A Plucking Model of Business Cycles

Stéphane Dupraz Emi Nakamura Jón Steinsson*

Banque de France UC Berkeley UC Berkeley

April 8, 2024

Abstract
In standard models, economic activity fluctuates symmetrically around a “natu-
ral rate” and stabilization policies can dampen these fluctuations but do not affect
the average level of activity. An alternative view—labeled the “plucking model” by
Milton Friedman—is that economic fluctuations are drops below the economy’s full
potential ceiling. We show that the dynamics of the unemployment rate in the US
display a striking asymmetry that strongly favors the plucking model: increases in
unemployment are followed by decreases of similar amplitude, while the amplitude
of a decrease does not predict the amplitude of the following increase. In addition,
business cycles last seven years on average and unemployment rises much faster dur-
ing recessions than it falls during expansions. We augment a standard labor search
model with downward nominal wage rigidity and show how it can fit the plucking
property. We then show that additional non-standard features are required to match
the level and asymmetry of the duration of contractions and expansions.

Keywords: Downward Nominal Rigidity, Stabilization Policy, Labor Search.

JEL Classification: E24, E30, E52
* We would like to thank Gadi Barlevy, Marco Bassetto, Pierpaolo Benigno, Fabrice Collard, Jordi Gali,
François Gourio, Robert Hall, John Leahy, Julien Matheron, Andreas Mueller, Ricardo Reis, Carlos Thomas,
Antonella Trigari, and Johannes Wieland for valuable discussions. Emi thanks the National Science Founda-
tion (grant SES-1056107) and the George S. Eccles Research Award in Finance and Economics at Columbia
Business School for financial support. Emi and Jon thank the Smith Richardson Foundation for financial
support. The views expressed herein are those of the authors and do not necessarily reflect those of the
Banque de France or the Eurosystem.
1 Introduction

In the workhorse models currently used for most business cycle analysis, economic ac-
tivity fluctuates symmetrically around a “natural rate” and stabilization policy does not
appreciably affect the average level of output or unemployment. At best, stabilization
policy can reduce inefficient fluctuations around the natural rate. As a consequence, in
these models the welfare gains of stabilization policy are trivial (Lucas, 1987, 2003).
An alternative view is that economic contractions involve drops below the econ-
omy’s full-potential ceiling or maximum level. Milton Friedman proposed a “plucking
model” analogy for this view of business cycles: “In this analogy, [...] output is viewed
as bumping along the ceiling of maximum feasible output except that every now and
then it is plucked down by a cyclical contraction” (Friedman, 1964, 1993).1 In the pluck-
ing model view of the world, improved stabilization policy that eliminates or dampens
the “plucks”—i.e., contractions—increases the average level of output and decreases the
average unemployment rate. Stabilization policy can therefore potentially raise welfare
by substantial amounts (De Long and Summers, 1988; Benigno and Ricci, 2011; Schmitt-
Grohe and Uribe, 2016).
We show that the dynamics of the US unemployment rate strongly favor the plucking
model of business cycles. An implication of the plucking model—highlighted by Fried-
man (1964)—is that the dynamics of unemployment should display the following asym-
metry: economic contractions are followed by expansions of a similar amplitude—as if
the economy is recovering back to its maximum level—while the amplitude of contrac-
tions are not related to the previous expansion—each pluck seems to be a new event. We
refer to this asymmetry as the plucking property. We present strong evidence that the US
unemployment rate displays the plucking property: The increase in unemployment dur-
ing a contraction forecasts the amplitude of the subsequent expansion one-for-one, while
the fall in unemployment during an expansion has no explanatory power for the size of

1
The term “plucking” originates in Friedman’s image of a string (output) attached to the underside of
a board (potential output): “Consider an elastic string stretched taut between two points on the underside
of a rigid horizontal board and glued lightly to the board. Let the string be plucked at a number of points
chosen more or less at random with a force that varies at random, and then held down at the lowest point
reached.” (Friedman, 1964)
1
the next contraction.
To match the facts about plucking that we document in the data, we introduce
downward nominal wage rigidity into a tractable version of the Diamond-Mortenson-
Pissarides model with endogenous separations analyzed by Fujita and Ramey (2012). We
show that the model can quantitatively match the plucking dynamics of US unemploy-
ment. Our model reproduces the plucking property because good shocks mostly lead
to increases in wages, while bad shocks mostly lead to increases in unemployment. The
dynamics of the unemployment rate thus become asymmetric; the unemployment rate
rises far above its steady state level in response to adverse shocks, but falls much less in
response to favorable shocks.
The plucking dynamics of our model imply that fluctuations in unemployment are
fluctuations above a resting point of low unemployment, not symmetric fluctuations
around a natural rate. As a consequence, a reduction in the volatility of aggregate shocks
not only reduces the volatility of the unemployment rate, but also reduces its average
level, as in the models of Benigno and Ricci (2011) and Schmitt-Grohe and Uribe (2016).
Eliminating all aggregate shocks in our calibrated model reduces the average unemploy-
ment rate from 5.7% to 3.1%.
An alternative interpretation of the “plucking” property of unemployment is that it
derives from exogenous shocks to the determinants of unemployment rate that them-
selves have this property. We have chosen to model the exogenous shocks in our model
as symmetric for two reasons. First, we are interested in exploring the ability of the DMP
model to generate asymmetry within the labor market. Second, prior empirical work
has found asymmetries to be more pronounced in the unemployment rate than in other
macroeconomic data suggesting that the source of asymmetry is the labor market (e.g.,
McKay and Reis, 2008).
While our baseline model fits the plucking property, it fails to fit some other salient
features of unemployment dynamics. Empirically, business cycles last around 7 years
from peak to peak, and the unemployment rate rises much more rapidly during down-
turns than it falls during expansions. In principle, search models such as the one we
introduce provide an intuitive mechanism for slow recoveries: firms can shed workers

2
rapidly, but it takes time due to search and matching frictions to expand employment.
In practice, however, the large number of workers who flow between employment and
unemployment in every month implies that the DMP model does not have appreciable
internal propagation (see, e.g., Cole and Rogersion, 1999). As a result, unemployment
falls quickly once negative shocks dissipate. Our baseline model inherits this feature of
standard search models, generating short business cycles.
Matching persistence is important in its own right. But it also interacts with explain-
ing plucking (since a high degree of persistence leads expansions to be cut off before they
have time to run their course). To match the persistence of unemployment cycles and
assess whether our earlier conclusions about plucking carry over to a setting with empir-
ically realistic persistence of unemployment cycles, we propose a new illustrative model
with several non-standard features. In particular, this model features insecure short-term
jobs and a hump-shaped driving process for productivity shocks (an AR(2) process).
The presence of insecure short-term jobs implies that most new matches turn out to be
poor matches and separate quickly; however, some survive and become stable matches.
As a consequence, workers who become unemployed cycle through several jobs before
finding stable employment, a pattern emphasized by Hall (1995). This, in combination
with the AR(2) shock process generates substantial internal propagation despite the large
amount of churning observed in the data. This model fits not only the plucking property
in the data but also the duration of business cycles and the asymmetry in the speed of
contractions versus expansions.
Our work is related to several strands of existing literature. Petrosky-Nadeau, Zhang,
and Kuehn (2018) show how a DMP model features asymmetries that can generate
business-cycles disasters—large drops in production—despite symmetric shocks. This
force generates some plucking. But when we consider business cycles of the size we have
experience in our sample period, it generates much less plucking than we document mo-
tivating the role of DNWR. Kim and Nelson (1999) and Sinclair (2010) are two of the very
few modern attempts to assess the specific asymmetry emphasized by Friedman. Ca-
ballero and Hammour (1998), Bordo and Haubrich (2012), and Fatás and Mihov (2015)
explore related ideas. Ferraro (2017) provides an alternative explanation for the speed

3
asymmetry in the unemployment rate.
The paper proceeds as follows. Section 2 presents our empirical results on the asym-
metric dynamics of the unemployment rate. Section 3 lays out our plucking model of
business cycles. Section 4 shows how this model can match the plucking property, and
demonstrates that stabilizing fluctuations can reduce the average level of unemployment.
Section 5 proposes a new model that can match the persistence of unemployment cycles.
We analyze the quantitative implications of this model in section 6. Section 7 concludes.

2 The Plucking Property and the Dynamics of Unemploy-

ment Cycles

We start by demonstrating Friedman’s plucking property for post-WWII US unemploy-

ment data; namely, the amplitude of a contraction forecasts the amplitude of the subse-
quent expansion, while the amplitude of an expansion does not forecast the amplitude
of the subsequent contraction. We also demonstrate a speed asymmetry in unemploy-
ment: the unemployment rate rises more quickly than it falls, implying that the duration
of recoveries is typically much longer than the duration of recessions, as emphasized by
Neftçi (1984) among others. Finally, we review the empirical evidence on wage rigidity,
and show that the incidence of wage rigidity is countercyclical.
To implement our empirical analysis, we define business cycle peaks and troughs such
that they line up exactly with peaks and troughs of the unemployment rate. This yields
business cycle dates that are very similar to but not identical to those identified by the
NBER Business Cycle Dating Committee. Figure 1 plots the unemployment rate over
our sample period—which runs from January 1948 to February 2020—with vertical lines
indicating the times that we identify as business cycle peaks and troughs. For details on
our algorithm, see Appendix A.2

2
We identify ten peaks and ten troughs. To these we add a peak at the beginning of our sample. One
might worry that the contraction at the beginning of our sample may have started earlier. We are however
reassured on this point by the fact that the NBER identified November 1948 as a peak. We end our sample
at the onset of the Covid-19 recession, another clear peak in the data.
4
2.1 The Plucking Property

Figure 2 presents scatter plots illustrating the plucking property for the unemployment
rate. The left panel plots the amplitude of a contraction on the x-axis and the amplitude
of the subsequent expansion on the y-axis. The amplitude of contractions is defined as
the percentage point increase in the unemployment rate from the business cycle peak to
the next trough. The amplitude of expansions is defined analogously. There is clearly a
strong positive relationship between the amplitude of a contraction and the amplitude of
the subsequent expansion in our sample period. In other words, the size of a contraction
strongly forecasts the size of the subsequent expansion.
Table 1 reports the estimated coefficient from an OLS regression of the size of the sub-
sequent expansion on the size of a contraction. The relationship is roughly one-for-one.
For every percentage point increase in the amplitude of a contraction, the amplitude of the
subsequent expansion increases by 1.1 percentage points on average. Despite the small
number of data points, the relationship is highly statistically significant. Furthermore, the
explanatory power of the amplitude of the previous contraction is large. The R2 of this
simple univariate regression is 0.59.
The right panel of Figure 2 plots the amplitude of an expansion on the x-axis and
the amplitude of the subsequent contraction on the y-axis. In sharp contrast to the left
panel, there is no statistically significant relationship in this case. The size of an expansion
does not forecast the size of the next contraction. One cannot reject that, in Friedman’s
language, each contractionary pluck that the economy experiences is independent of what
happened before. Table 1 reports the estimated coefficient from a linear regression of
the size of the subsequent contraction on the size of an expansion. The relationship is
actually slightly negative, but is far from statistically significant. Moreover, the R2 of the
regression is only 0.22.3

3
Jackson and Tebaldi (2017) suggest that the duration (not size) of an expansion is predictive of the size
of the following contraction. They motivate this idea by analogy to forest fires: the longer the expansion,
the more “underbrush” builds up—e.g., low quality matches and entrants—that becomes fuel in the sub-
sequent contraction. We find no evidence of the forest fire theory at the aggregate level: the duration of
an expansion is no more predictive of the size of the following contraction than the size of the expansion
is. The relationship is actually negative (but not significantly so), driven by the fact that the three longest
post-WWII expansions (1961-1968, 1982-1989, 1992-2000) were followed by relatively mild recessions. Tasci
and Zevanove (2019) confirm these results and also present state level results for the plucking model and
5
2.2 The Speed Asymmetry

Unemployment rises more quickly during contractions than it falls during expansions,
a point made quantitatively in early work by Neftçi (1984).4 Table 1 reports the average
speed of expansions and contractions to illustrate this asymmetry. We measure the change
in unemployment (in percentage points) over the spell and the length of time the spell
lasts for. The speed for a expansion or contraction is the ratio of those two numbers. We
then take a simple average across all expansions and separately a simple average across
all contractions.
We find that the unemployment rate rises roughly twice as quickly during contractions
(1.9 percentage points per year) as it falls during expansions (0.9 percentage points per
year). This difference is highly statistically significant. We run a regression of the absolute
value of the speed of expansions and contractions on a dummy variable for a spell being
a contraction and find that the p-value for the dummy is 0.002.
Looking back at Figure 1, we can clearly see that when the unemployment rate starts
falling, it usually falls relatively steadily for a long time. As a consequence, expansions
are quite long. The average length of expansions in our sample is roughly 59 months, or
almost five years. Contractions are also quite persistent, but less so. The average length
of contractions in our sample is roughly 27 months, a bit more than two years. Perhaps
most strikingly, in a few cases—the 1960s, 1980s, 1990s, and 2010s—the unemployment
rate has fallen steadily for six to ten years without reversal.

forest fire theory. Their state level results are similar to our results at the aggregate level: There is strong
evidence for the plucking property but no evidence for the forest fire theory.
4
Sichel (1993) refers to this asymmetry as the “steepness” asymmetry, while McKay and Reis (2008)
refer to it as the greater “violence” of contractions, in reference to Mitchell (1927). Given that contractions
and expansions are of about the same average size (3.7 percentage points), the fact that contractions are
“steeper” or “more violent” than expansions is equivalent to the fact that they are briefer. Awareness of
this asymmetry dates back at least to the 1920s. Mitchell (1927) notes that “business contractions appear to
be briefer and more violent than business expansions.”
6
2.3 Wage Rigidity

A large literature has presented microeconomic evidence of downward nominal wage

rigidity in US data.5 Olivei and Tenreyro (2010) present evidence that such wage rigidity
matters for real outcomes. A large theoretical literature has studied the implications of
downward nominal wage rigidity for business cycles.6
An intuitive fact about the data is that wage rigidity has been countercyclical in recent
US business cycles—i.e., wage freezes tend to occur during recessions. Figure 3 plots the
fraction of “Wage Freezes” from the San Francisco Fed’s Wage Rigidity Meter for the pe-
riod 1997-2019 along with the unemployment rate (Left Panel) and the non-employment
rate for the working age population (Right Panel).7 The correlation is striking. The frac-
tion of wage freezes rises rapidly in each of the three recessions that occur in this sample
period. Before 1997 the data are less complete, and the correlation is weaker because of
the confounding effects of changes in trend inflation.

