Wednesday, November 18, 2026

The Lucas critique at fifty

Robert Lucas published "Econometric Policy Evaluation: A Critique" in 1976. It runs to under 30 pages and contains one idea. That idea reorganised macroeconomics, and fifty years on it still sets the terms on which new ideas enter the field. In many ways it is the foundation stone of the New Keynesian synthesis, the framework central banks and mainstream academics still reason from.


What Lucas said

The large macroeconometric models of the 1960s estimated relationships between aggregates. Consumption depended on income. Investment depended on interest rates. Inflation depended on unemployment. Policymakers took the estimated coefficients and used them to simulate what a different policy would do.

Lucas argued that the coefficients were not fixed features of the economy. They recorded how households and firms behaved under the policy regime that produced the data. People form expectations about policy and act on them. Change the policy and you change the behaviour, so the coefficients move. A model fitted under one regime cannot tell you what will happen under another.


Where it came from

The worry was not new. Frisch and Haavelmo had asked in the 1930s and 1940s whether estimated relationships would survive a change in circumstances. Marschak made much the same point in 1953.

Lucas had two things his predecessors lacked. Muth had supplied rational expectations in 1961. Friedman and Phelps had argued in the late 1960s that the trade-off between inflation and unemployment would vanish once workers came to expect inflation. Lucas presented his paper in 1973, and the rest of the decade supplied the audience. Governments had treated the trade-off as a policy choice. Oil shocks and easy money then delivered rising inflation and rising unemployment together. The models' gravest failure came at the task Lucas said they could not reliably perform: predicting how the economy would behave under a different policy.


A warning and a blueprint

The critique carried two messages. The first was a caution about change: relationships shift when policy does, so be careful what you assume will hold still. The second was a building program. If estimated coefficients are unreliable, build models from things policy is less likely to alter. Lucas nominated tastes and technology. Derive behaviour from optimising agents who hold rational expectations and face explicit constraints, and the model should be more robust to a change of regime.

The goal was firmer foundations. Real business cycle theory followed in 1982. The New Keynesians added sticky prices to the same skeleton. The result was the DSGE model, which remains the workhorse of academic macroeconomics. I have a love-hate relationship with these models. They are conceptually beautiful, but beasts in practice.


Regimes are never fixed

The warning has held up completely. The blueprint assumed something the economy does not supply: a bottom layer that stays put.

Regimes are never fixed. Regulations change all the time. Clever people discover new ways to maximise utility. Financial innovation is a change in technology, and it arrives without notice. The money demand function looked stable until deregulation and new financial products broke it. Charles Goodhart saw it happen at the Bank of England in 1975 and gave the practitioner’s version of Lucas: a statistical regularity tends to collapse once policy leans on it for control. The popular form came later: when a measure becomes a target, it ceases to be a good measure. Australia floated the dollar in 1983, deregulated its banks, and abandoned its M3 targets in 1985 because the aggregate no longer meant what it had.

The supposedly deep parameters of DSGE models drift too when they are re-estimated over different periods. Some of that is the world changing for reasons unrelated to policy, which no model escapes. Some is more awkward. Firms reset prices more often when inflation is high, yet the standard models treat that frequency as a fixed structural parameter. It depends on the monetary regime. That is a failure of the models by Lucas's own test, not a failure of Lucas. "Structural" turns out to be a matter of degree.

David Hendry's argument is Lucas writ large, and sensible. Models break for many reasons, and a policy change is only one of them. Hendry's approach is to test whether a relationship stays constant instead of assuming it. Often the relationship held when the regime shifted. Ericsson and Irons surveyed the evidence in 1995 and found few documented cases of a relationship breaking in the way Lucas described. That does not dispose of Lucas. His claim was one of principle: any change in policy can shift the structure. How much, and in which relationships, you cannot know in advance.


Modellers must exercise judgement

If nothing is fully invariant, modellers sometimes have to exercise judgement. They always have. Every central bank forecast round adjusts the model's output for things the model cannot see.

The microfoundations program promised to discipline that judgement. Mostly it relocated it. Calibration is judgement about which features of the data to match. Bayesian estimation puts judgement in the priors, where at least it is stated openly. When the model cannot explain an episode, an unobserved shock absorbs the miss.

Central banks have hedged accordingly. The Federal Reserve's FRB/US and the Reserve Bank's MARTIN are semi-structural. They give up some theoretical DSGE purity for a better fit to the data.


A higher bar for ideas

The Lucas critique's largest effect was on how the discipline changes its mind. Before Lucas, a new idea had to fit the data. After him, it also had to say what agents know, what they want and what constrains them.

