Has the Business Cycle Changed and Why?

The standard deviation of four-quarter U.S. real GDP growth fell from about 2.0 percentage points in the 1960s to 2.7 in the 1970s and 2.6 in the 1980s, then down to only 1.5 in the 1990s, and this moderation is “widespread (but not universal),” appearing in both real and nominal series (Sections 1, 2.3). Across 168 quarterly macroeconomic time series, the median relative standard deviation (1990s versus full sample) is 0.73, and 78% of series had a relative standard deviation below 0.85 in the 1990s; employment growth volatility fell about 27% and the relative standard deviation of GDP-deflator inflation changes was 0.50. The moderation is largest, in GDP-share-weighted terms, for residential investment, output of durable goods, and output of structures, and the first published identifications of the break are credited to Kim and Nelson (1999) and McConnell and Perez-Quiros (2000), who independently date it to 1984:I.

The authors’ battery of break tests on time-varying autoregressions for 168 series finds that “the changes in variance are associated with changes in conditional variances,” not with changes in the autoregressive (conditional-mean) coefficients: for GDP specifically, the QLR test fails to reject a stable conditional mean but rejects a stable conditional variance at the 1% level, with an estimated break date of 1983:2 and a 67% confidence interval of 1982:4-1985:3 (Section 3.1). Across all 168 series, breaks in conditional means cluster in the 1970s while breaks in conditional variances cluster in the 1980s (or early 1990s for some series), matching the timing of the observed decline in unconditional variance to the variance breaks, not the mean breaks.

For GDP, a test nesting both hypotheses finds the time-trend coefficient statistically insignificant while the break (at the same estimated date, 1983:2) is significant at the 1% level, leading the authors to conclude “the sharp decline…is best described by a discrete reduction in the variance, rather than as part of a continuing trend” (Section 3.2), consistent with Kim and Nelson (1999) and McConnell and Perez-Quiros (2000) against Blanchard and Simon’s (2001) trend-decline reading. Some GDP components (nondurables consumption, residential investment, imports, goods production, construction) fit the break model well with estimated dates between 1983:2 and 1987:3; a few (durables consumption, nondurables production) fit a trend model better, and a few (exports) fit neither well. Multivariate VAR and dynamic-factor-model estimates (Section 3.3) tighten the break-date confidence interval to a range clustering around 1982:4-1984:1, consistent with the univariate GDP estimate.

Using a counterfactual VAR method that recombines estimated lag coefficients and innovation-covariance matrices from the pre-1984 and post-1984 subsamples, the authors find that substituting the second period’s dynamics with the first period’s shocks reproduces almost exactly the first period’s output volatility (2.63%, versus an actual first-period value near 2.7%), while substituting the first period’s dynamics with the second period’s shocks reproduces almost exactly the second period’s low volatility (1.48%, versus an actual value near 1.6%) – so that “had the shocks of the 1970s occurred in the 1990s, the 1990s would have been almost as volatile as the 1970s,” and vice versa (Section 4). They conclude changes in the covariance matrix of VAR innovations “account for virtually all of the reduction in the observed volatility of output,” a finding that is robust to most (though not all) specification changes they test, and that matches Primiceri (2002), Simon (2000), and Boivin and Giannoni (2002).

No: fixing sectoral output shares at their 1965 values rather than letting them evolve produces essentially the same decline in GDP growth’s standard deviation (from 3.1% to 1.8%) as using actual shares, because volatility fell “across the board” in every sector, including those (durables, structures) whose output shares stayed roughly constant (Section 5.1), a result the authors attribute to Blanchard and Simon (2001) and McConnell and Perez-Quiros (2000) and corroborate using employment-share calculations following Warnock and Warnock (2001). They also note a timing problem: the manufacturing-to-services shift occurred gradually over four decades, while the moderation is a sharp, mid-1980s event.

