Journal of Economic Behavior and Organization 229 (2025) 106827

Contents lists available at ScienceDirect

Journal of Economic Behavior and Organization

journal homepage: [Link]/locate/jebo

Synchronization of endogenous business cycles

Marco Pangallo 1
CENTAI Institute, Turin 10138, Italy

ARTICLE INFO ABSTRACT

Replication materials: [Link] Business cycles tend to comove across countries. However, standard models that attribute
rds/8003502 comovement to propagation of exogenous shocks struggle to generate a level of comovement
JEL classification:
that is as high as in the data. In this paper, we consider models that produce business cycles
C61 endogenously, through some form of non-linear dynamics—limit cycles or chaos. These models
E32 generate stronger comovement, because they combine shock propagation with synchronization
F44 of endogenous dynamics. In particular, we study a demand-driven reduced-form model in
which business cycles emerge from strategic complementarities within countries, synchronizing
Keywords:
their oscillations through international trade linkages. We develop an eigendecomposition that
Synchronization
Business cycles
explores the interplay between non-linear dynamics, shock propagation and network structure,
Non-linear dynamics and use this theory to understand the mechanisms of synchronization. Next, we calibrate the
Networks model to data on 24 countries and show that the empirical level of comovement can only
be matched by combining endogenous business cycles with exogenous shocks. Despite the
limitations of using a stylized model, our results support the hypothesis that business cycles
are at least in part caused by underlying non-linear dynamics.

1. Introduction

At one extreme, business cycles have been explained as the consequence of exogenous events that originate outside the economy,
such as political decisions, natural catastrophes or wars. According to this view, economies would live in stable stationary states
(up to a long-term growth trend), but their dynamics are perturbed by random shocks that pull them out of their steady state. At
the other extreme, business cycles have been explained as the consequence of forces that are endogenous to the economy, such
as debt dynamics, overinvestment and exuberant expectations. Under this endogenous view, the steady state of the economy is
fundamentally unstable, and macroeconomic fluctuations are mathematically described by some form of non-linear dynamics—limit
cycles or chaos. Of course, it is possible to combine these two extreme views by adding exogenous shocks on top of endogenous
macroeconomic dynamics. There are fundamental scientific and policy implications to understanding the contribution of exogenous
vs. endogenous forces, and so this question has received enormous empirical and theoretical attention.
In this paper, we attack this problem from a new angle, taking advantage of the fact that business cycles synchronize differently if
they are endogenous or exogenous. We consider a system composed of different countries that produce business cycles independently
but are connected through international trade linkages. Countries are hit by idiosyncratic shocks. In this situation, exogenous

E-mail address: [Link]@[Link].

1 I thank Paul Beaudry, Doyne Farmer, Thomas Peron and Alexander Teytelboym, as well as two anonymous reviewers, for useful comments at various stages
of this work. I also thank seminar participants at the Network Science in Economics Ninth Conference (Minneapolis), European Economic Association (EEA) 2020,
Computing in Economics and Finance (CEF) 2023, Progress on Difference Equations (PODE) 2023, Network Science Conference (NETSCI) 2020, Conference of
Complex Systems (CCS) 2022, EconophysiX Chair, Utrecht and Sant’Anna Pisa. Most of this work is included in two chapters of my DPhil thesis at the Mathematical
Institute of the University of Oxford, and was performed while I was also affiliated to the Institute for New Economic Thinking at the Oxford Martin School. I
acknowledge funding from INET, Baillie Gifford and from the James S. Mc Donnell Foundation. Replication files: [Link]

[Link]
Received 2 February 2024; Received in revised form 20 October 2024; Accepted 18 November 2024
Available online 29 November 2024
0167-2681/© 2024 Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

Fig. 1. Illustration of the main idea of the paper. Top: We consider two nodes (countries) whose economic activity depends 90% on domestic demand (self-loops
in the network representation) and 10% on demand by the other country (arrows). Bottom: Example dynamics of an index of economic activity for countries
1 and 2, obtained by running representative simulations for 70 time steps. Left: Exogenous shock processes. In this case, comovement originates from the
propagation of these exogenous shocks. Center: Endogenous business cycles. Here, comovement originates from the synchronization of the endogenous cycles.
Right: Combination of the shocks and cycles.

business cycle models lead to comovement, or positive correlation, among countries’ economic activity exclusively from the
propagation of the shocks. For example, when two countries are in a steady state and country 1 is hit by a positive shock, country
1 increases its demand to country 2 and makes country 2’s economy grow above the steady state, too. This produces a positive
correlation between the macroeconomic dynamics of the two countries.
According to the endogenous view, instead, comovement arises from the synchronization of the non-linear dynamics of the two
countries. Synchronization is a generic property of interacting components of a dynamical system to align their non-linear dynamics
in a way that they operate in synchrony, provided they are sufficiently ‘‘close’’. (In this paper, we use the term synchronization in this
technical sense, not as a synonym of comovement and positive correlation as usual in economics.) Synchronization is a fascinating
phenomenon that applies to very diverse systems such as oscillating pendula, flashing fireflies, firing neurons, and applauding
audiences (Strogatz, 2004). It has rarely been applied to economics.
To the extent that shock propagation and synchronization produce different empirical predictions on comovement, we can
indirectly test whether the exogenous or endogenous views are a better description of the business cycle.
We find that the propagation of exogenous shocks produces a much lower comovement than the synchronization of endogenous
business cycles. In particular, shock propagation falls short of generating a level of comovement that is as high as in the data, as in
standard models of trade and exogenous shocks (Kose and Yi, 2006). By contrast, the empirical level of comovement can be matched
by combining the synchronization of endogenous cycles with (relatively small) exogenous shocks.
The economic intuition for why synchronization can help generate higher comovement is given in Fig. 1. In this illustration,
we consider two countries whose economic activity depends much more on their domestic demand than on demand by the other
country, as is mostly the case empirically. Under exogenous business cycles, each country is hit by an idiosyncratic shock process
and produces a business cycle independently of the other country. As we can see, the correlation of the macroeconomic dynamics of
the two countries is virtually zero. Intuitively, this is because shocks in one country get transmitted to the other country only by the
trade linkage, which is small relative to domestic demand. On the contrary, if the two countries follow a deterministic endogenous
business cycle (in this case, a limit cycle), their non-linear dynamics perfectly synchronize. Combining the exogenous shock processes
with the endogenous business cycles gives a Pearson correlation of 0.29 between the dynamics of the two countries, in line with
empirical observations (Section 5.1).
These results can be interpreted as follows. The combination of persistent shocks and deterministic oscillatory dynamics implies
that, at any given time, recessions occur with a certain probability. For example, thanks to a series of positive shocks, a recession
can be delayed even after a peak of the underlying limit cycle has been reached, but recession probability increases as deterministic
dynamics move towards the next trough of the limit cycle. Because of synchronization, the two countries reach the peak of their
deterministic dynamics roughly at the same time. The first country may fall into a recession before the second country because it
is hit by a negative shock, while the second country may be enjoying a period of positive shocks. However, the second country is
already predisposed to start a recession soon, and so a reduction in international demand of the first country easily drags the second
country into a recession, too.
Our goal in this paper is to illustrate how synchronization theory can be useful to explain business cycle comovement. However,
the theory that we develop can be applied to explain comovement of any other economic or financial time series, such as commodity,
housing and stock prices. To highlight the generality of our approach, we avoid choosing models that focus on specific economic
forces causing endogenous business cycles. At the same time, we avoid choosing a detailed trade model, e.g. based on gravity
equations. To us, trade is just an example of an economic linkage, but other linkages could be relevant (De Haan et al., 2008).

