How has economics modeled what people expect about the future — and did rational expectations settle it?
Keynes put unstable expectations at the center of economics in 1936. The rational-expectations revolution declared that incoherent and replaced it with agents who know the true model. Then the surveys came in and the people in them turned out not to be rational expecters at all. This is one apparatus — the model of what agents believe about the future — followed across four eras, and the verdict it landed on: rational expectations was right to discipline, wrong to describe.
Animal spirits to adaptive expectations
“It is not a case of choosing those which, to the best of one’s judgment, are really the prettiest, nor even those which average opinion genuinely thinks the prettiest. We have reached the third degree where we devote our intelligences to anticipating what average opinion expects the average opinion to be.”
— John Maynard Keynes, The General Theory of Employment, Interest and Money, Ch.12, 1936
This is the most famous sentence ever written about expectations in economics. Keynes’s picture: beliefs about the future are conventional, self-referential, and unstable — you are not forecasting the world, you are forecasting other people’s forecasts. Everything that came after had to answer it. The first answer was not the one you would guess.
Keynes left expectations qualitative — “animal spirits,” conventions that hold until they don’t. That is evocative, but you cannot put it in an equation. The generation that built the macroeconometric models of the 1950s and 1960s needed a rule: a concrete formula for how a number called “expected inflation” gets updated as data arrives. The rule they reached for was adaptive expectations.
Phillip Cagan formalized it in 1956 studying hyperinflations: you expect next period’s inflation to be a weighted average of what you expected last period and what actually happened, leaning on the recent past. The error in your last forecast tells you how much to revise. Mechanically simple, and — this is the part the next stage attacks — entirely backward-looking. You never look forward; you only correct what you already got wrong.
The error-correction updating rule, with adjustment speed $0 < \lambda \le 1$:
$$\pi^e_t = \pi^e_{t-1} + \lambda\left(\pi_{t-1} - \pi^e_{t-1}\right)$$Each period you nudge your expectation toward last period’s actual inflation by a fraction $\lambda$ of the miss. The expectation is a geometrically-declining weighted average of all past inflation — nothing but history feeds in.
You expect next year’s inflation to be roughly what you’ve been seeing, weighted toward the recent past. If you under-predicted last year, you bump your forecast up a bit this year. It is the rule a sensible person uses when they have no theory of the economy — just a memory of the data.
Here is what makes this rung impossible to dismiss. Bolt adaptive expectations onto the Phillips curve — the apparent trade-off between inflation and unemployment — and you get the expectations-augmented Phillips curve of Milton Friedman (1968) and Edmund Phelps (1967). Its prediction was startling and correct: a government that tries to buy lower unemployment with higher inflation gets it only until people’s expectations catch up. Then the short-run curve shifts up, and you are left with the same unemployment and higher inflation. The long-run curve is vertical at the natural rate. This model predicted stagflation before stagflation arrived — the 1970s breakdown of the stable Phillips curve is exactly what Friedman and Phelps said would happen.
So the pre-rational-expectations world was not naive. It had a working theory of belief formation that scored one of the great forecasting successes in the history of macroeconomics. The formal apparatus — the adaptive updating rule and the expectations-augmented Phillips curve that rotates to vertical at the natural rate — lives in Ch 9 §9.7 (The New Keynesian Phillips Curve Preview) and Ch 15 §15.3 (The New Keynesian Phillips Curve). The intellectual origin — Keynes’s animal spirits, and the neoclassical-synthesis Phillips curve that adaptive expectations rebuilt — is in History of Economic Thought Ch.8 (The Keynesian revolution).
The model that predicted stagflation vs. the menu it replaced
“There is always a temporary trade-off between inflation and unemployment; there is no permanent trade-off. The temporary trade-off comes not from inflation per se, but from unanticipated inflation, which generally means, from a rising rate of inflation.”
