Chapter 11. What Can Be Undone

11.1Two kinds of undoing

This study's ledger codes a decision about risk in two rows, and the difference between those rows is what this chapter reads. On that coding a decision can act twice. One act puts a rule in place — a cap, a budget, a size, a permission — and under that rule something happens: a position is taken, an order rests, a trade is filled. The two rows are a coding choice rather than a proof that every decision about risk works this way; this study reads them as a way of asking what can still be changed once the rule is written, and once the exposure exists. The first act is changed by writing something else, though that can need approval and cost time inside an organisation before it takes effect. The second is changed in the market: an exposure is closed at whatever price offers itself, and only while the market still allows the trade.

This study reads those two acts as separate things, because the ledger reads them separately. Axis A4 is split into two rows. The first, called "reversal constraint", asks whether the decision itself can be changed back: how hard is it to raise the cap again, cancel the permission, rewrite the limit. The second, called "reversal action", asks about the exposure the decision produced: once the position exists, how hard is it to close, to unwind, to call back. As an illustration of the difference rather than a fact from any record: a kill-switch policy might be rewritten in an afternoon, while the orders it left standing could prove harder to recall.

The two rows do not agree often enough to carry much weight. The ledger's numbers report coder agreement at 11 of 53 cases, 21%, for the reversal constraint, and 13 of 53, 25%, for the reversal action. The coder table then gives, for the first row, 9 cases read as easy to reverse, 1 as costly and 1 as unknown, with 42 unsettled; for the second row, 6 costly, 5 unknown, 1 impossible and 1 easy, with 40 unsettled. Each easy, costly or impossible value carries the mark "implied". Agreement this low does not show why the two coders diverged, and the table cannot settle the question. Every value the coders gave the two reversal rows carries the mark "implied", so they were reading the settings as well as their words. What the table shows is that the reversal rows are where the two readings came apart; it does not show careless reading, and it does not name the cause.

This study therefore treats A4 as a question the records mostly leave open, and takes from the six cases in this chapter only what they actually write. Each of the six is present-tense and synthetic; none reports an outcome. The readings below are this study's, and the example judgments quoted are WSTiger's own, dated and revisable.

The six cases are these. Full Kelly from backtest and Stop distance and size are about sizing: what is fixed first, the quantity or the budget. Split entry or averaging and Unfilled order exposure are about positions that are not finished: an entry not planned in advance, and orders neither filled nor cancelled. Realized P&L account value is about a report and the exposure standing behind it. Simulation to limited live is about handing capital to a strategy whose live behaviour has not yet been seen. Each writes an example judgment and a condition for change, and each leaves the outcome open, which is what makes the six usable here: a decision can be read on its reversibility before anything has happened to it.

Reading them on A4 means asking two questions of each. What would have to be done to take back the rule, and what would have to be done to take back the exposure the rule produced. Where a record writes the answer, this chapter reports it. Where a record writes the mechanics of reversal but not its price, this chapter says so. Where the record writes nothing, the silence is left as it is, because the reversal rows of the ledger show how little these records carry on this axis.

11.2Sizing and ruin

Synthetic scene

A trader works from a backtest: the win rate and the payoff ratio it produced, and the perpetual-contract leverage that the Kelly calculation sets. The ratio is computed from those two inputs, and the number that comes back bets forty per cent of the account at once. What is waited for is the record's own test of the estimate — whether the live win rate matches the backtest over a long period — and no result of that test is in the setting yet. The desk's hands are on the sizing fields; the record does not say what else is open, how the calculation is run, or what the next live result will look like.

The two sizing cases in this chapter ask the same question in different clothes: what is fixed first, the quantity or the budget. Full Kelly calculated from the backtest win rate begins from a quantity produced by an estimate.

The record's example judgment keeps the distinction the setting needs:

Kelly's criterion, as its note gives it, is a rule for a gambler with a private source of information: the long-run growth of capital is maximised by staking a fraction set by the edge, and staking more than that fraction lowers long-run growth and raises the risk of ruin (J. L. Kelly Jr., A New Interpretation of Information Rate, 1956). The note gives the fraction as set by the edge, from the gambler's private source of information; the note does not say that the true edge is known without error of its own. What the record adds is the phrase "estimation error". This study reads the case as one where the fraction is computed from a number a backtest produced, and that number may be wrong. The calculation is still Kelly's, but its input is an estimate: if the estimate overstates the edge, the stake can exceed the fraction that would maximise growth from the true probabilities, and then the note's warning applies, since staking more than that fraction lowers long-run growth and raises the risk of ruin. If the estimate understates the edge, the stake is smaller than growth would favour and ruin is further away. The record's example judgment takes the second possibility seriously enough to write down what a slightly worse estimate would cost.