2
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

We instead adapt the reduced-form model proposed in Beaudry et al. (2020),2 in which endogenous cycles and connections have
more abstract nature. In particular, business cycles are caused by strategic complementarities, that is, the tendency of agents to
increase their action if other agents increase their action, too. In macroeconomics, this can be thought of as the tendency of firms
to increase production if their customers increase production themselves, or the tendency of households to increase consumption if
other households in their social network increase consumption (Cooper and John, 1988). This abstract framework can encompass
many causes for endogenous business cycles.
Our first set of results is theoretical. As we want to understand the interplay between synchronization of deterministic non-
linear dynamics, exogenous shocks and network structure, we adapt an approach first developed in physics, known as complete
synchronization theory (Pecora and Carroll, 1998), to our case. Here, links tend to align the non-linear dynamics of the nodes,
but idiosyncratic shock processes tend to pull them apart. Complete synchronization is an elegant mathematical formalism that
makes it possible to quantify the relative strength of synchronizing and desynchronizing forces, through an eigenvalue–eigenvector
decomposition that takes into account the network connecting the different components. The eigenvalues and eigenvectors give
information about how quickly the system may synchronize after some components are hit by idiosyncratic shocks, about which
network components have more synchronized dynamics depending on the presence of clusters, etc.
Next, we test empirically how far exogenous and endogenous business cycle models can go in terms of explaining comovement.
To do so, we specify the abstract model so that the nodes are 24 countries and a Rest of the World aggregate, and weighted,
directed, links represent international demand and domestic demand (self-loops). In our empirical application, we consider different
parameterizations under which the deterministic model either produces endogenous business cycles or converges to a stable steady
state. We then add idiosyncratic exogenous shocks to the dynamics. We show that only a combination of endogenous business cycles
and exogenous shocks can match the level of comovement that can be found in the data. Finally, we show how a few predictions
of complete synchronization theory have empirical relevance.

2. Literature review

Our paper relates first of all to the few papers that in the past have seen the appeal of synchronization theory to explain business
cycles comovement (Haxholdt et al., 1995; Selover and Jensen, 1999; Brenner et al., 2002; Matsuyama et al., 2014). Our paper
is different in several respects. The most important difference is that we use another concept of synchronization that makes it
possible to compare deterministic endogenous dynamics and random shocks. Previous work used the most well-known concept of
phase synchronization, that implies that each component of the synchronizing dynamical system has a relatively well-defined natural
frequency and phase; under synchronization, the frequencies would align to a common frequency and the phases would lock. The
problem with this concept is that in economics it is hard to characterize oscillations with a well-defined frequency and phase. In
this paper, we use the more general concept of complete synchronization. To our knowledge, this has not been applied to economics
yet. The second important difference is that the papers above are purely theoretical, whereas here we test if indeed synchronization
makes it possible to improve over shock propagation in explaining comovement. Other differences are that our paper (i) uses a
general model that can be adapted to several economic circumstances, rather than relying on a specific business cycle model; (ii)
studies synchronization on a network of an arbitrary number of agents, rather than just two agents or a trivial topology; (iii) produces
‘‘realistic’’ smooth business cycle dynamics, as opposed to e.g. limit cycles of period two.
Our paper contributes to the literature that explains international business cycles comovement. Most of this literature is empirical,
trying to make sense of which channels are more likely to lead to comovement (De Haan et al., 2008; Pangallo, 2019).3 The channel
that is unanimously found to have a positive effect on comovement is international trade (Frankel and Rose, 1998; Clark and
Wincoop, 2001; Imbs, 2004; Baxter and Kouparitsas, 2005; Kose and Yi, 2006; Calderon et al., 2007; Inklaar et al., 2008; Hsu et al.,
2011; Kleinert et al., 2015; Feldkircher and Huber, 2016; Di Giovanni et al., 2018; Hwang and Kim, 2021). Instead, the effect of
financial integration on comovement is ambiguous. For instance, Imbs (2006) find a positive effect and Kalemli-Ozcan et al. (2013)
find a negative effect. While it is true that financial integration helps spread financial shocks, free movement of capital implies that
capital may fly from countries in distress to countries that are doing well, thus reducing comovement.4 Following this evidence, we
consider trade as our main mechanism leading to synchronization.
Another literature attempts to explain comovement by building international real business cycle models in which comovement
originates from the propagation of exogenous shocks (Backus et al., 1992). These works generally struggle to obtain a level of
comovement that is as high as in the data (Arkolakis and Ramanarayanan, 2009; Johnson, 2014; Liao and Santacreu, 2015), and
this became known as the ‘‘trade-comovement puzzle’’ (Kose and Yi, 2006). As already outlined above, we contribute to this literature
by showing how synchronization of non-linear dynamics is a powerful way to generate higher international comovement.
More broadly, we contribute to the debate on the nature of business cycles as exogenous vs. endogenous. Over the last 80 years,
there have been cycles of preferences towards one view or another. Following the early theoretical work on endogenous business
cycles (Kaldor, 1940; Hicks, 1950; Goodwin, 1951), the real business cycles revolution (Kydland and Prescott, 1982) popularized

2 We build on the model discussed more thoroughly in an earlier working paper (Beaudry et al., 2015). Apart from some technical details, the main difference

between their model and mine is that I let agents have heterogeneous interactions and dynamics, which is a necessary ingredient to study synchronization.
3 There are also methodologically-focused studies that use novel techniques to measure comovement, such as wavelet analysis (Soares et al., 2011) and

random matrix theory (Guerini et al., 2023).

4 This mechanism is well-known since (Backus et al., 1992), and was recently explored in Cesa-Bianchi et al. (2019) together with country-specific heterogeneity

and common shocks.

3
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

the use of exogenous shocks to characterize business cycles. At the same time, the discovery of chaos theory revived the interest
in non-linear dynamic models, and in a few years a flurry of papers on general equilibrium models producing endogenous business
cycles came out (Boldrin and Woodford, 1990). This interest died out as it was not possible to find conclusive evidence on the
existence of non-linear dynamics (limit cycles or chaos) in economic time series,5 and was only recently revived (Beaudry et al.,
2020; Bonart et al., 2014; Asano et al., 2021). Our approach indirectly tests for non-linear dynamics in economic time series by
looking at the different implications of exogenous vs. endogenous business cycles on comovement.

3. A general framework for endogenous business cycles

Over decades, researchers have proposed several economic forces that could generate endogenous business cycles.6 Some scholars
focused on the role of finance, others on real factors such as investment and inventories, still others on bounded rationality and
adaptive expectations. Even more forces have been proposed, including overlapping generations effects, preference switching, search
and matching, technological progress, and wage bargaining. To discuss synchronization of endogenous business cycles, it would be
limiting to focus on a model that picks a particular economic force to produce business cycles in the first place, as results would
appear to depend on this choice. It would instead be desirable to use a macroeconomic framework that is as general as possible.
Luckily, Beaudry et al. (2015) have recently proposed such a framework. The authors focus on strategic complementarities (Bulow
et al., 1985; Cooper and John, 1988) between agents as the main force leading to endogenous cycles. In their framework, agents’
decisions self-reinforce through interactions, leading to an instability of the steady state, but at some point sufficiently far from the
steady state, stabilizing forces set in that prevent dynamics from exploding. Many of the narratives for endogenous business cycles
listed above can be cast within this framework (Section 3.3).
In this section, we start introducing the model in an abstract interaction network, and then consider a specification of the model
in which interactions correspond to international trade. We finally discuss how this model can generate endogenous cycles.

