Comments on "A New Measure of Monetary Shocks: Derivation and Implications"
📄 Summarized from the full manuscript · Human-reviewed for faithfulness before publication
In brief
Commenting on the Romers' new measure of monetary shocks, Cochrane argues it is better than its authors realize. Its key finding, that Fed staff forecasts summarize everything the Fed knew, lets shock identification and the output response collapse into simple single regressions. Redoing it his way, he gets essentially their result: output down 6 to 8 percent about 30 months after a one-point tightening, with almost no price response. He calls that long, price-inert response theoretically troubling. Asked instead about a move not followed by the Fed's usual further tightening, the output effect shrinks to roughly a tenth and fades faster, so he urges measuring the policy rule itself.
What this paper finds — and why it matters
These are John Cochrane’s discussant remarks, delivered July 17, 2004 at an NBER Economic Fluctuations and Growth meeting, on Christina Romer and David Romer’s “A New Measure of Monetary Shocks: Derivation and Implications” (published in the American Economic Review later that year). Cochrane opens by declaring he wants to argue the paper is “much better and deeper than the authors think it is,” identifying its two key innovations as extending the Fed’s intended (target) funds rate series back in time using historical “intentions” data, and, more importantly, establishing via narrative evidence that Federal Reserve staff Greenbook forecasts function as a “sufficient statistic” for the Fed’s information set at the time of each policy decision. Cochrane shows this sufficient-statistic property lets both halves of Romer-Romer’s procedure be radically simplified: identifying a valid policy shock requires only that it be orthogonal to the Fed’s own output forecast (not to price, exchange-rate, or other forecasts, since only reverse causality running through output needs to be purged), and once such a shock is in hand, the output or price response can be read directly off a sequence of single regressions rather than a full dynamic simulation with Monte Carlo standard errors. Re-estimating with his simplified shocks, Cochrane obtains essentially the same headline result as Romer-Romer – a 100-basis-point tightening produces a 6-8% decline in output peaking around 30 months out, with statistical significance only modestly above conventional thresholds (t-statistics “just above 2”) – alongside a “price puzzle” that he argues is not legitimately fixable by adding commodity prices to the shock regression, since the historical FOMC record shows no narrative evidence that the Fed actually responded to commodity-price movements in the way that fix presumes. He flags the response’s long delay and the near-total absence of any price response as jointly troubling for monetary theory, since most theories tie the two together. In a preliminary extension, Cochrane recomputes the output effect of a one-time, one-month target increase not followed by the Fed’s usual further tightening, and finds it roughly a tenth the size and considerably shorter-lived than the standard “full path” response – suggesting that whether anticipated policy actions matter is, in his view, at least as consequential a modeling choice as any other in this literature. He closes by suggesting the field’s next step should be estimating the Fed’s systematic policy rule directly (via narrative methods, since Fed statements arguably already summarize the rule) rather than continuing to refine shock-and-response estimates, and offers a conjecture that ordinary regression evidence cannot in principle distinguish a determinate (Taylor-rule-consistent) policy regime from an indeterminate one, because such regressions can only recover the stable root of the model’s dynamics, not the off-equilibrium explosive threats that actually enforce determinacy.
Summary of a classic paper, AI-assisted and human-reviewed. See the linked original for the authoritative claims and full conditions.
Questions & answers
Q1. What does Cochrane identify as the paper’s two most important innovations?
“Two big innovations”: first, extending the Fed’s “intentions” (target funds rate) data back in time, which Cochrane calls “surely the right variable to look at,” giving more data with less measurement error; second, and more important, establishing that “Greenbook forecasts are sufficient statistics” for the Fed’s information set (Section 1). He stresses that “historical evidence is crucial to both innovations,” and specifically credits narrative evidence – close reading of the actual FOMC/Greenbook record – with establishing the sufficient-statistic claim: “I thought the Greenbook was just a staff-amusement exercise, but the evidence is strong for it.” He calls the resulting shock “maximally informative” and says this property “gives us simpler, clearer, and more robust procedures.”
