The Debate on Rational Expectations vs. Psych Evidence
Muth's elegant hypothesis reshaped macroeconomics and disciplined a generation of models — but decades of psychological evidence on how people actually form beliefs have reopened the question of what expectations are rational.
An Idea That Disciplined a Discipline
In 1961, John Muth published a short paper in Econometrica titled “Rational Expectations and the Theory of Price Movements.” It proposed something that sounds, at first, almost trivially obvious: when people form expectations about the future, their expectations should be consistent with the model that actually describes how the economy works. If a model predicts that a price increase will follow a supply shock, then agents in the model should expect that price increase. They should not systematically predict too high or too low. Their errors should be random, not patterned.
Muth was writing about commodity markets — specifically, the cobweb model of agricultural prices, in which farmers base planting decisions on last year’s price, generating cycles of overproduction and underproduction. The cobweb model assumed adaptive expectations: farmers look backward, extrapolating from recent experience. Muth argued that this was unsatisfactory. If the cycle is predictable, someone will figure it out and profit from it. In equilibrium, expectations should incorporate all available information, including the structure of the model itself.
The paper was largely ignored for a decade. Then Robert Lucas picked it up and used it to reconstruct macroeconomics.
Lucas and the Rational Expectations Revolution
Lucas’s contribution, developed in a series of papers in the early 1970s, was to apply Muth’s hypothesis to macroeconomic policy. The Keynesian models of the 1960s assumed that policymakers could exploit a stable tradeoff between inflation and unemployment — the Phillips curve. If the government wanted lower unemployment, it could accept higher inflation. Lucas argued that this tradeoff was an artifact of adaptive expectations. Workers and firms based their wage and price decisions on past inflation, so a surprise burst of inflation would temporarily boost real output (because wages lagged behind prices). But if expectations were rational, agents would anticipate the inflation and adjust immediately. The tradeoff would vanish.
This was the famous Lucas critique: econometric models estimated under one policy regime would break down when the regime changed, because the model’s parameters were not structural — they depended on agents’ expectations, which depended on the policy. A model estimated during a period of stable monetary policy could not predict what would happen under an inflationary policy, because agents would change their behavior in response to the new policy.
The Lucas critique was devastating to the Keynesian consensus. It implied that large-scale macroeconometric models — the workhorses of policy analysis — were unreliable guides to the effects of policy changes. It launched a methodological revolution that demanded microfoundations: macroeconomic models had to be built from the optimizing behavior of individual agents with rational expectations. This program produced real business cycle theory, New Keynesian economics, and the dynamic stochastic general equilibrium (DSGE) models that dominate academic macroeconomics to this day.
The rational expectations hypothesis (REH) became the profession’s default assumption about how agents form beliefs. Not because anyone thought people literally solved complex optimization problems in their heads, but because it was seen as the only assumption that was internally consistent and resistant to the Lucas critique. Any other assumption — adaptive expectations, naive expectations, ad hoc rules of thumb — could be exploited by agents who did use all available information, driving the ad hoc forecasters out of the market or making their behavior unsustainable.
The Power of RE: What It Gets Right
The rational expectations hypothesis is not merely a modeling convenience. It has genuine intellectual power.
It eliminates free lunches. Under RE, there are no systematic, predictable errors that agents could exploit. If a central bank always increases the money supply by 5% per year, agents expect 5% money growth and set prices accordingly. The central bank cannot “fool” the public into producing more by creating surprise inflation, because there is no surprise. This is descriptively realistic in many financial markets, where professional forecasters and trading algorithms quickly incorporate public information.
It disciplines models. Before RE, modelers could assume almost any expectation formation process, giving them enormous freedom to fit the data. RE removes this degree of freedom: the expectations in the model must be consistent with the model’s own predictions. This is a demanding requirement, but it ensures that the model is internally coherent — agents within the model are not systematically wrong about the model they inhabit.
It formalizes the notion that expectations matter. Before Muth and Lucas, the role of expectations in economics was acknowledged but imprecise. Everyone agreed that expectations about future prices, incomes, and policies affect current behavior. But the models of expectation formation were ad hoc — last year’s inflation, an average of recent experience, a weighted combination of past observations. RE provided a unified, rigorous framework for modeling expectations as endogenous to the economic system.
It generates testable predictions. If agents have rational expectations, then forecast errors should be unpredictable given available information. This is testable using survey data on expectations or by examining whether asset prices fully reflect public information. The efficient market hypothesis in finance is essentially the RE hypothesis applied to stock prices.
The Challenge from Psychology
Against the intellectual power of RE stands a mountain of evidence from psychology and behavioral economics documenting systematic, predictable departures from rational belief formation.