3 A Plucking Model with Downward Nominal Wage

Rigidity

Next, we present a model designed to fit the plucking property discussed in section 2.
This model augments the workhorse DMP model with endogenous separation developed
5
See, in particular, McLaughlin (1994); Kahn (1997); Card and Hyslop (1997); Bewley (1999); Altonji and
Devereux (2000); Kurmann and McEntarfer (2017); Hazell and Taska (2018); Grigsby, Hurst, and Yildirmaz
(2018). Correcting for measurement error might reveal an even higher prevalence of wage rigidity. Early
work using data from the Panel Study of Income Dynamics (PSID) and the Current Population Survey (CPS)
includes McLaughlin (1994), Kahn (1997), and Card and Hyslop (1997). Altonji and Devereux (2000) report
larger amounts of downward nominal wage rigidity and virtually no wage cuts in the PSID after correcting
for measurement error. Gottschalk (2005) and Barattieri, Basu, and Gottschalk (2014) report similar findings
based on their adjustments for measurement error in the Survey of Income and Program Participation.
6
Prominent contributions include, e.g., Akerlof, Dickens, and Perry (1996); Kim and Ruge-Murcia
(2009); Benigno and Ricci (2011); Abbritti and Fahr (2013); Schmitt-Grohe and Uribe (2016); Chodorow-
Reich and Wieland (2018). The importance of wage rigidity for generating realistic fluctuations in unem-
ployment has been stressed by Shimer (2005), Hall (2005), Gertler and Trigari (2009), and Gertler, Huckfeldt,
and Trigari (2016).
7
The Wage Freeze measure reports the fraction of job-stayers whose wages are unchanged ver-
sus one year prior (Daly, Hobijn, and Wiles, 2011; Daly, Hobijn, and Lucking, 2012; Daly and Hobijn,
2014). See https://www.frbsf.org/economic-research/indicators-data/nominal-wage-
rigidity/.
7
by Fujita and Ramey (2012) with downward nominal wage rigidity.

3.1 Search and Matching with Endogenous Separation

The model consists of an infinite mass of atomistic firms, and a mass of workers with
inelastic labor supply normalized to one. At the beginning of a period, workers are either
already matched with a firm, or looking for a job.8 If a firm and a job-seeker match in
period t, we assume the worker starts working right away.9 To match with a new worker,
a firm must post a vacancy, at a cost c per period in which the vacancy remains open.
An open vacancy fills with probability qt , which is taken as exogenous by the firm. In the
aggregate, the probability qt is determined by a matching function q pθt q, where θt  Vt{U0,t
denotes labor market tightness. Labor market tightness is the ratio of the number of
vacancies posted Vt to the number of job-seekers U0,t at the beginning of the period. The
matching function also determines the probability for a job-seeker of finding a job. This
is equal to the ratio of matches q pθt qVt to job-seekers U0,t , f pθt q  q pθt qVt {U0,t  θtqpθtq.
When a worker and a firm are matched, they produce output At xt , where At and xt
are aggregate and match-specific productivity factors. We assume that both follow AR(1)
exogenous processes in logs,

logpAt q  ρa logpAt
1 q εat , (1)

logpxt q  ρx logpxt
1 q εxt , (2)

where εat and εxt are Gaussian shocks with standard deviations σεa and σεx . All new matches
start at the same match productivity level xhire , which we take to be average match pro-
ductivity x  1.10
8
Only unemployed workers look for jobs. Fujita and Ramey (2012) also develop a version of the model
with on-the-job search. We consider their model with endogenous separation but no on-the-job search.
9
This timing of the labor market follows, e.g., Blanchard and Gali (2010), but differs from Fujita and
Ramey (2012). It is not important for our results, but avoids the unpalatable assumption that the number
of job-seekers U0,t at the beginning of period t (which includes at least all the workers who separated from
their jobs at the end of t
1) is the same as the number of unemployed workers Ut in period t, even when
search frictions shrink to zero and the job finding rate is 1.
10
Because Fujita and Ramey (2012) also consider a model with on-the-job search, they assume that new
matches start at the highest productivity level to make sure all job offers are accepted. Since we do not
consider on-the-job search, job offers are accepted even if the productivity in new matches is not the highest
8
 An on-going match continues into the next period unless it is exogenously
terminated—which occurs with probability δ—or endogenously terminated. Let Jt pxt q
be the value to a firm of an on-going match with match-specific productivity xt .11 The
firm can terminate the match if it yields a negative value. This implies that Jt pxt q is given
by:

Jt pxt q  maxtJtc pxt q, 0u (3)

where Jtc is the value of the match to the firm if it is continued, which solves the recursion

Jtc pxt q  xt At
wt pxt q β p1
δ qEt pJt 1 pxt 1qq , (4)

where wt pxt q is the real wage paid in the match at period t.

Let Wt pxt q be the value to a worker of a match with match-specific productivity xt ,
and Ut be the value to a worker of being unemployed. Since the worker can terminate the
match if it yields it a negative value, the value of being employed in a match at xt is given
by

Wt pxt q  maxtWtc pxt q, Ut u (5)

where Wtc is the value if the match is continued, which solves the recursion

Wtc pxt q  wt pxt q
ζ βEt p1
δqWt 1pxt 1q δ p1
ft 1 qUt 1 δft 1 Wt 1 pxhireq , (6)

where ζ is the dis-utility cost of working relative to being unemployed. The value of being
unemployed solves the recursion

Ut b βEt p1
ft 1qUt 1 ft 1 W t 1 pxhireq , (7)

where b is unemployment benefits. Subtracting (7) from (6), the value of being employed
relative to being unemployed, Vt  Wt
Ut, solves the recursion

Vtc pxt q  wt pxt q
z β p1
δ qEt Vt 1 pxt 1q
ft 1 Vt 1 pxhireq , (8)

productivity level. Our process for match-specific shocks is also a more standard AR(1) process than the
memoryless Poisson process with infrequent large shocks assumed in Fujita and Ramey (2012).
11
We make explicit the dependence of Jt and other value functions only in the idiosyncratic state xt . The
dependence of Jt on the aggregate state is kept implicit in the t time-index of Jt .
9
where

Vt pxq  maxtVtc pxq, 0u (9)

and z b ζ is the flow value of unemployment in terms of both unemployment benefits

and leisure.
Assuming free entry, the cost of posting a new vacancy must in equilibrium be equal
to the benefit to the firm of posting a new vacancy. This equates the value of a new hire
to the firm to the expected cost of a new hire:

Jt pxhire q 
c
. (10)
qt

The model is closed with an assumption on wage-setting. Fujita and Ramey (2012)
assume that wages are set according to Nash-bargaining, as is standard in search mod-
els. Define the total value of a continuing match to be Stc pxt q  Jtcpxtq Vtc pxt q. Under
Nash-bargaining, Vtc,N ash  γStc,N ash, where γ P r0, 1s is the bargaining power of workers.
Combining equations (4) and (8) and the Nash-bargaining assumption gives

Jtc,N ash pxt q  p1
γ qpAt xt
z q β p1
δ qEt JtN ash
1 pxt 1 q
ftN ash
1 γJt 1 px
N ash hire
q . (11)

Equations (10) and (11) allow us to solve for qt and Jtc,N ash pxt q, as detailed in Appendix B.
Combining (4) and (11) allows us to recover the Nash wage as
 
wtN ash p xt q  γAt xt p 1
γ qz β p1
δ qEt γftN ash
1 Jt 1 px
N ash hire
q . (12)

3.2 Downward Nominal Wage Rigidity

We extend the Fujita-Ramey model to allow for downward nominal wage rigidity
(DNWR).12 The presence of match heterogeneity implies that wages differ for new and
ongoing matches. We assume DNWR for both groups. For an on-going match at time t,
the nominal wage is either the flexible nominal wage—which we assume to be the wage
12
We assume DNWR by directly specifying the wage-rule. Other work directly specifying a wage rule
in a search and matching framework includes Blanchard and Gali (2010); Shimer (2010); Michaillat (2012).
This approach has the advantage that we can more easily investigate what features the wage process needs
to generate realistic unemployment dynamics.
10
that would obtain if the firm set wages by Nash bargaining in this and future periods—or,
if this requires the nominal wage to fall, the nominal wage remains unchanged from the
previous period, i.e.:
" *
wt pxt , wt
1 q  max wtN ash pxtq, wΠ̄t
1 . (13)
t

For new matches, we assume that the hiring wage at time t is either the flexible
nominal wage—which we again assume to be the wage that would obtain under Nash
bargaining—or, if this requires the nominal hiring wage to fall, the nominal hiring wage
in the previous period, i.e.:
" new
*
wtnew px hire
, wtnew1

q  max wtN ash px hire
q, wΠ̄t
1 . (14)
t

To generate fluctuations in firms’ hiring decisions, wage rigidity in new matches is par-
ticularly important, as emphasized by Pissarides (2009). Assumption (14) is therefore es-
sential to introducing DNWR in a search and matching model with match heterogeneity.
Hazell and Taska (2018) document such DNWR in new hires’ wages.13
In both equations (13) and (14), inflation relaxes the constraint on downward real wage
adjustments: it greases the wheels of the labor market. We specify monetary policy as
directly setting a path for the inflation rate Πt , which we take to be constant at some
target value Π. Equations (10) and (11) still hold under DNWR, except that the firm’s
value function now depends on lagged wages. We first solve the model under Nash-
bargaining to recover the Nash wage (12), then use equations (10) and (11) to solve for Jtc
and qt , as detailed in Appendix B.

3.3 Worker Flow Accounting

Let st be the destruction rate, defined as the fraction of matches that get destroyed at the
beginning of period t. Because matches can be endogenously terminated, the destruc-
tion rate st depends on the cross-sectional distribution of employment across the state of
13
Our model does not feature preemptive wage moderation of the type that is present in the wage setting
models (e.g., Kim and Ruge-Murcia, 2009; Elsby, 2009; Benigno and Ricci, 2011). Yet this does not mean
firms in our model are myopic. They rationally maximize intertemporal profits. What they preemptively
moderate in anticipation of a fall in productivity is hires, not wages. Either wages or hires can respond to
concerns about the future. In our model it is hires that are moderated, because wages are not set by firms.
11
matches. Under Nash-bargaining, the state of a match reduces to match productivity xt .
Fujita and Ramey (2012) show how to keep track of the distribution of employment across
matches to calculate the destruction rate in this case. Under DNWR, the state of a match
includes both match productivity xt and the lagged wage wt
1 . Appendix B shows how
to keep track of the joint distribution of employment across the joint state pxt , wt
1 q and
calculate the destruction rate under DNWR.
Workers’ job-finding rate ft is a direct function of the vacancy-filling rate qt through
the matching function and can therefore be easily recovered from qt . Since we assume that
job-seekers who match with a firm at the beginning of period t start working at t, workers
who separate from a firm between t
1 and t and join the pool of job-seekers may find a
new job at t without spending time unemployed, instead transitioning directly from job
to job. The exit rate from employment to unemployment is therefore distinct from the
destruction rate st and equal to

s̄t  p1
ftqst. (15)

The law of motion of the unemployment rate Ut is

Ut  p1
ftqUt
1 s̄t p1
Ut
1 q. (16)

3.4 Calibration

Table 2 provides a summary of our calibration. We calibrate the model to a monthly fre-
quency. We set the discount factor β to correspond to an annual interest rate of 4%. We
assume a Cobb-Douglas matching function q pθq  µθ
η and set the elasticity of the match-
ing function to η  0.5, in the middle of the range reported in Petrongolo and Pissarides
(2001)’s survey. We calibrate the flow value of unemployment following Hagedorn and
Manovskii (2008) to z  0.95, so that the model generates significant fluctuations in the
job-finding rate under Nash bargaining, which we will consider as a benchmark.14 This
14
Fujita and Ramey (2012) show that their model under Nash bargaining can generate fluctuations in
the unemployment rate away from a high calibration of z by making unemployment fluctuate through
fluctuations in separations. As they show however, away from a high calibration of z, the model cannot
generate fluctuations in the job-finding rate.
12
calibration also generates the realistic prediction that firms’ surplus is increasing in pro-
ductivity under DNWR, as explained in Appendix F.1. We calibrate the exogenous de-
struction rate δ to match the average monthly share of quits in the non-farm sector in
JOLTS between January 2000 and February 2020 of 1.9%.
The parameters µ and c jointly determine hiring costs. One of the two is redundant

1
as only the composite parameter cµ 1
η is relevant for the equilibrium. (See Appendix
B.1 for further discussion of this point.) We normalize µ to 1. We calibrate c so that the
cost of hiring a worker c
q
 J is 10% of monthly wages in a steady state with u  5.7%
and s̄  2%, in line with what Silva and Toledo (2009) report based on the Employer
Opportunity Pilot Project survey in the US. This yields c  0.30.
We set the auto-regressive root of the aggregate productivity process ρa to 0.98 fol-
lowing Shimer (2010). We set the auto-regressive root of the match-specific productivity
process ρx to 0.98, following Foster, Haltiwanger, and Syverson (2008)’s estimates of the
persistence of plants’ TFP (0.8 on an annual basis). When we consider the model under
downward nominal wage rigidity, we set inflation to 2% per year. Inflation is immaterial
in the version of the model without DNWR.
This leaves γ, σεa and σεx . We pick them to match the average level of the unemploy-
ment rate (5.7% in the data), the standard deviation of the unemployment rate (1.6% in
the data), and the average of the rate of exit from employment s̄ (2% as reported by Fujita
and Ramey (2006, 2012)). We choose to match the standard deviation of unemployment
exactly (as opposed to calibrating to the standard deviation of productivity in the data) so
that we can apply our definition of expansions and contractions to our simulated samples
in the same way as we do to the real world data. These choices yield γ  0.57, σa  1.6%
and σx  2.1% under Nash bargaining, and γ  0.43, σa  1.5% and σx  1.5% under
DNWR.

4 Quantitative Analysis of the Baseline Model

Given the asymmetries and non-linearities our model is intended to capture, we rely on
global methods to numerically solve for the equilibrium. Appendix B discusses the algo-

13
rithm we use in detail. We simulate 5000 samples of 866 periods (the length of our sample
of real-world data) and calculate the statistics reported in the empirical Table 1 in each of
these simulated samples. We then report the median estimate across samples for each
statistic as a point estimate and the standard deviation of the estimates across samples in
parentheses below each point estimate.
Table 3 reports the results. The top panel presents results on the plucking property.
We first consider the model with flexible wages (Nash bargaining). This version of the
model generates non-trivial plucking, but substantially less than in the data. The regres-
sion coefficient for the size of expansions on the previous contraction is 0.39 versus -0.03
for contractions on the previous expansion. These coefficients are 1.09 and -0.38 in the
data. Next, we present results for the model with DNWR. This case produces substan-
tially more plucking that the model with flexible wages. The regression coefficient for the
size of expansions on the previous contraction is 0.61 versus -0.05 for contractions on the
previous expansion.
Figure 4 illustrates these results graphically. It presents scatter plots for the regressions
discussed in the last paragraph. These scatter plots reveal that there is a large amount of
noise in the relationship between the amplitude of expansions and contractions in the
model. This is particularly the case in the model with flexible prices. In this model, the
R2 of the size of expansions on the previous contraction is only 0.19, compared to 0.59 in
the data. The model with DNRW does substantially better on this metric with an R2 of
0.39.