— Milton Friedman, “The Role of Monetary Policy,” AEA Presidential Address, American Economic Review, 1968
Friedman’s 1968 address, with Phelps’s 1967 paper, is the advance. By letting expectations adapt to actual inflation, Friedman and Phelps derived a result the menu picture could not see: the inflation-unemployment trade-off is real only while inflation is accelerating and fooling people. Stabilize inflation at any level and unemployment returns to its natural rate. They published this years before the stable Phillips curve broke in the 1970s, and it broke the way they said. Backward-looking adaptive expectations was enough to win one of macroeconomics’s great predictive victories. The monetarist turn that carried it is traced in History of Economic Thought Ch.10 §10.1 (Friedman’s monetarism).
“In terms of our menu of unemployment and price stability, the situation can be described as one in which we can have less unemployment if we are willing to accept the consequence of a slightly higher price index.”
— the 1960s neoclassical-synthesis reading of the Phillips curve, after Samuelson & Solow, 1960
Take the predecessor at its strongest. The stable-Phillips-curve consensus was not a blunder — it was the best reading of two decades of data. Through the 1950s and early 1960s the inflation-unemployment trade-off really was stable, and a model that treated it as a policy menu fit the evidence and informed sensible policy. The consensus assumed, reasonably given the data it had, that the trade-off would hold. What it could not anticipate was that the trade-off would dissolve the moment policymakers tried to exploit it systematically — because that systematic exploitation is precisely what teaches expectations to move. Adaptive expectations did not refute a foolish theory; it added the one missing piece to a sensible one and showed why the menu would betray anyone who relied on it.
Where this leaves us
Adaptive expectations earns full credit. Backward-looking updating gave the expectations-augmented Phillips curve its teeth, and that model predicted stagflation when the reigning consensus could not. But it carries a flaw the next stage will exploit, and the flaw is not empirical — it is logical. If inflation rises persistently, an adaptive agent is wrong in the same direction year after year: she under-predicts, corrects too slowly, under-predicts again, and never learns to stop. A model in which intelligent people make the identical forecasting mistake forever is not describing an equilibrium. It is describing people who are leaving money on the table and refusing to pick it up. Sooner or later, someone was going to ask why they wouldn’t.
Adaptive expectations predicted the 1970s. But it also assumed something strange: that people who get fooled by rising inflation every single year never wise up. What happens to the model — and to the policymaker counting on fooling them — if they do?
The rational-expectations revolution
“Given that the structure of an econometric model consists of optimal decision rules of economic agents, and that optimal decision rules vary systematically with changes in the structure of series relevant to the decision maker, it follows that any change in policy will systematically alter the structure of econometric models.”
— Robert Lucas, “Econometric Policy Evaluation: A Critique,” 1976
Strip the jargon and Lucas is saying something lethal to the menu: the historical relationships a policymaker wants to exploit will change the instant the policy rule changes, because people re-optimize when the rules of the game move. You cannot ride a curve estimated under the old regime into the new one. This is the single most consequential methodological move in postwar macro — and the death of the idea that you can systematically fool people forever.
The flaw Stage 1 ended on — agents making the same mistake forever — is exactly what John Muth attacked in 1961. His proposal, almost offhand at the time, became the organizing principle of a generation: expectations are rational when they are the model-consistent forecast, using all available information. The subjective probabilities people hold equal the objective probabilities the economic model actually implies. People are not assumed to be geniuses; they are assumed not to make errors a smart observer could predict and they could fix.
Muth’s condition: the expectation agents hold equals the true conditional expectation given their information set $I_t$:
$$E^{\text{subjective}}\!\left[x_{t+1} \mid I_t\right] = E\!\left[x_{t+1} \mid I_t\right]$$Forecast errors $x_{t+1} - E[x_{t+1}\mid I_t]$ are then unpredictable given $I_t$: zero mean, uncorrelated with anything the agent already knows. No free, systematic, exploitable error survives.
People use the true model of the economy and don’t leave free money on the table. They can be unlucky — surprised by genuine shocks — but they are not predictably wrong. If there were a pattern in their mistakes, someone would trade on it and the pattern would vanish.
Lucas (1972) embedded this in a general-equilibrium macro model and drew out the bombshell: policy ineffectiveness. If a monetary expansion is systematic and anticipated, rational agents have already priced it in, so it moves nominal variables and nothing real. Only the surprise component of policy does anything — and you cannot systematically surprise people who know your rule. The Phillips-curve trade-off the menu sold is not just temporary; for anticipated policy it is zero.