Anchoring. When people estimate uncertain quantities, they are heavily influenced by arbitrary starting points — “anchors” — even when the anchors are obviously irrelevant. Tversky and Kahneman (1974) showed that spinning a wheel of fortune before asking people to estimate the percentage of African countries in the United Nations dramatically affected their answers. The anchor dragged the estimate toward it. In economic contexts, anchoring affects wage negotiations, house price estimates, damage awards in lawsuits, and inflation expectations.
Representativeness. People judge the probability of an event by how well it matches a prototype or stereotype, rather than by base rates or statistical reasoning. A sequence of coin flips that reads HTHHTTH “looks” more random than HHHHTTT, even though both are equally likely. In financial markets, representativeness leads to extrapolation from short streaks: a stock that has risen for three straight quarters “feels” like a winner, even though three observations provide almost no statistical information about future performance. De Bondt and Thaler (1985) documented that stocks with recent extreme gains subsequently underperformed, consistent with representativeness-driven overreaction.
Overconfidence. People are systematically overconfident in their own judgments. When asked to give 90% confidence intervals for uncertain quantities, people’s intervals are far too narrow — the true value falls outside the interval 50% of the time or more. In financial markets, overconfidence generates excessive trading (Barber and Odean, 2001) and excessive entry into competitive markets (Camerer and Lovallo, 1999). Overconfident agents do not form expectations consistent with the true model; they form expectations consistent with a model in which they are smarter, luckier, or more informed than they actually are.
Narrative bias. People form expectations through narratives — causal stories that organize and explain events — rather than through statistical models. Robert Shiller has argued that economic fluctuations are driven in part by “narrative epidemics,” in which compelling stories (about new technology, housing wealth, or financial collapse) spread through the population and shape expectations in ways that have nothing to do with rational information processing.
Inattention. As the rational inattention literature documents, people do not process all available information because doing so is costly. If forming rational expectations requires integrating all public information about the economy, monetary policy, fiscal policy, global trade, and technological change, then most people simply cannot do it. They attend to a few salient signals and ignore the rest. Coibion and Gorodnichenko (2012, 2015) show that information rigidity — the slow updating of expectations in response to new information — explains a large portion of the forecast errors in both professional and household surveys.
What the Surveys Say
If rational expectations hold, then survey measures of expectations should be unbiased and efficient: the average forecast should equal the average outcome (no systematic error), and forecast errors should be uncorrelated with any information available at the time the forecast was made.
The evidence is mixed. For professional forecasters — the economists who make a living predicting GDP growth, inflation, and interest rates — the evidence for RE is surprisingly weak. The Survey of Professional Forecasters, maintained by the Federal Reserve Bank of Philadelphia, reveals persistent patterns in forecast errors. Forecasters underestimate inflation when it is rising and overestimate it when it is falling. They herd — their forecasts are more clustered than their private information would warrant, suggesting that they anchor on the consensus. And their forecast errors are predictable from information that was available at the time of the forecast, violating the efficiency condition of RE.
For households, the departures from RE are larger. The Michigan Survey of Consumer Expectations shows that households’ inflation expectations are systematically biased, correlated with demographic characteristics (women expect higher inflation than men, older people expect higher inflation than younger people), and responsive to recent personal experiences (people who have recently seen large price increases at the gas pump or grocery store expect higher overall inflation) in ways that RE does not permit.
These findings do not prove that RE is “wrong” in the sense of being useless. They do suggest that it is wrong in the sense of being descriptively inaccurate as a characterization of how people actually form expectations. The question is whether this inaccuracy matters for the predictions of macroeconomic models.
Adaptive and Diagnostic Expectations: The Middle Ground
The dissatisfaction with both RE (too demanding) and ad hoc alternatives (not disciplined enough) has driven a search for expectation models that take psychology seriously while remaining tractable and empirically grounded.
Adaptive learning. In models of adaptive learning, agents do not know the true model of the economy. Instead, they estimate it from data, updating their beliefs as new observations arrive — much as an econometrician would. Over time, if the economy is stationary, their estimates converge to the rational expectations equilibrium. But along the way, their expectations can be wrong in systematic and persistent ways, generating dynamics that RE models cannot produce: boom-bust cycles driven by learning, overreaction to structural breaks, and slow convergence to new equilibria after policy changes.
The adaptive learning approach, developed by George Evans and Seppo Honkapohja, has become a mainstream alternative to RE in macroeconomics. It retains the discipline of RE as a limiting case while allowing for realistic deviations from it in the short and medium run.
