Chapter 6. Models Everyone Uses: Measures, Monoculture and Performativity

6.1Models that change what they measure

This chapter reads a shared measure as a number or a rule adopted for one purpose and then used for more than that purpose. A bank sets limits on it; a portfolio manager sizes positions by it; a supervisor asks banks to report against it. In that reading, a measure begins as a description and can end as part of the machinery it was built to describe. The four records do not establish that history for their own measures, so the definition stands as this chapter's own, offered for reading them rather than as a finding about them. Each record is written from inside an organisation deciding what a measure should cover, and only one of them has a measure proposed to several banks at once. The question the records put is narrower than "is the model good": it is what the measure covers, what the risk does anyway, and what would show the cover failing.

Three ideas from the reading notes sit behind the chapter. The first is performativity. Donald MacKenzie's An Engine, Not a Camera: How Financial Models Shape Markets (2006) holds that financial models, above all option pricing, can change the markets they describe, not only record them. The model is not a camera pointed at a market that would look the same without it; the title's image, an engine rather than a camera, marks the difference. This study reads the point as extending to any measure that becomes widely adopted: it is then part of the world it was built to measure. What the note does not give is a direction or a size for that change, so the four records below are read for the shape of the choice, not for a measured effect on a market.

The second idea is endogenous risk. Jón Daníelsson's The Emperor Has No Clothes: Limits to Risk Modelling (2002) holds that risk models such as value-at-risk perform worst in the crises they are meant for, and that when many firms use them under regulation their use changes the market, because risk is endogenous. The note's second clause is the one that matters here: it is not only that a model fails when the world turns violent, but that the shared use of the model changes the world. The note concludes that such use can deepen the crises the models measure. Two questions therefore sit apart in what follows: how well a measure describes the position it counts, and how the market responds to many parties counting the same way. Only the first is a question of accuracy.

The third idea is complexity and coupling. Its source is Richard Bookstaber, whose 2007 book is titled A Demon of Our Own Design; those words of the title belong to the work's own naming, not to the voice of this study. In that book, markets become complex and tightly coupled through their own innovations, and liquidity vanishes when it is most needed. The phrase tight coupling also appears in Charles Perrow's note, where it is one of two independent dimensions that make accidents normal in some systems; Chapter 5 sets out Perrow's two dimensions, and this chapter uses Bookstaber's version, which is about markets rather than about organisations. Bookstaber's addition is the disappearance of liquidity at the moment it is required, which is a claim about timing, not only about structure.

This study reads the three ideas together as two questions to bring to four records, not as a claim the works establish: what a shared measure counts in ordinary times, and whether its users move in the same direction when conditions change sharply. The records hold no comparison of accurate and inaccurate measures and no observation of many users at once, so neither question is settled by them; each can only be asked of a setting.

One historical example is available to the reading notes, and only at the level the notes give: Roger Lowenstein's When Genius Failed (2000) on Long-Term Capital Management. The note gives no model, no trades and no causes beyond the sequence of losses and recapitalisation it names, so this study mentions the case only at that level and adds nothing to it. This study reads that sequence as one in which the market-level question and the firm-level question cannot be separated, which is why it is worth one sentence in a chapter on shared measures. Beyond that, this study does not use the example.

This study's own place in the argument is narrow. The four records in 6.2 are not market histories and do not report what happened after any choice. Each is a synthetic setting in which one organisation decides, before any answer arrives, how to handle a measure. What they show is how such a choice is written down: what the measure counts, what the record names as evidence, and what condition would show the measure failing. What they cannot show is whether any shared measure raised or lowered systemic risk, because these are made settings rather than observed markets, and no outcome follows any choice in them.

6.2Four records of shared measures

Four records are read here, in the order the plan gives them: a past setting about a supervisory measure, a future setting about a learning mandate, and two present settings about diversification counted in two different ways. Each is read on two axes: A2, the grounds a decision rests on and what it leaves out; and A5, the condition for revisiting the judgment and whether a reader could observe it. Chapter 2 sets out the axes; this part applies them to four records. A2 is read here from the question, the options and the scenario text of each case, and A5 from the condition each record writes for itself. Each reading below states what an option buys and what it puts at risk, and who would be exposed while the answer is not in.

6.2.1 One measure all banks

Synthetic scene. The setting is a hypothetical bank supervisory authority and the banks it supervises, dated in the record as 1996. A recommendation has arrived: large banks should set trading limits using the same statistical risk measure. The banks that receive it must answer before any volatility spike arrives, and the record gives three answers to weigh: use the measure as it is, review in-house stress scenarios alongside it, or set breach responses differently for each bank. When volatility spikes, the measured values of all banks rise at the same time and the limits are breached; what is waited for is a decision taken before that moment, about a measure that every receiving bank would hold in common.

