Why Policy Changes Structurally Invalidate the Economic Models Used to Predict Them

Imagine a $1.5 trillion hole in a federal budget—roughly the size of the entire U.S. discretionary spend for a single year. That is the scale of the variance a government encounters when it assumes human beings will not change their behavior in response to a new law. When the Joint Committee on Taxation evaluated the 2017 Tax Cuts and Jobs Act, it produced two numbers: a $1.5 trillion deficit increase under 'static' scoring, which assumes behavior remains frozen, and a $1.1 trillion deficit under 'dynamic' scoring, which attempts to guess how people will react. The $400 billion difference between those two models illustrates the most persistent vulnerability in modern governance: the models used to predict the outcome of a policy are structurally invalidated by the policy itself.[3][4]

The argument here is straightforward: macroeconomic forecasting fails not because the math is flawed or the data is sparse, but because the parameters these models rely on are artifacts of the old rules. Once the rules change, the parameters change. This phenomenon is not a minor statistical friction; it is a fundamental paradox that guarantees any policy justified by historical data will produce outcomes that diverge from its projections. The strongest counter-argument is that modern dynamic models, which explicitly attempt to forecast behavioral shifts, have solved this problem. But as the historical record shows, even dynamic models struggle to anticipate the sheer creativity of human adaptation.[6]

In 1976, University of Chicago economist Robert Lucas Jr. published a critique that devastated the prevailing macroeconomic models of his era. The Lucas Critique argued that it is naive to predict the effects of a change in economic policy entirely on the basis of relationships observed in historical data. Why? Because the historical data was generated by individuals optimizing their choices under the old policy regime. If the government changes the regime, individuals will optimize for the new one, rendering the historical data obsolete.[1]

The lived-out evidence of Lucas’s argument is the catastrophic failure of the Phillips curve in the 1970s. Throughout the 1960s, policymakers observed a stable, inverse relationship between inflation and unemployment: when one went up, the other went down. Based on this historical data, central banks assumed they could permanently buy lower unemployment by tolerating slightly higher inflation. They treated the historical curve as a structural law of nature, failing to realize that it only existed because inflation had previously been low and stable, keeping public expectations anchored.

But the moment policymakers attempted to exploit this trade-off, the relationship collapsed. Workers and firms realized that inflation was being deliberately elevated, so they changed how they formed their expectations. Workers demanded aggressive cost-of-living adjustments in their contracts, and firms raised prices preemptively to protect their margins. The result was 'stagflation'—a simultaneous rise in both inflation and unemployment that the 1960s models had deemed mathematically impossible. The policy had altered the behavioral parameters that justified it in the first place.

This mechanism is not limited to central banking; it is a universal feature of human systems, captured by Goodhart’s Law. Coined by British economist Charles Goodhart in 1975, the law states that 'when a measure becomes a target, it ceases to be a good measure.' Or, in its stricter formulation: any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes. Goodhart recognized that proxy metrics fail the moment they are incentivized, because humans will always optimize for the metric rather than the underlying goal.[2]

Goodhart’s Law explains why seemingly rational targets produce absurd outcomes. If a factory is measured by the number of nails it produces, it will produce thousands of tiny, useless nails. If a hospital is measured by patient recovery rates, it will refuse to admit the severely ill to protect its average. The metric was a reliable indicator of performance only when it was not being targeted. Once it became the explicit goal, the subjects gamed the system, destroying the metric's predictive value entirely.[5]

The modern arena for this conflict is the ongoing debate over how the Congressional Budget Office (CBO) and the Joint Committee on Taxation (JCT) score proposed legislation. For decades, these agencies relied primarily on static scoring. If the government lowered a 50 percent tax rate to 25 percent on $100 of income, the static model projected a $25 loss in revenue. It assumed the taxpayer would continue to earn exactly $100, completely ignoring the fact that a lower tax rate changes the fundamental incentive structure of work and investment.[3][4]

But taxpayers are not static entities. A lower rate increases the incentive to work longer hours, invest more capital, and report income that might otherwise have been sheltered. Recognizing this reality, Congress increasingly requires dynamic scoring, which uses macroeconomic models to estimate how a policy will alter the overall size of the economy. Dynamic scoring attempts to capture the 'growth dividend'—the additional tax revenue generated because the tax cut stimulated more economic activity than the static baseline could foresee.[3][4]

Yet dynamic scoring is itself deeply vulnerable to the Lucas Critique. The models used to predict the behavioral response to a tax cut are built on historical data of how people responded to previous tax cuts. But the economy of 2026 is not the economy of 1986. The structure of global capital, the prevalence of remote work, and the sheer complexity of modern corporate supply chains mean that historical behavioral parameters are unlikely to hold. When the JCT models the macroeconomic feedback of a new law, it is still guessing at how human ingenuity will exploit the new rules.[6]

The transparency of these models is crucial to understanding their limitations. Proponents of dynamic scoring argue that it removes a structural bias against pro-growth policies, providing a more accurate picture than static models that pretend human behavior is frozen in time. Detractors counter that dynamic models require scorekeepers to make highly subjective assumptions about future behavior, risking the politicization of non-partisan agencies. Both sides are arguing over which set of flawed assumptions is less dangerous to the legislative process.[3][4]

The tension between static and dynamic forecasting reveals a hard limit on technocratic governance. Policymakers desperately want the certainty of an engineering schematic, where pulling a legislative lever produces a guaranteed, mathematically precise mechanical result. But an economy is not a machine; it is a complex adaptive system composed of millions of forward-looking agents who actively read the government's schematic and adjust their behavior to outmaneuver it. Every new regulation, tax bracket, or subsidy acts as a starting gun for accountants, lawyers, and citizens to find the most efficient path through the new maze.[1][2]

This does not mean that economic modeling should be abandoned entirely. It means that models must be treated as baseline scenarios rather than crystal balls. A robust policy is not one that perfectly predicts the future based on the past, but one that remains structurally sound even when the public inevitably changes its behavior to game the new regime. The true test of a policy is not how it scores on the day it is passed, but how it survives the adaptation it provokes.[6]

The next major test of this dynamic will arrive when the massive 2017 tax provisions finally expire. As Congress debates whether to extend them, the CBO and JCT will produce new scores, and lawmakers will again wield those projections as absolute, unassailable truths. But the numbers printed on those pages will only represent how Americans behaved under the old code. The moment the new code takes effect, the parameters will shift, the models will break, and the public will begin optimizing for a reality the scorekeepers could not possibly foresee.[6]