Chapter 3. The Measured and the Unmeasured: Knight, Keynes and Their Readers

3.1Knight's two uncertainties

Frank H. Knight's Risk, Uncertainty and Profit (1921) gave the risk literature the pair of words it still uses. Chapter VII of that book separates measurable uncertainty from unmeasurable uncertainty. Chapter VIII, whose title is Structures and Methods for Meeting Uncertainty, names both terms, the first for the measurable kind and the second for the unmeasurable one: "To preserve the distinction which has been drawn in the last chapter between the measurable uncertainty and an unmeasurable one we may use the term 'risk' to designate the former and the term 'uncertainty' for the latter."

The whole difference rests on one word: measurable. What follows from Knight's Chapters VII and VIII is a claim about what can be done with each kind. Insurance and grouping deal with risk. Profit, for Knight, is tied to bearing true uncertainty, which cannot be insured.

This study reads grouping as the practice of treating many instances of a case as instances of the same class, so that the losses of one are averaged against the losses of another. Insurance arranges the same idea as a contract: the exposure is transferred to a party that prices it from the class. Where instances sort, the count of the past carries the weight of the future. Where instances do not sort, no count carries that weight, and whoever acts does so on judgment rather than on a class behind the case. Knight ties profit to bearing true uncertainty, which cannot be insured; this study does not read that tie as a payment guaranteed to whoever bears the uncertainty.

Read this way, Knight's distinction is a distinction about what a decision-maker can stand on. Under his risk, a probability can be had for the case at hand, and this study reads the sorting of instances into classes as what makes that possible; such a class need not be built from past frequencies alone. Under his uncertainty, no class holds the case, and the decision is carried by judgment. A judgment of that kind can still be written as a number; what it lacks is a class of like cases behind it.

This chapter uses Knight's distinction to organise decision records, not to interpret his book for scholars.

The distinction also raises a question about arrangements. Insurance and grouping are the methods Knight names for dealing with risk, and this study reads the bearing of uncertainty as what remains for whoever decides when those methods do not reach the case. His treatment includes organisational methods for meeting uncertainty, so nothing in the distinction itself says that such a case is borne alone or that the two methods dispense with judgment. What the two words really turn on — measurement, or the market for the risk, or the ability to class instances — is what later readers of Knight have argued about, and Section 3.4 takes up that argument.

3.2Keynes on probability and weight

John Maynard Keynes's A Treatise on Probability (1921) is the other foundation this chapter uses, and it is built on a different question from Knight's. Not every relation between evidence and conclusion supports a number: some do, and others do not. Alongside probability, Keynes holds that evidence has weight, and that weight is not the same thing as probability. Two bodies of evidence can stand behind one and the same probability, one extensive and one thin, and the difference between them is a difference in weight: how much of the relevant evidence is at hand, not how strongly that evidence supports the conclusion.

The older problem underneath both Knight and Keynes is written down by David Hume in An Enquiry Concerning Human Understanding (1748). In Section IV, Part II, Hume states the foundation that every inference from experience carries: "For all inferences from experience suppose, as their foundation, that the future will resemble the past, and that similar powers will be conjoined with similar sensible qualities." Section V of the same enquiry answers the question Hume raises there: it is custom, not reasoning, that carries the mind from the past to the future. So the step from counted cases to the next case cannot be proved by the count itself. Whatever supports it supports it from somewhere else.

Put together, the three make one problem visible. Knight asks whether a case belongs to a class whose instances can be counted. Keynes asks how much evidence stands behind the relation drawn from what has been observed, and whether that evidence has any weight at all. Hume asks what licenses the move from the instances one has to the instance one is facing, and answers that the licence is habit rather than proof. For a decision record, the practical form of the problem is simple to state: what is this decision standing on — a number, a class, an argument, or the fact that things have gone this way before?

Weight gives that question a second edge. A record can carry a probability and still be thin: the evidence behind the statement may cover a short run of cases, one market, or one setting of prices, and the statement reads the same either way. Two records can agree on the number and differ in what stands behind it, and the difference matters to whoever must act on the number before the answer comes. This study reads weight as the part of Keynes's account a decision record handles worst: a field can hold a probability, but nothing in the form asks how much evidence stands behind it, and nothing in the four settings of this chapter says how large that body of evidence is.