The authors conclude this hypothesis “does not fare well” under closer inspection: the key supporting fact – that production volatility fell while sales volatility did not – is “not robust to the method of analysis,” since their own and others’ break tests (Ahmed, Levin and Wilson 2002; Herrera and Pesavento 2002; Kim, Nelson and Piger 2001) find significant breaks in sales variance too, and because the apparent production/sales divergence is a one-quarter-growth phenomenon that “disappears when one considers longer frequency fluctuations” (Section 5.2): at four-quarter growth rates, production and sales volatility fell by a similar 30%-40% across all sectors, and inventory-to-sales ratios actually rose for finished-goods and trade inventories even as they fell for raw-materials/work-in-progress inventories – the opposite pattern from what the information-technology-driven inventory story predicts.

Residential fixed investment’s instantaneous four-quarter-growth variance fell from an estimated 14.2 percentage points in 1981 to 6.0 in 1985, a decline comparable in share-weighted magnitude to that of durable goods output, and concentrated in residential (not nonresidential) construction and more pronounced for single-unit than multi-unit housing (Section 5.3). The authors consider, but do not resolve, several candidate explanations – mortgage-market deregulation reducing credit rationing, lower volatility of permanent income, and reduced volatility of expected real mortgage rates – noting each faces difficulties (e.g., mortgage deregulation occurred gradually, while the volatility drop is sharp; nondurables consumption volatility, their proxy for permanent income, shows no comparable 1980s decline) and conclude the question needs further work.

Using structural-VAR-identified shocks, the authors find that monetary policy shocks (Bernanke-Mihov and Christiano-Eichenbaum-Evans identifications) had a post-1984 standard deviation roughly 0.50-0.75 of their earlier value; Blanchard and Perotti’s (2001) tax shocks fell about 20%; Gali’s (1999) productivity shock fell about 25%; and commodity-price shocks fell 20%-30%, while oil-price shocks showed little change (using Hamilton’s 1996 asymmetric measure) or even increased in simple nominal/real growth-rate terms (Section 5.4). Weighting these shocks by their importance for GDP growth, productivity and commodity-price shocks make “modest contributions, in the neighborhood of 15%,” fiscal shocks make a negligible contribution, and oil shocks contribute negligibly or in the wrong direction; summing across shock categories is methodologically suspect because the shocks are not uncorrelated, but the authors judge that “most – perhaps two-thirds – of the reduction in volatility is not explained by the reduction in volatility of these shocks.”

Using a four-equation structural VAR (an IS equation, a hybrid New Keynesian Phillips curve, a forward-looking Taylor rule, and a commodity-price process) with benchmark parameters, the authors estimate that changes in the Taylor-rule (policy) coefficients account for about 14% of the reduction in GDP growth’s variance, while for most of the parameter combinations they examine the estimated contribution “falls in the range of 20% to 30%” (Section 5.5); the special case in which the IS curve’s interest-rate sensitivity is set to zero – so that monetary policy cannot affect output within the period by construction – yields an estimated policy contribution of only about 7%. They find corroborating evidence of a genuine policy shift (not just its output effect) in sharply higher estimated inflation and output-gap coefficients in post-1984 Taylor rules (Taylor 1999b; Clarida, Gali and Gertler 2000) and in a rise in the persistence of the Federal Funds rate (the largest autoregressive root rising from 0.96 in 1965-1978 to 1.00 in 1985-1998, per Watson 1999).

Combining their estimates, the authors attribute the moderation in GDP growth volatility to a combination of “improved policy (20-30%), identifiable good luck in the form of productivity and commodity price shocks (20-30%), and other unknown forms of good luck that manifest themselves as smaller reduced-form forecast errors (40-60%),” while stressing “these percentages have many caveats” (Abstract; Section 6). They explicitly rule out sectoral shift and inventory-management explanations, treat the housing-sector channel as an open question, and are candid that their search for causes “has not been completely successful,” leaving “much – perhaps half – of the decline in volatility unaccounted for.”

Because the authors judge that policy improvements are responsible for only a minority share of the moderation while “most of the reduction seems to be due to good luck in the form of smaller economic disturbances,” they conclude that while continued good policy might sustain part of the moderation, “we are left with the unsettling conclusion that the quiescence of the past fifteen years could well be a hiatus before a return to more turbulent economic times” (Section 6).