Monte Carlo Simulation in Backtesting

What is Monte Carlo simulation?

A single backtest is one draw from a much larger set of possibilities. The strategy earned the return it earned because the months arrived in the precise order they did, with a specific run of good and bad stretches. Shuffle that order, and the compounded ending value changes, sometimes dramatically, even though every individual month is identical. A strong-looking backtest may owe part of its result to a lucky sequence rather than a durable edge, and a single line on a chart gives you no way to tell the two apart. Monte Carlo simulation is the standard tool for separating them.

The method takes the strategy's realized returns and resamples them many times to build a population of alternative paths. Each path is a different plausible history, assembled from the same raw material, and compounding each one forward produces a different ending value. Collect thousands of those endings and you have a distribution: a sense of where the strategy tends to land, how wide the spread is, and how grim the unlucky tail looks. The familiar fan chart is just that distribution drawn over time, with the median path down the middle and probability bands widening as uncertainty accumulates.

The word "simulation" can suggest a crystal ball, so it is worth being precise about what Monte Carlo is and is not. It does not predict the future and it does not introduce any information the backtest did not already contain. It is a way of reading the uncertainty that was always present in a single backtested path, made visible by replaying the strategy's own history in thousands of different orders. The honest framing is a range of plausible outcomes consistent with the past, not a statement about what will happen next.

How the Projection Is Built

The naive version of Monte Carlo draws individual months independently, but markets do not behave that way: a bad month tends to sit near other bad months, and momentum and mean-reversion both leave fingerprints across adjacent periods. Sampling one month at a time destroys that structure and produces a cone that is too narrow, because it quietly assumes each month is unrelated to the last. The fix is to resample in blocks rather than single observations, so short runs of consecutive history travel together and their internal correlations are preserved.

SledgeKey uses a stationary block bootstrap, the method introduced by Politis and Romano in 1994. Each simulated path is stitched together from blocks of consecutive historical months whose lengths vary randomly, governed by a restart probability at every step. Because block lengths are random rather than fixed, the resampled series is stationary, free of the artificial seams a fixed block length would create. This keeps the short-run autocorrelation of the real strategy intact while still generating genuinely novel orderings, which is exactly the balance a credible cone requires.

Once the paths are generated, the cone is read off as percentiles at each point on the horizon. The median (P50) traces the central tendency, the P25 to P75 band holds the middle half of outcomes, and the P5 to P95 band spans the bulk of the distribution while leaving the extreme tails outside the shaded region. The bands necessarily widen with time, because uncertainty compounds: the spread of where a strategy might be in five years is far larger than where it might be in six months.

Why Monte Carlo Matters in Backtesting

The single most useful thing a cone does is expose path-dependence. Two strategies can post the same backtested return while having wildly different ranges of plausible outcomes, one tightly clustered around its result and the other spanning everything from a near-total loss to a windfall. The point estimate hides that difference entirely; the cone makes it the first thing you see. A wide cone is a warning that the headline backtest number is fragile and sequence-sensitive, while a narrow one suggests the result is more representative of the strategy's typical behavior.

A cone also reframes the downside in a more honest way than a single drawdown figure can. The lower band, the P5 path, is a concrete answer to "if things go poorly, roughly how poorly?" expressed in the same dollar or percentage terms as the rest of the report. That is a more decision-useful number than a backtest's realized maximum drawdown, because the realized drawdown is just one sample of bad luck, while the P5 band reflects a whole population of unlucky orderings. Sizing a position so that the P5 outcome is survivable is a far sturdier discipline than sizing to the single path history happened to hand you.

Finally, the cone is a sanity check against overfitting to a sequence. A strategy tuned until its one backtested line looks pristine has often been fitted to the particular order of history, not to a real and repeatable edge. If small reshufflings of that same history scatter the outcomes across an enormous range, the pristine line was substantially luck. Monte Carlo cannot prove an edge is durable, but it reliably flags when a result leans on a fortunate sequence, which is one of the most common ways backtests mislead.

How SledgeKey Implements Monte Carlo

SledgeKey runs the projection as post-processing on a completed backtest. It takes the strategy's realized monthly returns, resamples them into 5,000 alternative paths with the stationary block bootstrap, compounds each forward, and reports the P5, P25, P50, P75, and P95 bands as a fan chart alongside the backtest results. Because it reuses returns the backtest already computed, it adds only milliseconds and requires no extra data. The projection targets a five-year horizon but never extends further than the available history supports, and it needs at least twelve months of returns before it will render a cone, since a shorter series cannot describe a distribution with any credibility.

The simulation is seeded deterministically from the backtest's own inputs, so an identical backtest always yields an identical cone rather than drifting from run to run, while genuinely different strategies still produce different simulations. The Monte Carlo projection is a SledgeKey+ feature: free accounts see a locked preview of the cone with an upgrade prompt, and subscribers get the full computed distribution rendered on the results page, written into the research PDF as its own section, and exported as a dedicated sheet in the Excel workbook. The same percentile bands appear in all three places so the on-screen view, the report, and the spreadsheet always agree.

Common Pitfalls

The first pitfall is reading the cone as a forecast. It is a resampling of the strategy's own past, which means it inherits every assumption baked into that history and adds one of its own: that the return distribution is roughly stationary, that the future is drawn from the same statistical pot as the sample. Real markets violate that assumption across regime shifts, when interest rates, volatility, or correlations move to a new footing. The cone is most trustworthy as a description of variability within the conditions the strategy was tested in, and least trustworthy as a claim about an environment those conditions never contained.

The second pitfall is treating the median as the expected return. The P50 path is the middle of the simulated distribution, not a prediction and not an average; because compounded outcomes are skewed, the median and the mean rarely coincide, and neither is a promise. The value of the chart is the spread between the bands, not the line down the middle. A reader who fixates on the median and ignores the width of the cone has discarded the entire reason to run the simulation in the first place.

The third pitfall is over-trusting a cone built from thin history. The bootstrap can only reshuffle the months it was given, so a strategy tested over a single calm stretch will produce a reassuringly tight cone that simply reflects the absence of a crash in the sample, not the absence of crash risk in the world. No amount of resampling invents a market condition the history never saw. Treat a narrow cone from a short or placid window with suspicion, and lean on longer test periods that actually contain stress before letting the projection inform a sizing decision.

Related Concepts