Read on A4, the three options differ in what each can undo. Read on A4, the three options differ in what each can undo. Setting the leverage as calculated takes an exposure now; the record's ledger cell for this case reads "costly" on the undo axis, and changing the size later means closing part of that exposure at whatever price offers itself. The cost of that reversal is not written in the record, and nothing in the setting has yet been lost. Reducing to quarter Kelly keeps the same estimate and takes a smaller exposure; closing it would carry costs too, while changing the sizing rule before the next entry is a rewrite of the rule rather than a trade. A fixed risk limit sets the size from a budget the account chooses rather than from the estimate alone; the budget is a number that can be rewritten before entry, as the other two rules can, and the exposures all three produce are closed in the market. This study reads none of the three as the cheapest to undo, and the ledger's cells do not record what rewriting any of the rules costs.

What none of the three options fixes is when the estimate will be tested, and the record's condition for revisiting the judgment supplies the test: the judgment is to be taken again if the live win rate and the backtest rate match over a long period, and also if cutting the size is shown to harm the target return too much. The second half of that condition is the record's own account of what restraint costs, written in advance. The long period carries no length, and the ledger horizon reads "unknown" while the bearer of the waiting, coded "the-decider" in one reading of the ledger, is a name rather than a role someone could check. The example judgment names maximum drawdown and probability of ruin as recorded quantities for slightly reduced inputs, which is an act of preparation: the desk is writing down in advance what the estimate's error would look like. The record does not say who would read that record, or when.

Stop-Loss Distance and Position Size asks the mirror question, and this time the record itself says what to fix first.

Its example judgment is direct:

Here the restriction — the risk budget — comes first and the quantity follows from it. Read on A4, that ordering changes the shape of the decision's reversibility. A quantity fixed first is a commitment whose undoing is the market's business: closing the position happens at whatever price offers itself. A budget fixed first is a rule the desk can rewrite before entering, while the quantity is recomputed at each entry from whatever budget is then in force. This study reads the two as acts with separate undoings: recomputing a quantity changes nothing that has already been bought, and rewriting the budget changes what the next entry may be. The record forbids raising the budget to make a new quantity fit, which is a written restraint on one of the two acts rather than a cost of reversing either.

Both cases sit in the present and both leave the horizon unset. The ledger codes this record's horizon as unsettled and its undo cell as costly by implication, and the record's own next check asks that excess losses where the stop-loss fill is not guaranteed be checked separately. Nothing in either record names who waits while the estimate is tested. This study reads both records as writing a good deal about the mechanics of reversal and almost nothing about its timing or its bearer.

11.3Positions that wait

Synthetic scene

This study's illustration, drawn from the three records below and not a claim that they share an account: a desk where a position is to be increased after a loss, with the price, quantity and final risk of the additional order not fixed in advance, and where several entry orders remain outstanding in a fast market. In the second record the question is whether those orders can be left out of the account's risk calculation; in the third, only profitable trades are liquidated while loss positions are left open, so realized P&L and account value move in different directions. Hands in a setting like this would be on the order screen and the report, and eyes on the exposure figure. What is waited for across the three is a confirmation, a fill, and a valuation, and each of the three waits on a different clock.

The three cases in this section are about positions that are not finished. Their questions are asked at the moment when an exposure exists and a decision is pending, which makes them the clearest place in the chapter to read A4 and A3 together.

Split Entry and Loss-Averaging asks what makes a second entry a plan rather than a hope.

Its example judgment draws the line at the plan and the cap:

Read on A4, the two acts separate sharply. The second entry, once filled, is a position whose undoing is the market's price; the ledger's cell for this case reads "costly" on the undo axis. The plan and the total-risk cap are rules, and they can be written in advance. The record's condition for change is precise about timing: when a new plan meets independently verified split rules and the total-risk cap, application is considered from subsequent trades onward. This study reads that as a rule about when a new plan takes effect, and not about the exposure already taken; whether the existing position can be undone is a separate question, answered by closing it at whatever the market offers. The two kinds of undoing run on different schedules, and the record writes the plan's schedule without writing the position's.

Read on A3, the record names the evidence it wants — total exposure and average price after the additional order, total loss if the position is finally invalidated — but names no time. The record's horizon is unsettled. The waiting in this case is the interval between the first entry and the second, and the record's own scenario says the position size is to be increased after a loss occurs; the record does not measure that interval.

Potential Exposure of Unfilled Orders turns the same problem around: an exposure that has not happened yet.

