Record AR-SIM-052 · Exposure and sequencing
The same trades,
in a different order, are a different system.
Trade flow is the layer between having a signal and having a position. It decides how many trades may be open at once, how much risk they may collectively carry, how much of that risk is genuinely independent, and what happens to the qualifying signals there is no room for.
Sequencing is not
an administrative detail.
Expectancy is a property of a trade. Everything the account actually experiences — compounding rate, drawdown depth, recovery time, the worst week — is a property of the order the trades arrive in and how many of them are alive simultaneously.
Take a hundred trades with a fixed set of outcomes. Deal them out one at a time and the account rises and falls in small steps; the deepest decline is roughly the worst losing streak multiplied by the risk per trade. Now deal the same hundred trades four at a time. The number of trades has not changed, the expectancy has not changed, and the sum of the outcomes has not changed — but the account now moves in blocks of four, and the worst block is four losses arriving together.
That is the whole argument, and it has three consequences that are easy to state and hard to feel. Compounding is faster, because capital is working more of the time. Drawdown is deeper, because losses arrive in clusters instead of a queue. And the clustering is worse than independent arithmetic suggests, because the signals that fire together are usually responding to the same market condition.
There is a fourth consequence that almost never appears in trading literature: the opportunity you cannot take. Every concurrency limit and every open-risk ceiling rejects qualifying signals. Those trades never appear in the record, they contribute nothing to expectancy, and they are invisible in every performance report — which makes capacity the only major system property that is systematically under-measured because of how it fails.
What concurrency buys
Capital utilisation, faster compounding, and a larger sample per unit of calendar time — which is the single cheapest way to make any measurement on this site more reliable.
What concurrency costs
Deeper drawdowns, correlated clusters, harder attribution, and a governance problem: the moment several positions are open, the decision to take the next one is no longer about that trade alone.
The same edge, three flow designs
Identical trades and identical expectancy. Only the arrival structure differs, and it changes every number an account holder actually cares about.
Five ways to let
trades into the book.
Each mode is a different answer to one question: when a qualifying signal appears and something is already open, what happens? Select a mode to redraw the slot occupancy and re-evaluate the exposure arithmetic.
Flow tunnel
Slot occupancy over one trading week
Illustrative patternSix slots shown. Vertical rules mark cycle boundaries where the mode uses them. Hatched blocks are qualifying signals that arrived with no capacity to receive them.
Simulated demonstration data
This visualization illustrates proposed product behaviour and does not represent historical, live, or guaranteed trading performance. Each mode is the same preview engine with a different concurrency ceiling and correlation limit; the lane pattern is an illustration of the mode's shape, not a modelled week.
Open risk is the number
that caps a bad day.
Four controls decide how much of the account is exposed at once and how much of that exposure is genuinely diversified. Move any of them and watch which of the others stops mattering.
Exposure bench
Nominal open risk = max concurrent × risk per trade
Effective open risk = min(nominal, max open risk)
Independent bets = n ÷ (1 + ρ(n−1))
The third line is the one that does the damage. With four positions at a correlation limit of 0.35 you are not holding four bets; you are holding roughly two. Raise the limit to 0.7 and four positions behave closer to one and a third.
Live exposure arithmetic
Interactive demonstration — not a historical backtestAuthorised risk pool, allocated
Which ceiling is actually binding
The trades you could not take
are not in the record.
A blocked signal is a real cost with no accounting entry. It does not reduce expectancy, it does not appear in a drawdown, and the only way to know it happened is to count it deliberately at the moment capacity refuses it.
Weekly signal supply split into trades taken and trades refused for lack of capacity, at slot counts from one to eight.
Trades taken per week and effective open risk, plotted against the concurrency ceiling.
The two figures above describe the same sweep from opposite sides, and together they contain the most useful single observation on this page: throughput saturates long before exposure does. Once the slot count is high enough to absorb the signal supply, every additional slot adds open risk and adds nothing to the number of trades taken. The account is paying exposure for capacity it will never use.
The reverse failure is quieter. At a low slot count the system refuses a large share of qualifying signals — in the modelled sweep, more than two thirds at a single slot. Those refusals are not distributed randomly. They cluster during exactly the conditions that generate the most signals, which are frequently the conditions the strategy was designed for.
