Decision Under Uncertainty
decision theory · expected utility · risk and uncertainty
Choosing when the probabilities of the outcomes are not known. Frank Knight (1921) separated this from risk, where they are: risk can be priced and insured, uncertainty cannot, and on Knight's account profit is the return for bearing the second.
In practice
A pricing change has a knowable distribution — you have run it before across enough accounts. Entering a market nobody has served has no distribution at all, and the spreadsheet that assigns it one has manufactured the confidence it then reports.
The common mistake
Converting uncertainty into risk by assigning probabilities, then treating the output as though the probabilities were measured. The number is now auditable, defensible and no better informed, and the act of producing it has removed the one useful signal — that nobody knows.
The distinction that organises everything here is between a situation whose odds you know and one whose odds you do not. They look similar from inside and call for opposite behaviour.
Knight's distinction
Frank Knight (1921) drew the line that still carries his name. Risk is measurable uncertainty; true uncertainty is not measurableKnight, F. H. (1921). Risk, Uncertainty and Profit. Houghton Mifflin. Knight's economic conclusion follows from the distinction: measurable risk can be insured and its cost becomes an ordinary expense, so it cannot be a source of profit. Profit is the residual return to bearing uncertainty that could not be priced.. Where a distribution can be established — by frequency, by physical symmetry, by enough repetitions — the exposure can be insured, priced into costs, and eliminated as a source of advantage. Where no distribution exists, it cannot, and Knight argued that this is precisely where profit comes from. John Maynard Keynes reached a parallel position the same year in A Treatise on Probability, and put it bluntly in 1937: about the prospect of a European war or the price of copper in twenty years, 'we simply do not know.'
Why the distinction survives attempts to dissolve it
Leonard Savage (1954) built the standard reply: a rational agent always has subjective probabilities, derivable from their own consistent choices, so the Knightian category is empty. Daniel Ellsberg (1961) tested it and found people reliably violate Savage's axioms when the odds are unstated rather than unfavourable. Offered a bet on a 50-urn and an unknown-composition urn, most people pay to avoid the second on both coloursEllsberg, D. (1961). 'Risk, Ambiguity, and the Savage Axioms.' Quarterly Journal of Economics 75(4). Preferring red-from-known to red-from-unknown and also black-from-known to black-from-unknown is jointly inconsistent with holding any probability for the unknown urn — so the aversion is to ambiguity itself, not to unfavourable odds. — a preference no assignment of probabilities can rationalise. Ambiguity aversion is a distinct phenomenon, and the distinction Savage tried to dissolve turns out to be one people act on.
What to do instead
Herbert Simon (1955) established that finite agents do not optimise and cannot: they satisfice, taking the first option that clears a threshold. Gerd Gigerenzer and Peter Todd (1999) went further, showing that under genuine uncertainty simple heuristics often beat optimising procedures outright, because optimisation fits the sample and uncertainty is the condition of the sample not resembling the future. The practical consequences are specific: prefer decisions that can be reversed; size exposure so that being wrong is survivable rather than so that being right is maximal; buy information before buying commitment; and treat the absence of a distribution as itself a finding rather than a gap to be filled with an estimate.
The objections
The category is easy to abuse. Almost any decision can be declared Knightian, and doing so excuses the analysis that was available — base rates, comparable cases, small tests. In practice most business decisions described as facing 'radical uncertainty' have reference classes their owners have not looked for, and the label is doing the work of the search.
The Bayesian reply also retains real force. Even a poorly grounded prior, made explicit, can be updated and audited; a refusal to quantify cannot be. The defensible position is narrower than either camp states: quantify, and record how the number was arrived at, so the difference between a frequency and an assertion survives into the decision.
What it rules out
It rules out treating a confidence interval as informative when its inputs were asserted. It rules out insurance-style reasoning — expected value, diversification, the law of large numbers — where there is no distribution for them to operate on. And it rules out reading a good outcome as evidence of a good decision, since under uncertainty the two are only loosely connected.
It does not rule out acting. Knight's point is the opposite: bearing unmeasurable uncertainty is where returns come from, so the response is to structure exposure, not to wait for a distribution that will not arrive.
Sources
Ellsberg, D. (1961). 'Risk, Ambiguity, and the Savage Axioms.' Quarterly Journal of Economics 75(4). · Gigerenzer, G. & Todd, P. (1999). Simple Heuristics That Make Us Smart. Oxford University Press. · Keynes, J. M. (1921). A Treatise on Probability. Macmillan. · Keynes, J. M. (1937). 'The General Theory of Employment.' Quarterly Journal of Economics 51(2). · Knight, F. H. (1921). Risk, Uncertainty and Profit. Houghton Mifflin. · Savage, L. J. (1954). The Foundations of Statistics. Wiley. · Simon, H. A. (1955). 'A Behavioral Model of Rational Choice.' Quarterly Journal of Economics 69(1).
Concept web
Open the full webQuestions
What is the difference between risk and uncertainty?
Knight (1921) defined risk as measurable — the probabilities are known, so the exposure can be insured and priced — and uncertainty as unmeasurable, where no distribution exists. Because risk can be insured away, Knight argued profit is the return for bearing uncertainty.
What is the Ellsberg paradox?
Offered bets on an urn with known composition and one with unknown composition, most people pay to avoid the unknown urn on both colours — jointly inconsistent with holding any probability for it. Ellsberg (1961) showed ambiguity aversion is real and violates Savage's axioms.
How should you decide under uncertainty?
Prefer reversible decisions, size exposure so being wrong is survivable rather than so being right is maximal, buy information before commitment, and treat a missing distribution as a finding rather than a gap to fill with an estimate. Simon (1955) and Gigerenzer (1999) show simple rules often beat optimisation here.
Why is assigning probabilities to unknowns dangerous?
It converts uncertainty into something auditable without making it better informed, and it destroys the one useful signal — that nobody knows. The countervailing point is that an explicit prior can at least be updated and challenged, which an unstated one cannot.