Capstone: The System Card

What goes on the card

A system card is one page, and every line on it exists to be checkable by somebody else. It opens with the *edge hypothesis* in plain language: what the setup is, why it should pay, and who is on the other side of the trade and why they would keep doing it. That last clause is the one that separates a system from a pattern hunt, because a durable edge has a counterparty with a reason — an index fund that must buy regardless of price, a hedger who is indifferent to the level, a participant who is forced to sell into a decline. Then the *rules*, stated precisely enough that a stranger could execute them: the setup, the entry trigger, the stop at invalidation, the target or trail rule, the filters that refuse the setup, the instruments it applies to and the timeframe. Then the *sample*: the dates, the number of trades, and the conditions inside it — how many trending and ranging months, what happened to the index, whether a crash is included. The third block is the *statistics*, reported with their context. Expectancy in R with the break-even win rate beside it; the observed average win and loss; the win rate and how much room it has; the maximum drawdown in R and in account terms at the stated risk fraction; the risk-adjusted return; and the trades per optimised parameter, which is the honesty line on the whole document. If the parameter count is high relative to the sample, the card says so, and the expectancy is described as an in-sample estimate rather than a promise. Out-of-sample numbers, where any exist, are reported with the same formality — an out-of-sample collapse is a finding to record, not an embarrassment to omit. One page, six blocks — Edge hypothesis: The setup, why it should pay, and the counterparty with a reason to keep supplying it ← · Rules: Setup, trigger, stop at invalidation, trail or target, filters, instruments, timeframe · Sample: Dates, trades, and the regimes inside the window · Statistics: Expectancy versus break-even, average win and loss, drawdown in R and in account terms ← · Honesty line: Trades per optimised parameter, and whether the figures are in-sample estimates · Forward protocol: Fixed sample size, fixed rules, no changes until the sample completes The counterparty clause is the part that ages best. Patterns change and regimes rotate, but an edge explained by who is forced to trade survives longer than one explained by a shape.

The protocol, and why it is the hardest part

The forward-test protocol is three sentences: the rules are frozen until the sample completes, the sample size is fixed in advance, and the only permitted changes are those justified by a mechanism rather than by a result. It is hard to keep because the temptation arrives in the middle of the sample, when the numbers are disappointing and a small tweak feels like improvement. That tweak is the whole problem: it restarts the sample, resets the evidence and converts the forward test into a second backtest with a smaller dataset and the same biases. Writing the sample size down before the first trade is what makes the eventual answer interpretable — twenty trades cannot distinguish a 0.40R edge from nothing, and no amount of re-reading the same twenty trades will change that. The pragmatic answer to an underpowered sample is not a longer single test but a portfolio of tests: if one setup needs two hundred trades to be measurable at the size it trades, the trader who runs three unrelated setups gathers evidence three times as fast, which is an argument for breadth of methods rather than for larger positions in one. It is also an argument for taking the sizing from the risk budget rather than from the confidence, since the whole point of the protocol is that the size does not change while the evidence accumulates. The card closes with the *tolerance*, in the trader’s own terms. What is the drawdown at which this method stops being traded, what is the loss streak that would produce it at the stated risk fraction, and what does a month of ordinary variance look like so that it is not mistaken for a breakdown. That number belongs on the card because it is the one a trader cannot compute well in the moment: at a 40% win rate with a 3:1 payoff, six consecutive losses is a routine event whose frequency can be calculated in advance, and the version of the trader who planned for it is the one still trading the method when it starts working. • Freeze the rules until the sample completes, and fix the sample size before the first trade. • Permit changes only when a mechanism justifies them, not when a result disappoints. • Widen the evidence by running unrelated setups, not by increasing the size on one. • Pre-compute the loss streak that produces the tolerance drawdown, so variance is recognised as variance. • Record the out-of-sample result with the same formality as the in-sample one. The subtlest failure of a capstone is a card that describes the system the trader wishes they ran rather than the one being traded. The test is mechanical: take the last twenty journal entries and check each one against the written rules. Trades that fall outside the specification are either evidence against the rules or evidence that the rules are not being followed, and the two need different responses — but neither can be found on a card that was written from memory.

One budget, several systems

Everything above describes a single method, and the moment a trader runs two the arithmetic changes. Positions share one account and one risk budget, so the question stops being “what does this setup risk” and becomes “what does the book risk when several setups fire at once”. That question has no answer on a single-system card, which is why the card gets a portfolio annex once there is more than one method in it. Start with the obvious part. If each position risks half a percent and the rules permit five concurrent positions, the account is exposed to two and a half percent of loss on a day when every stop fills — and the extreme day is the day all of them gap together, because that is what a market-wide event does. A trader who wrote “0.5% per trade” and never multiplied it has not sized the book; they have sized each brick and ignored the wall. Then the part that is easy to get wrong even when the arithmetic is right: signals are correlated even when the instruments are not. Two setups that both trigger on a momentum breakout, or on a failed breakdown, or on strength in the same sector, are one bet expressed twice. The test is not the correlation of the stocks; it is the correlation of the trade *outcomes* — take the R-multiples from each system over the same period and correlate those. Two methods at 0.8 outcome correlation are one method with double size, and they should be budgeted as one. The arithmetic of combining genuinely unrelated methods is friendlier than intuition suggests and less friendly than hope. Independent systems do not add their risks linearly: combining k uncorrelated methods at the same per-trade risk raises the portfolio’s volatility roughly with the square root of k, so a portfolio of methods is better diversified than a portfolio of trades — but the drawdowns still overlap in time, because disorderly markets hurt trend, carry and mean-reversion together. The honest conclusion is that diversification between methods reduces, but does not remove, the account-level drawdown. That is why allocation between systems should not follow recent performance. Weighting by the last quarter is performance chasing one level up: the method that just ran hot is the one whose returns are most likely to have been flattered by the regime that is ending, and the method that has been quiet is the one whose setup is now being supplied. Fixed weights, set when the card was written, with a stated review date, keep the decision out of the moment where it is most expensive. The last piece is the kill rule at the portfolio level, and it is not the sum of the individual ones. Each method has its own tolerance — the drawdown at which it stops being traded — but the account also has one, and if four methods each tolerate an 11% drawdown, the account can reach a far deeper level while every system is inside its own limit. Decide the account tolerance first, then back out what each system may risk so that the combined worst case sits inside it. This is the same arithmetic as position sizing, applied one level up, and it is the last thing on the card for a reason: it is the constraint that makes the whole document survivable. • Per-trade risk × concurrent positions = the heat the book actually carries. • Correlate trade *outcomes*, not instruments — two methods that fire on the same event are one method with double size. • Independent methods diversify volatility sublinearly and overlap in drawdowns; they reduce the worst case, never remove it. • Fix weights when the card is written. Reweighting by recent performance is chasing noise at the system level. • Set the account tolerance first, then derive each system’s risk so the combined worst case fits inside it. A single ledger with one row per system — trades, expectancy in R, current drawdown — is enough to see when the *portfolio* drawdown is approaching its limit, which is the number no individual system can tell you.

Sources

• Pre-registration, the deflated Sharpe ratio and honest reportingBailey & López de Prado, “The Deflated Sharpe Ratio”
• Position sizing, expectancy and the drawdown a trader can holdVan Tharp, “Trade Your Way to Financial Freedom”
• Implementation intentions and written trading plansBehavioural finance literature on pre-commitment, as applied in Market Psychology PS10