4.1 Mechanisms for Plucking

Two features of the model contribute to the plucking property. First, non-linearity in
the worker-flow equations (15)-(16) generates plucking. Figure 6 plots the steady-state
relationship between the job finding rate and unemployment implied by the worker-
flow relationship (15)-(16) taken in steady-state.15 The relationship is convex because it

15
The analytical expression is u  s{ps pf {p1
f qq under our assumption that workers who separate
from a firm get a chance to find a new job right away and spend no time unemployed. Assuming instead
that workers must necessarily spend one period unemployed before finding a new job, the relationship
would be u  s{ps f q. In this case too, the steady-state worker-flow relationship is convex but the non-
14
gets harder and harder to lower the unemployment rate the lower it gets. Earlier work
has emphasized this non-linearity (Petrosky-Nadeau and Zhang, 2017; Petrosky-Nadeau,
Zhang, and Kuehn, 2018).16
The reason why this non-linearity generates only modest amounts of plucking is that
over the empirically relevant range of unemployment rates – 2.5% and 10.8% for the sam-
ple period 1948-2019 – the degree of non-linearity is modest. This is illustrated in Figure
6. In Table 3, we make sure not the overstate the strength of this non-linearity by present-
ing results for a “top-truncated” sample that only includes expansions and contractions
that are less than 6.5 percentage points in size—i.e., the size of the 2009-2019 expansion
which is the largest one in our sample.
The second feature of our model that contributes to plucking is DNWR. When wages
are downward rigid, negative shocks typically result in higher unemployment while pos-
itive shocks yield wage increases. As a result, unemployment sometimes rises far above
its steady state, but rarely falls below. Table 3 shows that this feature can generate a sub-
stantial amount of plucking.
But even the model with DNWR produces less plucking than the data. The top panel
of Figure 5 helps us understand why. This panel plots a simulated path of unemployment
in our model with DNWR. A prominent feature of this simulated series is that expansions
are often interrupted by a new contraction before unemployment has had enough time to
fully recover. This can introduce a great deal of noise in our measure of plucking (as in
Figure 4). Some expansions will be cut short. Others will appear excessively long because
their length is governed not only by the size of the previous contraction but also by earlier
contractions the recovery from which were cut short.
A second weakness of the model with DNWR is that the unemployment rate is much
less persistent than in the data. The low persistence in the model is documented in the
lower half of Table 3. The model generates unemployment cycles that are far too short.
The average duration of expansions is 25.1 months versus 59.1 in the data, and the average
duration of contraction is 16.1 months, versus 26.9 in the data.

linearities it induces are quantitatively small.

16
Hairault, Langot, and Osotimehin (2010), Jung and Kuester (2011), and Lepetit (2018) also emphasize
this source of non-linearity.
15
 The model with flexible wages does considerably better on these metrics. With our
Hagedorn and Manovskii (2008) calibration, movements in real wages are muted and
the model’s endogenous variables inherit the behavior of the shock process. This yields
quite a bit of persistence. With DNWR, however, it is nominal wages that are rigid in
downturns. Real wages fall due to increases in the price level. This implies that gaps
between At and wt get eroded faster than in the flexible wage model with the Hagedorn
and Manovskii (2008) calibration. In section 5, we present a model with several features
designed match both plucking and persistence simultaneously.

4.2 Stabilization Policy and Output Gaps in the Plucking Model

In a thought-provoking exercise, Lucas (1987, 2003) showed that the welfare benefits of
eliminating all economic fluctuations are trivial in a simple benchmark model. Crucially,
Lucas assumed in this thought experiment that the average level of economic activity
was unaffected by the elimination of business cycles. Our plucking model violates this
assumption.
In a plucking model, recessions are asymmetrically periods when the economy drops
below potential. Eliminating these business cycle fluctuations raises the average level of
economic activity. The top panel of Figure 5 shows that in our model, eliminating all fluc-
tuations reduces the average unemployment rate from 5.7% to 3.1%. Conversely, increas-
ing the standard deviation of aggregate shocks by 50% (from 1.5% to 2.25%) increases the
average unemployment rate to 12.2%. Figure C.1 in Appendix C plots the average level of
the unemployment rate in our plucking model as a function of the volatility of aggregate
shocks.
The plucking model also implies that standard measures of the output gap are biased.
Aiyar and Voigts (2019) show that common estimation methods of the output gap implic-
itly assume a zero-mean output gap. In the plucking model, however, output does not
fluctuate symmetrically around a natural rate. Standard methods systematically under-
estimate the amount of slack because the output gap is on average negative.

16
4.3 Entry and Exit over the Business Cycle

The bottom two panels of Figure 5 give simulated paths for the job-finding rate ft and
the rate of exit from employment s̄t in our model. Both the hiring and separation margins
contribute to the sharp rise in unemployment during recessions, consistent with evidence
documented by Elsby, Michaels, and Solon (2009) for US data.17 Negative shocks decrease
the job-finding rate and increase the rate of exit from employment, while positive shocks
mostly lead to increases in wages.
We quantify the relative importance of hiring versus separations by calculating their
contributions to the volatility of unemployment. For the contribution of the job-finding
rate f , we simulate the unemployment rate from equation (16) with s̄ fixed at its mean,
then calculate the standard deviation of the resulting counterfactual unemployment rate,
and divide it by the standard deviation of actual unemployment. Analogously, for the
employment exit rate, we do the same analysis with the job-finding rate fixed at its mean.
According to this metric, fluctuations in the job-finding rate explain 42% of fluctuations
in unemployment, while fluctuations in the employment exit rate explain 55%.
Table 4 reports additional statistics: the volatility, the auto-correlation, and the cor-
relation with productivity of the unemployment rate, of the job-finding rate, and of the
employment exit rate. We calculate these statistics for the HP-filtered log of simulated se-
ries from the model. Relative to the model with Nash bargaining, the model with DNWR
lowers the correlation of all three variables with productivity and brings it closer to the
data. The auto-correlation statistics are however substantially smaller than in the data—a
manifestation of the lack of persistence also captured with the speed and duration statis-
tics in Table 3.

17
The separation margin is not necessary to replicate the plucking property however. Appendix F
presents a simplified version of the model introduced in section 5 with a constant exogenous separation
rate, and shows that it, too, generates the plucking property (see Table F.5.)
17
5 A New Model with Decreasing Returns to Labor, Inse-
cure Jobs, and AR(2) Shocks

We next propose a model with several “non-standard” features, aimed at generating sub-
stantially more persistence in unemployment fluctuations relative to the model used ear-
lier in the paper. The non-standard features are: 1) decreasing returns to labor, 2) insecure
short-term jobs and 3) AR(2) shocks. Our model builds on Michaillat (2012), who also
emphasizes the importance of decreasing returns in understanding unemployment dy-
namics.
The new features come at a cost. We are no longer able to solve the model with match-
specific productivity shocks of the type we had in section 3. This limits the model’s re-
alism when it comes to the hiring and separation margins. We do, however, incorporate
sectoral heterogeneity in productivity which implies that downward nominal wage rigid-
ity binds in sectors with stagnant productivity, even if aggregate productivity growth is
positive.

5.1 Decreasing Returns to Labor and Heterogeneous Labor Inputs

The model consists of a continuum of sectors i P r0, 1s, each employing a distinct type of
labor i. In each sector, firms have access to a decreasing-returns production function that
uses labor i as its single input, Yti  AitF pNtiq, where Yti is output, Nti is employment, and
Ait is an exogenous productivity shifter. For simplicity, we restrict sectoral heterogeneity
to labor markets: consumers perceive goods produced in different sectors as identical
and therefore value them equally. All goods are sold in a competitive product market at
a common price Pt .
A given worker provides a particular type of labor, and can therefore only seek to
work at a firm in one sector. This implies that there is a distinct labor market for every
type of labor and workers cannot flow across labor markets. We think of these labor types
as occupations in a particular location, e.g., lawyers in Houston. Switching occupations
is difficult due to occupation-specific human capital. Mobility constraints limit the will-

18
ingness of workers to switch locations.18

5.2 Labor Demand with Insecure Short-Term Jobs

A firm in sector i starts period t with the stock of workers it inherits from the previous
period, denoted by Mti . These workers are securely attached to the firm: at the end of each
period only a small fraction δ P p0, 1q of them separate from the firm for exogenous rea-
sons. Before starting production, the firm hires Hti workers. These workers start working
at time t. The level of employment at the firm at t is then Nti  Mti Hti .
Newly-hired workers separate from the firm at the end of period t at a higher rate
d ¥ δ. If they do not separate at the end of period t, however, they join the pool of
securely attached workers and face the low exogenous separation rate δ from period t 1
onward. The stock of workers attached to the firm at the beginning of the next period is
then

Mti 1  p1
δqMti p1
dqHti if Hti ¥ 0. (17)

The difference between the separation rates of securely attached workers and newly-
hired workers captures the fact that many newly formed matches turn out to be poor
matches for various reasons and are therefore terminated quickly. An unemployed
worker will typically transition between several jobs with intervening unemployment
spells before one of these jobs turns out to be a good match and thus turns into a long-
term, secure position. The fact that newly hired workers face much higher separation
rates than workers with longer tenure has been emphasized by Hall (1995), Pries (2004),
Krolikowski (2017), Jung and Kuhn (2019), Jarosch (2021), and Hall and Kudlyak (2021),
among others. The fact that there are effectively only two types of jobs in our model—
very short-term jobs that last one period and very secure jobs—is a simplifying assump-
tion that we make for tractability. This assumption alleviates the need to keep track of the
stock of short-term workers working at the firm.

18
The assumption of such differentiated labor inputs is standard in the New-Keynesian literature. There,
differentiated labor inputs are an important source of strategic complementarity in price setting. See, e.g.,
Woodford (2003, ch. 3).
19
 We allow the firm not to hire in period t and instead endogenously lay off securely-
attached workers, above and beyond the fraction δ that exogenously left at the end of
period t
1. In this case, Hti 0 and all the workers Nti who work at the firm in period t
are securely attached, so that the firm starts the next period with a number of workers

Mti 1  p1
δqNti if Hti ¤ 0. (18)

We assume that the firm pays its securely attached and newly-hired workers the same
wage. One rationale for this assumption is that paying them differently would adversely
affect morale at the firm (Bewley, 1999). We denote the real wage in sector i by wti . To
hire workers, the firm must post vacancies. Posting a vacancy costs cAit units of goods,
where c is a constant.19 A vacancy translates into a hire if it matches with a job-seeker. A
match happens with probability qti , which firm i takes as given. Hiring one worker has
the expected cost Ait c{qti .
Firm i’s real profits at time t are revenues net of the cost of labor and hiring costs

Ait c i
Ait F pNti q
wti Nti
H I| i .
qti t Ht ¥0
(19)

Like all agents in the model, the firm is risk-neutral and discounts the future with a fac-
tor β P p0, 1q. The firm is forward-looking and chooses how many workers to hire to
maximize intertemporal real profits

Ait c i
Ωit p q  max
Mti
H
t i
Ait F p
Mti Hti q
p
wti Mti Hti q
qti t Ht ¥0
H1 i βEt pΩit 1 pMti 1qu, (20)
t

subject to the law-of-motion of its workforce Mti 1 pNtiq given by equations (17)-(18).
In this model with decreasing returns to labor, the level of employment at the firm Nti
affects the marginal productivity of all workers and therefore the bargaining position of
the firm when negotiating wages with all its workers. The firm can therefore in principle
internalize the effect of its chosen employment level on the wage it will be able to bargain
with its workers, leading to intrafirm bargaining (Stole and Zwiebel, 1996; Brügemann,

19
We make the cost of posting a vacancy proportional to productivity as in e.g. Blanchard and Gali
(2010), because it allows us to consider non-stationary sectoral productivity shocks without increasing the
size of the state-space beyond what is computationally feasible.
20
Gautier, and Menzio, 2019). With DNWR however, intrafirm bargaining becomes much
harder to solve. We therefore abstract from intrafirm bargaining and assume that firms
do not internalize the effect of their chosen employment level of the wage they will be
able to bargain.20
Let Jti pMti q  BΩit pMti q denote the equilibrium marginal value to the firm of a long-
BMti
term worker inherited from the previous period. For a level of employment Nti —not
necessarily equal to Mti —the marginal value to the firm of a worker already attached to
the firm is

Jti pNti q  Ait F 1 pNti q
wti β p1
δ qEt Jti 1 pMti 1 pNtiqq (21)

and the marginal value to the firm of a newly-hired worker is


Jti,new pNti q  Ait F 1 pNti q
wti β p1
dqEt Jti 1 pMti 1 pNtiqq . (22)

Each is equal to the marginal product of labor net of the real wage, plus the continuation
value discounted with the appropriate separation rate. When already-hired workers and
newly-hired workers face the same separation rate δ  d, then Jti,new  Jti and the model
reduces to its version without insecure short-term jobs which we consider in Appendix F.
Consider the demand for labor of a firm that starts off period t with Mti workers inher-
ited from the previous period. If at the inherited level of employment Mti , the marginal
value of a newly-hired worker is greater than the hiring cost Jti,new pMti q ¡ Ait c
qti
, the firm
will hire additional workers Nti ¡ Mti up to a point where
Ait c
Jti,new pNti q  . (23)
qti
If at the inherited level of employment Mti , the marginal value of an already-attached
worker is positive but the marginal value of a newly-hired worker is less than the hiring
20
Absent DNWR, the bargaining between the firm and its worker yields a differential equation for the
wage that can be solved in closed form for a constant-elasticity production function (Cahuc, Marque, and
Wasmer, 2008; Elsby and Michaels, 2013). Under DNWR however, the derivative of the wage is discontin-
uous at employment levels at which the DNWR constraint starts binding, preventing a simple closed-form
solution. Cahuc, Marque, and Wasmer (2008) and Elsby and Michaels (2013) show that the Nash-bargained
wage taking into account intrafirm bargaining only differs from the Nash-bargained wage that abstracts
from it (30) by a multiplicative coefficient in front of the marginal product of labor F 1 pN q.
21
cost, i.e., Jti pMti q ¥ 0 and Jti,new pMti q ¤
Ait c
qti
, the firm will freeze employment at

Nti  Mti. (24)

If at the inherited level of employment Mti , the marginal value of an already-attached

worker is negative Jti pMti q 0, the firm will endogenously fire some of its existing work-
ers Nti Mti up to

Jti pNti q  0. (25)

Appendix D.1 provides derivations. Note that firms cannot simultaneously hire and fire
endogenously at a sectoral level, since we do not allow for match heterogeneity.

