Monte Carlo for a Strategy: What You Are Actually Shuffling

In trading, Monte Carlo does not answer the question "how much will I make" but rather "how much of the result I have already seen in the backtest was down to the order of events". The method takes your strategy's statistics, reassembles them in many different ways and shows you not a single point but a spread of outcomes. The practical value is that you see in advance the drawdown that the historical run simply did not have time to show you, and you can decide whether you could live through it. The whole meaning of the procedure comes down to one choice: what exactly you randomise.

Four different Monte Carlos that go by one name

Trade reshuffling. You take the finished list of trades from the backtest and shuffle their order, building new equity curves. The question this answers is: "what could my drawdown have been if the losses had come in a row rather than scattered".

Bootstrapping with replacement. The same shuffling, but trades are drawn with repetition: some drop out, some appear twice. This is now a question about the sample: "how robust is my estimate of the average trade to the fact that history gave me this particular set rather than a neighbouring one".

Resampling daily returns. What gets shuffled is not trades but equity increments. This is convenient when a position is open almost all the time and the notion of a "trade" barely exists.

Synthetic price series. The heaviest option: you generate new price paths (a block bootstrap on historical bars, a model with a specified volatility, block permutation within a regime) and run the strategy over them again. Only here can the strategy make decisions different from the ones it made in history.

The first three variants reassemble the result. The fourth reassembles the market. These are fundamentally different claims, and passing one off as the other is the most common mistake in reports.

What trade reshuffling breaks rather than tests

When you shuffle the order of trades, you destroy everything that makes a market a market: return autocorrelation, volatility clustering, regime shifts. A trend-following system lives on streaks: the rare large profits come in clusters, in the same phases as the large losses. Once you mix them up, you get a distribution in which the streaks have disappeared, which means the drawdown estimate will turn out optimistic rather than conservative.

The second limitation is harsher: reshuffling does not change the set of trades itself. If the strategy found them by curve-fitting to history, Monte Carlo on those trades will neatly estimate the spread of an edge that does not exist. The method inherits the flaw in its input data and dresses it up as statistics. That is why it is no substitute for testing for overfitting; how that test works is better read separately, in the glossary entry, and it makes sense to analyse the Monte Carlo results only once the out-of-sample test has been passed.

How to read the distribution, not the median

The median of the simulations is almost useless: it roughly repeats the backtest. What you need to look at is the tail.

The maximum drawdown in the worst runs is your real risk limit, not the figure that appeared in the backtest report. Position size should be calibrated to the tail of the distribution, not to the average.

The share of runs that end in the red shows how distinguishable the edge is from noise. If a noticeable portion of the paths falls below zero, the strategy may well work, but the time it would take for that to become visible exceeds your willingness to wait.

Risk of ruin is the share of runs in which capital fell below the level after which you would stop trading. This is a psychological threshold, not an arithmetical one, and you have to set it yourself.

The duration of the drawdown in the worst scenarios is an underrated metric. People abandon systems not because of how deep a drawdown goes but because of how long it lasts.

The workflow

Step 1: lock the strategy and do not touch it again, since any change made after looking at the simulations takes you back to curve-fitting. Step 2: choose the type of randomisation to fit the question, not out of habit. Step 3: run the simulations on a sufficiently long history that covers different regimes; for the Russian market, this means including both periods of sharp repricing in equities and phases when OFZ yields set the entire risk picture. Step 4: build in costs and slippage not as a constant but as a random variable; in illiquid names this changes the conclusions more than the entry algorithm itself. Step 5: check the calendar; if the strategy holds positions through earnings releases and dividend record dates, shuffling the trades will smear these events across random dates and distort the tail.

What the method does not give you

Monte Carlo does not forecast returns and does not legitimise a strategy. It knows nothing about events that never occurred in history: any distribution assembled from the past is silent about structural shifts. Nor does it make up for a small number of trades: if there are few observations, the simulations will merely show, neatly, that there is nothing to say. That is an honest answer, and it is more useful than a confident one.