Monetary policy matters: Evidence from new shocks data

All four leading approaches to identifying monetary policy shocks – the Christiano-Eichenbaum-Evans (CEE) recursive VAR, the Bernanke-Mihov over-identified VAR, the Sims-Zha non-recursive VAR, and the Romer-Romer narrative measure – produce economically implausible results once applied to post-1988 US data: a contractionary shock is followed by a rise, not a fall, in output. The authors attribute this to the Fed becoming “more forward-looking” after 1988, which “invalidat[es] the identifying assumptions in conventional methods of measuring monetary policy’s effects, leading to spurious and unlikely results for this period” (Abstract, p. 950). Two structural breaks in the estimated reaction function, dated to 1979:10 and 1982:10, and the growing importance of forward-looking elements in the Fed’s information set are cited as the mechanism (Section 2.2).

Each of the four methods is run on both an earlier comparison sample and a post-1988 sample using its own original specification, and each flips sign in the later sample. The CEE recursive VAR shows output declining over 1960Q1-1992Q4 but rising significantly over 1988Q4-2007Q3 (Fig. 1); the Bernanke-Mihov over-identified VAR shows industrial production (IP) declining over 1965:01-1996:12 but rising over 1988:12-2007:11 (Fig. 2); the Sims-Zha non-recursive VAR shows GNP declining over 1964Q1-1994Q4 but rising over 1988Q4-2007Q4 (Fig. 3); and the Romer-Romer narrative measure shows IP declining over 1969:01-1996:12 but rising significantly over 1988:12-2008:06 (Fig. 4). The authors summarize: “the existing identification schemes lead to estimated impulse responses in the post-1988 sample that are both different from those found in the earlier samples, and counter to what most central bankers would find plausible” (p. 955).

The mechanism is an omitted-variable problem: when the Fed’s reaction function depends on anticipated future output and inflation (elements of its information set, Omega_t) that a VAR does not include, the VAR’s residual conflates the genuine policy shock with the Fed’s endogenous response to those anticipated conditions, biasing the estimated output effect upward. As the authors put it, “the key change is to the estimated effect rather than the actual effect, because identification problems become more pronounced when the Fed’s policy becomes more forward looking” (p. 962). This same logic implies a fourth identifying assumption for any shock measure: the measured innovation should not itself reveal information from the Fed’s private forecasts or reaction function beyond what is already priced in by the market.

Following Kuttner (2001) and Gurkaynak-Sack-Swanson (2005), the shock is the revision in the private sector’s expectation of the policy path implied by Fed Funds futures prices around each of 157 FOMC meetings from 1988:12 to 2008:06 (the futures market itself only began in October 1988), formalized as the pre-announcement expectation of the policy path minus its realized value (Eq. 6, p. 958), under the assumptions that private-sector beliefs prior to the announcement are correct on average and that a background “noise” term is unchanged around the announcement (Eqs. 4-5, p. 958). Rather than using a single futures contract, the authors combine daily price innovations from six Fed Funds futures contracts spanning the current month through five months ahead.

A maximum-likelihood two-factor model fit to the six futures-contract innovations yields a first factor explaining 92% of total variance, on which all six contracts load positively, interpreted as a “level” shift in the expected medium-term policy-rate trajectory; a second factor explains the remaining 9%, with loadings switching from positive to negative as contract maturity increases, interpreted as a “slope” or yield-curve effect (Eq. 11 and footnote 14, p. 959). Only the first (level) factor is used as the policy shock measure; the authors note the second factor plausibly captures forward-guidance content in FOMC communications but do not incorporate it into the shock series.

In a three-variable monthly VAR (log industrial production, log CPI, and the cumulated new shock, with 36 lags, a constant and linear trend, and the shock ordered last) over 1988:12-2008:06, industrial production falls significantly and persistently after a contractionary shock, “with a maximum impact at a horizon of around two years” – similar in shape to the pre-1988 CEE and Romer-Romer results, but the opposite of what those same methods find in the post-1988 period (Section 4.1, p. 962). Prices are more problematic: “the effect becomes significantly negative only after four years; the positive response over the medium term, although small, contrasts with the negative effect that has generally been found in the literature” (p. 962). This price puzzle is described as milder than in the conventional post-1988 results but is not resolved by adding a commodity price index or one-quarter-ahead Blue Chip inflation expectations (Section 4.2).

Forecast error variance decomposition shows the new shock accounts for roughly 40-50% of industrial production’s forecast error variance at a horizon of three years or more, “around twice the proportion using existing shock measures” over the same period, with the Romer-Romer and change-in-the-funds-rate measures each accounting for roughly 10-20% at the same horizon (Fig. 7, p. 964; Abstract, p. 950). The authors caveat that this comparison is made “in an era of low overall output volatility” (the Great Moderation), so part of the new measure’s relatively large share may reflect the low volatility of the denominator rather than a larger absolute effect (Section 4.1).

Regressing the new shock on the gap between Greenbook and Blue Chip forecasts for 17 variables in the Romer-Romer reaction function (113 FOMC meetings, 1988-2002) yields R-squared = 0.185 and F(17) = 1.50 (p = 0.132), so the authors cannot reject the hypothesis that the Fed’s exclusive private information has no joint explanatory power for the new shock at conventional levels (Table 1, p. 961). Two individual coefficients are nonetheless significant – current-quarter output growth (-2.37, significant at 1%) and current-quarter GDP deflator (2.34, significant at 5%) – suggesting some contamination by the Fed’s reaction to contemporaneous conditions. The authors argue any resulting simultaneity bias works toward attenuation (biasing estimated effects toward zero), so the paper’s reported effects “likely understate the true effect” (Section 4.3, pp. 963-965).

Across eight robustness checks – alternative VAR ordering (shock first), an alternative price index (PPI), alternative lag lengths (6, 12, 24 lags), subsample stability, adding a commodity price index, adding inflation expectations, excluding FOMC dates that coincide with the Employment Report, and a single-equation Romer-Romer-style specification – the qualitative results are described as essentially unchanged, with the price puzzle persisting throughout (Section 4.2). As an illustrative episode, the authors point to the June 27, 2001 FOMC meeting, where the Fed cut rates by 25 basis points, less than the 50 basis points the market had priced in: conventional VAR and narrative methods would record this as an expansionary shock, but the new measure records it as contractionary (roughly two standard deviations above zero), consistent with the subsequent rise in bond yields and the dollar’s climb to a ten-week high (Section 3.5, pp. 961-962). The authors also show that using only the current-month (spot) futures innovation, as in Kuttner (2001), produces an imprecise, positive output response, underscoring the value of combining information across the six-contract term structure rather than relying on the spot-month shock alone.