Olivier Coibion | Macro Paper Warehouse

Are the Effects of Monetary Policy Shocks Big or Small?

This 2012 American Economic Journal: Macroeconomics paper by Olivier Coibion asks why standard recursive VARs and the Romer-Romer (2004) narrative approach produce such different estimates of the real effects of monetary policy shocks, and argues the discrepancy is driven by three identifiable and largely correctable factors rather than genuine disagreement about the underlying policy shock. Using monthly US data (1970:1-1996:12, following Romer and Romer’s sample and variable choices), Coibion compares a standard recursive (Cholesky) VAR with the funds rate ordered last and 12 lags, the Romer-Romer single-equation approach (a macro variable regressed on 24 autoregressive lags and 36 lags of the R&R narrative shock series), and a “hybrid VAR” that replaces the funds rate in the VAR with the cumulative R&R shock series. At baseline lag specifications the three diverge sharply (Table 2, Panel A): a 100-basis-point shock produces a peak industrial production decline of about 0.7% and a 0.16 percentage-point rise in unemployment in the standard VAR, versus a 4.3% IP decline and 0.93 percentage-point unemployment rise in the R&R single-equation approach, with the hybrid VAR intermediate (1.6% IP decline, 0.40 points). Three mechanisms account for most of this gap: (1) an R&R shock corresponds to a larger, more persistent counterfactual path for the funds rate than a VAR Cholesky shock, and rescaling impulse responses to a common FFR path largely eliminates the difference; (2) R&R shocks during the 1979:10-1982:9 non-borrowed-reserves (NBR) targeting episode are statistically predictable from lagged macro variables (violating the shocks’ assumed exogeneity) and disproportionately drive the large R&R estimates; and (3) the R&R two-step procedure’s results rise nearly monotonically with the number of shock lags included, while AIC- or model-averaging-based lag selection picks shorter lags and shrinks the estimated effects (reducing the peak IP effect from -4.3% to -3.4%), whereas the VAR is comparatively insensitive to lag length. Once the NBR-targeting period is excluded and lag length is chosen by AIC or model averaging, the three approaches converge (Table 2, Panel B) on a “medium-range” peak effect of roughly a 1.8-2.8% industrial production decline and a 0.19-0.70 percentage-point unemployment increase; three additional Taylor-rule-residual shock measures constructed in Section III (GARCH, time-varying-coefficient, and Smets-Wouters DSGE-filtered shocks) corroborate this medium range and are markedly less sensitive to lag length or episode selection than the original R&R series, though the paper is explicit that its evidence covers only the pre-1997 sample and does not speak to the post-1996 or zero-lower-bound period.

Tell Me Something I Don't Already Know: Learning in Low- and High-Inflation Settings

This paper uses randomized control trials (RCTs) applied over time in multiple countries to study whether the economic environment — specifically the level of inflation — affects how agents learn from new information. The main finding is that as inflation rose in advanced economies, both households and firms became more attentive to and informed about publicly available news about inflation, causing them to respond less to exogenously provided information about inflation and monetary policy in the RCT treatments. When agents are already well-informed about the current inflation environment (because high inflation makes it salient), additional information provision moves their beliefs less — the marginal value of information is decreasing in prior attentiveness. Complementary evidence from Uruguay (persistently high inflation) and New Zealand (persistently low inflation) confirms the cross-sectional pattern: agents in high-inflation environments have stronger priors and respond less to information treatments. The results imply that central bank communication interventions are more effective during low-inflation periods when agents are less informed.

The Optimal Inflation Rate in New Keynesian Models: Should Central Banks Raise Their Inflation Targets in Light of the Zero Lower Bound?

