Rational expectations
Rational expectations is the hypothesis — more accurately, the axiom — that economic agents form predictions about the future using all available information and the correct model of the economy. In practice, this means that agents are assumed to know the true probability distributions of all relevant variables and to update their beliefs optimally as new information arrives. The hypothesis was introduced by John Muth in 1961 and became the cornerstone of neoclassical macroeconomics through the work of Robert Lucas, who used it to argue that systematic monetary policy cannot affect real output.
The Lucas critique — that policy evaluation must account for how agents' expectations change when policy changes — is formally correct and substantively trivial. It is correct that if agents know the true model and update perfectly, then policy effects will be anticipated and neutralized. It is trivial because no agent knows the true model, the true model does not exist in any stable form, and the information requirements of rational expectations are computationally infeasible for any organism with finite cognitive resources. Humans use heuristics, not Bayesian updating; they rely on narrative, not probability distributions; and they update their beliefs in response to stories, not signals.
The rational expectations hypothesis is not merely descriptively false. It is methodologically pernicious because it replaces the question of how agents actually form expectations with the question of how they would form expectations if they were infinitely computationally powerful and already knew the answer. It turns economics from the study of human behavior into the study of imaginary agents in imaginary worlds — and then uses the results to guide policy in the real one.
The Boundary with Bounded Rationality
The rational expectations hypothesis stands in direct opposition to bounded rationality, but the opposition is more illuminating if understood as a spectrum rather than a binary. At one extreme, rational expectations assumes agents possess the true model, infinite computational resources, and perfect information processing. At the other extreme, Herbert A. Simon's bounded rationality assumes agents use heuristics, satisfice rather than optimize, and operate under severe cognitive constraints. The question is not which extreme is correct — both are idealizations — but where real agents actually fall on the spectrum, and whether the location depends on the decision environment.
Recent work in ecological rationality suggests that the answer is context-dependent. In simple, stable environments with few variables and clear feedback, agents approximate rational expectations quite well. In complex, volatile environments with many interacting variables and delayed or noisy feedback, bounded rationality dominates. The MMLU benchmark debate in artificial intelligence provides a striking parallel: models perform well on structured, well-defined tasks but fail catastrophically on open-ended reasoning tasks that require the integration of disparate knowledge. Rational expectations is the MMLU of macroeconomics — it tests a narrow, artificial competence and generalizes poorly to the full range of human decision-making.
The middle ground — adaptive expectations — has been unjustly marginalized. Adaptive expectations treat agents as backward-looking learners who update gradually, using simple distributed lags rather than full Bayesian updating. In the rational expectations framework, this is a failure: agents are systematically wrong and exploitable. But in a world of model uncertainty and structural breaks, adaptive expectations may be more robust than rational expectations precisely because they do not overreact to new information. The agent who updates too quickly is vulnerable to noise; the agent who updates too slowly misses regime changes. Adaptive expectations, with its single free parameter λ, occupies a sweet spot that rational expectations cannot locate because it assumes the regime is known.
The rational expectations revolution did not solve the problem of expectation formation. It replaced one arbitrary assumption — that agents use simple adaptive rules — with another arbitrary assumption: that agents know the true model. The second assumption is not more rigorous; it is merely more computationally extravagant. Until macroeconomics develops a genuine theory of learning under uncertainty — one that accounts for computational costs, model ambiguity, and the heterogeneity of agent beliefs — it will continue to oscillate between these two poles without making progress. The bridge between rational expectations and bounded rationality is not a theoretical nicety. It is the only path to a economics that describes how humans actually think.