Its example judgment answers with what a cancellation request is worth until it is answered:

On A4 this is the cleanest case in the chapter. An unfilled order is an exposure that can be undone by cancelling it, and yet the record refuses to treat the request to cancel as the undoing. Only the confirmation counts. The reversal action here is neither cheap nor impossible in the abstract; what the record shows is that it is pending, and that a pending reversal is not a reversal. The ledger's cell for this record reads "costly" on the undo axis, and its horizon cell reads "days-to-months", by implication.

On A3 the record writes the shape of the waiting without giving it a length. The required evidence includes simultaneous execution scenarios, cancellation confirmation and available margin; the next check asks that current exposure and potential exposure if all orders were executed be displayed side by side. This study reads that display as the place where a confirmed cancellation would show, and as the record's way of keeping the two exposures apart; the record asks for them side by side, and does not say how quickly the display updates. The record does not say how long confirmation typically takes, and this study does not supply a figure the record does not have. What can be said is narrower and enough: the record treats the interval between request and confirmation as a period during which risk remains allocated, and allocates for it.

Realized P&L and Account Value is the case where the two kinds of undoing run in opposite directions.

Here the record's own words carry the reading:

The act being reviewed is not a trade but a report. The ledger's cell for this record reads "easy" on the undo axis, by implication, and this study reads a report as something rewritten at once; the open losses behind it cannot be undone by reporting them differently. Among the reversal values in the set this one sits at the easy end, which is a statement about this record and not a ranking of the others. This study reads the case as a demonstration of the chapter's opening distinction. The rule — what counts as performance — can be rewritten at once. The exposure the rule describes — open positions left unliquidated while profitable trades are closed — can only be closed at a price. The record's invalidation line keeps the same discipline: when net assets reflecting all exposures and costs increase, that is recorded as the performance of the overall account, and the next check asks for the mark-to-market net assets curve beside the realized P&L curve.

On A3 the waiting is the time open losses stay unrealised. The record does not give it a number, and its ledger horizon reads unknown. What it does give is a request for valuation at one point in time: the two curves it asks for would show the unrealised losses beside the account's growth as claimed. The record asks for that valuation to be simultaneous; it does not establish a measurement that cannot fall out of step with the exposure. The records in this section do not say who waits for those curves; the ledger's bearer cells for the three cases read "the-decider", "unsettled" and "unknown".

11.4From simulation to the real thing

Synthetic scene

A strategy has run in simulation, and the record's assumption is that the simulated environment does not fully reproduce the actual execution waiting, order rejection, and cost and margin constraints. The question on the table is how much of the target capital to hand over now that simulated trading is stable. The record's evidence list stands open beside it: the differences between simulated and actual environments, stage-by-stage loss and size limits, and conditions for expansion, stoppage and return. What is waited for is a sufficient number of operations cases, the record's own measure for considering the next stage.

Simulation to limited live trades is the record that makes the chapter's distinction operational.

The three options are the shape of the decision: full investment of the target capital, a restricted risk budget stage, and expansion after reviewing operations indicators.

The record's scenario states the difference the whole decision rests on: the simulated environment does not fully reproduce actual execution waiting, order rejection, and cost and margin constraints. The example judgment keeps strategy hypothesis verification and actual execution verification in separate stages, and makes expansion a later judgment rather than an earlier assumption.

Collingridge's dilemma of control is the idea this case puts on the table. The information that would settle the choice arrives after the technology is committed and after the commitment is costly to reverse, so waiting for perfect information does not resolve the dilemma. What is left is a question about how much to commit before that information is in, and the three options answer it differently.

Read on A4, the three options of this record take different positions on that dilemma. Full investment puts the whole target capital into a live account at once; the ledger's cell for this record reads "costly" on the undo axis, and pulling capital back out would be paid for in executions the market prices. What that option buys is scale from the start: the strategy runs at its target size immediately, nothing is held back for a later stage, and any edge the strategy has is exploited at full size from the first day. The restricted risk budget stage holds that scale back and keeps the decision corrigible, so what the live environment does differently can be observed before the exposure is large; the record's condition for change keeps the stage until, over a sufficient number of operations cases, execution slippage and risk fall within the criteria and control devices operate. What the restraint costs is the deployment it forgoes and the time the stages take, and the record does not say who bears that cost or how long they wait. Expansion after reviewing operations indicators makes the enlargement depend on what the review shows; the record's next verification warns that a stage is passed on operations indicators rather than on accidental initial profits. This study reads the three as a trade between scale now and the ability to change course later. The record gives the staged options a structure and the full-investment option a figure, and it prices neither the scale given up nor the freedom kept.