5.3 Workers

There is a fixed supply of workers in each sector, common across sectors. Workers supply
an exogenous quantity of labor which we normalize to 1. Workers in sector i can be in
i
one of three states at t. A number N0,t  minpNti, Mtiq are employed in a secure job. A
number Nti
N0,t
i
are newly employed. The remaining Uti  1
Nti are unemployed. We
i
denote Ut,0  1
Nt,0i the number of workers who are either unemployed or employed in
i
a short-term job, and refer to it as the “broad” unemployment rate. Ut,0 is also the number
of job-seekers (see below). The Nti employed workers—either in short-term jobs or secure
jobs—earn the real wage wti and forego leisure relative to being unemployed, which they
value as Ait ζ units of consumption. Unemployed workers earn unemployment benefits
Ait b.21
Unemployed workers transition to being newly employed with the job-finding prob-
ability fti , and remain unemployed with probability 1
fti . We assume that employed
workers who lose their jobs between periods t
1 and t get a chance to find a new job
at the beginning of period t and therefore to work in period t, spending no time unem-
ployed. Securely employed workers separate from their jobs with a probability δti , which

21
Making unemployment benefits and the utility of leisure proportional to productivity allows us to
consider non-stationary sectoral productivity shocks without increasing the size of the state-space beyond
what is computationally feasible.
22
may be higher than δ due to endogenous layoffs. They therefore transition to unemploy-
ment with a probability δti p1
fti q, transition to being newly hired with probability δti fti ,
and remain securely employed with a probability p1
δti q. Newly employed workers sep-
arate from their jobs with a probability dit , which can be higher than d due to endogenous
layoffs. They therefore transition to being securely employed with a probability 1
dit ,
transition to unemployment with a probability dit p1
fti q, and remain newly employed—
in a new job—with probability dit fti .
Workers, like firms, are risk-neutral. The values they derive from being unemployed
Uti , securely employed Wti , and newly employed Wti,new solve the recursive equations

Uti  Ait b βEt p1
fti 1qUti 1 fti 1 Wti,new
1 , (26)

Wti 
wti Ait ζ βEt δti 1 p1
fti 1qUti 1 δti 1 fti 1 Wti,new
1 p1
δti 1qWti 1 , (27)

Wti,new 
wti Ait ζ βEt dit 1 p1
fti 1qUti 1 dit 1 fti 1 Wti,new
1 p1
dit 1qWti 1 . (28)

5.4 Matching Function

The probability of filling a vacancy qti is determined in equilibrium by a matching func-

 maxp0, Hti{qtiq{U0,ti denotes labor market tightness in labor-market
tion q pθti q, where θti
i. Labor market tightness is the ratio of the number of vacancies posted maxp0, Hti {qti q to
the number of job-seekers at the beginning of the period. Our assumption that workers
who separate from their jobs at t
1 get a chance to find a new job at the beginning of
t implies that the pool of job-seekers at the beginning of period t is all workers not se-
i
curely employed at t, U0,t . The matching function also determines the probability for an
unemployed worker of type i of finding a job: it is equal to the ratio of hires to job-seekers
f pθti q  maxp0, Hti q{U0,t
i
 θtiqpθtiq.
i
When firms hire in sector i, they do not endogenously lay off workers so N0,t  Mti
i
and the number of job-seekers is equal to U0,t  1
Mti. The job-finding rate in sector i
is therefore the following function of employment Nti and the inherited stock of employ-

23
ment Mti

Nti
Mti
fti  max 0, 1
M i . (29)
t

Appendix D.2 provides details on worker flow accounting in the model.

5.5 Wage-Setting

Downward nominal wage rigidity is defined relative to a flexible-wage benchmark de-

fined as the wage bargained by the firm and a securely attached worker, assuming Nash
bargaining prevails today and in all subsequent periods. Appendix D.3 shows that this
wage satisfies
 
i
wN ash,t  Ait γF 1 pN i
N ash,t q p1
γ qz βEt p1
δqp1
γ qfNi ash,t i,new
1 VN ash,t 1 , (30)

where γ P r0, 1s parameterizes the bargaining power of workers.

We assume that the nominal wage is set to the flexible wage—given by equation (30)—
except if this requires the nominal wage to fall. Expressed in terms of real wages, the
wage-setting equation is
" *
wti
1
wti  max i
wN ash,t , . (31)
Πt

5.6 Equilibrium

We again specify monetary policy as setting inflation to some target value Π. An equi-
librium is given by a process for employment Nti , labor market tightness θti , and the real
wage wti for each sector i P r0, 1s such that in all sectors i, firm i is on its labor demand
schedule—equation (23), (24), or (25) depending on the situation—the job-finding rate
satisfies equation (29), and wages are set according to equation (31).
We assume that sectoral productivity logpAit q in sector i is the sum of a time trend g, an
aggregate component logpAt q, and an idiosyncratic component logpZti q: logpAit q  gt
logpAt q logpZti q, where all these processes are independent: logpAt q K logpZti q, logpZti q K
logpZtj q for i  j. We assume the idiosyncratic component follows an AR(1) in growth
rates ∆ logpZti q  ρ∆z ∆ logpZti
1 q ε∆z,i
t , with Gaussian innovations: ε∆z,i
t  N p0, σε∆z q.
24
We assume the aggregate component follows an AR(2) in levels logpAt q  pI
ρa1 Lq
1 pI

ρa2 Lq
1 εat , with Gaussian innovations: εat  N p0, σεaq. Here, we are motivated by the DSGE
literature, which often includes equations that yield AR(2) dynamics–e.g., investment ad-
justment costs, habits in consumption, and lagged terms in the price and wage Phillips
curves combined with AR(1) shocks. This feature of our model generates high persis-
tence at business cycle frequencies without extreme levels of persistence at very low fre-
quencies, and also helps fit the fact that the dynamic responses of economic activity to
many shocks is hump-shaped (e.g. Romer and Romer, 2004; Christiano, Eichenbaum, and
Evans, 2005).22

6 Quantitative Analysis of New Model

Given the complexity of the model presented in section 5, we focus on an illustrative cal-
ibration as opposed to a full quantitative analysis. Appendix E describes the calibration
we rely on. Like in the previous section, we rely on global methods to numerically solve
for the equilibrium. Appendix D.5 discusses the algorithm we use in detail.
Table 3 presents our results for the new model in the rightmost column. The main
way in which the new model improves on the DMP model discussed earlier in the paper
is that it generates realistically slow unemployment cycles. Moreover, the model also gen-
erates a substantial amount of asymmetry in the speed and duration of recessions versus
expansions, which the workhorse DMP model also failed to match. The average duration
of expansions is 64.0 months in the new model, while the average duration of contrac-
tions is 36.6 months. As in the data, expansions are about twice as long as contractions.
The unemployment rate rises much more quickly than it falls in the new model: on av-
erage, it rises by 1.44 percentage points per year and falls by 0.81 percentage points per
year. In the data, these number are 1.89 and 0.87. Figure 7 plots a simulated path for the
unemployment rate Ut , as well as the for “broad” unemployment Ut0 (i.e., all workers not
in secure jobs). The greater persistence and asymmetric persistence yields a simulated

22
Fujita and Ramey (2007) explore an alternative mechanism for increasing the propagation of shocks in
the labor market. They assume that the cost of opening a vacancy is non-zero and increasing in the number
of new vacancies opened. This makes vacancy creation sluggish.
25
unemployment rate that resembles the behavior of the real-world unemployment.
The new model also generates slightly larger amounts of plucking than the DMP
model discussed earlier in the paper. The overall fit of the new model is therefore sub-
stantially better. It can both generate realistic plucking and persistence, while the earlier
DMP model cannot.
In appendix F, we show that it is the addition of short-term insecure jobs and AR(2)
shocks that generate persistence in the new model. Plucking in the new model is gener-
ated by the same forces as in our earlier DMP model. In particular, the new model with
DNWR generates a great deal of plucking with and without the addition of short-term in-
secure jobs and AR(2) shocks, and it generates very little plucking when we drop DNWR
independent of whether we include short-term insecure jobs and AR(2) shocks.
The analysis in appendix F also shows that the asymmetry in persistence between ex-
pansions and contractions requires the combination of DNWR, short-term insecure jobs,
and AR(2) shocks. With DNWR, a string of negative labor demand shocks eventually re-
sults in significant endogenous separations of securely attached workers—i.e., a burst of
separations. These bursts of separations are short-lived and modest in size. But they con-
tribute to speeding up unemployment contractions.23 These laid-off workers then cycle
through several short-term jobs before they eventually find new stable employment. This
process of cycling through short-term jobs contributes to preventing the unemployment
rate from quickly returning to its steady state level.
As Benigno and Ricci (2011) emphasize, sectoral shocks are an important source of
volatility in labor demand. The fact that sectoral shocks are non-stationary allows the
Nash wage to rise for many consecutive periods in a sector, to a high level that then
durably constrains wages when sectoral productivity falls for many consecutive periods.
Decreasing returns to labor make firms able to withstand such large and persistent shocks
without laying off all their workers, as we explain in Appendix F.1. These factors together
imply that the DNWR constraint continues to bind in the presence of sectoral shocks

23
These bursts of separations do not arise in a model with symmetric real wage rigidity. In a version
of the model with symmetric real wage rigidity shown in appendix F, the monthly rate of separation from
secure employment rises above 0.5% once every 43,300 months (3,600 years). With DNWR, this rate rises
above 0.5% once every 253 months (21 years).
26
despite the combined effect of inflation (2% per year) and growth (2.3% per year) that
“grease the wheels of the labor market”.

7 Conclusion

We build a plucking model of the business cycle that captures the asymmetry in the pre-
dictive power of contractions and expansions emphasized by Milton Friedman. We show
that a workhorse labor search model augmented with downward nominal wage rigidity
can fit these facts. In this model, eliminating business cycles lowers the average unem-
ployment rate. Since output is more often below than above the natural rate, standard
methods systematically underestimate the amount of slack in the economy.
While our benchmark model model with match heterogeneity and downward nom-
inal wage rigidity succeeds in fitting the plucking property, it fails to match the overall
duration of unemployment cycles, and the fact that expansions are on average twice as
long as contractions. We therefore introduce a second model with several non-standard
features—decreasing returns to labor, insecure jobs and AR(2) shocks—on top of down-
ward nominal wage rigidity. These new features interact to slow down unemployment
cycles, and explain why unemployment rises faster than it falls. However, they come at
the cost of a less realistic hiring and separation margin since we are no longer able to solve
the model with match-specific productivity shocks. A full integration of these features is
a promising topic for future research.

References
A BBRITTI , M. AND S. FAHR (2013): “Downward Wage Rigidity and Business Cycle Asym-
metries,” Journal of Monetary Economics, 60, 871–886.
A IYAR , S. AND S. V OIGTS (2019): “The Negative Mean Output Gap,” IMF Working Paper
No. 19/183.
A KERLOF, G., W. D ICKENS , AND G. P ERRY (1996): “The Macroeconomics of Low Infla-
tion,” Brookings Papers on Economic Activity.
A LTONJI , J. AND P. D EVEREUX (2000): “The Extent and Consequences of Downward
Nominal Wage Rigidity,” in Worker Well-Being, ed. by S. Polachek, New York: Elsevier,
383–431.
27
B ARATTIERI , A., S. B ASU , AND P. G OTTSCHALK (2014): “Some Evidence on the Impor-
tance of Sticky Wages,” American Economic Journal: Macroeconomics, 6, 70–101.
B ENIGNO , P. AND L. A. R ICCI (2011): “The Inflation-Output Trade-Off with Downward
Wage Rigidities,” American Economic Review, 101, 1436–1466.
B EWLEY, T. (1999): Why Wages Don’t Fall During a Recession, Harvard University Press.
B LANCHARD , O. AND J. G ALI (2010): “AssociationLabor Markets and Monetary Policy:
A New Keynesian Model with Unemployment,” American Economic Journal: A New
Keynesian Model with Unemployment, 2, 1–30.
B ORDO , M. AND J. H AUBRICH (2012): “Deep Recessions, Fast Recoveries, and Financial
Crises: Evidence from the American Record,” NBER Working Paper No. 18194.
B RÜGEMANN , B., P. G AUTIER , AND G. M ENZIO (2019): “Intra Firm Bargaining and Shap-
ley Values,” Review of Economic Studies, 86, 564–592.
B RY, G. AND C. B OSCHAN (1971): Cyclical Analysis of Time Series: Selected Procedures and
Computer Programs, NBER.
C ABALLERO , R. AND M. H AMMOUR (1998): “The Macroeconomics of Specificity,” Journal
of Political Economy, 106, 724–767.
C AHUC , P., F. M ARQUE , AND E. WASMER (2008): “A Theory of Wages and Labor De-
mand with Intra-Firm Bargaining and Matching Frictions,” International Economic Re-
view, 49, 943–972.
C ARD , D. AND D. H YSLOP (1997): “Does Inflation Grease the Wheels of the Labor Mar-
ket,” Reducing Inflation: Motivation and Strategy.
C HODOROW-R EICH , G. AND J. W IELAND (2018): “Secular Labor Reallocation and Busi-
ness Cycles,” Working Paper.
C HRISTIANO , L. J., M. E ICHENBAUM , AND C. L. E VANS (2005): “Nominal Rigidities and
the Dynamic Effects of a Shock to Monetary Policy,” Journal of Political Economy, 115,
1–45.
C OLE , H. AND R. R OGERSION (1999): “Can the Mortensen-Pissarides Matching Model
Match the Business-Cycle Facts,” International Economic Revies, 40, 933–959.
D ALY, M. AND B. H OBIJN (2014): “Downward Nominal Wage Rigidities Bend the Phillips
Curve,” Journal of Money, Credit and Banking, 46, 51–93.
D ALY, M. C., B. H OBIJN , AND B. L UCKING (2012): “Why Has Wage Growth Stayed
Strong?” Federal Reserve Bank of San Francisco Economic Letter 2012-11.
D ALY, M. C., B. H OBIJN , AND T. S. W ILES (2011): “Dissecting Aggregate Real Wage
Fluctuations: Individual Wage Growth and the Composition Effect,” Federal Reserve
Bank of San Francisco Working Paper 2011-23.
D E L ONG , J. B. AND L. H. S UMMERS (1988): “How does macroeconomic policy affect
output?” Brookings Papers on Economic Activity, 1988, 433–480.
E LSBY, M. (2009): “Evaluating the economic significance of downward nominal wage
rigidity,” Journal of Monetary Economics, 56, 154–169.
E LSBY, M., R. M ICHAELS , AND G. S OLON (2009): “The Ins and Outs of Cyclical Unem-