3.1. Abstract formulation

Consider 𝑁 agents, indexed by 𝑖 ∈  = {1, … , 𝑁}. Denote by 𝑥𝑖 an accumulation variable representing a stock of agent 𝑖, and
by 𝑦𝑖 a decision variable representing a flow. We can think of 𝑦 as production and of 𝑥 as inventories; we can also think of 𝑦 as
investment or consumption of durable goods, and of 𝑥 as capital or net worth, respectively.
The time evolution of the accumulation variable 𝑥𝑖 is simple. At each time step 𝑡, it depreciates by a factor 𝛿, and increases by
the decision variable 𝑦𝑖,𝑡 . In formula, 𝑥𝑖,𝑡+1 = (1 − 𝛿)𝑥𝑖,𝑡 + 𝑦𝑖,𝑡 .
The dynamics of the decision variable 𝑦𝑖 is more involved and captures the effect of the interactions with the other agents. The
planned level of 𝑦 for agent 𝑖, 𝑦𝑖,𝑡+1 , depends most importantly on a term 𝑦𝑖,𝑡 , capturing the interactions among the decision variables
∑ ∑
𝑦 of the agents. The interaction term 𝑦𝑖,𝑡 is defined as 𝑦𝑖,𝑡 = 𝑁 𝑗=1 𝜖𝑖𝑗 𝑦𝑗 ,𝑡 , where 𝜖𝑖𝑗 ∈ [0, 1], such that 𝑗 𝜖𝑖𝑗 = 1. Each term 𝜖𝑖𝑗 , which
can be thought of as an interaction coefficient, is the weight that the decision variable of agent 𝑗 has in determining the value of 𝑦𝑖,𝑡
(self-interactions 𝜖𝑖𝑖 are also important). The values of 𝜖𝑖𝑗 , for all 𝑖 and 𝑗, define a weighted, directed, interaction network.
The effect of 𝑦𝑖,𝑡 is mediated by a non-linear function 𝐹 (⋅) that determines the effect of interactions. Suppose that the decision
variables of the agents with whom agent 𝑖 interacts increase, so that 𝑦𝑖,𝑡 becomes larger. If, as a consequence, agent 𝑖’s marginal
payoff goes up, one says that there are strategic complementarities between agent 𝑖 and the agents with whom she interacts (Cooper
and John, 1988).7 Intuitively, in case of strategic complementarities, agent 𝑖 decides to increase her decision variable if the agents
with whom she interacts do the same, i.e. 𝜕 𝑦𝑖,𝑡+1 ∕𝜕𝑦𝑖,𝑡 > 0.
We now complete the description of the model. We first assume that 𝑦𝑖,𝑡+1 also linearly depends on one-step lagged values of
𝑥 and 𝑦, with coefficients 𝛼1 and 𝛼2 (further to an intercept 𝛼0 ). We take 𝛼1 to be negative, and 𝛼2 to be positive and bounded
between zero and one (Beaudry et al., 2015, 2020). These assumptions reflect, respectively, decreasing returns to accumulation,
i.e. willingness to avoid excessive stocks of inventories, capital or net worth, and sluggishness in the adjustment of the decision
variable, i.e. difficulty to quickly modify the level of production, investment, or consumption of durable goods. The assumption on
𝛼1 is instrumental to avoid explosive dynamics, as it helps keeping the dynamics bounded when it wanders away from the steady
state. At the same time, the assumption about 𝛼2 introduces realistic smoothness in the dynamics.8
We finally add idiosyncratic shock terms 𝑢𝑖,𝑡 to the evolution of the decision variables. We parameterize this shock process as an
AR(1), i.e. 𝑢𝑖,𝑡+1 = 𝜌𝑢𝑖,𝑡 + 𝜄𝑖,𝑡 , where 𝜌 ∈ [0, 1] is a persistence parameter and 𝜄𝑖,𝑡 is white noise, normally distributed with mean zero
and standard deviation 𝜎.
Our model for endogenous business cycles is thus fully specified by the following equations, for all agents 𝑖:
𝑥𝑖,𝑡+1 = (1 − 𝛿)𝑥𝑖,𝑡 + 𝑦𝑖,𝑡 ,
( )
𝑦𝑖,𝑡+1 = 𝛼0 + 𝛼1 𝑥𝑖,𝑡 + 𝛼2 𝑦𝑖,𝑡 + 𝐹 𝑦𝑖,𝑡 + 𝑢𝑖,𝑡 , (1)
∑𝑁 ∑
𝑦𝑖,𝑡 = 𝑗=1 𝜖𝑖𝑗 𝑦𝑗 ,𝑡 , 𝜖𝑖𝑗 ∈ [0, 1], 𝑗 𝜖𝑖𝑗 = 1.

5 Barnett and Serletis (2000), Brock (1986), Barnett and Chen (1988), Scheinkman and LeBaron (1989), Barnett and Serletis (2000), Shintani and Linton

(2004), Hommes and Manzan (2006).

6 See, among many others, the contributions and review articles in: (Kaldor, 1940; Hicks, 1950; Goodwin, 1951, 1967; Shleifer, 1986; Boldrin and Woodford,

1990; Foley, 1992; Silverberg and Lehnert, 1993; Bullard, 1994; Matsuyama, 2007; Fazzari et al., 2008; De Grauwe, 2011; Nikolaidi and Stockhammer, 2017).
7 𝜕 2 𝑉𝑖,𝑡
Letting 𝑉𝑖,𝑡 be the payoff function for agent 𝑖 at time 𝑡, strategic complementarities correspond to the condition 𝜕 𝑦𝑖,𝑡 𝜕𝑦𝑖,𝑡
> 0.
8 A problem of many endogenous business cycle models is that they generate a ‘‘sawtooth’’ dynamics that has no persistence, in stark contrast to empirical
time series.

4
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

In the case in which all interaction coefficients 𝜖𝑖𝑗 are identical and equal to 1∕𝑁 (complete and homogeneous interaction
network), and all agents behave alike, 𝑥𝑖,𝑡 = 𝑥𝑗 ,𝑡 and 𝑦𝑖,𝑡 = 𝑦𝑗 ,𝑡 , for all agents 𝑖 and 𝑗 and for all times 𝑡 (and shocks are identical
across all agents), Eqs. (1) recover the model in Beaudry et al. (2015).9

3.2. International model

We now interpret the agents in the abstract model as countries. Following the discussion in the literature review section, we
interpret the interaction coefficients as trade linkages. As we stressed already, we do not aim at building a particular microfounded
trade model, but rather show how much comovement a general model with connections parameterized as trade linkages can produce
under exogenous vs. endogenous business cycles.
A practical difference with respect to the abstract model described in the previous section is that countries have diverse sizes.
This would imply that the accumulation and decision variables would be on different scales, making it necessary to consider a
different set of parameters for each country. To avoid this impractical situation, we work with ‘‘oscillation’’ variables 𝑥𝑖,𝑡 and 𝑦𝑖,𝑡
that define the oscillations independently of the scale, while moving all the effects of the scale to the interaction coefficients.
To do so, we let total investment (or production, or durable consumption) be given by 𝑌𝑖,𝑡 = 𝑌̃𝑖 𝑦𝑖,𝑡 and total capital (inventories,
stock of durables) be 𝑋𝑖,𝑡 = 𝑌̃𝑖 𝑥𝑖,𝑡 . In these expressions, 𝑌̃𝑖 denotes the steady state value of 𝑌𝑖,𝑡 , because with our parameter restrictions
𝑦𝑖,𝑡 oscillates with mean one. Then
𝑥𝑖,𝑡+1 ∶= 𝑋𝑖,𝑡+1 ∕𝑌̃𝑖 = (1 − 𝛿)𝑥𝑖,𝑡 + 𝑦𝑖,𝑡 ,
( ) (2)
𝑦𝑖,𝑡+1 ∶= 𝑌𝑖,𝑡+1 ∕𝑌̃𝑖 = 𝛼0 + 𝛼1 𝑥𝑖,𝑡 + 𝛼2 𝑦𝑖,𝑡 + 𝐹 𝑦𝑖,𝑡 + 𝑢𝑖,𝑡 .

In this application, we take the interaction term 𝑦𝑖,𝑡 to represent the oscillation of the total demand received by country 𝑖.
Letting total demand be denoted by 𝐷𝑖,𝑡 , we assume that its steady state 𝐷 ̃𝑖 is equal to the steady state of the flow variable 𝑌𝑖,𝑡 ,
̃𝑖 = 𝑌̃𝑖 . This corresponds to assuming market clearing in the steady state: depending on the interpretation for the decision
that is 𝐷
variable, it could be that total demand of intermediate and final goods equals production, or that total demand of capital goods
equals investment, or that total demand of durable goods equals realized consumption of such goods. Under this assumption, we
can write 𝐷𝑖,𝑡 = 𝐷 ̃𝑖 𝑑𝑖,𝑡 = 𝑌̃𝑖 𝑑𝑖,𝑡 , where again 𝑑𝑖,𝑡 is the oscillation of demand around its steady-state value. So, in this application
𝑦𝑖,𝑡 ∶= 𝑑𝑖,𝑡 .
The key behavioral assumption to compute total demand is that countries demand fixed proportions of imports from other
countries. Letting 𝑌̃𝑖𝑗 be the steady-state trade flow from country 𝑖 to country 𝑗, and being 𝑎𝑖𝑗 = 𝑌̃𝑖𝑗 ∕𝑌̃𝑗 the fraction of imports
∑
from country 𝑖 over country 𝑗’s output, we write the total demand that country 𝑖 receives as 𝐷𝑖,𝑡 = 𝑗 𝑎𝑖𝑗 𝑌𝑗 ,𝑡 . This means that,
whether country 𝑗 produces more or less than the steady-state value 𝑌̃𝑗 , it still demands the fixed proportion 𝑎𝑖𝑗 of imports from 𝑖.
∑
Equating the two formulations for 𝐷𝑖,𝑡 , 𝐷𝑖,𝑡 = 𝑌̃𝑖 𝑑𝑖,𝑡 and 𝐷𝑖,𝑡 = 𝑗 𝑎𝑖𝑗 𝑌𝑗 ,𝑡 , we can express the oscillations of demand 𝑑𝑖,𝑡 in terms of
the oscillations of the decision variables 𝑦𝑗 ,𝑡 , as:
1 ∑ 𝑌𝑖𝑗 ̃ ∑ 𝑌̃𝑖𝑗 ∑
̃
𝑑𝑖,𝑡 = 𝑌𝑗 𝑦𝑗 ,𝑡 = 𝑦𝑗 ,𝑡 = 𝜖𝑖𝑗 𝑦𝑗 ,𝑡 , (3)
𝑌̃𝑖 𝑗 𝑌̃𝑗 𝑗 ̃
𝑌𝑖 𝑗

