Deconstructing Monetary Policy Surprises—The Role of Information Shocks

Questions & answers

Q1. What is the paper’s central puzzle, and what motivating example illustrates it?

The paper starts from the observation that FOMC announcements “simultaneously convey information about monetary policy and the central bank’s assessment of the economic outlook,” so a naive interest-rate surprise around an announcement can reflect either a genuine policy action or the market learning something new about the Fed’s economic outlook — and conflating the two “biases the inference on monetary policy nonneutrality.” As a motivating example, on March 20, 2001 the FOMC cut rates by a larger-than-expected 50 basis points, yet the S&P 500 declined within 30 minutes, because the accompanying statement emphasized “substantial risks that demand and production could remain soft” — a pattern of positive interest-rate/stock-price co-movement (rather than the standard negative co-movement expected from a rate cut) that the paper finds characterizes roughly one-third of FOMC announcements since 1990.

Q2. How does the paper identify a “monetary policy shock” separately from a “central bank information shock”?

The identification exploits sign restrictions on the contemporaneous high-frequency co-movement of two surprises around each announcement — the change in the three-month federal funds futures rate and the change in the S&P 500 — within a Bayesian structural VAR: a monetary policy shock is defined as the combination in which the interest-rate surprise is positive and the stock-price surprise is negative (the standard asset-pricing prediction for a surprise tightening), while a central bank information shock is defined as the orthogonal combination in which both the interest-rate and stock-price surprises are positive. These sign restrictions are imposed only on the two high-frequency surprises themselves (not on the monthly macro variables), with the VAR set-identified via uniform-prior rotations (Rubio-Ramírez, Waggoner, and Zha 2010) — a less restrictive approach than Uhlig’s (2005) penalty-function method — using monthly U.S. data from February 1984 to December 2016 (high-frequency surprises available from February 1990).

Q3. How do the estimated effects of the two shocks differ?

On impact, the monetary policy shock is associated with the S&P 500 falling about 42 basis points (90% credible interval: −52 to −23), and over time real GDP and the price level decline persistently by roughly 10 and 5 basis points respectively while the excess bond premium rises about 5 basis points; the central bank information shock, in contrast, is associated with the S&P 500 rising about 28 basis points, and over time real GDP and the price level both rise (by roughly 5 and 3 basis points respectively) while the excess bond premium falls about 3 basis points. The one-year Treasury yield rises under both shocks, but reverts within about a year following the monetary policy shock versus taking more than two years to revert following the information shock — consistent with the information shock representing news about a more persistent, fundamentals-driven change in the economic outlook rather than a temporary policy action.

Q4. Why does the paper conclude that “standard” high-frequency identification is biased, and in which direction?

Because a standard high-frequency-identified shock (ordering the interest-rate surprise first in a Cholesky scheme, as in Barakchian and Crowe 2013 and Gertler and Karadi 2015) implicitly attributes every high-frequency surprise to monetary policy, it mixes together the monetary policy shock and the information shock, which move output, prices, and credit spreads in opposite directions — the paper shows this mixing attenuates (shrinks) the estimated output, price, and excess-bond-premium responses to a policy shock while simultaneously producing larger and more persistent interest-rate responses than the purified shock implies. The authors state that “standard HFI underestimates the effectiveness of monetary policy,” and note that the purified monetary policy shock’s more pronounced price-level decline “might account for the presence of the price puzzle in some relevant subsamples” documented in earlier work using the contaminated measure.

Q5. What does the paper find about how these two shocks are distributed over time, and what does the pre-1994 period reveal?

Central bank information shocks are not clustered in any particular period but occur throughout the sample, with notable episodes including a sequence of negative information shocks during the 2000-2002 dot-com bust and around August 2007 — consistent with then-Chairman Bernanke’s own later account of market uncertainty about what the Fed knew. In the period before 1994 (when the FOMC did not yet issue regular post-meeting statements), the monetary policy and information shocks are found to be roughly proportional and positively correlated, which the authors interpret as consistent with theoretical models (Melosi 2017; Nakamura and Steinsson 2018) in which markets observing only the funds-rate change — without an accompanying statement — cannot separate the two types of shocks from each other.

Q6. How does controlling for the information shock affect a previously documented “puzzle” about expected output growth?

The paper shows the central bank information shock raises both survey-based expected GDP growth and expected inflation, as well as market-based (break-even) inflation compensation, confirming that it genuinely conveys information about future demand conditions; crucially, once this information channel is controlled for, the counterintuitive finding in Nakamura and Steinsson (2018) that expected GDP growth rises following a (conventionally measured) contractionary monetary policy shock disappears — that puzzle, the authors argue, was itself a symptom of the contamination between the two shock types.

Q7. What does the structural (New Keynesian) model estimation reveal about the importance of financial frictions?

Matching the estimated impulse responses to a New Keynesian model with nominal and financial frictions (following Gertler and Karadi 2011), the standard (contaminated) monetary policy shock requires extreme nominal price stickiness (an estimated Calvo parameter of 0.94, implying prices reset on average only about once every four years, plus near-complete backward price indexation) and negligible financial frictions to fit the data — essentially reproducing the extreme-stickiness finding of Nakamura and Steinsson (2018) — whereas the purified monetary policy shock is consistent with more moderate, more empirically plausible price stickiness (Calvo parameter of 0.87, no backward indexation) and financial frictions roughly twenty-four times larger in the estimated model, with the output response substantially muted if financial frictions are switched off entirely. The authors conclude that “controlling for the presence of central bank information shocks can…modify our views on the importance of financial frictions in the transmission of monetary policy,” and separately find that the information shock itself is best matched by a model shock to capital quality (asset valuations) realized about two quarters ahead.

Q8. How does the paper’s evidence from the European Central Bank compare to the U.S. findings, and what do the authors flag as open limitations?

Constructing an analogous high-frequency dataset of 280 ECB announcements (1999-2016), the paper finds more than 40% show the “wrong-signed” positive co-movement pattern (versus about one-third for the Fed) — consistent with the ECB’s relatively transparent communication practices (press conferences and prompt publication of staff forecasts) — and shows that a standard (unpurified) high-frequency identification for the euro area produces an outright puzzle (financial conditions appearing to improve after a tightening), which the sign-restriction approach resolves. The authors are explicit about several limitations: the method “could not determine to what extent” the information shock’s effects reflect the central bank’s forecasts simply materializing versus a genuine causal effect of communication; sign restrictions deliver only set (not point) identification; the demand/supply decomposition of inflation compensation is “not completely innocuous” since compensation reflects both expected inflation and an inflation risk premium; and after the effective lower bound period (post-2008), reduced variation in short-term futures surprises limits the method’s ability to capture the effects of longer-rate or unconventional-policy communication, which the authors leave for future research.