This 2012 Review of Economic Studies paper by Olivier Coibion, Yuriy Gorodnichenko, and Johannes Wieland asks what rate of steady-state (trend) inflation maximizes welfare in a New Keynesian DSGE model once the zero lower bound (ZLB) on nominal rates is explicitly modeled, rather than assumed away. The authors build a medium-scale NK model with Calvo staggered price-setting, habit formation in consumption, and a Taylor rule truncated at the ZLB, solving for ZLB episodes’ endogenous duration using the Bodenstein-Erceg-Guerrieri (2009) nonlinear algorithm; they calibrate the model to standard U.S. moments and to the historical post-WWII frequency of ZLB episodes, and evaluate welfare via a second-order approximation to utility that decomposes into a steady-state term (from Calvo price dispersion) and variance terms in the output gap, inflation, and consumption. In the baseline calibration the optimal trend inflation rate is 1.5% per year – “close to the bottom end” of the 1-3% target ranges central banks commonly use – because, although each ZLB episode is individually costly (an 8-quarter ZLB spell costs the equivalent of a 6.2% permanent consumption loss at 2% trend inflation), such episodes are calibrated to occur only about once every 20 years at 2% inflation, so the unconditional expected cost of the ZLB is small (0.08% of permanent consumption) relative to the perpetual costs of higher trend inflation (steady-state price dispersion, and a previously unidentified channel by which higher trend inflation makes inflation volatility itself more costly). The optimal rate proves robust to a wide range of alternative calibrations and extensions – remaining under about 3% even when the output-gap loss weight is scaled up 100-fold, capital is added (2.1%), parameter uncertainty is incorporated (1.9%, 90% CI [0.3%, 2.9%]), or the historical ZLB frequency is tripled – with the risk-premium shock’s persistence being the single most sensitive parameter (raising optimal inflation from 1.5% to 3% when its autocorrelation rises from 0.947 to 0.96). The rate is highly sensitive to the assumed monetary and fiscal policy regime, however: optimal inflation falls to about 0.2% under commitment to a stabilization policy, rises to 2.7% under discretion, and falls to well under 0.3% under even a modest price-level-targeting response, and it falls further still, to 0.3%, if downward nominal wage rigidity is added to the model. The authors caveat that their cashless-economy setup ignores the Friedman optimal-deflation motive and seigniorage, and that omitting endogenous countercyclical fiscal policy during ZLB episodes likely overstates both the cost of the ZLB and the resulting optimal inflation rate.

Why Are Target Interest Rate Changes so Persistent?

This 2012 American Economic Journal: Macroeconomics paper by Olivier Coibion and Yuriy Gorodnichenko adjudicates between two observationally similar explanations for the federal funds rate target’s very high persistence (an estimated AR(1) coefficient of roughly 0.9-0.99 in typical Taylor-rule regressions): deliberate interest-rate smoothing (policy inertia, modeled as partial adjustment toward a Taylor-rule target) versus serially correlated monetary policy shocks (unexplained deviations from the Taylor rule that themselves carry over from meeting to meeting). Using real-time Federal Reserve Greenbook forecasts of inflation and the output gap – treated as the best available measure of the Fed’s actual information set at each decision, released only with a five-year publication lag – over a 1987:IV-2006:IV quarterly sample bounded by Greenbook availability and ending before the zero lower bound, the paper estimates a sequence of nested Taylor-rule specifications. A simple specification nesting a smoothing parameter and a persistent-shock parameter in a single AR(1) finds both statistically significant (smoothing coefficient about 0.81, shock persistence about 0.46), and the paper concludes that this simple nested specification cannot overwhelmingly differentiate between the two explanations. But once the paper generalizes to higher-order AR(p) smoothing and AR(q) shock lags and lets the Bayesian information criterion (BIC) select among the resulting grid of specifications, the selected model is robustly IS(2), AR(0) – two lags of interest-rate smoothing and no persistent-shock lags – across the full sample and pre-/post-2000 subsamples; under a richer AR(2)-smoothing-plus-AR(2)-shock specification the smoothing coefficients remain large and significant (summing to roughly 0.75-0.95) while the shock-persistence terms turn insignificant at the first lag and negative and significant at the second. The paper corroborates this with an instrumental-variables strategy that instruments the lagged rate with non-monetary shocks (oil prices, technology, fiscal policy) to address possible endogeneity, yielding smoothing estimates of 0.70-0.87; with a Rudebusch (2002)-style predictability test showing Greenbook forecasts (rather than Eurodollar futures) predict future target changes at 2- and 3-quarter horizons with R-squared of 20% and 12%, consistent with genuine smoothing; with checks showing financial conditions, forecast revisions, and time-varying-inflation-target measures cannot fully account for the estimated smoothing; and with narrative evidence – Greenspan’s 1994 stated preference for 25-basis-point increments and Bernanke’s 2004 “Gradualism” speech – that Fed officials themselves describe deliberate gradual adjustment. A further implication is that the interest-smoothing interpretation implies a long-run inflation response exceeding one (satisfying the Taylor principle), whereas the no-smoothing persistent-shock interpretation implies a contemporaneous inflation response that can fall below one, a determinacy-violating result the smoothing interpretation avoids.