28
 ployment,” American Economic Journal: Macroeconomics, 1, 84–110.
E LSBY, M. W. L. AND R. M ICHAELS (2013): “Marginal Jobs, Heterogeneous Firms, and
Unemployment Flows,” American Economic Journal: Macroeconomics, 5, 1–48.
FATÁS , A. AND I. M IHOV (2015): “Recoveries,” Working Paper, INSEAD.
F ERRARO , D. (2017): “The Asymmetric Cyclical Behavior of the U.S. Labor Market,”
Working Paper, Arizona State University.
F OSTER , L., J. H ALTIWANGER , AND C. S YVERSON (2008): “Reallocation, Firm Turnover,
and Efficiency: Selection on Productivity or Profitability?” American Economic Review,
98, 394–425.
F RIEDMAN , M. (1964): “Monetary Studies of the National Bureau,” in The National Bureau
Enters Its 45th Year, New York, NY: National Bureau of Economic Research, 7–25, 44th
Annual Report, available at http://www.nber.org/nberhistory/annualreports.html.
——— (1993): “The "Plucking Model" of Business Fluctuations Revisited,” Economic In-
quiry, 31, 171–177.
F UJITA , S. AND G. R AMEY (2006): “The cyclicality of job loss and hiring,” Working Papers
06-17, Federal Reserve Bank of Philadelphia.
——— (2007): “Job Matching and Propagation,” Journal of Economic Dynamics & Control,
31, 3671–3698.
——— (2012): “Exogenous versus Endogenous Separation,” American Economic Journal:
Macroeconomics, 4, 68–93.
G ALINDEV, R. AND D. L KHAGVASUREN (2010): “Discretization of highly persistent cor-
related AR(1) shocks,” Journal of Economic Dynamics and Control, 34, 1260–1276.
G ERTLER , M., C. H UCKFELDT, AND A. T RIGARI (2016): “Unemployment Fluctuations,
Match Quality, and the Wage Cyclicality of New Hires,” NBER Working Paper No.
22341.
G ERTLER , M. AND A. T RIGARI (2009): “Unemployment Fluctuations with Staggered
Nash Wage Bargaining,” Journal of Political Economy, 117, 38–86.
G OTTSCHALK , P. (2005): “Downward Nominal-Wage Flexibility: Real or Measurement
Error?” Review of Economics and Statistics, 87, 556–568.
G RIGSBY, J., E. H URST, AND A. Y ILDIRMAZ (2018): “Aggregate Nominal Wage Adjust-
ment: New Evidence from Administrative Payroll Data,” Working Paper, University of
Chicago.
H AGEDORN , M. AND I. M ANOVSKII (2008): “The Cyclical Behavior of Equilibrium Un-
employment and Vacancies Revisited,” American Economic Review, 98, 1692–1706.
H AIRAULT, O., F. L ANGOT, AND S. O SOTIMEHIN (2010): “The (un)importance of un-
employment fluctuations for the welfare cost of business cycles,” Review of Economic
Dynamics, 13, 759–779.
H ALL , R. (2005): “Employment Fluctuations with Equilibrium Wage Stickiness,” Ameri-
can Economic Review, 95, 50–65.
H ALL , R. E. (1995): “Lost Jobs,” Brookings Papers on Economic Activity, 1995, 221–273.

29
H ALL , R. E. AND M. K UDLYAK (2021): “Why Has the US Economy Recovered So Con-
sistently from Every Recession in the Past 70 Years?” NBER Macroeconomics Annual, 36,
forthcoming.
H ALL , R. E. AND P. R. M ILGROM (2008): “The Limited Influence of Unemployment on
the Wage Bargain,” American Economic Review, 98, 1653–1674.
H AZELL , J. AND B. TASKA (2018): “Posted Wage Rigidity,” Working Paper, MIT.
J ACKSON , M. AND P. T EBALDI (2017): “A Forest Fire Theory of the Duration of a Boom
and the Size of a Subsequent Bust,” Working Paper.
J AROSCH , G. (2021): “Searching for Job Security and the Consequences of Job Loss,”
Working Paper, Princeton University.
J ORGENSON , D., M. H O , AND J. S AMUELS (2012): “A Prototype Industry-Level Produc-
tion Account for the United States, 1947-2010,” Second World KLEMS conference, Har-
vard University.
J UNG , P. AND K. K UESTER (2011): “The (un)importance of unemployment fluctuations
for the welfare cost of business cycles,” Journal of Economic Dynamics & Control, 35,
1744–1768.
J UNG , P. AND M. K UHN (2019): “Earings Losses and Labor Mobility over the Life Cycle,”
Journal of the European Economic Association, 17, 678–724.
K AHN , S. (1997): “Evidence of Nominal Wage Stickiness from Microdata,” American Eco-
nomic Review, 87, 993–1008.
K IM , C.-J. AND C. R. N ELSON (1999): “Friedman’s Plucking Model of Business Fluctua-
tions: Tests and Estimates of Permanent and Transitory Components,” Journal of Money,
Credit and Banking, 31, 317–334.
K IM , J. AND F. R UGE -M URCIA (2009): “How much inflation is necessary to grease the
wheels?” Journal of Monetary Economics, 56, 365–377.
K ROLIKOWSKI , P. (2017): “Job Ladders and Earnings of Displaced Workers,” American
Economic Journal: Macroeconomics, 9, 1–31.
K URMANN , A. AND E. M C E NTARFER (2017): “Downward Wage Rigidity in the United
States: New Evidence from Administrative Data,” Working Paper.
L EPETIT, A. (2018): “Asymmetric Unemployment Fluctuations and Monetary Policy
Trade-offs,” Working Paper.
L UCAS , R. (1987): Models of Business Cycles, Oxford:Blackwell.
——— (2003): “Macroeconomic Priorities,” American Economic Association, 93, 1–14.
M C K AY, A. AND R. R EIS (2008): “The Brevity and Violence of Contractions and Expan-
sions,” Journal of Monetary Economics, 55, 738–751.
M C L AUGHLIN , K. (1994): “Rigid Wages?” Journal of Monetary Economics, 34, 383–414.
M ICHAILLAT, P. (2012): “Do Matching Frictions Explain Unemployment? Not in Bad
Timess,” American Economic Review, 102, 1721–1750.
M ITCHELL , W. C. (1927): Business Cycles: The Problem and Its Setting, National Bureau of
Economic Research.
30
N EFTÇI , S. (1984): “Are Economic Time Series Asymmetric over the Business Cycle?”
Journal of Political Economy, 92, 307–328.
O LIVEI , G. AND S. T ENREYRO (2010): “Wage-setting patterns and monetary policy: In-
ternational evidence,” Journal of Monetary Economics, 57, 785–802.
P ETRONGOLO , B. AND C. P ISSARIDES (2001): “Looking into the Black Box: A survey of
the Matching Function,” Journal of Economic Literature, 39, 390–431.
P ETROSKY-N ADEAU , N. AND L. Z HANG (2017): “Solving the Diamond-Mortensen-
Pissarides model accurately,” Quantitative Economics, 8, 611–650.
P ETROSKY-N ADEAU , N., L. Z HANG , AND L. K UEHN (2018): “Endogenous Disasters,”
American Economic Review, 108, 2212–2245.
P ISSARIDES , C. (2009): “The Unemployment Volatility Puzzle: Is Wage Stickiness the
Answer?” Econometrica, 77, 1339–1369.
P RIES , M. (2004): “Persistence of Employment Fluctuations =: A Model of Recurring Job
Loss,” The Review of Economic Studies, 71, 193–215.
R OMER , C. D. AND D. H. R OMER (2004): “A New Measure of Monetary Shocks: Deriva-
tion and Implications,” American Economic Review, 94, 1055–1084.
R OUWENHORST, K. G. (1995): “Asset Pricing Implications of Equilibrium Business Cycle
Models,” in Frontiers of Business Cycle Research, ed. by T. F. Cooley, Princeton University
Press, chap. 10, 294–330.
S CHMITT-G ROHE , S. AND M. U RIBE (2016): “Downward Nominal Wage Rigidity, Cur-
rency Pegs, and Involuntary Unemployment,” Journal of Political Economy, 124, 1466–
1514.
S HIMER , R. (2005): “The Cyclical Behavior of Equilibrium Unemployment and Vacan-
cies,” American Economic Review, 95, 25–49.
——— (2010): Labor Market and Business Cycles, Princeton University Press.
——— (2012): “Reassessing the ins and outs of unemployment,” Review of Economic Dy-
namics, 15, 127–148.
S ICHEL , D. (1993): “Business Cycle Asymmetry: A Deeper Look,” Economic Inquiry, 31,
224–236.
S ILVA , I. J. AND M. T OLEDO (2009): “Labor Turnover Costs and the Cyclical Behavior of
Vacancies and Unemployment,” Macroeconomic Dynamics, 13, 76–96.
S INCLAIR , T. (2010): “Asymmetry in the Business Cycle: Friedman’s Plucking Model
with Correlated Innovations,” Studies in Nonlinear Dynamics and Econometrics, 14, 1–31.
S TOLE , L. AND J. Z WIEBEL (1996): “Intra-firm Bargaining under Non-binding Contracts,”
Review of Economic Studies, 63, 375–410.
TASCI , M. AND N. Z EVANOVE (2019): “Do Longer Expansions Lead to More Severe Re-
cessions?” Federal Reserve Bank of Cleveland Economic Commentary.
TAUCHEN , G. (1986): “Finite State Markov-Chain Approximations to Univariate and Vec-
tor Autoregressions,” Economic Letters, 20, 177–181.
W OODFORD , M. (2003): Interest and Prices, Foundations of a Theory of Monetary Policy,
31
Princeton University Press.

32
 10

8
Unemployment (%)

1950 1960 1970 1980 1990 2000 2010 2020

Figure 1: Peaks and Troughs in the Unemployment Rate

Note: The unemployment rate is plotted in blue. Business cycle peaks are denoted by dashed red
vertical lines, while business cycle troughs are denoted by solid red vertical lines.

33
 Amplitude of the subsequent contraction
Amplitude of the subsequent expansion

7 7
2009

6 1982
6
2006
1949
1979
5 5
1973
4 1992 4 1957
1961 1953
1975
3 3 1989
1958 1968
1954 2000
1960
2 2003 2
1970

1 1

0 0
0 2 4 6 0 2 4 6
Amplitude of a contraction Amplitude of an expansion

Figure 2: The Plucking Property of the Unemployment Rate

Note: The points in the left panel are labeled with the year the contraction in question ended and
expansion in question began. The points in the right panel are labeled with the year the expansion
in question ended and contraction in question began. OLS regression lines are plotted in each
panel.

34
18 18
10

42

16 16
8
40

14 14

6 38

12 12

36
4
10 10

Share of zero wage-change (%) Share of zero wage-change (%) 34

Unemployment Rate (%) One minus employment rate (%)
2
2000 2005 2010 2015 2020 2000 2005 2010 2015 2020

Figure 3: San Francisco Fed Wage Rigidity Meter

Note: The figure plots the share of wage freezes of all job-stayers (paid by the hour or not) with
respect to the wage one year prior, with no correction for measurement errors. This series is con-
structed by the Federal Reserve Bank of San Francisco using data used from the Current Population
Survey.

35
 Amplitude of the subsequent contraction
Amplitude of the subsequent expansion
10 10

8 8

6 6

4 4

2 2
2 4 6 8 10 2 4 6 8 10
Amplitude of a contraction Amplitude of an expansion

(a) Fujita-Ramey Model under Flexible Wages (Nash Bargaining)

Amplitude of the subsequent contraction

Amplitude of the subsequent expansion

10 10

8 8

6 6

4 4

2 2
2 4 6 8 10 2 4 6 8 10
Amplitude of a contraction Amplitude of an expansion

(b) Fujita-Ramey Model under DNWR

Amplitude of the subsequent contraction
Amplitude of the subsequent expansion

10 10

8 8

6 6

4 4

2 2
2 4 6 8 10 2 4 6 8 10
Amplitude of a contraction Amplitude of an expansion

(c) New Model under DNWR

Figure 4: Plucking Scatter Plots

Note: The figure displays the scatter plots associated to the plucking regressions for the Fujita-
Ramey model under flexible wages (panel a), the Fujita-Ramey model under DNWR (panel b) and
the new model of section 5 under DNWR. The plots feature all the expansion/contraction pairs
obtained by pooling together 500 samples of 866 months. OLS regression lines are plotted in each
panel. 36
 Unemployment
15
Sample path
Steady-state level
10
(%)

0
100 200 300 400 500 600 700 800

Job-Finding Rate
60

40
(%)

20

0
100 200 300 400 500 600 700 800

Employment Exit Rate

6

4
(%)

0
100 200 300 400 500 600 700 800
Time (months)

Figure 5: Simulated Paths for the Unemployment Rate, Job-Finding Rate and Employ-
ment Exit Rate
Note: The figure plots sample path of 72 years (the same length as our empirical sample) for the
unemployment rate u, the job-finding rate f and employment exit rate s̄ in the Fujita-Ramey model
with downward nominal wage rigidity of section 3. The dashed lines indicate the steady-state level
of each variable.

37
 45

40

35
Unemployment u (%)

30

25

20

15

10

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Job-Finding Rate f

Figure 6: Steady-State Worker-Flow Relationship

Note: The figure plots the steady-state relationship between the job-finding rate and the unemploy-
ment rate implied by the worker-flow equation (15)-(16). In plain line the relationship is plotted
over the range of unemployment rates observed in the US between 1948 and 2019: from 2.5% to
10.8%.

38
 15

10
Unemployment (%)

Unemployment
Broad unemployment
Steady-state unemployment
0
100 200 300 400 500 600 700 800
Time (months)

Figure 7: Simulated Paths for the Unemployment Rate: New Model

Note: The figure plots sample paths of 72 years (the same length as our empirical sample) for the
unemployment rate and broad unemployment rate in our model with downward nominal wage
rigidity, secure and insecure jobs, multiple sectors and an AR(2) aggregate productivity process.
The solid vertical lines identify business-cycle troughs, while the dashed vertical lines identify
peaks.