The record's words for what makes a stage pass keep evidence and outcome separate: the stage is evaluated using operations indicators rather than accidental initial profits. This study reads that as a rule that keeps the decision corrigible, because the enlargement waits on the operations indicators and until they arrive the exposure stays where the stage put it. The record does not say how many operations cases are sufficient, so the rule is one of restraint rather than of measurement.

The waiting in this case is the interval between simulated and real fills, and the record names the elements of it — execution waiting, order rejection, cost and margin constraints — without giving any of them a duration. Its ledger horizon reads unsettled, and no decider is named. What the record does provide is a structure: stages, limits, and conditions for expansion, stoppage and return. The conditions for return are written before anything has happened, which is where this record does more than most in the chapter. They do not yet say what a return would cost, who could order it, or how long it would take.

11.5The value of waiting and H4

Two of the works this study uses treat waiting as worth something in itself: Kenneth J. Arrow and Anthony C. Fisher, "Environmental Preservation, Uncertainty, and Irreversibility" (1974), and Avinash K. Dixit and Robert S. Pindyck, Investment under Uncertainty (1994). What the first names is later called quasi-option value; the second gives an option value to delaying an investment that cannot be undone. Both put at the centre information that has not arrived yet, and this chapter uses them in one way only: to ask what these records write about the time between a commitment and the information that would test it.

Both arguments assume something the records in this chapter rarely supply: a decision that is both irreversible and genuinely waitable. Simulation to limited live trades can wait because there is a stage to occupy while the waiting happens, and the waiting is paid for by a smaller budget rather than by no activity. Unfilled orders rest in a state that resembles waiting, but the resemblance has a limit that the record itself supplies: an outstanding order remains executable, and the record allocates risk to it until cancellation is confirmed. This study reads the case as only partly like the option of delaying a commitment, since the exposure can arrive without anyone deciding again. In Full Kelly from backtest win rate, by contrast, the record does not establish that waiting is available to this trader at all. Its three options are the size as calculated, a smaller size, or a fixed limit; none of them is the size zero, and this study does not add it. The record's own test of the estimate needs live results, and not trading would not produce them.

Where the literature and the records part company is in what waiting costs and who pays it. The quasi-option value in Arrow and Fisher is the value of keeping an option open while information is still to come; this study reads the question of who holds the option as one these records leave open. A theory of the value of waiting that rests on information arriving later needs to know who will be there to use it. These records do not write that.

On A4 the records are uneven in an instructive way. Where the decision is a rule — the risk budget, the total-risk cap, the report of performance, the staged transition — the records write the rule but not its price, and every easy, costly or impossible cell in the two reversal rows is marked implied. Where the decision is an exposure — a filled position, a resting order, a live commitment of capital — the records describe the mechanics of reversal in more detail than they give its price. This study reads the contrast as a property of how these records are written: they are clearer about the shape of an act than about what unmaking it costs.

The ledger's own verdicts on H4 are the study's verdicts, and here they are unhelpful in a way the reader should see plainly. On the strict reading, H4 is not applicable: supports 0, against 0, not applicable 53, unclear 0. On the lenient reading it reads "against", with supports 1, against 3, unclear 3, and not applicable 46; the ledger gives the share as 25%.

Why so few records let the hypothesis be judged. Two reasons, both visible in the ledger's numbers rather than in anything this study has added. First, the form has no field for reversal, and every value the coders gave the two reversal rows is marked implied; that is where the two readings came apart, at 21% and 25% agreement on those rows. Second, H4 needs a pair in each record — a decider and a bearer, far enough apart — and the ledger's decider and bearer rows leave 20 and 29 cases unsettled respectively, so the records cannot be sorted by the distance the hypothesis asks about. The lenient verdict of "against" rests on one support and three disagreements among the cases the hypothesis applies to.

What the six cases of this chapter do offer, against that silence, is a set of instances where the two kinds of undoing have been written down in advance. The quantity that may be re-determined while the risk budget stays as it is. The plan that changes subsequent trades and not the last one. The cancellation request that is not a reversal until confirmed. The report that can be rewritten while the exposure behind it cannot. The stage that keeps a commitment corrigible until what it does differently has been observed. The ledger's own case-level readings put some of these against H4 on the lenient reading and leave others unclear where the hypothesis applies; this chapter claims no more for them than that. Each is a place where a decision record said something about reversal that its own ledger could not turn into a count, and that gap is what this chapter can stand behind.

The time inside these records stays open: the interval between request and confirmation, the time before live results, and the stages have no duration written into them. Chapter 12 reads the records for the units of time they do use, and for what changes when the same setting is written in the past, the present or the future.