, for example, you can put RAND() next to each result, sort by it, then use a running sum and track the lowest point below the previous peak. Copy that a few hundred times or use a small script. What most people find is that their real drawdown was somewhere in the middle of the range, and the bad end of the range is a lot worse. Better to see that in a spreadsheet than in the account.
It’s Fairman 
Re: Monte Carlo Simulation: Stress-Testing Your Own Results
The 95th percentile drawdown is a good number to size from. If you can't stomach that one, the risk per trade is too high, even if your actual history looked fine. Say the simulation shows a 95th percentile drawdown of 15R. At 1 percent risk that's about a 14 percent drop in the account, allowing for compounding. At 2 percent it's closer to 26 percent. Ask yourself honestly which of those you could sit through without changing the strategy or quitting. For most people it's the smaller one. Picking risk from the bad case rather than the average case is one of the simplest ways to stay in the game.
It’s Fairman 
-
FXS Prop Desk
- Posts: 1689
- Joined: Wed Oct 07, 2026 9:52 am
Re: Monte Carlo Simulation: Stress-Testing Your Own Results
One output worth adding to the distribution is the probability of hitting a specific drawdown limit, for example 10% for prop accounts. Counting how many of the simulated sequences reach that level before the end gives a direct estimate of the risk of failure at the current position size, which can then be lowered until that probability falls to an acceptable level.
FXS Official Team | The Idea Lab & Strategy
-
Shadow Trader
- Posts: 5594
- Joined: Wed Sep 09, 2026 5:19 pm
Re: Monte Carlo Simulation: Stress-Testing Your Own Results
Shuffling and resampling answer slightly different questions, and it is worth running both. A shuffle keeps the same trades, so every sequence ends at the same total, while drawing with replacement lets the ending value vary as well.
The chart is a made up example: 200 trades with 45 percent winners of 1.5R and losers of 1R, shuffled 5,000 times. The actual order had a worst drawdown of 8R, the median shuffle 12R, and the 95th percentile 19R.
Taking the prop desk point about a 10 percent limit, at 1 percent risk per trade that limit is roughly 10R. In this example 78 percent of the shuffled sequences reach it at some point, even though the actual run never got there.
The weak spot is the independence assumption. Scalping losses tend to cluster on news days and in dead sessions, so a block shuffle that moves whole days rather than single trades gives a more honest tail.
Has anyone compared single trade and block shuffles on the same trade list? I would expect the block version to show a noticeably worse 95th percentile.
The chart is a made up example: 200 trades with 45 percent winners of 1.5R and losers of 1R, shuffled 5,000 times. The actual order had a worst drawdown of 8R, the median shuffle 12R, and the 95th percentile 19R.
Taking the prop desk point about a 10 percent limit, at 1 percent risk per trade that limit is roughly 10R. In this example 78 percent of the shuffled sequences reach it at some point, even though the actual run never got there.
The weak spot is the independence assumption. Scalping losses tend to cluster on news days and in dead sessions, so a block shuffle that moves whole days rather than single trades gives a more honest tail.
Has anyone compared single trade and block shuffles on the same trade list? I would expect the block version to show a noticeably worse 95th percentile.
-
Shadow Trader
- Posts: 5594
- Joined: Wed Sep 09, 2026 5:19 pm
Re: Monte Carlo Simulation: Stress-Testing Your Own Results
A stress test worth adding on top of the shuffle is to degrade the edge before simulating. Live results usually come in below the backtest, so a simulation using the historical win rate is already on the optimistic side.
Take the same made up profile from my earlier post, 45 percent winners at 1.5R. Expectancy is 0.125R per trade, but at 42 percent it falls to 0.05R, and at 40 percent it is zero, so a three point slip in win rate cuts the edge by more than half.
Running the simulation at 42 percent instead of 45 shows how much of the comfortable tail depends on those few points. The worst drawdowns get deeper and, more importantly, the time spent below a previous peak stretches out.
That second number is the one people forget to look at. Say the median sequence spends 60 trades underwater, then at ten trades a day that is six sessions with no new high, which is the stretch where most traders start tinkering with the rules.
If the 95th percentile time underwater is longer than you could sit through without changing anything, the sizing question from Fairman's post comes back. Lower risk does not shorten the wait, but it makes the depth far easier to sit out.
Take the same made up profile from my earlier post, 45 percent winners at 1.5R. Expectancy is 0.125R per trade, but at 42 percent it falls to 0.05R, and at 40 percent it is zero, so a three point slip in win rate cuts the edge by more than half.
Running the simulation at 42 percent instead of 45 shows how much of the comfortable tail depends on those few points. The worst drawdowns get deeper and, more importantly, the time spent below a previous peak stretches out.
That second number is the one people forget to look at. Say the median sequence spends 60 trades underwater, then at ten trades a day that is six sessions with no new high, which is the stretch where most traders start tinkering with the rules.
If the 95th percentile time underwater is longer than you could sit through without changing anything, the sizing question from Fairman's post comes back. Lower risk does not shorten the wait, but it makes the depth far easier to sit out.
-
Shadow Trader
- Posts: 5594
- Joined: Wed Sep 09, 2026 5:19 pm
Re: Monte Carlo Simulation: Stress-Testing Your Own Results
A test needs to assume some price for entry and exit. The simplest is the mid price, halfway between bid and ask. It's convenient, and it ignores the spread. Adding a fixed cost per trade corrects for it, roughly.
The more precise method uses bid and ask at each moment. A buy fills at the ask, a sell at the bid. Costs emerge naturally, and they vary with the spread at the time.
The difference between the methods is small for a daily system and big for a scalper. If a strategy trades 30 times a day with a 1-pip average edge, the handling of the spread decides whether it's profitable.
Using mid prices without cost overstates results. Using a fixed cost smooths the tails. Only a bid/ask simulation captures the cases where the spread is large at the time of the signal, and those are typically the trades that perform worst.
Many open-source backtesting tools default to mid or close prices. Checking the documentation for what the engine assumes about fills is an essential first step.
Which assumption does your testing tool use for fills, and was it ever checked against live results?
The more precise method uses bid and ask at each moment. A buy fills at the ask, a sell at the bid. Costs emerge naturally, and they vary with the spread at the time.
The difference between the methods is small for a daily system and big for a scalper. If a strategy trades 30 times a day with a 1-pip average edge, the handling of the spread decides whether it's profitable.
Using mid prices without cost overstates results. Using a fixed cost smooths the tails. Only a bid/ask simulation captures the cases where the spread is large at the time of the signal, and those are typically the trades that perform worst.
Many open-source backtesting tools default to mid or close prices. Checking the documentation for what the engine assumes about fills is an essential first step.
Which assumption does your testing tool use for fills, and was it ever checked against live results?