Research lab · Interactive models · Monetary policy
The semi-structural model that the Banco Central do Brasil publishes in its inflation report, carried to the US: Phillips curves for core, goods, food and services, an IS curve on a clean output gap, a policy rule on a time-varying neutral rate and an expectations block. Run a scenario from the last observed quarter, trace impulse responses, read the estimates and the out-of-sample record.
The three experiments of the Banco Central do Brasil box, on the US estimates. Deviations from steady state, four-quarter accumulated inflation in points.
Posterior mode and 90% interval, 2003Q4 to 2026Q3, with the Brazilian reference values from the central bank's box. Pandemic variance scaling downweights 2008Q4 to 2009Q4 and 2020 in the IS curve, and 2020Q2 to 2022Q4 in the Phillips curves. Parameters estimated on the edge of their support are marked: there the prior and the bound, not the data, set the value.
Expanding-window walk-forward, 2010Q2 to 2026Q3, re-estimated at each origin. Forecast of core PCE inflation over four quarters. Relative RMSE below 1 beats the benchmark.
The gap is built from real GDP and private domestic demand with the unemployment gap, without the payroll cycle. Does the choice matter for the Phillips curve? Slope of core services and core PCE inflation four quarters ahead on each measure.
What the scenario engine is. The equations of the estimation, run forward from the last quarter. Core PCE: q/q = a1L · own lag + a1I · four-quarter mean of headline + (1 − a1L − a1I) · expectation/4 + a2 · imports + a3 · oil + a4 · gap. Gap: h = b1 · h(−1) − b2 · (real rate gap)/4 − b3 · fiscal drag. Rule: i = th1 · i(−1) + th2 · i(−2) + (1 − th1 − th2) · [r* + 2 + th3 · (expectation − 2) + th4 · h]. Expectation: phi1 · own lag + (phi2 + phi3) · past four-quarter headline + remainder · 2. The model-consistent term of the estimation is replaced by past headline inflation, and headline is core plus a gap that halves each quarter. Both are simplifications of mine.
Weak slope. The output gap coefficient in the aggregate Phillips curve (0.04) sits on the lower edge of its support, and in the slack test the slope is not statistically different from zero for services. Inflation in the scenarios therefore moves mostly through inertia and expectations. Import prices barely matter for US core (a2 on its lower edge), against 0.96 points per 10% in Brazil.
Record. At twelve months the aggregate model’s RMSE is 8% below the random walk and 9% below the AR(4) over the full sample, and 5% below the AR(4) outside 2021 to 2022. The gain is larger at three and six months. Table 6 has every sample and horizon.
What updates, and what does not. Every Monday a job refreshes the policy rate, the futures-priced path (quarterly averages, on the upper bound of the target range, held after the last listed meeting) and the neutral real rate (median of the r-star page, PCE basis, which replaces the panel’s suite median of 1.24% from the first refresh). The estimates, impulse responses, record, expected inflation, output gap and inflation history are as estimated on data through the last quarter of the panel: they need the quarterly panel, which carries terminal and survey series, and are re-estimated by hand. The page flags when the BEA has published a later quarter than the panel. The estimation uses the shadow rate in the IS curve and the rule.
Levels sourced from public providers and survey data through 2026Q3; expectations from a state-space aggregation of market and survey measures. Views, not trade recommendations. Not investment advice.