The most durable piece is the Lucas critique itself (1976). The behavioral relationships in an econometric model — including the Phillips curve — are not policy-invariant laws of nature. They are equilibrium outcomes of agents optimizing against the current policy rule. Change the rule and you change the relationship, so a curve fitted to historical data is worthless for predicting the effect of a new regime. This is permanent. It survived every later quarrel about rational expectations, and it is the reason no serious macroeconomist today reads a historical correlation as a stable policy menu. The apparatus — rational expectations inside the workhorse model, the Taylor rule, the policy logic — is in Ch 15 §15.6 (The 3-Equation NK Model) and Ch 15 §15.5 (The Taylor Rule). The Lucas-Sargent new-classical program is traced in History of Economic Thought Ch.10 §10.2 (Lucas and the rational expectations revolution); its passage into the modern monetary-policy consensus is in Ch.12 §12.2 (Woodford and the workhorse).
“There is no way that the government can fool people systematically. Anticipated changes in the money supply have no effect on real output or employment.”
— the strong-form policy-ineffectiveness reading of Sargent & Wallace, 1975–76
Did rational expectations prove systematic policy can’t move output?
The strong-form claim was thrilling: if people are rational, anticipated demand management does nothing real, and the whole Keynesian stabilization project collapses. Is that a theorem about the world, or about a benchmark?
The disciplining benchmark vs. the adaptive modeler it displaced
“Expectations, since they are informed predictions of future events, are essentially the same as the predictions of the relevant economic theory… the economy generally does not waste information.”
— John Muth, “Rational Expectations and the Theory of Price Movements,” Econometrica, 1961
Muth’s sentence is the entire revolution in compressed form. If beliefs about the future systematically deviate from what the model implies, that deviation is information lying unused, and a world of optimizers does not leave information unused for long. Lucas turned the principle into a method — the critique — that disqualified the whole practice of forecasting policy effects from policy-invariant correlations. This was not a fashion that came and went. It permanently raised the bar for what counts as a coherent model of expectations: any rule you write down has to survive the question “why don’t agents learn their way out of it?”
“The hypothesis of adaptive expectations had proved remarkably successful empirically — the demand for money in hyperinflations, the Phillips curve, the term structure — long before anyone asked whether agents were using the true model.”
— the adaptive-expectations / structural-macroeconometrics defense, after Cagan, 1956
Now the predecessor at full strength, because rational expectations is responding to it. The adaptive modeler has a serious case. Cagan’s backward-looking rule fit hyperinflation money demand beautifully; the adaptive Phillips curve predicted stagflation. The data did not demand that agents know the true structural model — a sensible rule that tracks the recent past fit fine across a range of episodes. Assuming agents carry the correct model of the entire economy in their heads is, the adaptive modeler argues, a stronger and less realistic assumption than anything the data forced on us. Rational expectations buys logical coherence at the price of cognitive plausibility, and it is not obvious, in 1976, that the trade is worth it. This is the live tension the revolution had to win on argument, not assume away — and the honest verdict is that it won the coherence point while the realism point stayed open.
Where this leaves us
Rational expectations was a genuine advance, not a fashion. It killed a real naivety — the belief that a stable Phillips menu could be exploited forever — and it handed macroeconomics a discipline it had lacked: no model of expectations may assume free, exploitable, systematic errors that go uncorrected for all time. The Lucas critique is permanent and remains true. But “agents use the true model and all available information” is a benchmark, not a documented fact about how human minds work. The revolution won the coherence argument cleanly. Whether it also won the description argument is a separate question — and it turns out to be answerable. You just have to ask the people doing the forecasting what they actually expect.
Rational expectations won the theory war. Then the surveys came in — decades of professional forecasters writing down what they expected inflation and output to be — and the people in them turned out not to be rational expecters at all. Their errors were not random. They were systematic. Which is exactly what rational expectations forbids.
The cracks in rational expectations
“In response to macroeconomic shocks, forecasters update their forecasts slowly. The average forecast underreacts to information… ex-post mean forecast errors are predictable from the forecast revisions — a direct violation of the full-information rational-expectations null.”