39
Table 1: Plucking Property of Unemployment and Speed Asymmetries
β R2
Subsequent expansion on contraction 1.12 0.59
(0.33)
Subsequent contraction on expansion -0.38 0.22
(0.27)

Speed of expansions (pp/year) 0.87

Speed of contractions (pp/year) 1.89
P-value for equal speed 0.002

Duration of expansions (months) 59.1

Duration of contractions (months) 26.9

Note: The first row reports the coefficient in an OLS regression of the size of the
subsequent expansion (percentage point fall in unemployment rate) on the size
of a contraction (percentage point increase in unemployment rate). The second
row reports the coefficent in an analogous regression of the size of the subsequent
contraction on the size of an expansion. The speed of expansions and contractions
in the third and fourth rows is measured in percentage points of unemployment
per year.

40
 Table 2: Calibration
β 0.961{12
η 0.5
c 0.30
z 0.95
δ 1.9%
ρa 0.98
ρx 0.98
Nash DNWR
Π – 1.021{12
γ 0.57 0.43
σεa st. σ  1.6% st. σ  1.5%
a a

σεx st. σ x  2.1% st. σ x  1.5%

Note: The abbreviation “st.” stands for “such that.”

41
 Table 3: Simulation Results: Plucking Property, Speed, and Duration
Data Fujita-Ramey Model New Model
Nash DNWR DNWR
Subsequent expansion 1.12 0.44 0.63 0.68
on contraction, β (0.40) (0.26) (0.42)
Subsequent contraction -0.38 -0.06 -0.11 -0.08
on expansion, β (0.49) (0.32) (0.67)
Subsequent expansion 0.59 0.19 0.45 0.53
on contraction, R2 (0.22) (0.25) (0.31)
Subsequent contraction 0.22 0.07 0.04 0.13
on expansion, R2 (0.19) (0.11) (0.29)
Speed of expansions 0.87 1.56 2.53 0.81
(pp / year) (0.57) (0.74) (0.22)
Speed of contractions 1.89 1.66 4.93 1.44
(pp / year) (0.69) (1.49) (0.35)
Duration of expansions 59.1 36.2 27.9 64.0
(months) (12.4) (7.0) (16.0)
Duration of contractions 26.9 36.6 18.1 36.6
(months) (11.7) (6.1) (12.6)
Note: The table compares data from the Fujita-Ramey model of section 3, both with Nash bargain-
ing and with downward nominal wage rigidity (DNWR), and from the new model of section 5.
The first (third) row reports the coefficient (R2 ) in an OLS regression of the size of an expansion
(percentage point fall in unemployment rate) on the size of the previous contraction (percentage
point increase in unemployment rate). The second (fourth) row report the coefficient (R2 ) in an
analogous regression of the size of a contraction on the size of the previous expansion. The next
two rows report the average speed of expansion and contractions, measured in percentage points
of unemployment per year. The final two rows report the average duration of expansions and con-
tractions, measured in months. For the models, the reported point estimate is the median value of
the statistic over 5000 samples of 866 periods each (the length of our sample of real-world data).
Expansions and contractions of more than 6.5 percentage points are excluded from the samples.
Appendix ?? provides results including all expansions and contractions. The standard error re-
ported in parentheses is the standard deviation of the estimates across the 5000 samples.

42
 Table 4: Second Order Moments
Xt ut ft s̄t
Data
σ pX q 0.096 0.077 0.058
corrpAt , Xt q -0.460 0.369 -0.535
corrpXt , Xt
1 q 0.926 0.803 0.631
Nash
σ pX q 0.115 0.063 0.081
corrpAt , Xt q -0.964 0.965 -0.877
corrpXt , Xt
1 q 0.814 0.767 0.510
DNWR
σ pX q 0.131 0.072 0.116
corrpAt , Xt q -0.80 0.779 -0.525
corrpXt , Xt
1 q 0.654 0.555 0.191
Note: The table reports the second order moments of the Fujita-Ramey model of section 3 under
Nash bargaining and Downward Nominal Wage Rigidity. The moments are reported after data
has been turned quarterly, logged, and HP-filtered. The data panel uses the data from Fujita and
Ramey (2012) and is therefore identical to the results they report. For the model panels, the table
reports medians of each statistics over our 5000 samples of 866 months each.

43
A Defining Expansions and Contractions

A.1 Defining Expansions and Contractions

Since our empirical analysis is based on the amplitude and speed of cyclical movements
in unemployment, we define business cycle peaks in troughs in such a way that they line
up exactly with peaks and troughs of the unemployment rate. This yields business cycle
dates that are very similar to but not identical to those identified by the NBER Business
Cycle Dating Committee (because the NBER Business Cycle Dating Committee uses a
wide variety of cyclical indicators beyond unemployment to date turning points).
We develop a simple algorithm that defines business cycle peak and trough dates for
the unemployment rate. The basic idea is to find local minima and maxima of the un-
employment rate. However, we ignore small “blips” or “wiggles” in the unemployment
rate and focus instead on delineating substantial swings in the unemployment rate in a
similar manner as the peaks and troughs identified by the NBER Business Cycle Dating
Committee.
Table A.1 presents the peak and trough dates we identify and compares them with the
peak and trough dates identified by the NBER.

A.2 An Algorithm for Defining Expansions and Contractions

Let ut denote the unemployment rate at time t. The algorithm begins by taking the first
month of our sample as a candidate for a business cycle peak, cp. If, in all the following
months until unemployment becomes X percentage points higher than ucp , unemploy-
ment is higher than ucp , we confirm that cp is a business cycle peak. If, instead, the un-
employment rate falls below ucp before it is confirmed as a peak, the month in which this
happens becomes the new candidate peak. Once we have identified a peak, we switch to
looking for a trough (in the analogous manner) and so on until we reach the end of the
sample. Formally, starting with t  1 the algorithm is:

1. Set cp  t and set t  t 1 (i.e., move to the next time period).

2. If ut ucp go back to step 1

44
 3. If ucp ¤ ut ucp X set t  t 1 and go back to step 2

4. If ut ¥ ucp X add cp to the set of peaks

5. Set ct  t and set t  t 1

6. If ut ¡ uct go back to step 5

7. If uct ¥ ut ¡ uct
X set t  t 1 and go back to step 6

8. If ut ¤ uct
X add ct to the set of troughs, and go back to step 1

We set X  1.5 percentage points. With this value, our algorithm generates the same
set of expansions and contractions as the NBER Business Cycle Dating Committee with
one exception: Our algorithm considers the 1979-1982 double-dip recession as a single
contraction as opposed to two contractions interrupted by a brief and small expansion
(unemployment decreased by 0.6 percentage points in 1980-1981). Values for X between
0.8 and 1.5 percentage points identify exactly the same cycles. Values of X larger than 1.5
drop the 1970-1973 expansion.
An advantage of our algorithm is that it does not impose a duration upon expansions
and contractions but only a size X, in contrast to other algorithms based on turning points
like the Bry and Boschan (1971) routine. Our algorithm can therefore also be used to
define expansions and contractions in our model simulations, even for models that do
not match the duration of expansions and contractions in the real-world data.

A.3 Peak and Trough Dates from 1948 to 2020

Table A.1 presents the peak and trough dates we identify. For comparison purposes, we
also present the peak and trough dates identified by the NBER. We identify the same
set of expansions and contractions as the NBER Business Cycle Dating Committee with
one exception: we consider the 1979-1982 double-dip recession as a single contraction as
opposed to two contractions interrupted by a brief and small expansion (unemployment
decreased by 0.6 percentage points in 1980-1981). The exact timing of the NBER peaks
and troughs do not line up exactly with ours for the reasons discussed above. However,
45
 Table A.1: Business Cycle Peaks and Troughs
Unemployment NBER
Peak Trough Peak Trough
1 [1/1948] 10/1949 11/1948 10/1949
2 5/1953 9/1954 7/1953 5/1954
3 3/1957 7/1958 8/1957 4/1958
4 2/1960 5/1961 4/1960 2/1961
5 9/1968 12/1970 12/1969 11/1970
6 10/1973 5/1975 11/1973 3/1975
7a 5/1979 1/1980 7/1980
7b 11/1982 7/1981 11/1982
8 3/1989 6/1992 7/1990 3/1991
9 4/2000 6/2003 3/2001 11/2001
10 10/2006 10/2009 12/2007 6/2009
11 9/2019 2/2020
Note: Business cycle peaks and troughs defined solely based on the unemployment
rate and, for comparison, business cycle peaks and troughs as defined by the Busi-
ness Cycle Dating Committee of the National Bureau of Economic Research.

in most cases, our dates are quite similar to theirs. The NBER peaks tend to lag our peaks
by a few months and the NBER troughs tend to precede our troughs by a few months.
This implies that our estimate of the average duration of contractions is about one year
longer than what results from the NBER’s dating procedure. We identify September 2019
as a peak as opposed to February 2020 because the unemployment rate first hit 3.5%
in September 2019. When several months are tied for the lowest unemployment rate at
the end of an expansion, our algorithm picks the first of these months as the peak (and
similarly for troughs). Table A.2 lists the duration of all expansions and contractions over
our sample period.

B Solution Method

B.1 Normalization of µ

Recall that the matching function is Cobb-Douglas. The vacancy-filling rate is therefore
qt  µθ
η . Furthermore, the job finding rate is ft  θtqpθtq. Combining these equations

46
 Table A.2: The Duration of Expansions and Contractions
Dates Length in Months
Peak Trough Expansion Contraction
1 [1/1948] 10/1949 21
2 5/1953 9/1954 43 16
3 3/1957 7/1958 30 16
4 2/1960 5/1961 19 15
5 9/1968 12/1970 88 27
6 10/1973 5/1975 34 19
7 5/1979 11/1982 48 42
8 3/1989 6/1992 76 39
9 4/2000 6/2003 94 38
10 10/2006 10/2009 40 36
11 9/2019 119
Mean 59.1 26.9

allows us to express the vacancy-filling rate as a function of the job-finding rate:

 µ
pftq

1 η
qt 1 η 1 η . (B.1)

We can now see that there is a one-to-one mapping between the cost of hiring a worker
Ct  c{qt and the job-finding rate ft:


1
 c
pftq

η
Ct cµ 1
η 1 η . (B.2)
qt

This mapping can be used to write the equilibrium conditions of the model in terms of
either the cost of hiring a worker or the job-finding rate, without reference to the other
(and without reference to labor market tightness). When the model is written in this
way (e.g., in terms of the cost of hiring a worker), the parameters c and µ only enter the

1
model though the composite term cµ 1
η . This implies that we can normalize either c or
µ without loss of generality. We choose to normalize µ  1. Intuitively, only the cost of
hiring a worker matters to a firm. It is immaterial to the firm whether this cost consists
of posting few vacancies that fill with a high probability but are expensive to post, or of
posting many vacancies that fill with a low probability but are inexpensive to post.

47
B.2 Solving for the Policy Functions

To solve for the policy function under Nash-bargaining, we follow the solution method
of Fujita and Ramey (2012) to solve for the functions J N ash pA, xq and q pAq. The state-
space consists of the two exogenous states A and x. We discretize the AR(1) process for
At using the Rouwenhorst (1995) method with 11 grid points, and the AR(1) process for
xt using the Tauchen (1986) method with 201 grid points. Combining equations (10) and
(11), we can solve for the functions J N ash and 1{q by iteration on the policy functions.
Specifically, given guesses on the functions J and 1{q (and therefore f ), we use these
guesses to calculate the expected terms on the RHS of equation (11) and update J N ash .
We then update 1{q using equation (10). We iterate until convergence.
Under DNWR, we first solve for the Nash wage as a function of the state pA, xq by
solving the model under Nash-bargaining. This gives the Nash wage under the assump-
tion that wages will be flexible at all future dates, including wages in new matches that
are relevant to determine the outside option of workers. Under DNWR, the value func-
tion J depends on the two exogenous states x and A, and the new endogenous state of the
lagged real wage w
1 , J pA, x, w
1 q. Under the assumption on the Nash-bargained wage,
new
J is independent of the state w
1 , so that for numerical considerations, the state-space is

only three-dimensional.
The recursion on J is the same as (4), up to the new dependence of the value function
on the new state w
1 :

J pA, x, w
1 q  maxtJ c pA, x, w
1 q, 0u, (B.3)

where J c is the value if the match is continued, which solves the recursion

J c pA, x, w
1 q  xA
wpA, x, w
1 q β p1
δ qE pJ pA1 , x1 , wqq , (B.4)

where the real wage w is given by equation (13). Equations (B.3)-(B.4) allow to solve for
J by iteration. We again use 11 points on the A dimension, 201 on the x dimension, and
401 grid points on the new endogenous dimension w
1 . When iterating on equation (B.4),
calculating the expected term on the RHS requires to evaluate the value function J at
values of the endogenous state w that are not on the grid. We rely on linear interpolation
to do so.
48
 Once J is solved for, we can obtain 1{q pA, w
1 q from J and the free-entry condition
new

new
(10) which now depends on the new state w
1 ,

J pA, xhire , w
1 q 
new c
q pA, w
1 q
new
. (B.5)

B.3 Calculating Separation

Because matches can be endogenously terminated, the destruction rate st depends on

the cross-sectional distribution of employment across matches’ states. Calculating the
destruction rate in simulations of the model therefore requires us to keep track of the dis-
tribution of employment across matches’states. Under Nash-bargaining, matches’ states
reduce to match productivity xt . We follow the method in Fujita and Ramey (2012) to
keep track of the distribution of employment across xt and calculate the destruction rate,
only adapting it to any Markovian process for x so that it can accommodate our AR(1)
assumption (2) on xt .
Define nt pxq the number of employed workers at productivity x, and nt the vector of
nt pxq. (Note that our nt is the density of the distribution of employment, while Fujita-
Ramey’s et on p.75-77 is the CDF.) We therefore have:
¸
ut 1
nt pxq. (B.6)
x

Define n0t pxq the number of workers employed at productivity x at the beginning of period
t, after shocks and exogenous separation have occurred, but before endogenous separa-
tion has occurred. Denote n0t the vector of n0t pxq. We have:

n0t  p1
δqpT xq1nt
1 (B.7)

where T x is the transition matrix of the Markovian process of x. Define f iredt the number
of workers fired at t. It solves:
¸
f iredt  n0t pxq1Jt pxq0 . (B.8)
x

The job-destruction rate st is given by:

st δ f iredt
1
ut
1
. (B.9)

49
The new distribution of employment at t solves the recursion:

nt pxq  n0t pxq1Jt pxq¡0 ft pst p1
stqut
1q1xx hire . (B.10)

Under DNWR, calculating the destruction rate requires to keep track of the distri-
bution of employment along both match productivity x and wages w. We do so in the
following way. Let mt
1 px , w q P pxt
1  x, wΠ
¤ wq be the number of matches at
t 1

t
1 with idiosyncratic productivity xt
1  x and a real wage less than Πw . Considering
the number of real wages below Πw instead of below w is for convenience: This way
it gives the number of matches with real wages below w at the beginning of period t,
after inflation from t
1 to t has eroded lagged real wages. Note that mt
1 px , 8q is the
number of firms with idiosyncratic productivity xt
1  x at t
1, and °x mt
1px, 8q is
employment at t
1.
Denote m0t px , w q  P pxt  x, wΠ
¤ wq the number of matches with idiosyncratic
t 1

productivity xt  x and inherited real wage less than w at the beginning of period t,
after match-specific productivity shocks and exogenous separation shocks have hit but
before any wage-adjustment. It is given by

m0t  p1
δqTx1 mt
1 (B.11)

where Tx is again the transition matrix of the Markovian process of x.