Hume's answer reaches the same settings from the other side. Each of them pairs a rule or a plan fitted on what came before with a present case that may not join it. The rule was fitted where the future resembled the past enough for the fitting to work, and the question inside the setting is whether that resemblance still holds at the moment of deciding. Custom is what carries the mind from the earlier cases to this one, and custom does not certify that this case belongs with them. Whether it belongs with them is the question each of the four settings poses in its own way.

3.3Four records at the edge of the measurable

Each of the four settings below is a synthetic setting published by WSTiger for study; none reports a real desk, firm or trade. Each example judgment is WSTiger's reading, dated and revisable. The scenes are this study's rendering of the setting the record states, and the reading that follows each scene is this study's, not the studio's.

3.3.1 Confidence score as size

Synthetic scene.

The setting is a synthetic future risk engine published by WSTiger for study, standing between agents and capital. Each agent submits a confidence score, and the engine reads the number and sets the order size in proportion to it. The agents know from the training process that the more confidence they report, the more capital they receive, so the hands that submit the numbers have an interest in how large the numbers are. There is nothing else in the setting for an eye to rest on: no hit record is named in it, only the scores the agents report and the sizes the engine hands back. What is waited for is the post hoc record of hits, against which a reported score could be set beside an actual hit rate.

WSTiger's own reading in the record separates two things that look alike on a screen.

The record's thread adds the point that governs the time question.

The record also names the conditions under which the matter would be judged again.

On A2, what the record rests on is the gap between a reported score and a verified probability, and the incentive that makes an agent inflate the score. What it leaves out is anything a reader could use to plan the check: no threshold for what counts as inflation, no statement of how many outcomes would make a hit rate mean something, no decider named, and no item anywhere in the record's 53 siblings listing what is missing. The ledger's A2 row records that: cases that list what is missing, 0 of 53.

On A3, the ledger's horizon cell for this record reads unknown, and the bearer cell is unsettled. The answer here is not an event that arrives on a calendar. It arrives when enough of an agent's calls have resolved for the score to be lined up against what actually happened. The record does not say how long that is, or how many cases it needs.

The confidence score also puts Keynes's weight into a plain form. A hit rate is a probability statement with a body of evidence behind it, and that body can be long or short: a small run of calls and a long run of calls both produce a rate, and the two rates do not stand on the same evidence. Calibrating scores against post hoc hit records is a way of letting the body grow before the rate is used for size; nothing in the setting says how large it has to grow first.

Which of Knight's two kinds does this setting describe? If an agent's reported scores and its outcomes can be lined up, and if its calls are alike enough to form a class and to say something about the calls still to come, the score becomes a measurable thing — instances that can be counted, a class that can be grouped. But whether one agent's calls form a class with another's, and whether the calls of the future resemble those of the past, is left open. The record does not let a reader tell whether the count will ever reach the point where the score is more than a reported number. Chapter 7 asks what an outcome, once it comes, is evidence of.

3.3.2 Timeframes in collision

A strategy in 2026 runs on a system that carries more than one timeframe. One screen holds the longer structure, which is in an uptrend; the other holds the execution timeframe, which is in a downtrend. The strategy document states what each timeframe is for, and the entry rules live there. A person at a desk looks from one screen to the other and back. The expected holding period and the stop-loss criteria sit in the same document. What is waited for is the condition that finalises the execution timeframe.

On A2, the record states its grounds as required evidence: the timeframe roles stated in the strategy, the conditions that finalise the execution timeframe, the expected holding period, and the stop-loss criteria. What it leaves out is who fixes the roles when two of them claim the same case, what counts as the higher structure changing, and any deadline. The ledger's decider cell for this record is unsettled, and its horizon cell is unsettled.

On A3, the bearer cell reads the-decider: where entry is suspended, the cost of the waiting falls on whoever decided to suspend. The answer comes when the execution timeframe's conditions are met, and the record says entry is re-reviewed at that point, with the analysis timeframe and the order timeframe recorded separately in the trade log.

Which of Knight's two kinds does this setting describe? Each timeframe's rules refer to structures that recur and can be described; if those structures form a class, each rule stands on counted cases. The collision is the case in which two classes claim the same instance. The record does not let a reader tell whether the higher timeframe's evidence is a count of similar structures or a judgment about one structure, and the difference decides which of Knight's two kinds is in play.