— Olivier Coibion & Yuriy Gorodnichenko, “Information Rigidity and the Expectations Formation Process,” American Economic Review, 2015
Rational expectations makes a sharp, testable promise: the average forecast already incorporates all available information, so forecast errors should be unpredictable. Coibion and Gorodnichenko ran the test on decades of Survey of Professional Forecasters data. The errors are predictable. If a forecaster just revised her inflation forecast upward, you can bet she revised too little — the error runs the same way as the revision. That is information sitting unused, period after period, which is precisely what rational expectations says cannot happen.
The survey evidence is not a one-off. Forecasters under-react to their own information at short horizons and over-react at long ones; asset prices swing far more than rational-expectations fundamentals can justify (Shiller’s excess-volatility result). Full rational expectations, taken as a literal description of belief formation, is false. The question for the field was the one that defines the whole modern settlement: abandon rational expectations, or repair it without throwing away what it got right? The field chose repair — a family of disciplined deviations, each keeping the no-free-lunch core and relaxing the full-information assumption in a single, micro-founded way.
Rational inattention (Christopher Sims, 2003). Attention is a scarce resource. Agents have limited capacity to process information and allocate it optimally — tracking what matters most and economizing on the rest. The deviation from full rational expectations is not assumed; it is derived from an information-cost optimization. You would track inflation perfectly if attention were free; it isn’t, so you optimally track it imperfectly, and the imperfections look exactly like the sluggish under-reaction the surveys show.
The agent chooses a signal about the state $x$ to maximize expected payoff net of an information cost measured in bits (Shannon mutual information $\mathcal{I}$), subject to a capacity constraint $\kappa$:
$$\max_{\text{signal}} \; E\!\left[u(a, x)\right] - \theta\,\mathcal{I}(x; \text{signal}), \qquad \mathcal{I}(x;\text{signal}) \le \kappa$$As capacity $\kappa \to \infty$ (or the price of attention $\theta \to 0$), the optimal signal becomes perfect and beliefs collapse to full rational expectations. The deviation is the priced-in cost of thinking, not a psychological glitch.
You would track inflation perfectly if paying attention were free. It isn’t — attention is scarce and you have a life — so you optimally track it a bit late and a bit coarsely. Your forecasts lag the news not because you are irrational but because watching every release closely is not worth the effort. Make attention free and you become a perfect rational forecaster again.
Adaptive learning (Thomas Sargent, The Conquest of American Inflation, 1999). Drop the assumption that agents already know the true model. Instead they estimate it, like an econometrician, and update their estimates as data arrives. Under broad conditions the estimates converge toward the rational-expectations equilibrium — so rational expectations is the limit, not the premise — but the transition matters, and occasional “escapes” from the equilibrium generate the actual inflation history. Rational expectations is where you end up, not where you start.
Level-k reasoning rounds out the family: agents reason a finite number of strategic steps rather than iterating all the way to the fixed point. All three share one move — keep rational expectations’ discipline (no free systematic errors a smart agent would eliminate) while explaining, from an explicit constraint, why real forecasts deviate. The new-Keynesian apparatus that absorbs these informational frictions is in Ch 15 §15.8 (The Zero Lower Bound), where expectations and forward guidance do the heavy lifting; the lineage into the modern monetary consensus is in History of Economic Thought Ch.12 §12.5 (Labor and the credibility revolution).
“Rational expectations was never meant as a literal account of how people think. It is an equilibrium discipline — a restriction on models, not a claim about psychology.”
— the modern defensive reframe of rational expectations
Is rational expectations just a benchmark, not a description?
Confronted with the survey violations, the defense reframes: rational expectations was always an equilibrium discipline, never a psychological claim, so the deviations don’t embarrass it. True — but does it let the discipline off too easily?
Discipline the deviations vs. defend the benchmark
“Because information-processing capacity is limited, people react to economic data slowly and imperfectly — and that single constraint, optimally handled, reproduces the sluggishness we see in the data.”