The record's question is long because it asks three things at once:

The example judgment separates two things that a shared measure tends to blur:

The record's scenario states that when volatility spikes, the measured values of all banks rise at the same time and the limits are breached. On A2, what the shared measure leaves out is therefore not accuracy but coverage of other banks' behaviour: a bank can measure its own position correctly and still be standing in a room where everyone's limits fire together. The three options divide accordingly. Using the recommended measure as it is buys simplicity, comparability and compliance with a supervisory expectation; it puts at risk the bank's ability to see an event the measure was not built to capture. Reviewing in-house stress scenarios in parallel buys a second view of the same exposure; the record names no cost in time or staff and no deadline, so the conditions under which the second view would be set aside are left open here. Differentiating breach responses buys the ability to act differently from the rest of the market; it puts at risk the uniformity a recommendation written for large banks alike is built on. The record gives no reason to think that a differentiated response could not be stated in advance, and this chapter claims none.

On A5, the record's own revisit condition names what would make the judgment be revised: records accumulating that measurement methods differ enough among banks, or that simultaneous limit breaches did not lead to market sales. Those tests bear on the example judgment's concern with synchronised responses; the second would weaken the claim that simultaneous breaches produce market sales, and neither shows the statistical measure failing on its own terms. This study reads the condition as naming the gap the example judgment names, a market-wide one that no single bank's internal data could settle, and it is worded so that a reader could watch for it. The record also carries an Insight line, which states the general form of the problem:

6.2.2 Self-retraining mandate

Synthetic scene. The setting is hypothetical and dated 2031. An execution agent retrains overnight within a bounded mandate, and its feature weights have drifted toward a concentrated regime. Nothing in the record shows a person watching that drift as it happens, and no desk, document or console is named beside it. What is waited for is a decision about the night's learning: whether the mandate covers the learning rule, whether the drift needs a bound, and whether any human being is required to look at it before the next night's job runs.

The record asks a short question with three options:

The example judgment treats the mandate as something wider than a set of position limits:

On A2, what a position-limit reading leaves out is the time dimension of the measure. The example judgment states that a mandate covers the learning rule as well as the position limits, and the scenario shows the feature weights drifting toward a concentrated regime inside the bounded mandate. This study reads that drift as a change in the rule that produces the positions, which a measure aimed only at position limits would not show. The options divide on who bears that. Freezing retraining stops the drift and holds the weights as they are; the record has already described the profile as drifting, so this chapter does not claim that a freeze keeps an approved profile in force. This study reads what the freeze gives up as the learning the mandate was granted for. The record names neither a regime the frozen weights could miss nor any examination of the frozen state, so both are left open, as is who pays for holding still. Constraining the drift buys a bounded learning process that keeps running; it costs the organisation the work of defining the bound and monitoring whether it is being approached. Requiring nightly human approval buys a human check on each night's update. The record gives no cost in speed and no repeated situation in which nothing appears different, so the conditions under which that check would fail are left open; what the record does fix is that the check would come each night, before the next job runs.

On A5, what would show the measure failing is written in the record's revisit condition: whether bounded retraining with observable weight limits restores the approved profile. This study reads that condition as naming two requirements at once, a bound and observability, and notes that the second is what makes the first testable; without a definition of the approved profile, a reader could not tell a restoration from a drift. The record does not supply that definition, and this study does not add one.

6.2.3 Strategy count diversification

Synthetic scene. The setting is hypothetical and dated 2026. A set of trading strategies is in view, each carrying its own name. The record names the evidence a decision would rest on: common factor exposure, concurrent losses over periods of sharp moves, and portfolio-level collateral and liquidation requirements, and its next check asks that the correlation table of normal periods be separated from that of periods of concentrated losses. What is waited on is whether the number of named strategies is to be counted as risk spread, or whether behaviour under sharp moves is to be examined before any number is treated as diversification.

The record's question is phrased as a question about a measure:

The example judgment answers by changing the unit of observation:

The record's scenario states the assumption plainly: correlation of returns is usually low, but on days of sharp moves most losses occur in the same direction. On A2, what the count-based measure leaves out is exactly that pattern. A count is easy to compute, easy to report and compatible with a budget allocated by strategy name; it cannot show that strategies holding the same liquidity factor are one exposure under several names. The option of recognising diversification by strategy count buys speed and a clean budget line; it puts at risk the portfolio in the one condition the scenario names. A stress correlation check buys a view of behaviour under sharp moves; it costs time and produces a table that may be set aside when the count is already on the report. Risk aggregation management buys the view that the portfolio's risks are to be managed together. The record does not say that this costs the ability to attribute performance and loss to individual strategies, and a portfolio can be aggregated and still attributed strategy by strategy, so this chapter names no such sacrifice.