Neither failure is visible in a performance report. Both are trivially visible in a capacity log, which is why the intake specification for the planned application records the refusal alongside the fill.
Idle capital is not free
A sequential design leaves most of the authorised pool unused for most of the week. That is a legitimate and often correct choice — it is the cheapest way to make attribution clean while a strategy is still under evaluation. It should be a decision, though, not an accident of a concurrency limit nobody revisited.
Concurrency decides
which branch gets starved.
Branch weights describe how a week's trades should be distributed. Slots decide how many trades there are. Cut the slot count and the weights become a queueing policy, whether or not anyone wrote one.
The same baseline weights under a full sixteen-trade week and under a halved eight-trade week, resolved two different ways. Proportional cutting preserves the architecture and gives every population a sample too small to read; core-first cutting preserves the sample for the populations that carry the count and removes the supplemental one entirely.
Four slots and four cycles a week produce sixteen trades, which is exactly enough to express weights of 45, 33.75, 11.25 and 10 as seven, six, two and one. That tidiness is not an accident — the allocation was designed against the throughput, not derived independently of it.
Halve the capacity and the tidiness disappears. Cut proportionally and the no-partial population receives about one trade a week, which over a quarter is a sample from which nothing can be concluded. Cut core-first and that population keeps a readable sample while the supplemental one is switched off — a harsher decision that leaves the record more informative.
Neither answer is universally right. What is universally wrong is having no rule, because then the branch that loses its slot is whichever one happened to signal last, and the realised weights become a record of arrival timing rather than of design.
| Cycle | Slot 1 | Slot 2 | Slot 3 | Slot 4 |
|---|
One illustrative assignment satisfying the baseline weights across a sixteen-trade week. Many others do.
Eight positions can be
one position eight times.
The correlation limit is the least intuitive control in the exposure group and the one that most often converts an apparently diversified book into a single concentrated bet.
Effective independent bets as the permitted correlation between concurrent positions is swept from zero to one hundred percent, at two concurrency ceilings.
The decay is steep and it is front-loaded. Most of the diversification available from four positions is gone by the time permitted correlation reaches forty percent, and the curves for four and eight positions have almost converged by sixty. Doubling the concurrency ceiling in a correlated book buys almost nothing except exposure.
The practical reading is that concurrency and correlation must be set together. A high concurrency ceiling with a loose correlation limit is not an aggressive configuration — it is a concentrated one wearing a diversified costume, and it will produce a worst week far outside what the position count suggests.
| Correlation limit | 4 positions | 8 positions | Reading |
|---|
Review at a boundary,
not continuously.
A cycle is a block of trades treated as one commitment. The block completes, capital authority is reviewed, and only then does the next block open. The pause is the entire mechanism.
Example cycle architecture
Example, not recommendation
Reviewing continuously sounds more responsive and is usually worse. Capital authority derived from a drawdown reading that updates on every fill will step down in the middle of a block, leaving some positions sized under one regime and the rest under another — and the resulting record cannot be attributed to either.
A boundary review makes the decision point structural rather than emotional. It arrives on a schedule, it arrives when nothing is open, and it arrives at a moment when the last block's outcome is complete and therefore actually informative. That is a governance property, not a performance one, and it is the reason cycle architectures survive contact with a losing week.
Every win/loss ordering of a four-trade block
Arithmetic illustration| # | Ordering | Wins | Block result | Distinct outcomes at this win count |
|---|
Why sixteen orderings collapse to five outcomes
Inside a concurrent block the ordering is irrelevant to the block's arithmetic — all four positions are alive at the same time, so the sum is the same whichever order they resolve in. Sixteen orderings produce only five distinct block results, distributed one, four, six, four, one. Ordering matters between blocks, not within one, and that is precisely why the review point is placed at the boundary: it is the only place where the sequence has actually done something.
Five modes,
six axes, no free choice.
Every column trades against another. There is no row in this table that is best on all six, and any claim that one exists should be treated as a marketing statement.
| Mode | Utilisation | Compounding | Drawdown | Concentration | Missed opportunity | Governance difficulty |
|---|
Five ways a flow design
fails in production.
