1.1The question
An average is a promise about a group. It says that if enough cases are put together, their separate fortunes will settle into something steady enough to plan on. Insurance lives on that promise, and so does any portfolio built from many positions. The promise has conditions. The cases must be numerous, they must not all go wrong at once, and the person relying on the average must last long enough for it to appear.
This article asks what happens to the promise when the cases pooled are not alike. Frank Knight's Risk, Uncertainty and Profit (1921) gives the classic account of pooling in its eighth chapter, and it makes a bolder claim than its reputation suggests: that gains and losses offset each other even among ventures that resemble each other in nothing obvious, provided they are many and competently judged. The question is whether a decision record, written before anyone knows how the decision turns out, says what such pooling leaves out. Put plainly: when grouping needs like cases, and the cases at hand are unlike, does the record name what the average hides?
The records come from WSTiger, a studio in the RISCON family that publishes synthetic decision settings for study. Two are read here. In the first, a strategy earns small gains most of the time and suffers rare, large losses, and the question is whether its average figures justify adopting it. In the second, a portfolio holds many differently named strategies whose returns usually move apart but which lose together when markets move sharply, and the question is whether more names mean more diversification. Both records are synthetic. They are modern summaries written for the platform, not archival documents, and the firms, desks and strategies in them are invented.
1.2The contribution
The dispute over Knight that most scholarship has pursued concerns the line between his two kinds of uncertainty: whether it falls where insurance markets fail, as LeRoy and Singell (1987) read it, or where cases cannot be sorted into classes of like cases, as Langlois and Cosgel (1993) answer. That dispute is about where grouping stops. This article takes up a different and narrower matter: what grouping assumes while it is still working. Knight's Chapter VIII states those assumptions as it goes. It notes that fire losses are not independent and come in clusters; that the offsetting of a business person's estimates appears only after the fact and is limited by the brevity of a working life; and that the quality of an estimate is not independent of how many estimates must be made. These remarks are scattered and brief, and the chapter's main line runs the other way, toward the claim that pooling succeeds even across unlike cases.
The contribution is to set those scattered conditions beside a later debate about rare large losses and to test them on two records written at the moment of decision. The records were not written with Knight in mind. Each nonetheless refuses an average offered as a substitute for a class: in one case an average return figure, in the other a count of named strategies. The grounds each gives for refusing correspond to one of Knight's two failure conditions. The first record turns on survival before the average arrives; the second on losses that arrive together. Neither record says where the rare loss comes from, and that silence is what the later debate is about.
The two accounts of the rare loss used below are this article's construction, not positions the works take in so many words. Assigning a work to an account, saying that an account places the loss outside or inside a class, and linking a work to Knight or to a record are all this article's own. Each work is credited only with what its reading note in the parent study records.
1.3The debate joined
The later literature offers two accounts of the rare large loss, and they sit uneasily together. The first account places the loss outside any class that could have been formed in advance. It is built from two works. Taleb (2007) is concerned with rare events of extreme impact that are explained after the fact as if they had been predictable. Kay and King (2020) argue that many decisions face a radical uncertainty that probabilities cannot capture, and recommend strategies robust and resilient to what cannot be foreseen rather than optimisation on invented numbers. On this account the tail is a gap in knowledge. Grouping fails at the tail because the tail was never in the group.
The second account places the loss inside the pooled system. It is built from three works. Daníelsson (2002) argues 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, so that risk becomes endogenous and the models can deepen the crises they measure. Bookstaber (2007) describes markets made complex and tightly coupled by their own innovations, in which liquidity vanishes when it is most needed. Rajan (2005) argues that managers paid on returns measured against their peers have reason to take risks that pay off most of the time and lose heavily only rarely, and to herd. On this account the tail is a property of the pool. Grouping fails at the tail because the act of grouping, measured and rewarded in a certain way, put it there.
Between the two stands a third body of work, about what organisations do with such losses once they are named. Power (2007) describes risk management as the way organisations turn uncertainty into objects they can manage and account for. Both records end by proposing such objects: a pre-set budget and an independent check in one, two separate correlation tables in the other.
Knight's Chapter VIII bears on both accounts. Its remark on revolutionary occurrences that no one expected sits near the first; its remark on fires that spread sits near the second. The article proceeds as follows. Section 2 lays out the debate side by side. Section 3 describes the primary text, the records and the parent study's ledger. Sections 4 to 6 are three case sections. Section 7 discusses what the records can and cannot carry, and Section 8 concludes.