Record AR-MC-021 · Survivability Core
The history you have is one ordering of the outcomes you got. Reorder the same trades and the experience changes completely: the drawdown deepens or halves, the objective arrives sooner or never, and the point at which you would have stopped moves. A Monte Carlo study replaces that single sample with a distribution — and then refuses to tell you which path you are on.
Demonstration simulation based on selected assumptions. Not a forecast.
Monte Carlo is the most over-claimed instrument in retail trading and one of the most useful when its boundaries are respected. Both lists below are short and neither is negotiable.
What it can tell you
What it cannot tell you
Interactive · Level 2
Every control below is an assumption you are making, not a measurement anyone has taken. Change one and the entire distribution moves. The governance toggle is the only control that changes the rules of the simulation rather than its inputs.
Simulation assumptions
DeterministicPresets
Not a forecast
Survival status
Demonstration simulation based on selected assumptions. Not a forecast.
| Percentile | Ending equity | × start | Max drawdown | Reading |
|---|
These are two separate marginal distributions ranked independently. The path that ends at the P90 equity value is not necessarily the path that experienced the P10 drawdown, and no row below describes a single simulated history.
Time to objective is conditional
The path field shows shape. These three figures show mass: where terminal equity actually lands, how deep the decline goes on the way there, and how much of the simulated trading life is spent in each capital state. All three redraw with the controls above.
Demonstration simulation based on selected assumptions. Not a forecast.
Demonstration simulation based on selected assumptions. Not a forecast.
Demonstration simulation based on selected assumptions. Not a forecast.
Reading the dwell record
Dwell is measured in trade-time, not calendar time: the share of all simulated trade decisions taken while the account sat in each capital state. It answers a question the equity curve cannot — not how deep did it go, but how much of the working life was spent compressed.
A system that spends most of its trade-time below the growth band is not simply volatile. It is operating at reduced size for most of its existence, which means the compounding assumption behind every projection you have made for it is wrong.
The gate ladder in full →Demonstration simulation based on selected assumptions. Not a forecast.
Demonstration simulation based on selected assumptions. Not a forecast.
Section 05
A distribution is only useful if you can locate yourself inside it. This is the doctrine the Foundry applies when live results are read against a simulated band — and it is deliberately asymmetric. Being unusually good is treated as a question, not a reward.
The band must be built before the live period it is used to judge. A benchmark fitted after the fact is not a benchmark; it is a description.
One rule above all others
A reading below the band never authorises an increase in risk. Under-performance against a simulated benchmark is information about the model or the execution, never a licence to size up in pursuit of the median.
The simulation on this page draws outcomes from an assumed distribution. That is resampling, and it is the weaker of the two available tests. Resequencing — reordering the trades you actually took — assumes far less and is far harder to argue with.
| Resequencing | Resampling | |
|---|---|---|
| What is drawn | Your own realised outcomes, reordered without replacement. Every trade appears exactly once in every path. | Synthetic outcomes generated from a stated win probability and payoff, as on this page. |
| What is preserved | The exact outcome distribution, including the fat tails, the outliers and the trades you would rather forget. | Only the summary statistics you chose to specify. Everything else is invented by the generator. |
| What is destroyed | Serial structure: streaks, regime clustering and the tendency of losses to arrive together. | Serial structure and the true shape of the distribution simultaneously. |
| What it answers | How much of my drawdown was the edge, and how much was the order? | What range of outcomes does an edge of this shape produce over this many trades? |
| Main weakness | Requires a real, adequately sized trade record — and it can never produce an outcome worse than your worst actual trade. | Requires you to assert an edge. If the assertion is wrong, every number downstream is decoration. |
| Honest use | The primary test once you have a record. Run it first. | A design instrument for reasoning about risk before a record exists. |
Both share one blindness
Neither test preserves serial correlation. Real losing runs are longer than either method produces, because the conditions that cause one loss frequently persist into the next trade. Record AR-EXP-063 finds that trade order alone moves modelled maximum drawdown by more than a factor of two — and that is with the clustering removed. Treat every drawdown figure on this page as a floor rather than a ceiling.
Sample size gate
Resequencing a record of thirty trades produces confident-looking output from nothing. The reordering cannot manufacture information that the sample never contained, and the narrow bands it returns are an artefact of the small sample, not evidence of stability. Treat any record under roughly one hundred trades as a description of a period rather than a measurement of a system.
Section 07
A simulation is a set of assumptions with a chart attached. Most published Monte Carlo work states the assumptions that flatter it and omits the rest. The list opposite is complete for the model running on this page, and five of its entries are known to be wrong.
They are not wrong by accident. Each one is a simplification that makes the model tractable, and each one biases the result in the same direction — towards outcomes that are better than reality. That is the important part: the errors do not cancel.
Direction of the bias
Independent outcomes, a stationary win probability, no regime change, no correlation between concurrent positions and no execution failure all push in the optimistic direction. The real distribution has a longer left tail and a deeper drawdown percentile than anything this page will draw.
Stated assumptions
The moment a simulated median becomes something to chase, the instrument has been inverted and is now actively dangerous. Three rules govern how the Foundry treats a simulated benchmark, and the MARS layer enforces them structurally rather than relying on the operator to remember them.
