Structural Econometric-Time Series Analysis

Definition

Structural Econometric-Time Series Analysis (SEMTSA) is a modeling strategy introduced by Zellner and Palm (1974, 1975) that links structural economic models to their implied reduced-form time-series representations. The key idea is to start from a theoretically grounded structural model (demand-supply, Investment-Savings/Liquidity-Money (IS-LM), real business cycle), derive the AutoRegressive Integrated Moving Average (ARIMA) or vector autoregression (VAR) forecasting equations that the structure implies, test whether those implied equations fit the data, and use them for forecasting. This provides a discipline on model specification that neither pure atheoretical time series nor purely structural estimation achieves alone.

Key Ideas

• Theory-to-time-series mapping: any structural economic model implicitly implies a set of AutoRegressive Moving Average (ARMA)/VAR representations for observed variables. Zellner-Palm (1974) showed how to derive these implied representations from the structural form; mismatches between the implied and empirically fitted ARMA processes reveal specification errors in the structural model.
• Sequential model-building: begin with a simple, theory-consistent ARIMA model; check compatibility with the structural model; enrich with additional variables (leading indicators, world income) only when theory or forecast evaluation supports it.
• ARLIWI model (Autoregressive + Leading Indicator + World Income): the workhorse SEMTSA equation for gross domestic product (GDP) growth forecasting. For country $i$: $\Delta y_{it} = \delta_0 + \sum_j \delta_j \Delta y_{i,t-j} + \delta_4 \Delta SR_{t-1} + \delta_5 \Delta m_{t-1} + \text{world-income terms}$, where $SR$ = real stock prices and $m$ = real money. Derived from aggregate supply-demand, Hicksian IS-LM, and generalized Real Business Cycle (RBC) models (Hong 1989; Min 1992). Applied to 18 industrialized countries; pooled root-mean-square error (RMSE) 1.17–2.53%, median 1.74% one-year ahead.
• Turning-point forecasting: using the ARLIWI predictive density, compute $P = \Pr(\Delta y_{t+1} < \Delta y_t \mid \text{data})$. Forecast "downturn" if $P > \tfrac{1}{2}$ under symmetric loss. Achieves ~70% accuracy across 211 turning-point episodes.
• KISS principle: Keep It Sophisticatedly Simple (Zellner). Complicated models fail systematically against simple AR benchmarks in macro forecasting; SEMTSA favours parsimony guided by theory rather than complexity.

How It Works

1. Specify a structural model, e.g. demand $y_d = f(p, z_1)$, supply $y_s = g(p, z_2)$, entry $dN/dt = h(p, N)$.
2. Solve the structural model for the observed variables in terms of exogenous shocks and lagged endogenous variables.
3. Derive the implied ARMA or VAR representation of each observed variable.
4. Fit the implied ARMA/VAR and test whether the derived representation matches the data.
5. Use the resulting equations — augmented with leading indicator variables if supported empirically — for point and turning-point forecasting with Bayesian posterior/predictive densities.

The Marshallian Macroeconomic Model (MMM, Zellner-Chen 2000) is SEMTSA applied at the sectoral level, combining the three Marshallian equations per sector to produce a logistic-form GDP growth equation, then fitting and summing across 11 U.S. sectors.

Why It Matters

SEMTSA provides a principled answer to the Sims (1980) critique that large structural models are "incredible" due to arbitrary exclusion restrictions. Rather than abandoning structural models, SEMTSA uses them selectively — only to discipline the choice of variables and lag structure — while retaining the flexibility to fit the implied time-series representation. The result is a middle path between the atheoretical VAR and the fully specified simultaneous-equations model. Empirically, SEMTSA-derived ARLIWI models outperformed benchmark VARs and many large-scale macro models in the 1980s-1990s for GDP point and turning-point forecasting.

Open Questions

• The SEMTSA program pre-dates modern Dynamic Stochastic General Equilibrium (DSGE) models; it is unclear how the approach would handle forward-looking expectations or stochastic general equilibrium structures.
• The ARLIWI model was calibrated to 18 industrialized countries in the 1970s-1990s; generalizability to emerging markets or post-2008 data is untested.

Related

• Bayesian VAR
• Vector Autoregression
• Bayesian Method of Moments
• Zellner (c.2000)
• Zellner-Chen (2000)
• Arnold Zellner