Fully atheoretical models like VARs treat every variable symmetrically and let the data decide everything. Fully structural models like DSGE impose general equilibrium from first principles. Central banks found both extremes unsatisfying. Reduced-form VARs forecast well at short horizons, and identified (structural) VARs can estimate the effects of policy shocks, but neither can say what happens under a systematically different policy rule without assuming the estimated relationships stay put (the Lucas critique). DSGE models can answer counterfactuals but forecast poorly at horizons that matter for monetary-policy decisions. The semi-structural approach splits the difference: impose a small number of structural equations grounded in economic theory (IS curve, Phillips curve, policy rule, exchange rate arbitrage), but estimate their coefficients from the data rather than deriving them from household optimization. The Bank of Canada's Quarterly Projection Model (QPM, 1994) and the Reserve Bank of New Zealand's FPS were early examples. The IMF's FPAS framework (Berg, Karam, and Laxton 2006) generalized the approach for member-country surveillance.
A semi-structural core typically contains four to six behavioral equations. An IS-type equation links the output gap to the real interest rate gap, the real exchange rate gap, and foreign demand. A Phillips-curve equation links inflation to the output gap, expected inflation, and import prices. A monetary-policy reaction function (Taylor-type rule) links the nominal interest rate to the inflation gap and the output gap. An uncovered interest parity condition links the exchange rate to the interest differential. Trend-cycle decomposition equations separate each observed variable into a stochastic trend (potential output, equilibrium real rate, trend inflation) and a cyclical gap. The gaps, not the levels, drive the behavioral equations.
What distinguishes the semi-structural core from reduced-form VARs is the a priori identification of structural shocks. What distinguishes it from DSGE is the absence of explicit intertemporal optimization or cross-equation restrictions derived from micro foundations. Parameters are estimated or calibrated equation by equation. Expectations can be model-consistent (forward-looking, iterated from the model's own solution) or adaptive (backward-looking weighted averages). Most implementations use a mix: forward-looking expectations in the Phillips curve and policy rule, backward-looking dynamics in the IS curve.
The Bank of Canada's LENS, the IMF's Global Projection Model (GPM), and the Reserve Bank of India's QPM all use semi-structural cores. (The Czech National Bank's g3/g3+, by contrast, is a micro-founded DSGE that replaced its semi-structural QPM in 2008.) These models are the workhorse for medium-term forecasting (4-8 quarter horizon) and scenario analysis at inflation-targeting central banks. They are updated quarterly with judgment overlays - staff can adjust trends, impose conditioning paths for commodity prices or fiscal policy, and override model-implied gaps when the filter produces implausible estimates.