Shocks

Cochrane opens from a 1993 AEA session in which prominent macroeconomists tried to explain “what caused the 1990 recession?” and “came up empty-handed,” reviewing a long list of candidates and largely disagreeing (Section 1). He organizes his own review around four themes: (1) identifying and evaluating a shock’s effect is never atheoretical – “one’s view of the propagation mechanism…is crucially important” and results “can change drastically” across theoretical perspectives; (2) the choice of statistic (variance decompositions, Hodrick-Prescott-filtered variance, Beveridge-Nelson-filtered variance, etc.) is “crucially important” and can give “drastically different results”; (3) economic agents have far more information than econometricians, so “what is a shock to us may be known by them”; and (4) certain “level” variables – the consumption/output ratio, M2 velocity, term spreads, hours – forecast long-horizon output with R-squared values of 60% or more (Section 1).

Five: (1) genuinely exogenous shocks are rare, since policy responds to the economy and even “technology” and prices are increasingly treated as endogenous in the growth and RBC literatures, leaving “only the weather” truly exogenous; (2) shock and propagation-mechanism analysis cannot be separated, since “shocks are only visible if we specify something about how they propagate to observable variables”; (3) identifying which shock “accounts for” fluctuations often has limited policy content, since a shock (e.g., oil prices) can operate entirely through another actor’s response (e.g., the Fed), making the causal story about the second actor, not the first; (4) shock-identification procedures are vulnerable to agents’ superior information (his analogy: “the weather forecast Granger-causes the weather, but shooting the weatherman won’t produce a sunny weekend”); and (5) most of the paper implicitly treats booms and busts as draws from the same linear process, not a qualitatively distinct nonlinear phenomenon (Section 2).

In simple VARs (e.g., an M2-output-price VAR), monetary shocks appear to explain dramatic fractions of output variance – up to 82% at a three-year horizon for M2 – but “the output response is surprisingly drawn out,” peaking two to three years after the shock and appearing nearly permanent, which does not match qualitative monetary theory (Section 3.1). As more “level” variables (consumption, hours) are added, orderings are refined, and cointegrating relationships (M2 velocity, the consumption/output ratio) are imposed so the impulse responses start to look like genuine, transitory monetary dynamics, the estimated output-variance share falls steadily – to roughly 15-25% at two-to-three-year horizons in Cochrane’s preferred error-correction specification, and below 10% once hours are included or the specification is otherwise tightened (Sections 3.1.3-3.1.6). Across M1, federal funds, and nonborrowed-reserve-ratio shocks, Cochrane finds the same qualitative pattern, and notes that all specifications explain “very little output variance at horizons less than a year, where…short-run non-neutrality is most likely to show up” (Section 3.6). His overall verdict: “monetary policy shock[s] account for at most 20% of the variation in output,” and likely much less once sampling and specification uncertainty are taken into account (Section 1, Section 3.6).

In horse-race regressions forecasting three-year output growth, “the consumption/output ratio has the highest t-statistic and…R-squared, 0.63,” and when federal funds or M2/output are entered jointly with the consumption/output ratio, “consumption drives out federal funds [and M2] as a forecaster of output” – their coefficients become insignificant and much smaller once consumption is included (Section 3.5, Table 3.7). Cochrane notes the consumption-based forecasts are also more contemporaneous with actual output growth, whereas the federal-funds-rate forecast leads it – evidence, in his reading, that the “level variable” property common to both consumption and monetary aggregates, not something specifically monetary, is doing much of the forecasting work attributed to money in the simpler VARs (Section 3.1.2, 3.5).

He distinguishes “how much output variance is due to monetary policy shocks?” from “how much output variance is due to monetary policy?”, noting variance decompositions “cannot be negative” and so are poorly suited to capturing systematic policy’s stabilizing role – for example, if the Fed has learned to systematically offset real shocks, “a negative fraction of output variance is due to monetary policy,” a possibility the standard VAR toolkit cannot even express (Section 3.7). He argues the persistent output responses found throughout his monetary VARs are “hard to swallow” as purely delayed responses to unanticipated shocks, and are more plausibly consistent with a world where anticipated, systematic monetary policy has short-lived real effects – in which case “the study of systematic monetary policy…may be more important to macroeconomics than an assessment of how much output can be further stabilized by making monetary policy more predictable” (Section 3.7).