Q2. What is “Proposition 1,” and why does Cochrane say a valid shock need not be orthogonal to price or exchange-rate forecasts?
“To measure the effects of monetary policy on output it is enough that the shock is orthogonal to output forecasts. The shock does not have to be orthogonal to price, exchange rate, or other forecasts… All the shock has to do is remove the reverse causality from output forecasts” (Section 2). He illustrates with examples: a Fed move to offset the exchange rate, or a move responding to unemployment (conditioned on the output forecast), both still count as valid “shocks” for measuring output effects, because “the Fed never rolls dice; every move is a response to something” – the identification problem is specifically about removing feedback from output expectations, not about finding truly random Fed behavior.
Q3. What is “Proposition 2,” and how does it simplify Romer-Romer’s estimation of the output response?
“One can measure the output (price) response by running single regressions of output (price) on the shocks, with no other variables” rather than Romer-Romer’s dynamic simulation of a full autoregressive-distributed-lag model with Monte Carlo standard errors (Section 3). Cochrane’s reasoning: because the identified shock is by construction uncorrelated with everything else that determines output, regressing output growth at horizon k directly on the shock recovers the moving-average coefficient at that horizon exactly as a multiple regression would “in population,” with two practical advantages in finite samples – single regressions “are more stable” (avoiding a large matrix inversion) and, “most importantly,” the estimate is “tied directly to the fact you want to measure: what is the average change in output following a one percent change in intentions, orthogonal to output forecasts?”
Q4. When Cochrane re-estimates the model his way, how do his results compare to Romer and Romer’s, and what does the accompanying scatterplot analysis show?
“For output we have just about the same results. A 100 bp tightening gives a 6-8% decline at about 30 months,” significant “with a t a bit over 2” (Section 4). Cochrane’s price response shows a larger “price puzzle” than Romer-Romer’s, which he attributes to not cleaning unemployment and output forecasts out of his shock regression’s right-hand side. He supplements the standard impulse-response plots with scatterplots of 30-month output and price growth against the shock, showing the result is driven by a modest number of observations (with several large outliers from the early-1980s Volcker disinflation, whose removal, he checks, strengthens rather than reverses the results) – and dated versions of the same scatterplot to show which historical episodes are doing the work. He calls this “a nice way to show just what evidence we have,” making visible “a summary of history” that is otherwise hidden inside a fancy impulse-response function and Monte Carlo confidence bands, and cautions “the scatterplots may seem depressing, but that’s what t just above 2 in 30 years of data looks like!”
Q5. Why does Cochrane consider the paper’s finding of a large output response with almost no price response (“the dog that did not bark”) to be “a huge puzzle”?
He gives three reasons: (1) the “price puzzle” is a robust feature of his own re-estimated shocks and of the 30-month delay in Romer-Romer’s own results, and he argues the standard fix – adding commodity prices to the shock-identification regression – is illegitimate on narrative grounds, since reading FOMC minutes shows no evidence the Fed actually tightened in direct response to rising commodity prices as that fix presumes: “either RR give a dreadfully wrong reading of FOMC minutes, or adding commodity prices to fix the price puzzle is just a mistaken fishing expedition”; (2) “all theories tie output and price responses together,” so finding one without the other – “it’s like ordering eggs sunny side down and don’t turn it over” – requires an unusual theory in which interest-rate policy affects output with literally no effect on the price level; and (3) since current policy wisdom holds the Fed should control the price level via interest rates, but the data show “no evidence that fed interest rate policy has any effect on the price level,” Cochrane concludes this belief “must [be] regard[ed]…as matters of faith, a-priori beliefs, rather than matters supported by the historical experience” (Section 4).
Q6. What is Cochrane’s “one-month blip” recalculation, and what does it imply about anticipated versus unanticipated policy effects?
