Reading the fan: Monte Carlo without self-deception.

What the fan is

Your evidence, reshuffled into thousands of futures.

The Monte Carlo Lab resamples your recorded trade distribution — real R outcomes, real hit rates, real branch mix — into tens of thousands of alternative sequences. The fan that emerges is not prediction; it is the range of histories your current system could plausibly write. P50 is the median story. P90 is the flattering one. P10 is the one an operator plans around, because a system that survives its P10 path survives its future. The fan therefore describes the system, not the future — it is a portrait of what this edge is capable of producing, favourably and otherwise.

The governing read

Size to the adverse band, not the median.

The lab's most consequential output is drawdown distribution: how deep the adverse paths cut before recovering. If the P10 drawdown breaches the gate ladder's guardrails at your current tier, the tier is wrong — no matter how attractive the median looks. This is how simulation feeds governance in MARS: the benchmark ties tier authority to what the evidence says can go wrong, before real capital finds out. Sizing to the median is the most common way an otherwise sound system arrives at a drawdown it was never authorised to survive.

Sequence risk

The same trades in a different order write a different account.

Shuffling an identical set of outcomes changes nothing about expectancy and everything about the experience of trading it. A run of losses early, while the account is small, is survivable; the same run after a period of compounding removes a far larger absolute sum. This is path dependency, and it is the specific thing a fan of simulated sequences exists to expose. A trader who has only lived one ordering of their own history has seen exactly one sample from the distribution — and has usually mistaken it for the distribution itself.

Reading a live result against the fan

The question is which band you are in, not whether you are up.

The benchmark's operational use is comparison: a live cycle is placed inside the simulated distribution, and its percentile position is the verdict. Sitting near P50 means the system is behaving as its own evidence predicts, which is the outcome to want — not the flattering one. Sitting below P10 is the signal that matters, because it says live results have left the range the sample can explain, and something has changed that the model does not know about. A result above P90 deserves the same suspicion for the same reason: unexplained is unexplained in both directions.

• Near P50: behaving as evidence predicts. Change nothing.
• Below P10: outside what the sample can explain. Investigate before sizing.
• Above P90: also unexplained. Do not promote luck to skill.

What the fan cannot see

The simulation inherits the blind spots of its input window.

A resampled distribution can only recombine what it was given, which makes the sample window the quiet determinant of everything downstream. Draw the sample from a period containing no volatility shock and the fan will contain no volatility shock, no matter how many paths are generated — fifty thousand simulations of a benign year produce fifty thousand benign years. The same limitation applies to structural change: if the edge has drifted since the sample was taken, the entire distribution is describing a system that no longer exists. Neither failure announces itself inside the output, which is why the fan is cross-examined by regime classification and structural diagnostics rather than trusted alone. Simulation answers how this evidence distributes. It cannot answer whether the evidence is still current.

Where authority ends

Simulation assumes tomorrow resembles the sample.

Monte Carlo inherits every property of its input window — including its blind spots. A sample drawn entirely from a friendly regime produces an optimistic fan; a structural drift in edge quietly invalidates the whole distribution. That is why the fan is cross-examined by regime classification and structural diagnostics rather than trusted alone. The simulation is evidence, never oracle.