where the coefficients 𝜖𝑖𝑗 = 𝑌̃𝑖𝑗 ∕𝑌̃𝑖 are the fraction of the output of country 𝑖 that is exported to 𝑗. Importantly, this includes the
fraction of output that remains in country 𝑖, 𝜖𝑖𝑖 , representing domestic demand. These coefficients satisfy the conditions laid out in
∑
Eq. (1): 𝜖𝑖𝑗 ∈ [0, 1] and 𝑗 𝜖𝑖𝑗 = 1.
In sum, the abstract model introduced in Section 3.1 can be interpreted as a model of business cycles of countries coupled by
an international trade network. Under strategic complementarities, an increase in domestic and international demand may prompt
firms in several countries to increase their decision variables, further increasing domestic and international demand. This could
make the steady state unstable, and, as we discuss next, the model could produce endogenous business cycles.

3.3. Model dynamics

We formally show under which conditions the steady state loses stability in Appendix A.1, where we use the strength of strategic
complementarities at the steady state as a bifurcation parameter.10 We also present numerical simulations representing this dynamics,
with and without shocks, in Appendix A.2. Here, we just provide the intuition for endogenous dynamics in the abstract framework,
and then discuss its economic interpretation.
Consider a country whose dynamics is purely driven by internal demand by identical agents (and thus drop the subscript 𝑖).
Starting from a situation below the steady state, the regime of strategic complementarity makes the agents increase their decision

9 A minor difference with respect to Beaudry et al. (2015) is that 𝑦

𝑖,𝑡+1 depends only on variables at 𝑡, while in Beaudry et al. (2015) it depends on both
variables at 𝑡 and (contemporaneous) variables at 𝑡 + 1. As there did not seem to be much difference in the implied dynamics, we chose this form as it was
computationally simpler. Economically, it corresponds to assuming that the decision variables of the agents with whom agent 𝑖 interacts have a lagged effect on
her decision.
10 A possible criticism of this paper is that we are using strategic complementarities both as a way to get nonlinear dynamics and as a way of generating

comovement, thus making it obvious that nonlinear dynamics corresponds to high comovement. This criticism is misplaced because using other bifurcation
parameters while keeping the strength of strategic complementarities fixed leads to the same results (Appendix D.2.5).

5
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

variables 𝑦 when the other agents are doing the same, in a way that they overshoot the steady state. However, as the agents start
to over-accumulate 𝑥 because the value of 𝑦 more than offsets the depreciation (1 − 𝛿)𝑥, the rise of 𝑦 slows down. This is both
because the agents dislike accumulation and because, as 𝑦 increases, the function 𝐹 (⋅) stops increasing or even decreases (Fig. 6B in
Appendix A.1). In this situation of large 𝑦, strategic complementarities are very weak, or we may even be in a regime of strategic
substitutability, where an increase of 𝑦 by one agent makes the other agents decrease their decision variables. Soon, 𝑦 reverts and
starts to decrease, leading to a prolonged recession. At some point, the decrease in 𝑦 slows down as agents liquidated excessive
stock, and we are back to a regime of strategic complementarity. The process restarts.
Depending on the interpretation for the accumulation variable 𝑥, the decision variable 𝑦, and the function 𝐹 , the narrative above
could correspond to several economic forces causing endogenous business cycles. Beaudry et al. (2020), for example, consider a
microfounded model in which 𝑥 is net worth, 𝑦 is consumption of durable goods, and 𝐹 models banks’ willingness to give loans.
In good times, lending is perceived safe, and so agents can borrow to consume more durable goods. This further strengthens the
economic boom, making lending to be perceived even safer (strategic complementarity). When the economy slows down because
of overaccumulation, lending instead starts to be perceived less safe, making banks cut back on credit (strategic substitutability).
This behavior of banks can easily trigger a recession, which lasts until agents have liquidated assets in excess, at which point the
cycle starts again. This narrative (Beaudry et al., 2020) is similar to many of the previously proposed narratives for finance-based
endogenous business cycles (Nikolaidi and Stockhammer, 2017). However, the strategic complementarity–substitutability framework
easily lends itself to other narratives, such as ones based on overinvestment in the Keynesian tradition (Kaldor, 1940; Hicks, 1950;
Goodwin, 1951), and ones based on temporal clustering of technological innovations (Judd, 1985; Shleifer, 1986; Matsuyama, 1999).

4. Complete synchronization

Synchronization theory is about the alignment of dynamics of weakly-interacting self-oscillating units (Pikovsky et al., 2003). It is
important that units are self-oscillating, that is, they would still display some form of self-sustaining non-linear dynamics (limit cycles
or chaos) if they were uncoupled. Synchronization is not about whether stable units, when coupled, would become unstable (Gualdi
et al., 2015); it is about whether the dynamics of unstable units, when coupled, would align. The coupling is usually ‘‘weak’’ in
the sense that the dynamics of the nodes are only weakly affected by the other nodes at a given time. This characterization of
synchronization theory seems ideally suited to our case study, featuring different countries, weakly coupled through international
trade (Section 3.2), displaying endogenous business cycles on their own (Appendix A.1).
A large part of synchronization theory is about the alignment of the frequency and phase of the oscillations (phase synchroniza-
tion). To us, this is not so relevant for economics, as economic fluctuations do not generally display clear periodicity. We instead
turn to the other part of synchronization theory, that goes under the name of complete synchronization (also known as chaotic, or
full, or identical synchronization). Under this theory, different units are described by the same parameters and so, under coupling,
individual periodic dynamics would completely align. However, desynchronizing forces such as chaos and noise tend to separate
individual dynamics. Complete synchronization theory is an elegant mathematical formalism that makes it possible to quantify the
relative strength of synchronizing and desynchronizing forces, through an eigenvalue–eigenvector decomposition that takes into
account the network connecting the different units.

4.1. Theory

We build on the master stability approach originally proposed by Pecora and Carroll (1990, 1998). Appendix B.1 derives all the
equations, while here we give some intuition into how the approach works.
Master stability analysis starts by assuming that all agents are in a synchronized state 𝒔𝑡 = (𝑥𝑠𝑡 , 𝑦𝑠𝑡 ), in which they all behave
alike. We now express the dynamics of individual agents by 𝑥𝑖,𝑡 = 𝑥𝑠𝑡 + 𝑥𝜉𝑖,𝑡 and 𝑦𝑖,𝑡 = 𝑦𝑠𝑡 + 𝑦𝜉𝑖,𝑡 , where 𝝃 𝑖,𝑡 = (𝑥𝜉𝑖,𝑡 , 𝑦𝜉𝑖,𝑡 ) denotes a small
deviation from the synchronized state, for example due to idiosyncratic shocks hitting agent 𝑖. By performing a Taylor expansion of
the dynamical equations around the synchronized state, it is possible to derive the evolution of 𝝃 𝑖,𝑡 , for all agents 𝑖. As we show in
Eq. (11) in Appendix B.1, 𝝃 𝑖,𝑡+1 depends on all the deviations of the other agents 𝑗, 𝝃 𝑗 ,𝑡 . So the dynamics of each agent are coupled
to those of all other agents. The coupling is mediated by the normalized Laplacian, a matrix derived from the interaction network
whose eigenvalues and eigenvectors give key information about the structure and dynamics of the network.
The key idea of the master stability approach is to perform a diagonalization that uncouples the deviations from the synchronized
state into orthogonal independent components 𝜻 𝑖,𝑡 = (𝑥𝜁𝑖,𝑡 , 𝑦𝜁𝑖,𝑡 ), called eigenmodes. These components correspond to the eigenvalues
and eigenvectors of the normalized Laplacian. By numerically computing the Lyapunov exponents of the eigenmodes dynamics, for
each possible eigenvalue, one derives the so-called master stability function. It is then possible to see that eigenmodes corresponding
to smaller eigenvalues have less negative Lyapunov exponents, and so take longer to return to zero after a shock (Fig. 8 in
Appendix B.1). The eigenvectors corresponding to each eigenmode relate the eigenmodes to the dynamics of the agents.
To illustrate this, consider the simplest possible case of a network of two agents (Appendix B.2.1), and assume that each agent
follows limit cycle dynamics. The smallest eigenvalue of the normalized Laplacian of this simple network is always null, and the
corresponding eigenvector is (+1, +1). The Lyapunov exponent corresponding to this eigenvalue is zero: therefore, any shock hitting
the agents does not decay in this eigenmode. Moreover, the eigenvector shows that the shock affects both agents in the same way,
so this eigenmode can be interpreted as a shift of the common dynamics of the two nodes. Conversely, the second eigenvalue is
positive. It corresponds to negative Lyapunov exponent, so any shock reverts to zero after a few time steps. The eigenvector is
(+1, −1), so this eigenmode can be interpreted as the relative shift in the dynamics between the two agents due to the shock. In