None of these is a signal problem. All of them are discovered as one, which is why they take so long to fix.
How the exposure bench works.
Stated in full, because every number on this page is produced by a model whose assumptions are the interesting part.
All five modes and every figure on this page are the same preview engine, evaluated with different values for two controls: the concurrency ceiling and the correlation limit. The strategy is held constant — a pullback continuation archetype on a four-hour trigger with a daily authority timeframe, a single confirmation and a loose signal filter, chosen deliberately so that the signal supply is large enough for capacity to bind at low slot counts.
Effective open risk is the smaller of nominal open risk and the open-risk ceiling. Effective concurrent positions is effective open risk divided by risk per trade, capped at the concurrency ceiling. Effective independent bets applies the standard variance-of-a-correlated-sum result: n positions at pairwise correlation ρ carry the variance of n ÷ (1 + ρ(n − 1)) independent ones. That is a deliberately simple treatment — real correlation is not a single number, is not stable, and is not symmetric across regimes.
Trade frequency is capped by throughput, not by preference. The engine computes how many trades a given slot count and holding period can physically carry per week and refuses everything above it; the refused count is what Figure 1 draws. Modelled drawdown comes from seeded outcome sequences in which trades are grouped into concurrency blocks and correlation pulls the outcomes inside a block toward a shared sign — which is how the model reproduces the clustering effect described at the top of this page rather than merely asserting it.
A model of relationships, not a measurement
This bench answers directional questions — if the correlation limit rises, what else moves and which way. It reads no market data and no broker history, and it cannot tell you what concurrency is right for your account. Every figure is deterministic and seeded, so the same controls always produce the same output; that is a property of the demonstration, not evidence of stability.
Held constant across every figure
Capacity logging — recording refused signals alongside filled ones so that blocked opportunity becomes a measured quantity rather than an inference — is planned production functionality and is not part of this preview.
What this bench
cannot tell you.
The limitations of an exposure model are mostly limitations in how correlation is treated, and they all point the same way: toward underestimating the bad case.
Correlation is one number
The bench applies a single pairwise correlation to every open position. Real books have a structure — clusters that move together and pairs that do not — and a single average conceals exactly the concentration that matters.
Correlation is not stable
The number that matters is correlation during the worst week, not the average. In a risk-off shock it converges toward one across almost everything, which means the diversification this page models is at its weakest precisely when it is needed.
Arrivals are not modelled
Signals are treated as a smooth weekly rate. Real signals arrive in bursts, and a burst is what actually tests a concurrency limit. A model with smooth arrivals understates both the blocking rate and the clustering.
Holding periods are uniform
Every trade is assumed to occupy a slot for the same period. In practice the trades that hold longest are usually the winners, which means slots are disproportionately occupied by the population you most want room for.
No execution queue
A refused signal is simply discarded. Nothing here models a waiting list, a re-entry attempt, or the very common behaviour of taking the trade anyway at reduced size — which is a different system and should be evaluated as one.
Calibration is not published
The cycle sizes, review rules and pool percentages used by the MARS workbook family are part of that product. The structure is described here; the numbers attached to it are not.
Where flow connects
to the rest of the system.
Branch architecture
Slots are what branch weights are spent on. The branch module defines the four management populations whose allocation this page has to fit inside a concurrency ceiling.
Risk and drawdown
Concurrency and correlation are the two controls that most reliably change drawdown depth without changing expectancy at all. The risk observatory is where that shows up.
Monte Carlo core
Sequencing is exactly what a Monte Carlo resequencing test attacks. If a flow design only survives one ordering, it has not survived anything.
MARS intelligence
Open exposure capacity is rank four in the authority stack: active risk is subtracted from the authorised pool before any new trade is sized, and cycle boundaries are where that authority is reviewed.
Development register
Capacity logging is the part of this module that does not exist yet.
Measuring blocked opportunity requires recording the refusal at the moment it happens. That is a data-intake decision, not an analysis one, and it is the design question founding researchers are being asked to weigh in on first.
What is not promised
No launch date, no guaranteed invitation and no performance claim of any kind. Registration records interest and nothing more.