— Christopher Sims, “Implications of Rational Inattention,” Journal of Monetary Economics, 2003
Sims and Sargent built the repair, not a rejection. Rational inattention derives the deviation from an optimization — agents are still doing the best they can, just against a capacity constraint — so it inherits rational expectations’ discipline rather than discarding it. Adaptive learning makes rational expectations the destination agents converge to rather than the assumption they start from. Both keep the core that rational expectations got right (no free, exploitable, perpetual errors) while matching the survey facts rational expectations cannot. This is what “disciplined deviation” means: the no-free-lunch principle survives; the literal full-information assumption is replaced by an explicit, estimable friction.
“Survey forecasts are noisy, possibly non-incentivized proxies for the latent expectation that actually drives behavior. Measured violations of rational expectations may be artifacts of measurement, not failures of the equilibrium concept.”
— the new-classical defense of full rational expectations as the equilibrium concept
Full rational expectations deserves its strongest defense here, because this is the benchmark the deviations are responding to. The new-classical rejoinder is not foolish: elicited survey forecasts may not be the expectations that govern actual spending and pricing decisions; respondents have weak incentives to report carefully; and the “true” market expectation, revealed in prices and quantities, could still satisfy rational expectations even when stated forecasts wobble. On this reading, rational expectations survives as the right equilibrium concept and the surveys are a contaminated window onto it. There is genuine force here — and a striking historical wrinkle: Thomas Sargent, who built the adaptive-learning program that relaxes rational expectations, is also one of rational expectations’ founders. He sits on both sides of this debate, which is not a contradiction but a feature: the person who most rigorously defended the benchmark is the one best placed to see precisely how, and how far, to relax it.
Where this leaves us
Full rational expectations is empirically false as a literal account of belief formation — the survey evidence is systematic and decisive, and the measurement-error defense softens it without overturning it. But the field’s response was not surrender; it was discipline. Rational inattention, adaptive learning, and level-k reasoning all keep rational expectations’ core insight — no free systematic errors a smart agent would eliminate — while explaining why real forecasts deviate, each from a single explicit constraint. This is the method of the modern settlement: rational expectations as benchmark, deviations as disciplined. The excess-volatility face of this argument — whether asset prices wander too far from rational-expectations fundamentals — is the markets-side question walked in Are markets efficient? And it sets up the rung that pushes hardest on the deviation side: what if you deviate not because information is costly, but because of how your mind is built?
Rational inattention says you deviate because information is costly. But what if you deviate because of how your mind is built — because a recent boom makes you over-believe the future is bright, and a recent bust makes you over-believe it is dark? The last rung borrows its engine from a psychologist’s 1979 model of how people misjudge risk.
Behavioral expectations
“But how do we know when irrational exuberance has unduly escalated asset values, which then become subject to unexpected and prolonged contractions?”
— Alan Greenspan, “The Challenge of Central Banking in a Democratic Society,” December 5, 1996
Greenspan’s phrase named, in two words, a phenomenon rational expectations cannot easily generate: expectations that detach from fundamentals and drive asset values up and then down. A psychologist’s 1979 model of how people misjudge risk became, four decades later, the cognitive engine of the most active frontier in macro-finance expectations — over-reaction to recent news, credit cycles, and bubbles. This is the last rung, and it is not where rational expectations is defeated.
The cognitive foundation is prospect theory (Daniel Kahneman and Amos Tversky, 1979). People evaluate outcomes as gains and losses relative to a reference point, feel losses roughly twice as keenly as equivalent gains, and weight probabilities non-linearly — over-weighting the rare, under-weighting the near-certain. Used here strictly as the input to belief formation: this is the cognition that, pointed at the macroeconomy, warps which futures people expect.
Prospect theory replaces the expected-utility objective with a value function $v(\cdot)$ over gains and losses around a reference point $r$, and a probability-weighting function $\pi(\cdot)$:
$$V = \sum_i \pi(p_i)\, v(x_i - r), \qquad v \text{ kinked at } 0 \text{ with } v'(\text{loss}) > v'(\text{gain})$$The kink at the reference point is loss aversion; the weighting $\pi(p) \ne p$ is the probability distortion. (Why this violates the expected-utility independence axiom — the Allais and Ellsberg paradoxes — is the decision-under-risk story, treated separately; here it is only the engine of distorted beliefs.)
