Chapter 5. Case two: two classes for one trade

5.1The claim

Knight's Chapter VII says that working inference groups things by one property or another according to the problem at hand, and does not rely on a fixed set of kinds. If that is so, a single case can belong to more than one class, each with its own frequency, and which class is the right one cannot be read off the case. Chapter VIII adds that for a single decision taken alone it makes no practical difference whether the uncertainty is measurable. This case reads those two arguments and then a synthetic record in which one trade belongs to two classes that point opposite ways. It adds to the classing reading a case where the difficulty is not that no class exists but that two do. It adds to the market reading a case with no market in sight: the decider bears the outcome.

5.2Shifting groupings

The passage in Chapter VII on kinds and properties, summarised in Section 4.2, has a consequence Knight does not spell out for probability, though his argument implies it. If the mind groups things now by one property and now by another, then the same object belongs to as many classes as it has properties of interest, and a frequency counted in one class need not agree with a frequency counted in another. Knight comes close to this when he discusses mixed groups. Statisticians warn against drawing conclusions from groups that are not homogeneous, and Knight's illustration is a comparison of death rates between a group of soldiers serving abroad and the general population at home, where the general population mixes ages, sexes and occupations with very different rates. The lower death rate of the soldiers says nothing about the climate where they served. The fault lies in the choice of class.

Knight then turns the warning into a paradox. If perfectly homogeneous groups could be formed, the result would be uniformity, not probability: identical people in identical conditions would all live or all die. Probability needs groups that are alike in the respects that can be measured and unlike in respects that cannot. A frequency is therefore always relative to a class, and the class is always partly chosen.

The urn passage in the same chapter makes the point from the side of knowledge. One person knows only that an urn holds red and black balls; another knows the proportion is three to one. Knight grants that for a wager the first person must act as if the chances were even, while the second has a definite figure. He then adds a middle case: a person who may glance inside the urn without counting the balls. That person forms an estimate in the proper sense, he says, different from both ignorance and exact knowledge. Ellsberg's (1961) experiment later gave this sort of difference an experimental form: people choose differently between known and unknown odds.

5.3The single case

Chapter VIII begins by naming the two kinds, in the one passage this article quotes: "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." Knight observes that ordinary speech uses risk for the chance of loss and uncertainty for the chance of gain, and that this usage hides the real difference. He allows that the pair subjective and objective probability, already in use, could serve as well.

The practical difference, he goes on, is that with risk the spread of outcomes across a group of cases is known, by calculation or from past counts, while with uncertainty it is not, generally because no group can be formed for a situation that is highly unique. Uncertainty's clearest instance, he says, is the forming of opinions about the future, which guide most conduct.

Then comes a reservation that matters for the second record. Where only one case is at stake, Knight says, a lone decision is acted on in the same way whether its uncertainty can be measured or not. The person puts the estimate in the form of so many successes in so many trials and feels about it as about any other probability. He qualifies this twice. If a true probability can be calculated, uniqueness does not matter, since the odds hold whether one bets many times in one game or once in each of many games; but in business this is rare. And a sensible person seldom treats a case as wholly isolated, unless the whole of a fortune or a life is at stake. The second qualification goes back to Chapter VII: a person's decisions can be grouped, so that errors tend to offset one another even without a measurable probability. Each judgment, Knight suggests, has a steady part, the quality of the person's judging, and an accidental part that varies from one decision to the next. What makes an opinion different from a true probability is that the two parts cannot be separated and valued, by calculation or by sorting.

Knight then turns to how people actually behave when outcomes are uncertain. Here he says that people regularly depart from what sound logic would advise. They will stake a little for a long chance at a large gain at odds much worse than fair, and refuse a small chance of a large loss for a near-certain small gain even when the odds favour them; he notes that Adam Smith had discussed the pattern. Many believe in their own luck, a belief that is strongest when what is uncertain is the quality of their own judgment. And any coincidence that catches the eye can be raised into a sign, so that a hunch is acted on as if it were a law.