3.3.3 Event window conditions

A strategy in 2026 was validated on data gathered under normal conditions. Before an important indicator announcement the record names three things the validation did not sufficiently address: spread, latency, and price gaps before and after the announcement. A calendar on a desk screen carries the official announcement schedule with its time zone. A supervisor eyes the clock approaching the announcement and a set of positions already open. What is waited for is the announcement itself, and, alongside it, whatever the pre-established operating rules say about entries and about positions that are already running.

WSTiger's reading in the record is the example judgment.

On A2, the record rests on an announcement schedule with a time zone, the execution cost of similar events, and rules for holding positions and for new entries. It leaves out what makes an announcement important, the risk budget that its invalidation line refers to, who sets it, and any deadline. On A3, the ledger's horizon cell for this record reads days-to-months, while the record's own clock is the time around a scheduled announcement; the check its invalidation line asks for is whether a change in announcement time and daylight saving time is reflected in the restriction window. That check needs a calendar, not a market outcome, and the record does not say how long such a check takes or who performs it.

Which of Knight's two kinds does this setting describe? The strategy was validated on a class of cases called normal conditions, and the count produced by that validation does not extend to the event window. This is the shape of a case that cannot be sorted into the class from which the rule was drawn, and whether that is Knight's risk or his uncertainty is exactly what Section 3.4's readers disagreed about. What the record does not let a reader tell is whether similar events occurred before in a form that could have been counted, or whether each announcement stands alone.

3.3.4 Missed entry price

A plan in 2026 carries a target, an invalidation point, and an entry price. The entry price has moved before the order was placed, so the remaining expected profit has shrunk and the risk attached to the trade has grown. The target and the invalidation point have not moved. A person at a desk works at a recalculation on the executable price, adding costs to the reward and to the risk. What is waited for is that recalculation, and then the trade's own outcome, which the record holds apart from it.

WSTiger's reading is the example judgment.

On A2, the record rests on the currently executable price, a valid target and invalidation point, and the changed reward measured against cost and risk. It leaves out the content of the criterion that decides whether the new structure is good enough, who holds that criterion, and anything about the class of trades the original plan was drawn from. The ledger's decider cell reads individual.

On A3, the horizon cell is unsettled and the bearer cell reads the-decider: the person who waits while the trade resolves is the person who chose it. The record separates two clocks without naming either. One is immediate — the recalculation, which decides whether entry happens at all. The other is the trade's own duration, which decides whether the recalculation was well made. The invalidation line points at the second clock: recalculate the plan at the price that can be executed now, not at the missed price.

Which of Knight's two kinds does this setting describe? The record treats the new structure as measurable: a calculation with costs, run against a criterion. Whether the missed entry belongs to the class of trades the plan was drawn from is not settled anywhere in the record. What the record does not let a reader tell is what that class was, or whether a plan drawn from past instances can stand on a price that has moved past them.

3.4The debate over what Knight meant

The two words Knight named have been read in two incompatible ways, and both readings come after his book.

Stephen F. LeRoy and Larry D. Singell Jr., writing in the Journal of Political Economy in 1987 under the title "Knight on Risk and Uncertainty", argue that Knight's distinction does not turn on whether people can form probabilities. They argue that Knight took subjective probabilities for granted. What the distinction turns on, on their reading, is whether insurance markets can exist for the exposure. Uncertainty is the state in which moral hazard and adverse selection keep a risk from being insured. Moral hazard is the change in behaviour that follows once a risk is covered; adverse selection is the concentration of the most claim-prone parties among the buyers of a policy. On this reading, the wall between risk and uncertainty is a wall in the market.

Richard N. Langlois and Metin M. Cosgel, writing in Economic Inquiry in 1993, answer LeRoy and Singell. On their reading, uncertainty arises where instances cannot be sorted into classes of like cases, so that no frequency can be had. Such situations call for judgment, and the entrepreneur's judgment is what the firm is built around. On this reading, the wall between risk and uncertainty is a wall in the classing of instances: one side is where cases line up, the other where they do not.

Daniel Ellsberg's 1961 article in the Quarterly Journal of Economics gives the distinction an experimental form. Ellsberg reports a pattern of choice: people prefer a bet whose odds are known to an otherwise equal bet whose odds are unknown. The pattern reported in Ellsberg's 1961 article is known as ambiguity aversion, and it breaks the axioms of subjective expected utility that Savage set out. The result shows that people's choices do not respect the rule that a person's preferences should correspond to a single consistent set of probabilities.

