Zellner-Chen (2001) Bayesian Modeling of Economies and Data Requirements

Summary

Presented as a keynote at the 2000 International Institute of Forecasters conference, this paper introduces the Marshallian Macroeconomic Model (MMM) — a sectoral structural model grounded in Marshall's demand, supply, and entry equations — and uses it to forecast U.S. annual real GDP growth rates for 1980–1997. The central finding is that disaggregated sectoral forecasts summed to aggregate GDP substantially outperform aggregate-only benchmarks, especially when combined with Bayesian shrinkage. The paper also develops the restricted reduced-form seemingly unrelated regression (SUR) interpretation of the structural sectoral equations and discusses data improvements needed for extending the MMM.

Key Claims

• The MMM solves Marshall's three-equation sector system (demand, supply, entry) to yield a logistic differential equation for sectoral output $S$: $(1/S)\,dS/dt = a(1 - S/F) + g$, where $F$ is long-run capacity and $g$ is a linear function of growth rates of wages, capital prices, and demand shifters (real income, real money). Both the rate of change and level of $S$ appear, building in a cointegration-like correction mechanism.
• Writing the structural equation in restricted reduced form and allowing for correlated reduced-form errors produces a nonlinear SUR system with an error covariance restriction; Pagan (1979) earlier noted this triangularity-SUR connection.
• Disaggregating across 11 U.S. sectors and summing sectoral forecasts improves aggregate GDP forecast precision over aggregate-only models in all cases. Best result: MMM(DA)III with SUR achieves root mean squared error (RMSE) = 1.40 and mean absolute error (MAE) = 1.17 vs AR(3) RMSE = 2.26, MAE = 1.65 (1980–1997, currency variable).
• Currency (real) is a better leading indicator variable than M1 in this context; models using currency achieve roughly 0.3–0.5 RMSE improvement over those using M1.
• Four sectoral MMM(DA) variants (I–III include 1–3 lagged level variables; IV uses lagged level and lagged level squared) are compared against AR(3), AR(3)LI, and Distributed Lag benchmarks; MMM(DA)III consistently dominates across estimation methods.
• Among estimation methods — ordinary least squares (OLS), two-stage least squares (2SLS), Extended Minimum Expected Loss (MELO; Zellner 1986/1998 balanced loss with weight $w$: $L = w\cdot\text{fit}^2 + (1-w)\cdot\text{precision}^2$), SUR, complete shrinkage, gg-shrinkage, hh-shrinkage — differences are small, suggesting that endogeneity of wages and income is mild in this sector-level setting.
• gg-shrinkage (g-prior based, shrinks toward zero) with g = 0.25–0.5 performs well; further shrinkage (larger g) mildly deteriorates precision. hh-shrinkage (toward sector means) performs best at h = 0.5–0.75 in several models.
• Bayesian and non-Bayesian point forecasts perform similarly under diffuse priors; Bayesian advantage is the predictive density, which enables turning-point probability statements.
• Direct Monte Carlo (Zellner 1998) provides exact finite-sample posterior and predictive densities for the simultaneous equations system without Markov chain Monte Carlo (MCMC) burn-in.
• MMM(DA) MAEs of 1.17–1.46 are competitive with professional GDP forecasters (Zarnowitz 1986 compilation: MAEs of 1.0–1.3 across various institutions, 1953–1984).

Concepts Introduced or Extended

• Bayesian Method of Moments
• Seemingly Unrelated Regression
• Bayesian VAR

Entities Mentioned

• Arnold Zellner

Quotes

"The empirical evidence indicates that it pays to disaggregate, particularly when employing Bayesian shrinkage forecasting procedures."

"Bayesian and certain non-Bayesian point forecasts performed about equally well for our disaggregated MMM models. However, the Bayesian approach provides exact finite sample posterior densities for parameters and predictive densities."

My Take

An interesting attempt to revive structural micro-founded macro forecasting at a time (2000) when VARs and atheoretical time-series models dominated. The logistic sectoral model has real theoretical appeal (Marshall's entry equation prevents the "representative firm shuts down" pathology of many macro models), and the disaggregation result is convincing — 11 sectors provide both more observations and sector-specific variation that aggregates wash out. The practical limitation is data: sectoral stock price indices, firm count data, and quality-adjusted output prices remain difficult to obtain at the frequency needed for forecasting. The paper also predates modern Dynamic Stochastic General Equilibrium (DSGE) models with multiple sectors, which have since become the main vehicle for Marshallian-style disaggregated macro analysis. The Extended MELO estimator is an interesting connection between Bayesian and frequentist estimation under asymmetric loss but is rarely cited outside Zellner's own work.