6
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

Fig. 2. Master stability analysis applied to international synchronization. A: Countries in our sample colored by their component of the Fiedler vector (also
annotated). B: Dynamics of the decision variables 𝑦𝑖,𝑡 in the synchronized state (solid black line), and in case of shocks (dashed lines representing individual
countries, colored as in panel A). C: Deviations from the synchronized state after a shock is applied at time 𝜏. Each line represents a country and is colored like
in panel A. D: Deviations projected in the eigenbasis. Each line represents an eigenmode; the darker the color, the smaller the corresponding eigenvalue is.

this case, the master stability analysis indicates that the synchronized dynamics is stable, in the sense that agents synchronize again
after a shock.
In a more general network, the eigenmodes illustrate how long different components of the network take to synchronize after a
shock. Consider a network composed by six nodes, with two cliques of three nodes each, connected by a link between two of the
nodes of the cliques (Appendix B.2.2). In this case, the second smallest eigenvalue is small, and the corresponding eigenvector (Fiedler
vector) is positive at the nodes in one clique, and negative at the nodes of the other clique. All other eigenvalues are much larger.
Here, synchronization occurs first within cliques and last between cliques. The second eigenmode corresponds to between-clique
synchronization.
In conclusion, master stability analysis is a principled method to understanding when synchronized dynamics are stable, and, if
they are, how quickly the agents revert to the synchronized state after they are hit by an idiosyncratic shock. In the case of a complex
network structure, the analysis quantifies how quickly different parts of the network synchronize. The approach is semi-analytical,
in the sense that the relative strength of synchronizing and desynchronizing forces is obtained by numerically computing Lyapunov
exponents, but once these are computed, the key insights are obtained analytically.

4.2. Application: Complete synchronization of real-world countries

To conclude this section, we now discuss a real-world example. We first explain how we build the international trade network
that specifies the coupling between countries, and then discuss the insights that can be gained by applying complete synchronization
theory to these data.

4.2.1. Trade network

To build the network of interaction coefficients 𝜖𝑖𝑗 described in Section 3.2, we merge the recently released Long-Run World
Input-Output Database (LR-WIOD), covering the period 1965–2000 (Woltjer et al., 2021), with the established 2013 release of the
WIOD, spanning the years 1995–2011 (Timmer et al., 2015). We consider 24 countries (Fig. 2A), comprising more than 80% of
World GDP, and a Rest of the World aggregate. Fig. 12 in Appendix C shows a visualization of the network in 2011.
We use international input–output tables instead of trade data because they make it easier to consistently estimate the share of
internal demand vs. international demand. At the same time, we consider a long time span to test the robustness of our assumption of
fixed interaction coefficients, which turns out to be a good approximation at business cycle frequencies (Appendix C). The data also
show that the share of internal demand is generally high (almost always above 0.7, often above 0.9), suggesting that the standard
assumptions under which synchronization theory can be applied (weakly coupled self-sustaining oscillators) holds.

7
 M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

4.2.2. Master stability analysis

Fig. 2A shows the countries in our sample, colored by their component of the Fiedler vector of the matrix 𝑰𝑁 − 𝝐, where 𝑰𝑁 is
the identity matrix of size 𝑁 and 𝝐 is the matrix with components 𝜖𝑖𝑗 (see Appendix B.2.3). The Fiedler vector is commonly used
in graph partitioning because it is an approximate solution to the problem of dividing a graph into two components such that the
sum of the links between the two components has minimal weight. When partitioning a graph using the Fiedler vector, one puts all
nodes with positive components in one cluster and all nodes with negative components in the other cluster. Components that are
close to zero in either the positive or negative direction are less clearly in one cluster than components that are far from zero.
Here, we see that Brazil’s component is strongly negative, and no other country has so strongly negative values (presumably
because Brazil trades a lot with other South American countries, which are included in our sample as ‘‘Rest of the World’’ and not
shown here). Next, European countries have components ranging between −0.19 and −0.06 (with the exception of Ireland, which
has −0.01), with Southern European countries having more negative values than Northern European countries. Furthermore, Asian
countries have slightly positive components (with the exception of South Korea which is slightly negative). Finally, the components
associated to North American countries range between 0.15 and 0.18. Thus, under graph partitioning one cluster would mostly
comprise Brazil and Europe, and the other Asia and North America.
To explore how the theory developed in Section 4.1 helps understand synchronization, we let the model reach the synchronized
state, and then apply at a given time 𝜏 an i.i.d. shock to each country, drawn from a Gaussian distribution with mean zero and
standard deviation 0.1. Such shocks are about 10% of the amplitude of the endogenous limit cycles.
As we can see in Fig. 2B, the dynamics of individual countries diverge from the synchronized state, but then after about 60 time
steps they synchronize again. Fig. 2C zooms into the deviations from the synchronized state, coloring each line by the same color
of the corresponding country in panel A. Even if the shocks are independent across countries, we see that lines get roughly divided
into two groups, along the lines of the graph partitioning mentioned above.
Fig. 2D shows the deviations from the synchronized state in the eigenbasis. We see that the black line, corresponding to the null
eigenvalue, keeps fluctuating around zero, suggesting a small permanent shift of the shocked dynamics from the unshocked one.
Moreover, dark lines corresponding to eigenmodes with small eigenvalues show damped fluctuations that take long to return to
zero. These lines include the second eigenmode, implying that it takes longer to reach synchronization between the clusters defined
by the Fiedler vector. Light-colored lines instead quickly return to zero, suggesting that synchronization within clusters happens
fast.
In conclusion, the community structure discovered by our spectral analysis, combined with the theory of complete synchro-
nization that we developed, makes it possible to understand analytically which countries are more likely to experience higher
synchronization when each country follows an endogenous business cycle and is hit by idiosyncratic shocks.