John Kay and Mervyn King, in Radical Uncertainty: Decision-Making Beyond the Numbers (2020), go past the laboratory. Their claim is that many decisions face radical uncertainty that probabilities cannot capture at all. Their remedy is a change of question: rather than inventing a number to optimise on, ask what is going on here, reason with narratives, and choose strategies that are robust and resilient to what cannot be foreseen.

Where do these four records bear on that debate? Take each reading in turn.

Under LeRoy and Singell's reading, the decisive question is whether the risk can be moved to a counterparty who prices it. In the four records read in this chapter, none of the options transfers the exposure. The options are to maintain a rule, recalibrate it, cap it; to fix a rule's roles or suspend an entry; to apply normal conditions, restrict an event window, or validate it separately; to enter, recalculate, or treat an entry as missed. The debate is about what happens when a risk cannot be insured, and these records are silent on that point by the shape of their own option sets.

Under Langlois and Cosgel's reading, all four records turn on the same question: does the present case belong to the class the rule was drawn from? The event window was not part of the data under which a strategy was validated. Two timeframes claim one case between them. An agent's reported score and an agent's outcomes are two things that must be lined up to make a count. A plan drawn from past instances is asked to stand on a price that has moved. On this reading, the records turn on sorting: whether the case at hand can be classed with the cases the rule came from. Whether it can be so classed is what none of them settles.

On Ellsberg, the records bear almost nothing. Ambiguity aversion is a pattern of choice made under known and unknown odds, and these records contain no observed choices and no odds stated anywhere. The coders found the decider unsettled in three of the four records and individual in the fourth, so there is no set of agents whose preferences could be read. WSTiger's example judgments do read as limits chosen in the face of unknown magnitudes — a size cap, a restriction, a recalculation — but that is this study's reading of the wording, not evidence that anybody prefers known odds to unknown ones.

On Kay and King, the fit is partial. Three of the four records carry lists of required evidence, the notes a decision-maker would need before acting: announcement schedules, execution costs of similar events, the executable price, the changed reward against cost. That list is a cousin of the question Kay and King would have a decision-maker ask first: what is going on here. The remedy they propose is missing here. The options in these records are limits and recalculations rather than strategies chosen for robustness across outcomes that cannot be foreseen, and the records never argue in a narrative. Where the fit is bad, the difference is worth naming: Kay and King write about how organisations should reason before a decision; the records write about what a rule should do.

Where the synthetic records cannot settle the debate is the larger part. The debate is about what the term "uncertainty" means, and these records are settings written for study, each with an example judgment that is WSTiger's reading and an invalidation line saying when the judgment would be re-opened. No record carries an outcome, so no instance can be added to a count, and the counting that would decide Langlois and Cosgel's reading cannot begin. No record states whether a market exists for the exposure, because the form has no field for it, and no record lists what is missing. Ellsberg's result is about choices people make under stated odds, and a record is not a choice. The ledger's decider cell is unsettled in three of these four, and its horizon cell is unsettled in two; where the cells are unsettled, this chapter reports the cell rather than filling it. Chapter 2 sets out the coders, their agreement, and the frame rule that keeps studio wording out of the findings.

3.5What carries forward

Both readings of Knight describe states of the world. A decision record describes a moment. Put together, the distinction reads as a question about time: how long before a kind of case can be counted? This study reads it that way, and the four records in this chapter put different things on that scale.

Under Knight's risk, the count exists or is close to existing. The cases recur in recordable form, the average of the past can be leaned on, and the answer to a decision arrives as a comparison with what was expected. Under Knight's uncertainty, the count has not begun or cannot be had. The decision is carried by judgment while the waiting goes on, and the person who waits is not named by any frequency.

The four records show how much of that time each option leaves open. Restricting entries during an event window removes those cases from the count being gathered; validating them separately starts a new count. Fixing timeframe roles stops one class from claiming cases the other should hold; suspending an entry waits for the execution timeframe's conditions to arrive. Calibrating confidence scores against hit records begins the count that inflating scores defers; a size cap reduces the exposure before the count can run. Recalculating a P/L structure on the executable price decides in minutes what would otherwise be decided by waiting for the trade to resolve.

The records carry no outcome that would settle among these options. What the records do give is the shape of the gap: reasons are written, a condition for revisiting is written, and the time before the answer comes is not. The institutional form of the same problem is the next chapter's subject. In a panic, the question of whether a case can be counted is answered by what collateral is accepted and what rate is charged, and the time before the answer is measured in minutes. Chapter 15 returns to where the ledger's rows bear on the Knight debate and where they do not.