People over-weight small probabilities, feel losses about twice as hard as gains, and judge everything against a reference point. Aim that cognition at the future and it shapes what you expect: a recent run of good news makes the rosy scenario feel far likelier than it is. The mind that misjudges a gamble also misjudges the economy.
Diagnostic expectations (Pedro Bordalo, Nicola Gennaioli, and Andrei Shleifer, 2018) turn this into a macro belief-formation model. Drawing on representativeness — the tendency to over-weight outcomes that recent news makes feel typical — agents over-react to good news, so the future looks too bright after a boom and too bleak after a bust. Credit expands when diagnostic optimism runs ahead of fundamentals, then contracts when reality disappoints: booms and busts generated endogenously from belief formation, not from external shocks.
Diagnostic beliefs distort the rational forecast by exaggerating the recent surprise. With distortion parameter $\theta \ge 0$:
$$E^{\theta}_t[x_{t+1}] = E_t[x_{t+1}] + \theta\left(E_t[x_{t+1}] - E_{t-1}[x_{t+1}]\right)$$The bracket is the news — how much the rational forecast just moved. At $\theta = 0$ the model is exactly rational expectations; $\theta > 0$ over-reacts to recent news. One parameter, it nests rational expectations, and the data identify it.
A recent boom makes the boom feel like the new normal — until it isn’t. Good news doesn’t just raise your forecast; it raises it too far, because the fresh good news feels more representative of the future than it should. That over-shoot, repeated across a credit market, is how an expectations model manufactures a bubble and the bust that follows.
Sparsity (Xavier Gabaix, 2014–2020) completes the frontier with a tractable behavioral-macro framework: agents simplify by ignoring variables that barely matter (a “sparse max”), and rational expectations re-emerges as the zero-inattention limit. Like diagnostic expectations, it nests rational expectations and adds a small number of estimable parameters. The decision-theoretic foundation — prospect theory, loss aversion, probability weighting — is in Ch 19 §19.2 (Prospect Theory), with the expected-utility benchmark it departs from in Ch 11 §11.1 (Choice Theory: axioms and utility representation). The behavioral-expectations turn in the history of thought — from Simon’s bounded rationality through prospect theory into the modern departures — is in History of Economic Thought Ch.13 §13.2 (Prospect: the descriptive alternative).
“Diagnostic expectations over-react to news, generating excess optimism in good times and excess pessimism in bad times. This single mechanism produces credit cycles — booms followed by busts — from the inside.”
— Pedro Bordalo, Nicola Gennaioli & Andrei Shleifer, “Diagnostic Expectations and Credit Cycles,” Journal of Finance, 2018
Is over-reaction to recent news a free parameter, or a disciplined mechanism?
Diagnostic expectations explains bubbles and credit cycles that rational expectations cannot generate endogenously. The objection writes itself: is “people over-react” a real mechanism, or a knob you turn until the data fit?
Behavioral expectations vs. the rational-inattention benchmark
“A psychologically founded account of expectations — over-reaction to representative news — explains the predictable forecast errors, the excess volatility, and the credit cycle within a single, estimable framework that contains rational expectations as a limit.”
— the behavioral-expectations program, after Bordalo-Gennaioli-Shleifer 2018 and Gabaix 2014–2020
The behavioral-expectations case is that the deviations the surveys document are not merely costly-information artifacts — they have a cognitive shape that information costs alone cannot reproduce. Over-reaction to recent, representative news produces the long-horizon over-reaction in forecasts, the boom-bust asymmetry of credit, and the bubbles Greenspan worried about, all from one documented psychological regularity. And it does so as a disciplined deviation: diagnostic expectations and sparsity nest rational expectations, carry a small number of estimable parameters, and make falsifiable predictions. The frontier is not a retreat from rigor; it is rigor applied to a richer model of the believing mind.
“Before reaching for a psychological primitive, ask whether the deviation is already derivable from optimal behavior under constraints. Much of what looks like over-reaction is what a rationally inattentive agent should do — and adding biases risks an epicycle for every anomaly.”