4.3. Comovement implications

Further to providing intuition into the mechanisms driving synchronization, the theory of complete synchronization helps explain
how synchronization of endogenous business cycles could improve on exogenous business cycle models in terms of generating higher
comovement between macroeconomic time series.
Consider two countries following two deterministic identical limit cycles in isolation. When coupled, absent shocks, after some
time their dynamics perfectly align in the synchronized state 𝒔𝑡 = (𝑥𝑠𝑡 , 𝑦𝑠𝑡 ). Suppose now that at time 𝜏, the decision variables 𝑦1,𝜏
and 𝑦2,𝜏 of the countries are hit by idiosyncratic shocks 𝑦𝜉1,𝜏 and 𝑦𝜉2,𝜏 respectively. The comovement between the time series of the
decision variables, calculated for the time steps 𝑡 immediately following the shocks, is given by the Pearson correlation coefficient
( )
cor 𝑦𝑠𝑡 + 𝑦𝜉1,𝑡 , 𝑦𝑠𝑡 + 𝑦𝜉2,𝑡 =
( ) ( ) ( ) ( )
cov 𝑦𝑠𝑡 , 𝑦𝑠𝑡 + cov 𝑦𝑠𝑡 , 𝑦𝜉2,𝑡 + cov 𝑦𝜉1,𝑡 , 𝑦𝑠𝑡 + cov 𝑦𝜉1,𝑡 , 𝑦𝜉2,𝑡 (4)
( ) ( ) .
std 𝑦𝑠𝑡 + 𝑦𝜉1,𝑡 ⋅ std 𝑦𝑠𝑡 + 𝑦𝜉2,𝑡

In the above equation, cov denotes the covariance, std indicates the standard deviation, and we have used linearity of the covariance
( 𝑠 𝑠)
to decompose it in various
( terms.
) The term( cov
) 𝑦𝑡 , 𝑦𝑡 indicates the perfect synchronization of the unperturbed deterministic
dynamics; the terms cov 𝑦𝑠𝑡 , 𝑦𝜉2,𝑡 and cov 𝑦𝜉1,𝑡 , 𝑦𝑠𝑡 indicate the comovement between the deterministic dynamics and the deviations
( )
due to the shocks; and the term cov 𝑦𝜉1,𝑡 , 𝑦𝜉2,𝑡 indicates the correlation of the deviations (i.e., shock propagation). Compare this
equation with the correlation between time series of two countries that are in the same stable steady state and are hit by shocks
(𝑦𝜉1,𝑡 , 𝑦𝜉2,𝑡 ):
( )
( ) cov 𝑦𝜉1,𝑡 , 𝑦𝜉2,𝑡
cor 𝑦𝜉1,𝑡 , 𝑦𝜉2,𝑡 = ( ) ( ). (5)
std 𝑦𝜉1,𝑡 ⋅ std 𝑦𝜉2,𝑡

If we assume that the variance of endogenous and exogenous fluctuations is the same, the correlation coefficient is only determined
by the numerators of these expressions. Comparing (4) and (5), it is clear that the endogenous component of dynamics potentially
increases correlation. In particular, when the variance of the shocks is small relative to the variance of 𝑦𝑠𝑡 (i.e., the typical amplitude
of the endogenous cycles), the correlation coefficient tends to one when business cycles are purely endogenous.

8
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

Fig. 3. Mean Pearson correlation across countries as a function of the standard deviation of idiosyncratic shocks, for three parameterizations of the model and
for two empirical variables. For the simulations, each line represents the mean across 100 replications, and error bands represent one standard deviation. For
the data, lines represent the mean across 18 detrending procedures, and error bands represent one standard deviation (Section 5.1).

5. Empirical application

We now test quantitatively to which extent exogenous and endogenous business cycle models are able to match the empirical
level of international comovement that can be found in real-world time series.

5.1. Time series data

To do so, we must measure empirical comovement in the first place. This is not a trivial task. Papers measuring business cycle
comovement have been using all sorts of variables that measure economic activity and all sorts of filters to detrend the corresponding
time series (De Haan et al., 2008).
Here, we consider annual data on employment and value added to measure economic activity. We obtain these data from the
Penn World Tables, version 10.0 (Feenstra et al., 2015). More specifically, we use the number of persons engaged as a measure of
employment, and real GDP at constant 2017 national prices to measure value added. For the countries considered in this paper,
data are available from 1950 to 2019.
We cannot use these raw data to measure comovement, because they show clear common trends (confirmed by an augmented
Dickey–Fuller test) that would bias upward the estimate of comovement at business cycle frequencies. To address this issue, we
consider various modifications. First, we may divide employment or value added by total or working age population.11 Second, we
filter the resulting data so as to remove long-run fluctuations from the sample. To do so, we use two filters. One is the well-known
Hodrick–Prescott filter, with both the standard parameter 𝜆 = 100 that is used for annual data, and the value 𝜆 = 6.25 suggested
by Ravn and Uhlig (2002). The other is the band-pass filter proposed by Christiano and Fitzgerald (2003), alternatively keeping
fluctuations between 2–15 or 2–25 years.12 For all these options for the filters, we either consider the cyclical component or the ratio
between the cyclical and trend components. In conclusion, from all these combinations we get 27 alternative detrending procedures
for both employment and value added.
We find that the mean pairwise Pearson correlation among countries’ employment is 0.23, with a standard deviation of 0.03
across detrending procedures. When considering value added, the correlation is 0.34, with a 0.05 standard deviation. These results
are in line with estimates in the literature (Kose and Yi, 2006).

5.2. Main result

We now check whether endogenous business cycles are indeed necessary to reproduce the empirical level of comovement, or if
instead shock propagation in an exogenous business cycle specification is sufficient. To do so, we simulate the model under three
different parameterizations, and compare the comovement of the decision variables 𝑦 to empirical comovement as measured in the
previous section. We perform this comparison at the aggregate, country and country-pair level.
We simulate the model under three baseline sets of parameters (many more parameter combinations are checked in Appendix D.2.3).
These correspond to the deterministic skeleton of the model producing limit cycles, or converging back to the steady state with

11 Total population is available in the Penn World Tables, while working age population is downloaded from the World Bank.
12 We prefer the filter by Christiano and Fitzgerald (2003) over the (Baxter and King, 1999) filter because the latter loses some data points at the extremes
of the series.

9
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

Fig. 4. Each point represents how much a given country is correlated to the other countries in the sample. The position on the horizontal axis indicates this
country-specific correlation in the empirical data as measured from GDP, and is the same in the three panels. The vertical axis shows the same country-specific
correlation in simulated data, as measured from the decision variables 𝑦, and is different in each panel depending on the model parameterization. All simulated
results are averaged across 100 replications. The color code follows the position of the countries in the interaction network as exemplified by the Fiedler
eigenvector, as in Fig. 2. The identity line is plotted for reference.

damped fluctuations (focus) or monotonically (node). We choose the parameters in Appendix A.2 as a baseline. Moreover, we
parameterize the shock process as an AR(1) with persistence parameter 𝜌 = 0.3 and standard deviation 𝜎 varying between 0.01
and 0.2. We simulate the model for 280 time steps after removing an initial transient. It makes sense to interpret time steps as
quarters (Appendix A.2), so this leads to a 70-years period that is as long as the data we compare against (1950–2019). To calibrate
the interaction network, we consider year 1990, which is in the middle of the available data (we will also consider time-varying
interaction coefficients as a robustness test).
Fig. 3 shows our key result. When the model produces limit cycles, the mean correlation coefficient across countries ranges from
0.99 in the case of idiosyncratic shocks with standard deviation 𝜎𝑢 = 0.01 (or 1% of the steady state value of the decision variables)
to 0.03 when the standard deviation of idiosyncratic shocks is 𝜎𝑢 = 0.20 (20% of the steady state). By contrast, when the model
produces fluctuations that converge to a stable steady state, the average correlation coefficient is never larger than 0.15, for any
value of the standard deviation of idiosyncratic shocks. More specifically, when the steady state is a focus, the average correlation
varies between 0.08 and 0.15, whereas when the steady state is a node it is close to zero. As empirical comovement is 0.23 when
considering employment and 0.34 when considering value added, only in the limit cycle case the model can reproduce this level
of comovement. This is confirmed by a t-test for the difference in means, with high statistical significance for all values of 𝜎𝑢 (𝑡-
statistic < −11). In conclusion, under the assumptions in this section the model can only match the empirical level of comovement
with endogenous business cycles, for standard deviation of idiosyncratic shocks between 5 and 10% of the amplitude of the limit
cycles. This confirms the theoretical intuition in Section 4.3.
Next, Fig. 4 zooms in at the level of countries. Here, we select a standard deviation of idiosyncratic shocks of 0.08, under which
the limit cycle model produces a correlation that is close to the data (Fig. 3). This figure conveys three results. First, for every
country and not only for some, comovement is higher in the limit cycle case than in the focus and node regimes. Second, under
limit cycles, empirical correlation coefficients are more similar to simulated ones (Pearson empirical-simulated coefficients: 0.69).
This means that countries whose dynamics are more correlated to those of other countries in the data are also more correlated in the
model. For instance, the mean correlation coefficient of the Netherlands is 0.46 in the data and 0.40 in the model, while the mean
correlation coefficient of Brazil is 0.04 in the data and 0.12 in the model. This similarity between empirical and simulated correlation
coefficients extends to the focus case, although it is slightly lower (Pearson 0.62). It is however completely absent when the steady
state is a node (Pearson −0.15). The third result is that, in the limit cycle and focus cases, both empirical and simulated correlations
are clearly linked to the position of the countries in the international trade network, as exemplified by the Fiedler eigenvector
discussed in Section 4.2. Countries that are more correlated to the other countries in the sample tend to have intermediate values
for their eigenvector components, whereas countries with extreme values such as Brazil tend to have lower average correlations.
These results are confirmed in Fig. 5, which shows pairwise correlations. In the data, pairwise correlations are high within
Europe, North America and Asia, but they are lower across continents. The simulations, both in the cycle and focus cases, generally
reproduce these patterns, although they fail to yield the few negative correlations that can be seen in data. By contrast, when the
model is parameterized as a node, pairwise correlations do not show a clear pattern (Fig. 13 in Appendix D.1).
In conclusion, the results in this section support the main claim of the paper, namely that endogenous business cycles are needed
to explain the level of international comovement that can be found in the data. They also show that the theory developed in Section 4
can be fruitfully employed to interpret synchronization.