We now keep track of how wage adjustments change the distribution of wages under
DNWR. Denote m1t px , w q the number of wages with xt  x and inherited real wage
less than w after wage adjustments. It is the same as m0t , except that all wages below
wtN ash px q are reset to wtN ash px q, i.e.

m1t px , w q  0 for all w ¤ wtN ashpxq. (B.12)

We now calculate the number of endogenously terminated matches, and keep track of
how it affects the distribution of wages. Denote wthresh px q the threshold on wages above
which matches with productivity x are terminated. It is defined as the lowest wage w
such that J px , wq 0. The number f iredtpxq of matches with productivity x that are
terminated is m1t px , 8q
m1t px , wthresh px qq. Knowing the number of exogenously and
50
endogeneously separated matches we can calculate the separation rate as:
°
f iredt pxq
st  x
Nt
1
δ. (B.13)

Denote m2t px , w q the number of wages with xt  x and inherited real wage less than
w of wages after endogenous separation. It is the same as m1t , except that it no longer
includes wages above wthresh px q, i.e.

m2t px , w q  m1t px , wthresh px qq for all w ¥ wthreshpxq. (B.14)

We now keep track of how new hires affect the distribution of wages. Denote
m3t px , w q the number of wages with xt  x and inherited real wage less than w after
hiring. It is the same as m2t , except that it adds the number of new hires at productivity
xhire and hiring wage wthire , i.e.

m3t pxhire , w q  m2t pxhire , w q ft p1
p1
st qNt
1 qfor allw ¥ wthirepxhireq. (B.15)

This is the distribution of effective real wage in period t. Employment at t is therefore

°
given by x m3t px , 8q.
Finally, we keep track of the eroding effect of inflation and growth from t to t 1 to
get mt px , w q and be able to start the whole process in period t 1. We have that

mt px , w q  m3t px , Π  w q. (B.16)

We calculate it by linear interpolation.

C The Volatility of Aggregate Shocks and the Average

Level of Unemployment

Figure C.1 plots the average level of the unemployment rate in our plucking model as a
function of the volatility of aggregate shocks. Both models have the property that average
unemployment increases with the volatility of the aggregate shocks, from a steady-state
level of 2.9% in the model of section 3 and 4.2% in the model of section 5. Average unem-
ployment increases less steeply with the volatility of shocks in the new model of section
5 because decreasing returns to scale make firms able to withstand larger shocks under
DNWR without being willing to lay off all their workers, as explained in Appendix F.1.
51
 15 8

10
5
E(u) (%)

E(u) (%)
4

3
5

0 0
0 0.5 1 1.5 2 0 1 2 3 4 5 6 7 8

(a) Fujita Ramey Model of Section 3 (b) New Model of Section section 5

Figure C.1: Average Unemployment and the Volatility of Aggregate Shocks

Note: The figure gives the average rate of unemployment as a function on the standard deviation
of aggregate shocks in the DNWR model of section 3 (panel a) and the the DNWR model of section
5 (panel b).

D Appendix for the Model in Section 5

D.1 Labor Demand

The employment decision of the firm is

Ait c i
Ωit pMti q  max tAitF pMti Hti q
wti pMti Hti q
βEt pΩit pMti 1qu,
qti t Ht ¥0
i
H1 i 1 (D.1)
Ht

subject to the law-of-motion of its workforce Mti 1 pNtiq given by equations (17)-(18).
The envelope condition is

Jti pMti q  Ait F 1 pNti q
wti β p1
δ qEt Jti 1 pMti 1 pNtiqq . (D.2)

where we define Jti pMti q  BΩit

BMti pMt q, the equilibrium marginal value to the firm of a
i

securely attached worker inherited from the previous period.

The first-order conditions give:

• If Hti ¡ 0 (i.e., if the firm hires):

Ait c 
qti
 AitF 1pNtiq
wti β p1
dqEt Jti 1 pMti 1 pNtiqq . (D.3)


This happens when Ait F 1 pMti q
wti β p1
dqEt Jti 1 pMti pMtiqq ¡ Aq c . i
t
1 i
t

52
 • If Hti 0 (i.e., if the firm fires):

0  Ait F 1 pNti q
wti β p1
δ q Jt 1 pMti 1pNtiqq . (D.4)

This happens when Ait F 1 pMti q
wti β p1
δ qEt Jti 1 pMti 1 pMtiqq 0.

• When neither condition is satisfied, i.e. when

 i
Ait F 1 pMti q
wti β p1
dqEt Jti 1 pMti 1 pMt qq ¤
i At c
, (D.5)
qi t

Ait F 1 pMti q
wti β p1
δ qEt Jti 1 pMti 1 pMtiqq ¥ 0, (D.6)

then the firm freezes hires Hti  0.

To simplify and interpret this labor demand schedule, equation (B.3) defines the
marginal value to the firm of a worker already attached to the firm and equation (22)
defines the marginal value to the firm of a newly-hired worker. The labor-demand sched-
ule can then be rewritten as in equations (23)-(25).
In equilibrium, because all firms in sector i are identical, if firms are freezing employ-
ment then Ait c{qti  0. Therefore, the aggregate labor demand schedule in sector i is:
• If J i,new pMti q ¡ 0, then firms hire up to:

Ait c
J i,new pNti q 
q pNti q
. (D.7)

• If J i pMti q 0 then firms fire up to:

J i pNti q  0. (D.8)

• If J i pMti q ¥ 0 ¥ J i,new pMti q firms freeze employment:

Nti  Mti. (D.9)

D.2 Worker Flow Accounting

From economy-wide worker flows, we can define the economy-wide rate of job destruc-
tion st , the economy-wide rate of inflow from employment into unemployment s̄t , and
53
the economy-wide rate of outflow from unemployment to employment ft (equal to the
i
economy-wide job-finding rate). In sector i at time t, there are N0,t  minpNti, Mtiq workers
who work in the same job as at time t
1. The number of workers who find a new job
at the beginning of time t in sector i is therefore Nti
N0,ti . In the overall economy it is
³
Nt
N0,t , where the aggregates Nt and N0,t are defined as sums across sectors, Nt  i
Nti di
³
and N0,t  i
i
N0,t di. The economy-wide job-finding rate is therefore

ft  N1t

NN0,t . (D.10)
0,t

The number of workers who separate from their jobs between t
1 and t in sector i is
Nti
1
N0,t
i
. In the total economy it is Nt
1
N0,t . The economy-wide job destruction rate
is therefore:

st  Nt
N1
N0,t . (D.11)

t 1

Our assumption that a worker who separates from his job at the end of period t
1
has a chance ft of finding a new job at the beginning of period t implies that the job-
destruction rate st is not equal to the rate of inflow from employment into unemployment
s̄t , or employment exit rate. Among the workers who separates from their jobs at the end
of t
1, the fraction ft that starts a new job at the beginning of t does not transition from
employment to unemployment but from job to job. Only the fraction 1
ft transitions to
unemployment. The economy-wide rate of inflow from employment into unemployment
is therefore:

s̄t  p1
ftqst. (D.12)

From combining equations (D.10), (D.11), and (D.12), the law of motion of economy-
wide unemployment Ut  1
Nt is still given by equation (16).

D.3 The Nash-Bargained Wage

We solve for the Nash-bargaining equilibrium in order to obtain the Nash wage. We
take the Nash wage to be the one that prevails when wages are Nash-bargained between
the firm and securely attached workers, both today and in all subsequent periods. For
54
workers, we get from equations (26)-(28) that the values of having a secure job relative
to being unemployed Vti  Wti
Uti and the value of having a short-term job relative to
being unemployed Vti,new  Wti,new
Uti solve:
 
Vti,new 
wti Ait z βEt p1
dit 1 q Vti 1
fti i,new
1 Vt 1 , (D.13)
 
Vti 
wti Ait z βEt p1
δti 1 q Vti 1
fti i,new
1 Vt 1 , (D.14)

where z  b ζ is the flow value of unemployment in terms of both unemployment

benefits and more leisure.
For firms, we get from equations (B.3)-(22) that:

Jti,new  AitF 1pNtiq
wti βEt pp1
dqJti q, 1 (D.15)

Jti  Ait F 1 pNti q
wti βEt pp1
δ qJti 1 q. (D.16)

Summing up to get the total surpluses of a secure job Sti  Jti Vti and of a temporary
job Sti,new  Jti,new Vti,new :24

Sti,new  AitpF 1pNtiq
zq βEt pp1
dqpSti
fti 1Vti,new

11 qq, (D.17)

Sti  Ait pF 1 pNti q
z q βEt pp1
δ qpSti 1
fti 1 Vti,new

1 qq. (D.18)

Nash-bargaining between the firm and long-term workers implies that Jti  p1
γ qSti
and Jti 1  p1
γ qSti 1. Combined with equation (D.18) this implies:

JNi ash,t  p1
γ q Ait pF 1pN i N ash,t q
zq βEt p1
δqJNi ash,t 1
p1
γ qp1
δqfNi ash,t i,new
1 VN ash,t 1 q .

(D.19)

Combining equations (D.16) and (D.19) to eliminate JNi ash,t gives the expression of the
wage given in equation (30).
Injection of the expression for the Nash wage—equation (30)—into equation (D.15)
gives:

JNi,new
ash,t  p1
γ q Ait pF 1pN i N ash,t q
zq βEt p1
dqJNi ash,t 1
p1
γ qp1
δqfNi ash,t i,new
1 VN ash,t 1 .

(D.20)

We use the fact that whenever δti 1  δ and dit 1  d there are endogenous layoffs so Sti
24
1  Jti 1  0
and fti 1  0, so we can replace δti 1 with δ and dit 1 with d.
55
 Injecting the expression for the Nash wage (30) into equation (D.13) gives:
 
VNi,new
ash,t  γAit pF 1pN i
N ash,t q
zq βEt p1
dq 1
γ
γ
JNi ash,t 1
γ p1
δq δ
d fNi ash,t 1 VNi,new
ash,t 1 .

(D.21)

An equilibrium under Nash bargaining is then processes for Nti , Jti , Jti,new and Vti,new that
solve (D.19), (D.20) and (D.21), and the labor demand schedule (23)-(25), and where fti is
the function of Nti given by equation (29). This can be calculated recursively as follows.
Given Jti 1 , fti 1 and Vti,new
1 the expressions (D.19) and (D.20) solve for employment Nti
when intersected with the labor demand schedule (23)-(25), and so for fti . Equations
(D.19) and (D.21) then allow us to calculate Jti and Vti,new . The Nash wage can then be
recovered, e.g., by equation (D.16).

D.4 Steady-State

In a non-stochastic steady-state equilibrium with with At  1 and ∆ logpZtiq  0, equa-

tions (B.3) and (22) reduce, for J˜  J i {Ai , J˜new  J i,new {Ai , and w̃  wi {Ai , to

J˜pN, w̃q  pF 1 pN q
w̃q,

1
1
p1
δq (D.22)
βeg
J˜new pN, w̃q  p1
βeg pd
δ qqJ˜pN, w̃q. (D.23)

In steady-state, firms hire workers such that the law of motion of the stock of attached
workers is given by equation (17). Combined with the definition of hires—Nti  Mti Hti —
this gives the steady-state relationship between M and N

1
d
M pN q 
1
d δ
N. (D.24)

Combined with equation (29), this gives f pN q as a function of N alone, and through
equation (B.1) it gives the hiring costs c{q pN q as a function of N alone.
In steady-state, firms hire workers such that labor demand is given by equation (23).
Combined with equation (D.23), this gives

c
q pN q
 p1
βeg pd
δqqJ˜pN, w̃q. (D.25)

56
To calibrate the model, we assume that hiring costs c ˜ and obtain the steady-
are 10% of J,
q

state wage as the only one consistent with equations (D.25)-(D.22) and a level of employ-
ment N  1
U  1
5.7%.
In steady-state, the wage is equal to the Nash-bargained wage. (Under downward
nominal wage rigidity, we consider cases where logpΠ̄q g ¥ 0 to make sure this can be
the case.) We back out the value of a workers’ bargaining power γ as the only value that
make the steady-state Nash wage equal to the steady-state wage we have obtained.

D.5 Solution Method

The hiring decision of a firm in sector i is a function of four or five state variables depend-
ing on the process for aggregate productivity. These are: aggregate productivity A (and
lagged aggregate productivity A
1 if aggregate productivity follows an AR(2)), idiosyn-
cratic productivity growth ∆Z i , the lagged wage w

i
1 , and the stock of workers inherited

from the previous period, M i . We have introduced sectoral heterogeneity in such a way
that a firm does not need to forecast any endogenous aggregate variable in order to decide
how many workers to hire. Therefore, we do not need to keep track of the endogenous
aggregate state of the economy in order to solve for the hiring decision of a firm.
A solution to the model can be described as a pair of policy functions for N i and
J˜i  J i{Ai over this (four or) five dimensional state space. We make the following change
of variables. First, we define the AR(1) process:

ηt  p1
ρa2 Lq
1εat, (D.26)

so that:

logpAt 1 q  ρa1 logpAtq ηt 1 . (D.27)

Second, we define ζ
i 1  i

w
1 {A the ratio of the lagged wage to current total pro-
i

ductivity. Given these definitions, the five-dimensional state can be parameterized as

(logpAq, η, ∆ logpZ i q, ζ
i 1 , M i q.
We form a discrete grid of the state-space with 11 points along the exogenous dimen-
sions (logpAq, η, ∆ logpZ i q) and 21 along the endogenous dimensions (ζ
i 1 , M i q). We ap-
proximate the AR(1) processes for the exogenous variables ηt and ∆ logpZti q using the
57
Rouwenhorst (1995) discretization method. The Rouwenhorst method is more accurate
than the Tauchen (1986) method for persistence processes. Petrosky-Nadeau and Zhang
(2017) emphasize this point in the context of the DMP model. Our approach to adapting
the Rouwenhorst method to an AR(2) process is close to the one used by Galindev and
Lkhagvasuren (2010), who consider the more general case of a VAR(1).
We solve for the policy functions at each point on the grid by policy function iteration.
Specifically, we start from an initial guess for the policy functions for N i and J˜i . At each
point of the grid, we then use this guess to calculate the expectation term in equation (B.3)-
(22). In calculating the expectation term, we need to evaluate the policy function for J˜i at
points that are not on the grid. We do so through linear interpolation. Having calculated
values for the expectation term, we compute the values of N i and J˜i that solve equations
(23)-(25), (D.10), and (B.1), and store the resulting values in new policy functions. Once
this has been done for all points on the grid, we have a new set of guesses for the policy
functions. We repeat this process until the policy functions converge. In simulating the
heterogeneous model, we assume 1000 sectors.
In solving for the model under wage-rigidity, we need to first solve for the Nash wage
for each point of the grid. We do so by first solving the model under the assumption
of Nash-bargaining using the same iterative method. The policy function under Nash
bargaining does not depend on ζ
i 1 .