On A5, the record writes its own test: the change condition says the offsetting effect would be recognised if it is maintained in both independent periods and stress conditions, and the next check asks for the correlation table of normal periods to be separated from the correlation table of periods of concentrated losses. This study reads that as one of the more concrete conditions in the chapter, because it names two tables and two periods rather than a general improvement.

6.2.4 Exposure across accounts

Synthetic scene. The setting is hypothetical and dated 2026. The same asset is exposed in the same direction across several exchanges, and the reports available are by account, showing quantity, direction and notional exposure. The record names as further evidence collateral and risk limits by exchange, and transfer, withdrawal and access constraints. Its next verification asks for an aggregate risk table for all accounts in addition to the reports by account, so the table of the whole is named as something to be produced rather than as something already missing. What is waited on is whether the split across exchanges counts as risk spread.

The record asks:

The example judgment keeps two kinds of diversification apart:

The record's scenario states the situation as an assumption: accounts and exchanges differ, but the same asset is exposed in the same direction; and it names required evidence, being quantity, direction and notional exposure by account, collateral and risk limits by exchange, and transfer, withdrawal and access constraints. On A2, what the account-level view leaves out is the direction of the whole. The option of recognising exchange diversification buys attention to the risks the record asks to be assessed separately by venue: collateral segregation by exchange, transfer constraints and access disruption. How far those protections differ by venue is not in the record, and recognising a description does not create them. It puts at risk the count of the price exposure: the record states that the same asset is exposed in the same direction across the exchanges, and this study reads a split counted as diversification as leaving that exposure out of the count. An aggregate exposure calculation buys a single number for the direction. The record's next verification keeps the reports by account beside the aggregate table, and its change condition keeps total risk by exchange and collateral demand, so the aggregate is written as an addition to the detail rather than a replacement for it. Checking whether offsets apply buys an answer to the only condition under which splitting could reduce price risk; it costs time before the answer arrives, and in the meantime the exposure stands as it is.

On A5, the record's change condition asks for economically offsetting exposure actually to be verified, and the next verification is a second table: an aggregate risk table for all accounts, in addition to the reports by account. This study reads the next verification as naming an object a reader could check, since the table either exists or it does not. The change condition asks for economically offsetting exposure actually to be verified, and a table could exist without settling that, so the two are different tests of different things.

Taken together, the four records share a shape. In each, a measure is already in use, and the question is not whether the measure is accurate but what it counts. In each, the options divide over what the measure should count, and the records do not write down what each option gives up. Where this chapter names such a cost, it is the chapter's own reading, marked as such, not a fact the record carries. In none of them does the record write who waits while the choice is made, and in none does it write a deadline.

6.3The debate: do shared models stabilise or destabilise?

The question this chapter's readings leave open is whether one model used by many parties makes the system steadier or less steady. It is a question about the direction of an effect, and the four records hold no outcome with which to measure one, so what follows is a reading of the notes beside the records rather than a test. The reading notes give the tools for one side of that question clearly, and a difficulty for the other side that this study must state rather than paper over.

Daníelsson's note gives a direction of its own: when many firms use risk models under regulation their use changes the market, because risk is endogenous, and such use can deepen the very crises the models measure. MacKenzie's note gives a possibility without a direction: a model can change the market it describes, not only record it. This study reads the two as meeting at one point, that a shared adoption is not a neutral act of measurement, and marks that reading as its own rather than as either note's. Bookstaber's note supplies a mechanism at the level of markets: innovations make them more complex and more tightly coupled, and liquidity disappears when it is most needed. The step from that to shared measures is this study's own reading: a shared measure is most load-bearing at the moment the shared environment is least able to bear it. Taken together, this chapter reads the three notes as a position worth putting to the four records, that a widely shared measure can change both the market and the moment when the measure is relied on most.

The other side of the debate is named in the plan as Andrew W. Lo's adaptive markets. A reading note is the limit of what this study may claim, and the note held for Lo supplies no adaptation argument, so none is put in Lo's mouth here. The idea of adaptation can still be used in a narrow form that the records themselves raise: a market in which users of a shared measure respond to it, so that measure and users change together. That form is this chapter's own construction, not Lo's. The plan places the adaptive-markets debate in Chapter 4, where this study reads Lo as set beside Minsky and Kindleberger on the question of whether instability is made inside markets or shows that they adapt; that chapter is the place where the reader will find the fuller treatment. What this chapter can do is use the idea of adaptation as the records themselves raise it, and see whether the records bear on the question at all.