10
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

Fig. 5. Pairwise Pearson correlation coefficients between the dynamics in different countries, both in the model and in the data, averaged across 100 replications.

5.3. Robustness tests

In Appendix D.2 we check if the main result of the paper is robust to alternative specifications. We perform seven robustness tests.
First, we investigate how higher persistence of the shock processes 𝜌 may influence the results. We find that, with high persistence,
the model becomes more noisy for a given level of the standard deviation 𝜎. So, in the limit cycle case, smaller values of 𝜎 are needed
for the model to match the empirical level of comovement. Still, the focus and node cases cannot match empirical comovement for
any value of 𝜌 and 𝜎. The second test is to relax the assumption that the interaction coefficients 𝜖𝑖𝑗 have to be fixed. In particular, we
consider time-varying interaction coefficients 𝜖𝑖𝑗 ,𝑡 , with 𝑡 = 1965, … , 2011, representing varying shares of international trade over the
data availability period. We find virtually identical results to the baseline. The third robustness test considers alternative parameters
for the baseline specification of the interaction function 𝐹 , which is a generic quartic function. It shows that in no case the model
can match empirical comovement under focus or node parameterizations, whereas it can always do so with limit cycles, for an
appropriate level of noise. The results are similar in the fourth robustness test, which considers an alternative logistic specification
of the function 𝐹 . The fifth robustness test considers other bifurcation parameters than the strength of strategic complementarities at
the steady state. It still shows that node and focus parameterizations cannot produce a level of comovement that is as high as in the
data, confirming that it is the type of dynamics and not the strength of strategic complementarities that substantially changes the
level of comovement. The sixth robustness test shows that values of the 𝛼1 and 𝛼2 parameters that are heterogeneous across countries,
although relatively close to the baseline, do not lead to almost any difference in the results. The seventh, and last, robustness test
considers common shocks. We find that for sufficiently strong common shocks the focus parameterization does lead to sufficiently
high comovement, but the node parameterization always fails to do so.

6. Conclusion

This paper contributes to the debate on whether business cycles are exogenous or endogenous. On an empirical level, it has
not been possible to find conclusive evidence for limit cycles or chaos in economic time series, but it has not been possible to rule
them out either. One contribution of this paper is to propose an indirect empirical test, as comovement of economic time series has
different origins when business cycles are exogenous vs. endogenous, originating from shock propagation in the first case and from
synchronization in the second case. Our main empirical result is that a limit cycle model, buffeted by relatively small idiosyncratic
shocks, comes closest to explaining the level of comovement that can be found in macroeconomic time series.
This result comes with several caveats. First, we consider a reduced-form model that interprets the linkages between countries
as international trade, rather than a full-fledged structural model. This prevents us from estimating deep parameters and shock
processes. Second, the unit of our model is a country, but an interesting extension would be to consider sector-country pairs, as
it would allow to consider the impact of sector-specific shocks. Third, we mostly neglect common shocks, as it would be trivial to
obtain a level of comovement that matches the data if the standard deviation of the common shocks is sufficiently large; in our view,
to properly evaluate the role of common shocks, they would need to be estimated with a structural model. Despite these limitations,
we expect our main result to hold in more complicated models: when deterministic underlying dynamics synchronize, comovement
is bound to be larger.
On a theoretical level, we extend complete synchronization theory to deal with an economic problem. The idea that synchro-
nization can help explain comovement has been considered (Krugman, 1996), but never tested empirically. Moreover, almost
everyone thinking about synchronization considers phase synchronization. Using complete synchronization, we do not require that
the underlying time series have a clear periodicity, and we investigate the relative role of exogenous shocks and endogenous
non-linear dynamics in more detail. Beyond macroeconomics, we hope that the tools that we offer in this paper can find wider
applicability, as any disaggregate economic and financial model that can be described by some form of non-linear dynamics can be
studied with the tools of synchronization theory.

11
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

Fig. 6. Stability of dynamics for homogeneous agents. (A) Diagram for stability of 2-dimensional maps. The blue lines correspond to 𝐷 = 1, 𝐷 = 𝑇 − 1, 𝐷 = −𝑇 − 1
and 𝐷 = (𝑇 )2 ∕4. The red line corresponds to varying 𝐹 ′ (1) with the other parameters fixed. (B) Specifications 𝐹1 (⋅) and 𝐹2 (⋅) for 𝑦𝑡 ∈ [0, 2] (see Appendix A.2).

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared
to influence the work reported in this paper.

Appendix A. Analysis of uncoupled dynamics

A.1. Homogeneous dynamics

Here we mathematically describe how the general framework for endogenous fluctuations laid out in Section 3.1 may generate
cyclical dynamics (see Beaudry et al. (2015) for more details). For simplicity, we focus on a single country 𝑖 whose dynamics is
purely driven by internal demand, namely 𝜖𝑖𝑖 = 1, and 𝜖𝑖𝑗 = 0, for all 𝑗 ≠ 𝑖 (autarchic case). This can be interpreted as the country
being composed by a large number of agents with a homogeneous interaction network and we consider the symmetric solution in
which all agents behave alike, as in Beaudry et al. (2015). For ease of notation, in this section we drop the subscript 𝑖.
We first look for the steady state (𝑥, ̂ Following the characterization of 𝑥 and 𝑦 as oscillation variables, as discussed in
̂ 𝑦).
Section 3.2, for convenience we constrain parameters so that 𝑦̂ = 1 is the unique steady state. This means that the steady state
value of the accumulation variable is 𝑥̂ = 1∕𝛿 and the following relation must hold: 𝛼0 = 1 − 𝛼1 ∕𝛿 − 𝛼2 − 𝐹 (1). In the following, we
always select 𝛼0 so that this condition is satisfied. For the steady state to be unique, the slope of the line −𝛼0 + 𝑦𝑡 (1 − 𝛼1 ∕𝛿 − 𝛼2 ) must
be larger than the derivative of 𝐹 at 𝑦̂ = 1, namely 1 − 𝛼1 ∕𝛿 − 𝛼2 > 𝐹 ′ (1), suggesting that strategic complementarities at the steady
state should not be too large. We always choose parameters so that the steady state is unique.
The Jacobian of this system is
( )
1−𝛿 1
 = ′ . (6)
𝛼1 𝛼2 + 𝐹 (1)
The stability( of the steady
) state is completely characterized in terms of the trace 𝑇 = 1 − 𝛿 + 𝛼2 + 𝐹 ′ (1) and determinant
𝐷 = (1 − 𝛿) 𝛼2 + 𝐹 ′ (1) − 𝛼1 of the Jacobian. Following the standard conditions for 2-dimensional maps, stability obtains if 𝐷 < 1,
𝐷 > 𝑇 − 1 and 𝐷 > −𝑇 − 1 (the gray region in Fig. 6A, representing the usual diagram for the stability of 2-dimensional maps). If
𝐷 > (𝑇 )2 ∕4, the eigenvalues of the Jacobian are complex, and so the system admits either damped oscillations or sustained cycles.
If the line 𝐷 = 1 is crossed above 𝐷 = (𝑇 )2 ∕4, the system undergoes a Hopf (Neimark–Sacker) bifurcation. Beaudry et al. (2015)
show that the Hopf bifurcation is supercritical, i.e. the resulting limit cycle is attractive.
To build some intuition into the transition between stability and instability, we fix the parameters 𝛼1 , 𝛼2 , and 𝛿, and vary
the values of 𝐹 ′ (1) so that it varies in the interval 𝐹 ′ (1) ∈ [−1, 1]. In other words, we consider the level of complementarity or
substitutability at the steady state as a bifurcation parameter. Point P in Fig. 6A corresponds to 𝐹 ′ (1) = −1. This corresponds to
strategic substitutability, and indeed under this parameterization the steady state is a stable node, meaning that eigenvalues are
real and of magnitude smaller than one. When the steady state is a node, dynamics converge without oscillations. Next, point Q
corresponds to 𝐹 ′ (1) = 0, i.e. to a transition between strategic complementarity and substitutability. Under these parameters, the
system has a stable focus, with complex eigenvalues of magnitude smaller than one. In this setting, dynamics converge to the steady
state producing damped oscillations. As the level of complementarities increases, the steady state loses stability through a Hopf
bifurcation, up to point R, corresponding to 𝐹 ′ (1) = 1.
The analysis so far has focused on local stability of the steady state. To build some intuition into global dynamics, in Fig. 6B
we plot the function 𝐹 (⋅) for two specifications 𝐹1 (⋅) and 𝐹2 (⋅) detailed in the next section. Under both parameterizations, the
function 𝐹 increases at the steady state, 𝑦𝑡 = 1, but flattens out or even decreases when 𝑦𝑡 becomes large, which leads to smaller
complementarities or even to strategic substitutability and prevents explosive dynamics.