E Calibration of the Model in Section 5

Table E.3 provides a summary of the calibration we use for the new model presented in
section 5. In appendix F, we present a few variants on this model to understand which fea-
tures drive which results. The bottom half of Table E.3 breaks out a case with symmetric
real wage rigidity along side the baseline assumption of DNWR for certain parameters.
We calibrate the model to a monthly frequency. We set the discount factor β to corre-
spond to an annual interest rate of 4%. We set the growth rate of productivity g to 2.3%
annually, the average growth of US labor productivity from 1948 to 2018. We set the rate
of inflation to 2% per year.

58
 We assume a constant-elasticity production function F pN q  N α and set α  2{3.
We assume a Cobb-Douglas matching function q pθq  µθ
η and set the elasticity of the
matching function to η  0.5, in the middle of the range reported in Petrongolo and
Pissarides (2001)’s survey. We calibrate the flow value of unemployment z to 70% of the
wage (specifically, 70% of the labor share α) following Hall and Milgrom (2008): 25%
through unemployment benefits b and 45% through less foregone leisure ζ.
We calibrate the monthly separation rates δ and d so that, based on steady-state re-
lationships, the average separation rate is s  4.95%. This is the separation rate con-
sistent in steady-state with an unemployment rate of 5.7% and an average job-finding
rate f of 45%, as estimated by Shimer (2012) based on CPS data, since in steady-state
1{s p1{u
1qp1{f
1q. The average separation rate in steady-state satisfies s  δ 1δ
d .
We set δ  0.2%, implying d  96.2%. This calibration allows us to generate a slow re-
building of long-term firm-worker relationships during expansions, even though insecure
short-term jobs last only a month.
The parameter µ is still redundant and normalized to 1 (See Appendix B.1 for further
discussion of this point.) We set c so that the cost of hiring a worker c
q
 J is 10% of the
monthly steady-state wage w̄ in a steady state with u  5.7%, in line with what Silva and
Toledo (2009) report based on the Employer Opportunity Pilot Project survey in the US.
This yields c  6.2  10
3 .
We estimate the two roots of the aggregate productivity process from the BLS quar-
terly series on labor productivity. We first apply a three-period moving-average filter to
smooth out high-frequency variations including measurement errors. We then detrend
the series by removing a quadratic trend. The quadratic trend allows us to capture the
productivity slowdown from the 1970s onward. We then estimate an AR(2) on the cycle
component of labor productivity. After converting the roots to a monthly frequency, this
yields ρa1  0.985 and ρa2  0.88.25 We calibrate the volatility of the aggregate shocks to
match the volatility of the unemployment rate—see below.
We calibrate the persistence of the idiosyncratic productivity process based on KLEMS

25
The autoregressive coefficients of the AR(2) estimation are ϕa1  1.64 and ϕa2 
0.65. They are related
to the roots ρa1 and ρa2 through the equation I
ϕa1 L
ϕa2 L2  pI
ρa1 LqpI
ρa2 Lq. This gives roots 0.96 and
0.68 on a quarterly frequency. The monthly roots are the quarterly roots raised to the power 1{3.
59
 Table E.3: Calibration Full Model
β 0.961{12
α 2/3
η 0.5
z 0.47
δ 0.2%
d 96.2%
c 6.2  10
3
ρ1a 0.985
ρ2a 0.88
ρ∆z 0.96
z
σ∆ε 7  10
4
SRWR DNWR
g 0.023{12 0.023{12
Π – 1.021{12
ρ 0.9 –
γ 0.99 0.97
σεa st. σ  3.4% st. σ  5.6%
a a

Note: “DRL” stands for “Decreasing Returns to Labor”. The abbreviation

“st.” stands for “such that.”

annual data on US sectoral productivity from 1947 to 2010 (Jorgenson, Ho, and Samuels,
2012). The KLEMS dataset provides labor productivity series (value added per hour)
for 31 sectors. We take logpZti q to be the log difference between the sectoral labor pro-
ductivity series and the BLS series for aggregate labor productivity. Here again, we
first apply a three-period moving-average filter to the level of these series to smooth
out high-frequency variations in logpZti q. We then first-difference the resulting series
and estimate AR(1) models for ∆ logpZti q in each sector. The average estimated autore-
gressive root across sectors is ρ∆z  0.62 at an annual frequency. We therefore calibrate
ρ∆z  0.62  0.96 in our monthly calibration. We calibrate the volatility of idiosyncratic
1
12

productivity growth σε∆z to roughly match the average of the fraction of constrained firms
in the data as measured by the San Francisco Fed’s Wage Rigidity Meter: 13% (see Figure
3). The value we use is σε∆z  7  10
4.
We are left with calibrating the bargaining power of workers γ and the standard devi-

60
ation of innovations to the aggregate productivity process σεa . We calibrate them so that
the average and standard deviation of the unemployment rate in simulations of the model
match their values in the data (5.7% and 1.6 percentage points). Like for the model of sec-
tion 3, we choose to match the standard deviation of unemployment exactly (as opposed
to calibrating to the standard deviation of productivity in the data) so that we can apply
our definition of expansions and contractions to our simulated samples in the same way
as we do to the real world data. The resulting bargaining power of workers is γ  0.97.
The resulting value for the standard deviation of the productivity process is σ a  0.056.

F Decomposing Plucking in the Model of Section 5

The new model presented in section 5 has several non-standard features in addition to
DNWR: decreasing returns to labor, insecure short-term jobs, and AR(2) aggregate shocks.
Here we shed light on the role of these different features by presenting results for simpli-
fied versions of this model that more closely approximate the model used in section 3.
In particular, we present results for a version of our new model in which we eliminate
the insecure short-term jobs (by setting d  δ) and AR(2) aggregate shocks (replacing
them with AR(1) shocks (1)). In this version we also eliminate sectoral shocks and trend
growth. This version of our new model then only differs from the model in section 3 by
having decreasing returns to labor, and not having match-specific shocks. The calibration
for this simplified model is presented in Table F.4.26
We compare both the full model of section 5 and this simpler version to the same
models under symmetric wage rigidity instead of DNWR. Since we calibrate the model
away from the Hagedorn-Manovskii calibration, we cannot assume Nash bargaining as
a benchmark, as it would fail to generate fluctuations in unemployment (Shimer, 2005).
We therefore instead assume symmetric real wage rigidity. Following Shimer (2010), we

26
We set δ  d  4.95% to match the steady state separation rate. We set c  0.15 so that the cost of
hiring a worker is 10% of the monthly steady-state wage. We set the auto-regressive root of the aggregate
productivity process ρa1 to 0.98 following Shimer (2010). The parameters γ and σεa are set to match the
average and standard deviation of the unemployment rate as in the full model.
61
 Table F.4: Calibration Simple DRL Model
β 0.961{12
α 2/3
η 0.5
z 0.47
δ 4.95%
d 4.95%
c 0.15
ρa 0.98
SRWR DNWR
g 0.023{12 0
Π – 1.021{12
ρ 0.9 –
γ 0.87 0.80
σεa st. σ  2.0% st. σ  3.1%
a a

Note: “DRL” stands for “Decreasing Returns to Labor”. The abbreviation

“st.” stands for “such that.”

assume the real wage is a weighted average of the past real wage and the present flexible-
wage target given by equation (30):

logpwti q  ρ logpeg wti
1 q p1
ρq logpwNi ash,tq, (F.1)

where ρ P r0, 1s is a weight that we set to 0.9 following Shimer (2010), and g is the rate
of trend growth in productivity.27 Again, we set the parameters γ and σεa to match the
average and standard deviation of the unemployment rate.
Table F.5 presents results that are analogous to Table 3 in the main paper for these
cases. This table shows clearly that the addition of insecure short-term jobs and AR(2)
shocks has very little bearing on plucking. These features primarily add persistence to
the model. This can be seen by comparing the two columns for the “Simple DRL” model
to the two columns for the “Full” model. The plucking statistics are large with DNWR
in both models and small with SRWR in both models. What differs strongly across these
27
Since we assume a symmetric process for the logarithm of Ait , we take the average to be geometric—
arithmetic for the logarithm of wages—in order not to introduce an ad hoc source of asymmetry in the
model. Also, in some cases wages in a sector may be so low relative to productivity in that sector in the
model with symmetric real wage rigidity that firms are willing to hire more workers than there are in the
sector. In these cases, we assume that firms hire all workers but no more.
62
models is the persistence statistics in the bottom half of the table. The Full model gener-
ates a great deal more persistence than the Simple DRL model.
In addition to this, we see from Table F.5 that the sources of plucking are similar in our
new model as in the DMP model we analyze in sections 3-4. The new model with DNRW
produces slightly more plucking than the DMP model with DNWR. However, the new
model with SRWR actually produces somewhat less plucking than the DMP model with
Nash bargained wages.

F.1 The Role of Decreasing Returns to Labor

Table F.5 makes clear what the role of insecure short-term jobs and AR(2) shocks are in the
new model. But what about decreasing returns to labor (DRL)? Why do we include this
feature in our new model? The main reason is to allow the model to replicate the fraction
of wage freezes in the data (on average 13%). With constant returns to labor, replicating
this turns out to be impossible, as firms react to large negative shocks by laying off all
their workers rather than maintaining them at a frozen wage.
With constant returns to labor, the firm’s flow value of labor is A
w, which is in-
dependent of the number of employees working for the firm. If the firm faces a large
negative shock to A that is quite persistent (as shocks are in the new model) the firm will
want to lay off not just a few workers but all its workers. With DRL, the marginal product
of workers rises as the firm lays off workers. DRL therefore leads firms to lay off a portion
of its workforce after negative shocks as opposed to all workers.
This implies that with constant returns to scale, the average unemployment rate in-
creases sharply with the volatility of sectoral shocks, but the share of matches in which
the DNWR constraint binds remains far below our calibration target. In contrast, with
DRL, most workers are retained even when the DNWR constraint binds and it is possible
to match a 13% fraction of wage freezes in steady-state.
Decreasing returns to labor also allow us to move away from the (Hagedorn and
Manovskii, 2008) calibration while keeping hiring pro-cyclical. Recall that in the standard
DMP model low values of the flow value of unemployment z imply that the firms’ value

63
 Table F.5: Plucking Property, Speed, and Duration: New Model, Additional Results
Data Simple DRL Full
SRWR DNWR SRWR DNWR
Subsequent expansion on contraction, β 1.12 0.24 0.80 0.28 0.68
(0.22) (0.20) (0.37) (0.42)
Subsequent contraction on expansion, β -0.38 0.13 -0.04 0.12 -0.08
(0.22) (0.21) (0.47) (0.67)
2
Subsequent expansion on contraction, R 0.59 0.06 0.62 0.11 0.53
(0.10) (0.24) (0.20) (0.31)
Subsequent contraction on expansion, R2 0.22 0.03 0.02 0.08 0.13
(0.08) (0.06) (0.17) (0.29)
Speed of expansions (pp/year) 0.87 3.59 3.76 1.45 0.81
(0.37) (0.52) (0.22) (0.22)
Speed of contractions (pp/year) 1.89 3.41 3.96 1.45 1.44
(0.33) (0.57) (0.21) (0.35)
Duration of expansions (months) 59.1 14.2 12.8 34.9 64.0
(1.8) (1.7) (6.8) (16.0)
Duration of contractions (months) 26.9 14.8 16.0 34.2 36.6
(1.8) (2.6) (6.5) (12.6)
Note: The table compares real world data with data from four versions of the model of section 5. The first
column—labeled “Data”—reports empirical results based on data from the U.S. economy from section
2. The second and third columns—labeled “Simple DRL”—report results for the simple version of the
model with decreasing returns to labor but a single type of job (δ  d), no sectoral heterogeneity, and
aggregate productivity following an AR(1) process. The second column reports results under symmetric
real wage rigidity (“SRWR”) and the third column reports results under downward nominal wage rigidity
(“DNWR”). The fourth and fifth columns—labeled “Full”—report results for the full model with decreas-
ing returns to labor, two types of jobs (δ d), sectoral heterogentiy, and aggregate productivity following
an AR(2) process. The fourth column reports results under symmetric real wage rigidity (“SRWR”) and
the fifth column reports results under downward nominal wage rigidity (“DNWR”). The first (third) row
reports the coefficient (R2 ) in an OLS regression of the size of an expansion (percentage point fall in unem-
ployment rate) on the size of the previous contraction (percentage point increase in unemployment rate).
The second (fourth) row report the coefficient (R2 ) in an analogous regression of the size of a contraction
on the size of the previous expansion. The next two rows report the spell-weighted average speed of
expansions and contractions, measured in percentage points of unemployment per year. The final two
rows report the average duration of expansions and contractions, measured in months. For the models,
the reported point estimate is the median value of the statistic over 5000 samples of 866 periods each (the
length of our sample of real-world data). Expansions and contractions of more than 6.5 percentage points
are excluded from the samples. The standard error reported in parentheses is the standard deviation of
the estimates across the 5000 samples. The full model is simulated with 1000 sectors.

64
function increases little with productivity, and hiring is therefore close to acyclical—the
Shimer (2005) unemployment volatility puzzle. With DNWR this problem can become
even worse, making hiring counter-cyclical. In this case higher productivity has two ef-
fects: it increases the current flow value to firms A
w, but it also increases wages,
increasing the probability that the DNWR constraint will bind in the future. For a low
value of z, the first effect is small—the root of the Shimer puzzle—and the second effect
can dominate.
Decreasing returns bring a third effect into play. While higher productivity today still
increases wages and makes it more likely that the DNWR constraint will bind in the fu-
ture, the flow value of firms is now AF 1 pN q
w. If the DNWR constraint will be binding
in the future, employment N will also be lower at that point, raising the marginal pro-
ductivity F 1 pN q of workers. The job-finding rate is therefore procyclical in our calibration
of the model of section 5 with decreasing returns to labor, even though we calibrate the
value of unemployment to 70% of wages as in Hall and Milgrom (2008).

65