Noting that after a target increase, “the target keeps rising for quite a while,” Cochrane asks whether output responds to the initial shock itself or to the later, anticipated further increases that typically follow it – and computes the response to a hypothetical one-month, 1-percentage-point increase in the target NOT followed by the Fed’s customary further tightening, by dividing the estimated target-response process out of the estimated output-response process (simulating b(L)y_t = a(L)ff_t) (Section 5). The result: “the effect of a one-month 1% increase in ff target, not followed by customary further target increases, is much shorter than the usual response. It ends at t=30, rather than peaking at t=30. More importantly, it’s 1/10 smaller, peaking at about -0.4-0.5% not -6-8%.” He is explicit this is a “preliminary calculation” and flags that “the anticipated/unanticipated question has more impact on the results than most other specification questions” – directly echoing the same identification concern Cochrane raises in his own “What do the VARs mean?” (1998).
Q7. What broader methodological pivot does Cochrane propose for future work in this area?
“Perhaps we are asking the wrong question. The rule, not shocks and responses, may be more interesting” (Section 6). He offers two reasons: a minor one – that the shock-defining regression is estimated over the whole sample despite the likelihood that the Fed’s actual policy rule changed over time, a problem narrative methods applied directly to the rule could sidestep – and a major one, that Taylor-rule-style questions about whether the Fed’s response coefficient on inflation exceeds one (a determinacy condition) are what most current monetary theory actually cares about. He closes with “Conjecture 3”: that ordinary regressions cannot answer the determinacy question, illustrating with a simple example in which solving a model’s unstable root forward means an econometrician estimating the reduced form will recover only the model’s stable root (0.9 in his numerical example) regardless of whether the true policy coefficient implies determinacy (phi>1) or not, because “these models work by specifying off-equilibrium threats (explosive inflation) that we never see in equilibrium. We can never measure off-equilibrium threats” – a problem he suggests narrative evidence might be better positioned to address than regression evidence.
Key terms in this paper
Definitions below follow the paper's own usage.
- Greenbook forecast as sufficient statistic
- Cochrane's formalization of what he sees as Romer and Romer's central empirical discovery: that the Federal Reserve staff's Greenbook forecast of future output (and other variables), E(y_(t+k) | Fed info at t), captures essentially all of the information relevant to the Fed's own policy decisions, so that a policy shock need only be constructed to be orthogonal to this one forecast rather than to a whole VAR's worth of variables; "narrative evidence is important for this conclusion," and "this property results in a 'maximally informative' shock" (Section 1).
- Proposition 1: orthogonality only to output forecasts is required
- Cochrane's restatement of what a valid monetary policy shock, for the purpose of measuring output effects, actually requires: "it is enough that the shock is orthogonal to output forecasts. The shock does not have to be orthogonal to price, exchange rate, or other forecasts... All the shock has to do is remove the reverse causality from output forecasts." A Fed move that responds to inflation, unemployment, or the exchange rate still counts as a valid shock for measuring the output response, so long as it is unrelated to the Fed's own output forecast (Section 2).
- Proposition 2: single regressions recover the impulse response
- Cochrane's simplification of Romer and Romer's dynamic-simulation/Monte-Carlo procedure for estimating the output (or price) response to a shock: because the identified shock is by construction uncorrelated with everything else on the right-hand side, "one can measure the output (price) response by running single regressions of output (price) on the shocks, with no other variables," regressing output growth at each horizon k directly on the current shock to read off the impulse response coefficient by coefficient (Section 3).
- One-month policy "blip" response (vs. the full anticipated path)
- Cochrane's preliminary re-estimate of the effect of a one-month, 1-percentage-point increase in the federal funds target that is NOT followed by the Fed's customary further tightening, obtained by dividing the estimated target-rate response process out of the estimated output response process (simulating b(L)y_t = a(L)ff_t); the result is an output effect that ends by month 30 rather than peaking there, and that peaks at roughly -0.4 to -0.5% rather than the -6% to -8% found for the full, anticipated-policy-inclusive response -- "about 1/10 smaller" (Section 5).