Finally, Knight sets the economic case in its usual form. A person who ventures holds an opinion about the outcome within some range. The choice is between a smaller reward expected with more confidence and a larger reward expected with less.

5.4The record: a long trend and a short decline

WSTiger's record Timeframes in collision is set in 2026. The setting is synthetic: no real desk, strategy or market lies behind it. Its scenario is written as three labelled parts.

The question follows.

Read beside Knight, the trade in this record belongs to two classes at once. As an entry during a long-term uptrend, it is one of a class of such entries, with whatever share of them succeeded. As an entry during a short-term downtrend, it is one of another class, with another share. The record does not give either share, and nothing in it suggests either is known. But the difficulty it names is not the absence of a class. It is the presence of two, and a strategy that wants to use the more favourable one to skip the rule that the other imposes. The phrase "on the grounds of a long-term uptrend" is a choice of reference class made to justify an act.

This is the soldiers-and-climate problem from Chapter VII in a trading room. A frequency borrowed from the wrong class can make a case look better than it is. Knight's response in the insurance setting is twofold: divide the classes more finely, and adjust within a class by judgment. The record's options take a different route.

The first option lets the more favourable class decide. The third declines to choose and waits. The second assigns each class a job.

5.5Assigning each class its job

The example judgment takes the second path and explains it.

The first sentence does not resolve the conflict between the two classes by finding the true one. It splits the decision into two decisions, each tied to one class. The higher timeframe says which trades are allowed at all: a story about what is happening, of the kind Kay and King (2020) recommend asking for. The execution timeframe says when to enter and when to admit error: a rule, checkable against prices. Neither class is asked to supply a probability for the whole trade.

The second sentence is about time. A trade that fails in the short term can be redescribed afterwards as a trade in the long-term class, where failure is only a temporary setback. The record forbids that. The class that governs the exit must be the class named at entry. The rule guards against something close to what Baron and Hershey (1988) studied, in which the outcome of a decision changes how the decision is judged, but runs the other way: here the danger is that a bad outcome leads the decider to change the description of the decision so that it no longer counts as bad.

Knight's Chapter VIII gives a reason this matters for a single trader. Where one case is at stake alone, measurable and unmeasurable uncertainty feel the same, and the trader acts on a felt ratio. A trader who believes in a long-term trend, and in the trader's own luck at reading it, will find the long-term class the more comfortable one, and Knight names that comfort among the ordinary departures from sound logic. The fixed roles take the choice of class away from the moment after the outcome, when the trader has most reason to choose the flattering one.

5.6Who bears it

The parent study's two readers agreed on one thing about this record that they did not agree on for the first: the outcome falls on the decider. The tag "prop system" and the request for a stop-loss suggest a trader trading the firm's or the trader's own account, and no other party is named. On the market reading, this record therefore shows no market that has failed. There is nobody outside the decider to whom the risk could be passed, and nobody to suffer moral hazard. If the trade's uncertainty is Knightian, it is so in the classing sense: the case belongs to two classes, and the trader must judge which is in charge.

Yet the record still writes as if someone were watching. It requires the roles to be stated in the strategy, the conditions that fix the execution timeframe, the expected holding period and the stop-loss criteria. Those are the instruments by which another party could check that the decider did what the decider said. In a setting where the decider bears the outcome, the watcher the record writes for appears to be the decider at a later time.

5.7What the case adds

Chapter VII of Knight's book implies that a case can belong to several classes, chosen by the purpose at hand, and Chapter VIII that a single decision taken alone is met in the same way whether or not its uncertainty can be measured. Timeframes in collision shows a difficulty that neither reading of Knight is built for: not a missing class but a surplus of classes, with an interest in picking the favourable one after the fact. For the classing reading, the case shows that the act of classing is itself a decision, and one the record fixes in writing before the outcome. For the market reading, the case is an awkward fit: there is no market here to fail, and still the decider's own later self behaves like a party whose description of the case cannot be trusted.