Uncertainty, learning, and rational expectations
29 Aug, 2025 at 08:04 | Posted in Economics | Leave a comment
The rational expectations hypothesis presupposes — largely for reasons of internal consistency — that agents possess complete knowledge of all relevant probability distributions. When economists attempt to incorporate learning into these models, it is always in a very restricted sense. Nothing genuinely unanticipated ever occurs; instead, learning is reduced to a mechanical process of updating — refining the precision of existing information sets and probability functions.
In these ergodic models, nothing truly new happens. Their statistical representation of learning is little more than a caricature of how information and adaptation actually unfold in the real world. This follows from the assumption that people’s decisions can be represented as if they were based on a known probability distribution — which by definition presumes knowledge of every possible event. Otherwise, in the strict statistical sense, it is not a probability distribution at all.
Yet in reality, as behavioral and experimental economics repeatedly show, people often mistake a conditional distribution for a full probability distribution. Such mistakes are impossible in the kinds of models built on the rational expectations hypothesis, which mainstream economists defend so adamantly. In those models, rational expectations agents are always correct on average. But truly new information does not simply reduce estimation error; it can overturn the estimation itself and thereby alter the decisions made. To be genuinely new, information must be unexpected. If it were already anticipated, it would simply be inferred from the existing information set.
In rational expectations models, new information is typically presented as reducing the variance of the parameter estimated. But truly new information can instead increase uncertainty — expanding the information set from (A, B) to (A, B, C). Such information gives rise to new probabilities, revised plans, and different decisions — phenomena that the rational expectations hypothesis, with its finite-sample representation of incomplete information, cannot capture.
In the world of rational expectations, learning is like becoming ever better at reciting the complete works of Shakespeare by heart, or at hitting the bull’s-eye in darts. It presupposes a complete list of possible states of the world and assumes that mistakes are always non-systematic. This is a narrow and rather trivial conception of learning — a closed-world learning process where adaptation improves performance only in an unchanging environment. In real, open-world situations, by contrast, learning is about adapting to and coping with genuinely novel phenomena.
The rational expectations hypothesis assumes consistent behaviour in which expectations never display persistent errors. In its world, we always, on average, hit the bull’s-eye. In the more realistic open-systems view, however, there is always the possibility — indeed, the danger — of systematic mistakes. This, presumably, is why modern knowledge societies place so much emphasis on learning.
So where does this leave us? As John Kay and Mervyn King aptly put it:
The disregard of radical uncertainty by a generation of economists condemned modern macroeconomics to near irrelevance … Keynes’ critique of ‘getting on the job’ without asking ‘whether the job is worth getting on with’ would prove to be as true of the new macroeconomic theorising as of the older econometric modeling …
Over forty years, the authors have watched the bright optimism of a new, rigorous approach to economics dissolve into the failures of prediction and analysis which were seen in the global financial crisis of 2007-08. And it is the pervasive nature of radical uncertainty which is the source of the problem.
