By the early 2000s, macroeconomics faced a credibility and usability problem that was as much institutional as academic. Central banks needed quantitative models that could speak to policy counterfactuals—What if we raise rates by 50 basis points? What if oil prices spike?—while remaining disciplined by the data. Yet the profession had spent decades oscillating between richly estimated reduced-form econometrics (notably VARs) that were hard to interpret structurally, and microfounded dynamic general equilibrium models that were elegant but often weakly connected to empirical time series. The post-1980 “Great Moderation” seemed to vindicate rule-like monetary policy, but it also raised the stakes: small misspecifications could produce large policy errors. Against this backdrop, the 2003–2005 work of Frank Smets and Rafael Wouters—culminating in their widely cited 2003 and 2007 papers and heavily influencing practice by 2005—provided a practical blueprint for “modern” DSGE modeling inside central banks: estimate a medium-scale New Keynesian model with Bayesian methods, evaluate it with formal likelihood-based tools, and use it for forecasting and policy analysis.
The core idea was to make DSGE models empirically competitive without abandoning structural discipline. Smets and Wouters built a “medium-scale” framework: large enough to fit key macro time series, but structured enough to interpret shocks and policy trade-offs. The model combines (i) microfounded intertemporal decisions by households and firms, (ii) nominal rigidities that give monetary policy real effects, and (iii) real frictions and multiple shocks to match observed persistence. Canonical ingredients include habit formation in consumption (so current consumption depends partly on past consumption), investment adjustment costs (so investment responds sluggishly to shocks), variable capital utilization, and wage and price stickiness à la Calvo. These mechanisms were not new individually; the innovation was their integration into a coherent, estimable system that could compete with benchmark statistical models in predictive performance.
Monetary policy modeling sits at the center. Policy is typically represented by an interest-rate rule, most commonly a Taylor-type rule with inertia: the central bank smooths interest rates and reacts to inflation and the output gap. A standard formulation is $$ i_t = \rho\, i_{t-1} + (1-\rho)\big(\phi_\pi \pi_t + \phi_y y_t\big) + \varepsilon^i_t, $$ where (i_t) is the nominal policy rate, (\pi_t) inflation, (y_t) (often) the output gap or output growth, (\rho\in(0,1)) captures interest-rate smoothing, and (\varepsilon^i_t) is a monetary policy shock. The rule is not merely an empirical fit; within a DSGE it closes the model and allows counterfactual experiments under alternative policy regimes.
The distinctive methodological contribution—especially salient to central banks by 2005—was Bayesian DSGE estimation. Traditional calibration fixed many parameters by assumption. Classical maximum likelihood was possible but brittle in high-dimensional nonlinear systems with weak identification. Bayesian methods offered a disciplined compromise: combine prior information (from micro evidence, long-run ratios, or earlier studies) with the likelihood from macro data to obtain a posterior distribution over parameters. Formally, for parameters (\theta) and data (Y), $$ p(\theta\mid Y) \propto p(Y\mid \theta)\, p(\theta), $$ where (p(\theta)) is the prior and (p(Y\mid\theta)) the likelihood implied by the DSGE after linearization and solution.
Operationally, the model is log-linearized around a steady state and written in state-space form: $$ x_{t} = A(\theta) x_{t-1} + B(\theta)\varepsilon_t,\qquad y_{t} = C(\theta) x_{t} + D(\theta)\eta_t, $$ with (x_t) latent states (e.g., technology, capital, expected inflation), (y_t) observed variables (GDP growth, inflation, interest rates, hours, wages), structural shocks (\varepsilon_t), and possible measurement errors (\eta_t). The likelihood (p(Y\mid\theta)) is evaluated via the Kalman filter, and the posterior is explored using Markov chain Monte Carlo (MCMC), typically Metropolis–Hastings. This machinery transformed DSGEs from “theory with simulations” into an estimable class of structural econometric models with standard diagnostics: marginal likelihoods for model comparison, posterior intervals for uncertainty, and impulse responses conditional on estimated shocks.
Historically, the impact was immediate within policy institutions. By 2005, the logic of Smets–Wouters-style estimation had become a reference point for model-based forecasting and scenario analysis at major central banks. It helped legitimize DSGEs as practical tools, not just academic constructs, and encouraged a common language between researchers and policymakers: decompositions into interpretable shocks, policy counterfactuals under alternative rules, and probabilistic statements about parameters and forecasts. It also raised the bar for empirical validation. Medium-scale DSGEs began to be judged against VARs on out-of-sample forecasting, and Bayesian model comparison made such “horse races” methodologically coherent.
Criticism came from multiple directions. Some econometricians worried that strong priors could “do the work,” masking weak identification; Bayesian discipline does not eliminate the possibility that the data are uninformative about key parameters. Others argued that the proliferation of shocks—technology, preference, investment-specific, price markup, wage markup, risk premium—risked turning DSGEs into “structural VARs with tastes,” where fit is purchased by adding disturbances rather than improving mechanisms. A deeper critique, later sharpened by the 2007–2009 global financial crisis, was that standard Smets–Wouters-era models had limited financial sectors and downplayed balance-sheet constraints, default risk, and nonlinear crisis dynamics. Even before the crisis, skeptics questioned whether linearization around a stable steady state could capture rare but consequential events.
The legacy, nonetheless, is durable. Modern central-bank DSGEs—whether in the euro area, the United States, or small open economies—largely inherit the Smets–Wouters template: estimated, shock-rich, New Keynesian cores with Bayesian inference and state-space implementation. Today’s variants often add financial frictions (e.g., spreads and borrower constraints), heterogeneous agents, occasionally binding constraints (like the effective lower bound), and richer expectation formation. But the procedural spine remains: explicit structural equations, likelihood-based estimation, and quantified uncertainty.
In the longer line of continuity, Smets and Wouters sit at the junction of three traditions. They build on real business cycle theory of the 1980s (Kydland and Prescott’s emphasis on microfoundations and explicit shocks), incorporate New Keynesian nominal rigidities and policy rules developed in the 1990s, and adopt Bayesian methods popularized in macroeconometrics by the late 1990s and early 2000s. They also respond to the Lucas critique by ensuring policy analysis is grounded in preference and technology primitives rather than reduced-form correlations. After 2005, their approach becomes the platform from which newer paradigms depart: financial DSGEs after the crisis, nonlinear and regime-switching methods for the zero lower bound, and, more recently, hybrid models that blend DSGE structure with data-driven forecasting tools. In that sense, the “modern DSGE” is less a fixed model than an institutionalized workflow—one that Smets and Wouters did much to standardize for central banking and structural econometrics.