On that point the records do say something, and it is the more interesting direction. Two of the four records write revisiting conditions that concern change in the measure itself. One measure all banks names a condition under which its judgment would be read again: records accumulating that measurement methods differ among banks, or that simultaneous limit breaches did not lead to market sales. Strategy count diversification names another: the offsetting effect maintained in both independent periods and stress conditions. Self-retraining mandate writes a condition about a learning rule restoring an approved profile, and exposure across accounts writes a condition about an aggregate table being produced where only per-account reports exist. None of these conditions predicts that its trigger will occur; each states what would have to appear for the judgment to be revisited. This study reads them as the records' way of setting a later reading rather than as evidence that a market adapts to a model, because a conditional reconsideration rule does not forecast its own trigger, and a later reassessment and a market adaptation answer different questions.

Where the records bear on the debate is therefore at the level of one organisation's choice inside a shared measure. Each of the four is written as one organisation deciding how a measure should be used: the banks answering a supervisory recommendation, the party that sets a learning mandate, the party that counts portfolio risk, the party that spreads one asset across exchanges. Only the first establishes that others use the same measure, since the recommendation is addressed to large banks alike; in the other three, whether anyone else uses the same measure is not in the record. That is the level at which a decision record can frame a question about performativity, because the choice being made is about how a measure will be applied locally. Whether the measure then changes the market it describes cannot be observed here, since these records hold no market response and no outcome. The example judgment in One measure all banks makes the point directly, and its Insight line states the general form.

Where the records cannot bear on the debate is the level above. These are synthetic settings rather than observed markets, and none of them reports an outcome after any choice. One measure all banks does name parties beyond a single organisation, a supervisory authority, the banks it recommends to and possible market sales, but only inside the setting. Nothing here can show whether a shared measure raised or lowered systemic risk, whether use under regulation deepened or damped a crisis, or whether markets learned around a model over time. The one record that speaks at market level, One measure all banks, does so through an example judgment that estimates rather than observes: it estimates the scale of market sales that would emerge when limit breaches occur simultaneously. This study reads that sentence as the record's own modelling of the question, and treats it as a claim inside a synthetic setting, not as evidence about markets. Two further limits deserve stating. The four records were selected by topic for this study and are not a sample of how organisations use shared measures; and none of the four admits a gap of its own, so every gap named in 6.2 was read from the question, the options and the scenario text rather than taken from a record's own admission.

A final point about the framing of the debate, which this chapter will not resolve and does not try to. The two sides are not symmetric in what they claim about time. Daníelsson's note puts the model's failure in the crisis, the moment the models are meant for; Bookstaber's note puts the liquidity loss at the moment it is most needed; the records' conditions set a later reading of the same question. What differs across the three is where the stress falls, not when an answer is available, since crisis-time model weakness, crisis-time liquidity loss and a record's later reconsideration condition can stand together. The question of when an answer is available remains the one this study asks of every record. A record that names a later reading is not thereby evidence that a later reading will arrive in time to matter.

6.4Hands on: judging a decision before and after its answer

The exercise for this chapter uses the four records in a form a reader can do without any market data. It has three parts.

First, before looking at the example judgment, take the question of one record, Strategy count diversification, and write down what would count as the measure failing, in terms a reader could check. The record's own condition asks for offsetting effects to be maintained in both independent periods and in stress conditions, and for a separation of the normal-period correlation table from the concentrated-loss period correlation table. Write down how much of that a reader could check, and what would have to be defined before the rest became checkable. Then do the same for one of the other three records.

Second, write what the measure leaves out. Do not write what the model misses in the abstract; write what the record itself names. For One measure all banks, the record's scenario names the simultaneous rise in measured values; for Self-retraining mandate, the drift of feature weights; for Exposure across accounts, the direction exposure by account and the constraints on transfer and access. This part of the exercise is a check on whether a reader can tell coverage from accuracy, which is the distinction the example judgments in this chapter keep making.

Third, write the two readings apart. Before the answer, write what each option buys and what it puts at risk, as 6.2 has done for each of the four. Then imagine that an answer has arrived, in whichever direction, and write again how you would assess the choice now. Notice what changes and what does not. Chapter 7 treats exactly that shift: how an outcome changes the judgment of a decision made on the same information, and what would count as evidence about a decision rather than about its result. This chapter cannot settle that question, and the records here have no outcomes to settle it with.

Two further exercises are available to a reader who wants to connect this chapter to the rest of the study. The first is a timing exercise: for each of the four records, write down who would need to see the evidence named in the revisit condition, and how long before that evidence appears. Chapter 10 reads the ledger's row on who waits and pays for a decision. The second is a comparison: take the four records and ask what a record would have to hold for the shared-measure question to be answerable at all, meaning a deadline, a named decider, a second party, or an outcome. The ledger cannot answer that question in advance. Its fields carry no deadline and no named decider for any of the 53 cases, which is a limit of the record form rather than a finding about the settings, and the settings themselves name roles in their text. These four records in any case report no outcome after any choice, because they are written as settings decided before an answer arrives. What a reader is left judging, in this chapter and in the one that follows, is a decision written before its answer, with the measure already in use and the condition for revisiting already set.