Performing under the simulated median tells you that the assumed edge, the execution, or the market conditions differ from the model. Every one of those is a reason to investigate and none of them is a reason to increase risk. The instinct to size up towards the benchmark is the single most reliable route from underperformance to failure.
Underperformance is a diagnostic, not a debt to repay.
Returns above the simulated P90 with drawdown also above the simulated P90 are not alpha. They are over-risk that has not yet been billed. The correct response to unusually good results is to check the drawdown percentile, the position sizing and the concurrency — in that order — before drawing any conclusion about skill.
Return above benchmark with worse drawdown is leverage, not edge.
A benchmark constructed after the results are known will always fit them. The assumptions must be recorded, dated and frozen before the live window opens, and the comparison must then be made against that frozen band even when it is unflattering. A revised benchmark is a new experiment, not a corrected one.
Freeze the ruler before you measure with it.
Where this leads
Monte Carlo tells you what should happen under stated assumptions. Expectancy tells you what is actually happening. Neither has authority on its own. The governance layer is what converts the difference between them into a permitted risk envelope — and it is the reason the benchmark can be read honestly instead of being chased.
Each of these is a specific, common error rather than a general caution. If you take one thing from the Monte Carlo Core, take the corrections rather than the numbers.
Methodology
Each path is an independent sequence of trades. On every trade a uniform draw decides a win or a loss; the result is the stated average win or average loss in R, less the stated friction, applied to equity as a fraction of the current risk allocation. Equity compounds, so a fixed percentage risk means a shrinking absolute risk in decline and a growing one in expansion.
Capital authority is reviewed on cycle boundaries rather than after every trade — four trades per cycle in this model. At each review the decline from the running equity peak selects a gate, and the gate supplies a multiplier applied to the risk fraction until the next review. With governance disabled the multiplier is fixed at one and no lock exists.
A path stops when it reaches the objective equity — completion is absorbing, so the completion probability counts paths that ever touched the objective, not paths that ended above it. With governance enabled a path also stops if decline from the peak reaches the lock boundary, at which point capital authority is withdrawn and the path is recorded as locked.
Every draw comes from a seeded generator. The same seed index and the same assumptions always produce the same distribution, on every visit and every device. No unseeded randomness is used anywhere on this site.
Limitations
Two-outcome trades
Real outcome distributions are continuous, skewed and fat-tailed. This model collapses them to a single win value and a single loss value, which removes both the outliers that carry trend systems and the oversized losses that end them.
One position at a time
There is no concurrency in the model, so there is no correlation either. An operator running four correlated positions is exposed to a single bad draw at roughly four times the modelled size, and the drawdown percentiles here do not describe that at all.
Perfect compliance
Every path assumes the gate schedule is followed exactly, that no trade is skipped in a drawdown and none is oversized in a recovery. The historical failure rate of that assumption among human operators is the largest unmodelled term on the page.
Calibration is not published
The gate boundaries, pool percentages and tier tables used in the MARS workbook are part of that product and are not published here. The structure and the principle are described in full; the calibrated numbers are not, and the values in this preview are illustrative.
Drawn paths are a subset
Statistics are computed over every path. The field in Figure 1 draws a thinned selection so the chart reads as texture rather than noise; the count of computed versus drawn paths is stated in the caption on every redraw.
09Monte Carlo Lab ZonesAR-MC-001·Z
The planned Monte Carlo Lab carries 225 distinct configurable field definitions across fourteen zones. They are not all sold at every tier — but the parts that keep a simulation honest are.
Never sell an unsafe simulator
At every tier including the free one: pre-flight validation runs and blocks incomplete or unsafe configurations; the payoff, risk-governance and friction models apply at their calibrated defaults; and the core risk outputs are shown. What is tiered is the operator's control over those models — never whether they are in force. A free run is a correct simulation with fewer knobs, not a cheaper simulation with the safeguards removed.
What a locked zone does and does not cost you
Every setting a locked zone contains is inherited from your saved strategy configuration rather than replaced by a house default. That is what makes a restricted run honest: the model is running your exits, your correlation assumption, your friction and your risk policy, and a locked zone costs you the ability to vary them inside the study — not the guarantee that they are yours. Every applied value is printed beside the result.
We will not make the stronger claim that a locked zone always leaves the answer correct, because it is not true in general. If a hidden exit, correlation, friction or risk assumption differed from the strategy you actually configured, the result would not be merely less detailed — it could be wrong, and you would have no way to see it. Inheritance is what removes that risk, which is why it is a property of every tier rather than a feature of the paid ones.
Two catalogues, tiered on different principles
The 288-control general parameter catalog is tiered by percentage — 5, 40, 70 and 100 per cent — with twelve correctness-critical controls held outside that ladder and visible at every tier. The Monte Carlo Lab zones above are tiered by zone instead. They are cut differently because they fail differently: a missing strategy parameter is a decision the engine makes silently on your behalf, which is exactly why the twelve that would make an answer wrong rather than merely smaller are never withheld. The general catalog is for any trader; these zones are the granular, MARS-specific instrument, and their controls become considerably more detailed as the roadmap is released.
Continue
Knowing that the tenth percentile of your own strategy involves a decline you have never experienced is worth nothing unless a rule exists that changes your behaviour before you reach it. That rule is the drawdown gate, and it is the subject of the next two chambers.