— the rational-inattention / new-classical defense of the benchmark
The benchmark gets the last and strongest word, because behavioral expectations is the rung responding to it. The defender’s point is methodological discipline, not denial of the facts. Apparent over-reaction may be exactly what an agent optimally processing limited information would produce — in which case rational inattention explains the data with no psychological primitive, and Occam favors the leaner model. The deeper worry is the epicycles problem: once you license cognitive biases, there is a bias available to fit any anomaly, and a framework that can fit anything predicts nothing. The benchmark’s demand is therefore a good one: every proposed distortion must be disciplined — nested in rational expectations, governed by few parameters, and falsifiable — or it does not earn its place. Behavioral expectations is welcome at the table precisely to the extent that it meets that demand, and the strongest versions (diagnostic expectations, sparsity) do. The disagreement that remains is not whether to discipline, but which friction — information cost or cognitive over-reaction — does more of the work.
Where this leaves us: the thread’s verdict
Behavioral expectations is not the chapter where irrationality wins and rational expectations loses. It is the disciplined-deviation rung pushed to the cognitive frontier: diagnostic expectations and sparsity nest rational expectations as a limiting case, carry a small number of estimable distortion parameters, and are falsifiable. They keep rational expectations’ no-free-systematic-error discipline while explaining the over-reaction, credit cycles, and excess volatility that full rational expectations forbids. The terminus is integration, not triumph.
So the thread’s verdict, in full: rational expectations was right to discipline, wrong to describe, and the modern frontier is rational-expectations-as-benchmark plus disciplined deviation. Right to discipline — the Lucas critique is permanent, and you cannot model people as fooled the same way forever. Wrong to describe — surveys and asset prices reject full rational expectations as a literal account of belief formation. The frontier — rational inattention, adaptive learning, level-k reasoning, and behavioral expectations, each a constrained relaxation rather than a license for ad-hoc irrationality.
Name the layer precisely: this is method-settled, magnitude-contested. The field agrees on the modeling discipline — keep the benchmark, relax it only through explicit, estimable, rational-expectations-nesting frictions. What stays open is parameter-magnitude and mechanism-choice: how much deviation, which friction (information cost versus cognitive over-reaction versus learning), and how much it matters for any given question. That is not a “you decide” punt. It is a committed position with a live research frontier sitting exactly where the position says it should.
The asset-price face of all this — whether prices reflect rational-expectations fundamentals or diagnostic over-reaction — is the markets-side debate walked in Are markets efficient?
The thread, end to end
Follow the one apparatus across four eras. Keynes (1936) put expectations at the center — conventional, self-referential, unstable — but left them qualitative. Adaptive expectations made them concrete and backward-looking, and the expectations-augmented Phillips curve it powered predicted stagflation before stagflation arrived; it was retired not for failing the data but for the logical flaw of assuming people get fooled the same way forever. The rational-expectations revolution exploited that flaw: agents who don’t leave free money on the table, the Lucas critique, the death of the exploitable Phillips menu — a genuine and permanent advance. Then the surveys showed that real forecasters violate rational expectations systematically, and the field repaired rather than abandoned it, with disciplined deviations — rational inattention, adaptive learning, level-k — each keeping the no-free-lunch core. The last rung, behavioral expectations, borrowed prospect theory’s account of the distorting mind and built diagnostic expectations and sparsity: models that nest rational expectations, carry estimable parameters, and generate the credit cycles and excess volatility rational expectations cannot.
The thread did not end in “anything goes,” and it did not end in a coronation of either side. It landed on a settlement: rational expectations was right to discipline, wrong to describe, and the modern frontier is rational-expectations-as-benchmark plus disciplined deviation. The method is settled — no model of expectations may bank on free, exploitable, perpetual errors; deviations must be explicit, estimable, and benchmark-nesting. The magnitude is contested — how much deviation, which friction, and how much it matters is the live research frontier. The next time someone tells you “rational expectations proved Keynes wrong” or “behavioral economics proved people are irrational,” you have the thread to push past both slogans to the settlement they each ignore.