Not robust at all: Eichenbaum’s bootstrap of the calibration procedure finds the estimate of technology’s explained variance share is “0.78 with a standard error of 0.64,” and Cochrane shows the figure is also highly sensitive to which moments are matched in calibration, to whether output-technology correlation is imposed structurally (real business cycle models are “stochastically singular”), and to measurement error (Section 4). Reworking the calculation using long-horizon output-growth forecastability instead of Hodrick-Prescott-filtered variance, Cochrane finds that because the standard King-Plosser-Rebelo model’s output is nearly unforecastable while actual U.S. output has substantial forecastable (“business cycle”) variation, “technology shocks explain 0.002% or less of business cycle variation in output”; using Beveridge-Nelson detrending instead of Hodrick-Prescott gives a similarly tiny figure, “0.009% or less” (Sections 4.1.3-4.1.4).

Following Plosser and, more explicitly, Hansen and Prescott, “anything that causes output to vary given capital and labor will result in a Solow residual, and hence will be identified as a ’technology shock’” – including labor or capital hoarding, taxes, regulatory frictions, or “changes in the inefficiencies induced by policy” (Section 4.4, quoting Hansen and Prescott 1993). Cochrane’s conclusion is double-edged: this broadening is “good news” for RBC methodology, since it lets the framework eventually incorporate explicit tax, monetary, and credit distortions within dynamic general-equilibrium models, but it is “obviously bad news for the view that technology shocks, narrowly defined, are the source of fluctuations,” since under this interpretation “it is vacuous to say that technology shocks cause fluctuations” (Section 4.4, 4.5).

Simple VARs using the crude-petroleum producer price index show oil-price innovations do produce sustained output declines, “however, the magnitude of the declines is much smaller than the declines produced by output or consumption shocks,” and “less than 10% of the variance of output is explained by oil price shocks” (Section 5.1); related general-equilibrium estimates cited by Cochrane (Kim and Loungani; Finn) find comparable ranges of roughly 7-19%. He notes the standard “small input” objection – imported oil is too small a share of GDP for classical production theory to generate large output effects from oil-price changes – applies with equal force to money, motivating ongoing work on explicit-friction models for both. On credit, Cochrane finds “credit shocks do not seem to explain a large part of postwar US output fluctuations”: most empirical credit research targets an amplification channel for other shocks rather than an independent credit shock, the estimated financing-constraint effects are concentrated in small firms too minor to explain aggregate swings, and Ramey’s survey shows “monetary aggregates drive credit indicators out of VARs similar to those discussed above” (Section 5.2).

Simulating the King-Plosser-Rebelo model’s response to news that a 1% permanent technology improvement will occur in one year, Cochrane shows “consumption rises instantly,” but because output has not yet changed and wages have not risen, workers respond to their now-higher wealth by working less; with labor down and technology/capital unchanged, “current output…also goes down,” and investment “declines so much I couldn’t fit it on the graph” – so “news of a future improvement in technology sets off a recession…in the standard real business cycle model” (Section 6.1). He reports this qualitative result is robust across variations including investment adjustment costs, variable labor effort, and variable capital utilization – meaning the “obvious” story of good news raising both consumption and investment together does not survive standard general-equilibrium logic, since in these models an increase in consumption today must be met by a corresponding decline in investment or output absent an actual current change in productive capacity.

Adding an explicit news-shock process – a small but very persistent signal of future technology improvement, alongside a standard random-walk technology shock – to the King-Plosser-Rebelo model, Cochrane shows the resulting simulated consumption-output VAR reproduces the empirical pattern of a large transitory component in output (75-95% of output forecast-error variance across one-to-three-year horizons, comparable to the roughly 63-89% found in the actual U.S. consumption-output VAR) alongside a mostly-permanent component in consumption (Section 6.2, Tables 6.1-6.2, compared with Table 4.1). He also shows that feeding a VAR-estimated (rather than hand-specified) technology-forecast process through the same RBC model, following a method suggested by King and Watson, produces a similarly data-consistent pattern (Section 6.3) – though he cautions that news shocks cannot resolve every discrepancy, since the model’s first-order conditions still impose an exact relationship among output, labor, and consumption that a three-variable VAR of the actual economy does not share (stochastic singularity, Section 6.4).

“I find that none of the popular candidates for observable shocks robustly accounts for the bulk of business cycle fluctuations in output” (Section 7). He lays out three live possibilities going forward: newer candidates (oil reallocation, credit, nonlinear dynamics) or new propagation mechanisms (noncompetitive models, lending channels) might eventually be fleshed out to rehabilitate traditional shocks; real-business-cycle theorists might refine their models to generate more forecastable dynamics and amplification; or consumption and output may in fact move on news economists cannot observe. He is explicit that this last possibility “at least explains our persistent ignorance, but it means that we may forever be ignorant of the true shocks that drive fluctuations” – while stressing this is not a vacuous or unfalsifiable claim, since “models that explain business cycle dynamics with news shocks must be constructed and matched to data just like other models,” and standard RBC models, as Section 6.1 shows, do not do so automatically (Section 7).