Diagnostic expectations. Pedro Bordalo, Nicola Gennaioli, and Andrei Shleifer have proposed a model in which expectations are formed by a process that overweights information that is representative — in the Kahneman-Tversky sense — of the current state of the economy. If the economy has recently experienced good news, agents overweight the likelihood of continued good news, generating over-optimism. If the economy has experienced bad news, agents overweight the likelihood of continued bad news, generating over-pessimism.
The model produces boom-bust cycles, excess volatility in asset prices, predictable reversals in expectations (over-optimistic forecasts are followed by negative surprises), and credit cycles in which lending expands during booms and contracts during busts — all patterns that are documented in the data and difficult to explain under RE.
Diagnostic expectations are not adaptive expectations in the old sense. They do not simply extrapolate from the past. Instead, they use the past to assess which scenarios are most “representative” of the current situation and then overweight those scenarios. The model is disciplined — it has a single free parameter that governs the degree of overweighting — and it generates quantitative predictions that can be tested against survey data and asset prices.
Natural expectations. Fuster, Laibson, and Mendel (2010) proposed a model in which agents use a simplified statistical model to forecast the future — for example, using an AR(1) process when the true process is more complex. The simplification generates predictable forecast errors that match features of the survey data, including over-extrapolation from recent trends and insufficient response to long-run mean reversion.
The “As If” Defense and Its Limits
Defenders of RE have long invoked the “as if” argument: even if individual people do not literally form rational expectations, market outcomes may behave as if they do, because the agents who are closest to rational will dominate markets (through arbitrage, competition, or natural selection). A noise trader who systematically overestimates future stock prices will buy high and sell low, losing money until they are driven from the market. In the long run, the rational agents remain, and prices reflect rational expectations.
The “as if” defense has force in some contexts. In liquid financial markets with low barriers to entry and rapid feedback, it is plausible that prices approximately reflect rational expectations, at least on average. The efficient market hypothesis, while imperfect, has been a reasonable description of many asset markets for many decades.
But the defense has clear limits. In many economically important settings — household savings decisions, labor supply choices, voting, health care decisions — there is no selection mechanism that drives out irrational agents. A household that under-saves for retirement does not go “bankrupt” in a way that removes them from the economy. A voter who forms expectations based on narrative bias continues to vote. The “as if” argument requires a selection mechanism, and in much of the economy, no such mechanism operates.
Moreover, even in financial markets, the limits of arbitrage — documented by Shleifer and Vishny (1997) — constrain the ability of rational agents to correct mispricing. Arbitrage requires capital, bears risk, and operates over time horizons that may be shorter than the mispricing persists. The dot-com bubble and the 2008 financial crisis both demonstrated that rational agents can know that prices are wrong and still be unable — or unwilling — to bet against the crowd.
Can We Have Disciplined Models Without RE?
The central question is whether the discipline that RE provides — internal consistency, robustness to the Lucas critique, the elimination of free parameters in expectation formation — can be preserved without the assumption that agents are perfectly rational forecasters.
The answer, increasingly, is yes — or at least, we are getting closer. Models of adaptive learning, diagnostic expectations, and natural expectations all impose structure on how agents form beliefs. They are not ad hoc; they are derived from psychological evidence and disciplined by the requirement that they match observed patterns in survey data, asset prices, and macroeconomic dynamics. They are robust to the Lucas critique in the sense that the psychological mechanisms they model (overweighting representativeness, simplifying statistical models, anchoring on recent experience) are plausibly stable across policy regimes, even if the specific forecasts they generate are not.
The profession has not abandoned RE, nor should it. In many applications — especially in financial markets and in models where the primary goal is qualitative rather than quantitative prediction — RE remains a useful and powerful benchmark. It tells you what would happen if agents were perfectly informed and perfectly rational, and deviations from this benchmark are informative about the role of bounded rationality.
But the era in which RE was the only acceptable assumption is ending. The behavioral and survey evidence is too strong, and the alternative models are now sophisticated enough, to sustain the pretense that all agents form expectations as if they had a PhD in economics and access to the central bank’s internal model. The frontier of macroeconomics is a search for expectation models that are psychologically realistic, empirically grounded, theoretically disciplined, and capable of generating the rich dynamics — the booms, busts, bubbles, and panics — that characterize the actual economy.
That search is far from over. But the fact that it is underway, and that it is taken seriously by mainstream macroeconomists and not only by behavioral economists, represents a significant shift in the discipline. Muth gave economics a benchmark. The next generation is building beyond it — not by abandoning discipline, but by grounding it in a more accurate picture of how people actually think.