That is a higher bar, but not an impossible one. Sticky prices cleared it in the 1980s. Unemployment cleared it through search theory, which showed that workers and vacancies take time to find each other even when everyone behaves rationally. Limited attention cleared it through rational inattention and sticky information. Each idea was familiar long before it arrived in a form the profession would accept. The bar was also applied with judgement. Calvo pricing, the standard device for sticky prices, simply imposes the chance that a firm can change its price. Firms optimise when the chance comes, but nobody derived the chance itself. It passed because it was tractable and it worked. The cost of the bar is time. Ideas that are plainly true can wait a decade or more at the door.

The same bar protects against fashion. Modern Monetary Theory had its moment around 2019 and 2020, and the profession did not reorganise around it. A discipline that demands a good account of behaviour before it admits an idea does not change course on enthusiasm.

The same logic weighs on policy frameworks. If behaviour depends on the regime, data from inflation targeting cannot tell you how the economy would perform under nominal GDP targeting. A framework tends to change when two things coincide: the old one fails and a replacement stands ready. Bretton Woods broke down between 1971 and 1973, and floating rates, long argued for, took its place. Monetary targets failed in the 1980s and inflation targeting followed within a few years.

Nominal GDP targeting has stood ready for decades. It works within DSGE models, so the bar for ideas is not what holds it back. But those models were built and estimated under inflation targeting. Their verdict on a different regime is the kind of evidence Lucas warned against. 

Advocates of nominal GDP targeting say inflation targeting turned the banking crisis of 2008 into a deep recession. The crisis is better explained as primarily a failure of banking regulation than a failure of monetary policy. Monetary policy may have added to its intensity, but it is a big step from adding to the intensity to causing it. Some advocates say the failure is judgement itself: a rule would do better than a committee. But rules are built on history, and they often fail when the system goes outside it. The trend growth rate behind a nominal GDP target is an estimate, and estimates shift. What remains is an argument that inflation targeting is sub-optimal, which falls well short of failure. Failure is plain when it comes. It is a calamity, like the Great Depression or the stagflation of the 1970s, or it is a regime that has lost its usefulness, as monetary targets did. Inflation targeting is neither.

Inflation targeting works the way chemotherapy works. Slowing the economy stops demand-driven inflation, and the treatment hurts. We accept it because the disease is worse: inflation left to fester more often than not takes a deeper recession to cure. Most of the complaint is about the cost, not about whether the cure works.

None of this is permanent. The bar is high, not closed. Nominal GDP targeting lacks a failure to replace, and Modern Monetary Theory lacks the evidence. The right calamity could supply the first. The right data could supply the second.


Adjustment in 2008 and 2020

The two great shocks of this century produced adjustment, not replacement. That is telling.

After 2008 the profession added financial frictions to the existing frame. The modelling already existed. Bernanke, Gertler and Gilchrist had published the financial accelerator in 1999. Heterogeneous households came next. Those models built in an old observation: people with little liquid wealth spend transfers quickly. The pandemic payments of 2020 gave further evidence for it. The method survived while the assumptions inside it changed.

The inflation that followed is a mixed exhibit. In the United States, long-run expectations held through the 2021 to 2023 surge, and inflation fell a long way without a deep recession. But inflation never quite returned to target. Australia did worse. The Reserve Bank tightened less to protect employment. Headline inflation dipped into the target band, but government electricity subsidies were holding it down. Underlying inflation is the better guide, and it never got back on track. It returned briefly to the upper part of the band in 2025, never reached the midpoint, and was rising again before the year ended. Then war in the Middle East lifted oil prices. Inflation in both countries is back around 4 per cent and both central banks are raising rates. Households expect higher prices over the coming year, which any oil shock would produce. The long-run view is the real test. The University of Michigan survey puts long-run expectations at 3.4 per cent, against a range of 2.8 to 3.2 per cent in 2024. The anchor is under test a second time, and the result is not in.

DSGE models have a poor record on forecasting turning points. They missed 2008 and they missed the inflation of 2021 on the way up. But forecasting turning points was never Lucas's claim, nor what DSGE models were designed to do. His claim was that expectations and regimes govern how an economy responds to policy. Nothing in 2026 contradicts that. If expectations slip now, one explanation will be that central banks tolerated inflation above target for too long. People will have read the regime as it was practised, not as it was announced, which is what Lucas said they would do.

We are not back to normal. Inflation is unresolved, and we may well have a global AI bubble readjustment in the next year or two. A final judgement can wait for the end of the cycle. So far, the paradigm has held.

No comments:

Post a Comment