12
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

Fig. 7. Qualitatively different types of dynamics (node, focus, limit cycle and chaos), depending on the strength of strategic complementarities and other
parameters, for both the deterministic and stochastic cases.

A.2. A numerical example

We now show representative time series for the various cases discussed above: stable steady state of type node and focus, and
limit cycle.
To perform numerical simulations, we must specify the function 𝐹 . As our main specification, we consider a generic quartic
function, as it is a very flexible functional form that can have weak or strong complementarity, or even substitutability, at the
steady state. Likewise, it may flatten out or even decrease far from the steady state. Thus, our first specification 𝐹1 , in the most
general terms, is given by
𝐹1 (𝑦) = 𝛽0 + 𝛽1 𝑦 + 𝛽2 𝑦2 + 𝛽3 𝑦3 + 𝛽4 𝑦4 . (7)

Depending on the values of the 𝛽 parameters, the steady state 𝑦̂ = 1 can be a node, a focus, or may be unstable. (In all cases, we
select the 𝛽 parameters so that, conveniently, it is 𝐹1 (1) = 0, as in Beaudry et al. (2015)).
For robustness, we also consider a logistic function as an alternative specification:
1 1
𝐹2 (𝑦) = − . (8)
1 + 𝑒−𝛽(𝑦−1) 2
In the above equation, the term (𝑦 − 1) ensures that the steady state is 𝑦̂ = 1, and the shift 1∕2 makes sure that 𝐹2 (1) = 0, as in
Eq. (7). The larger 𝛽, the stronger the complementarities at the steady state.
Moreover, we select the 𝛼1 , 𝛼2 and 𝛿 parameters for the limit cycle specification so that every cycle lasts about 36 time steps.
If we interpret time steps as quarters, this is in line with the empirical evidence that the US economy tends to undergo a business
cycle every about 9 years (Beaudry et al., 2020). Experimenting with larger values of 𝛼1 (higher dislike for accumulation), we found
out that the model could exhibit chaotic dynamics.
In conclusion, we consider the following baseline parameterizations:

• Node: 𝛼1 = −0.04, 𝛼2 = 0.4, 𝛿 = 0.1, 𝛽0 = −0.19, 𝛽1 = −0.11, 𝛽2 = 0.4, 𝛽3 = 0.2, 𝛽4 = −0.3.

• Focus: 𝛼1 = −0.04, 𝛼2 = 0.4, 𝛿 = 0.1, 𝛽0 = −0.2, 𝛽1 = −0.1, 𝛽2 = 0.1, 𝛽3 = 0.3, 𝛽4 = −0.1.
• Cycle: 𝛼1 = −0.04, 𝛼2 = 0.4, 𝛿 = 0.1, 𝛽0 = −0.5, 𝛽1 = 0.1, 𝛽2 = 0.2, 𝛽3 = 0.5, 𝛽4 = −0.3.
• Chaos: 𝛼1 = −0.35, 𝛼2 = 0.4, 𝛿 = 0.1, 𝛽0 = −0.5, 𝛽1 = 0.1, 𝛽2 = 0.2, 𝛽3 = 0.5, 𝛽4 = −0.3.
We show the time series for each case in Fig. 7, showing both the corresponding deterministic and stochastic dynamics.
(In all cases, the stochastic terms 𝑢𝑖,𝑡 follow an AR(1) process with autocorrelation 𝜌 = 0.3 and standard deviation 𝜎 = 0.1.)
When the steady state is a node, deterministic dynamics converge to the steady state without oscillations, and the corresponding
stochastic dynamics are quite irregular. When the steady state is a focus, convergence happens under damped oscillations, and the
corresponding stochastic dynamics show more structure. Under limit cycles, deterministic dynamics are perfectly periodic, while
stochastic dynamics retain the shape of the cycle with clearer periodicity than under a focus. Finally, under chaotic dynamics, even
in the deterministic case there is no perfect periodicity, and it is even more so in the stochastic case.
In conclusion, the dynamical system in Eq. (1) of the main paper exhibits a variety of dynamics, some that converge towards
stable steady states, in line with the usual narrative of business cycles driven by exogenous shocks, and others that instead follow
limit cycles or chaos, resulting in endogenous business cycles. When adding noise, all dynamics could qualitatively resemble business
cycle fluctuations, which are clearly not perfectly periodic but do show some structure.

Appendix B. Master stability analysis

In this section, we provide the technical details to the master stability analysis discussed in the main text.

13
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

B.1. Master stability functions

In the following, we first derive our key equations, and then provide some examples that clarify the working of our approach.
To be as general as possible, we mostly focus on the abstract formulation of the model (Section 3.1).
Let 𝒔𝑡 = (𝑥𝑠𝑡 , 𝑦𝑠𝑡 ) denote the synchronized state in which all agents behave alike. Let the dynamics of individual agents be given
by 𝑥𝑖,𝑡 = 𝑥𝑠𝑡 + 𝑥𝜉𝑖,𝑡 and 𝑦𝑖,𝑡 = 𝑦𝑠𝑡 + 𝑦𝜉𝑖,𝑡 , where 𝑥𝜉𝑖,𝑡 and 𝑦𝜉𝑖,𝑡 denote a small deviation from the synchronized state, for example due to
( )
idiosyncratic shocks hitting agent 𝑖. Let further 𝝃 𝑡 = 𝑥𝜉1,𝑡 , 𝑦𝜉1,𝑡 , 𝑥𝜉2,𝑡 , 𝑦𝜉2,𝑡 , … , 𝑥𝜉𝑁 ,𝑡 , 𝑦𝜉𝑁 ,𝑡 be the 2𝑁-dimensional vector of deviations.
To compute the evolution of this vector, we take advantage of the small deviations to Taylor-expand the non-linear terms in the
dynamical equations. We illustrate this with a specific example of a star network with three nodes. This network is composed by a
central node, which we denote as node 1, that is linked to two leaf nodes, which we denote as nodes 2 and 3. The leaf nodes are
only connected to the central node. We assume that the dependence on other nodes is 𝜖𝑖𝑗 = 𝜖∕𝑘𝑖 , where 𝑘𝑖 is the degree of node 𝑖,
while the dependency on oneself is 𝜖𝑖𝑖 = 1 − 𝜖. This defines a weighted interaction network with adjacency matrix
⎛1 − 𝜖 𝜖∕2 𝜖∕2 ⎞
⎜ 𝜖 1−𝜖 0 ⎟. (9)
⎜ ⎟
⎝ 0 𝜖 1 − 𝜖⎠

Replacing 𝑥𝑖,𝑡 = 𝑥𝑠𝑡 + 𝑥𝜉𝑖,𝑡 and 𝑦𝑖,𝑡 = 𝑦𝑠𝑡 + 𝑦𝜉𝑖,𝑡 in Eq. (1) with the interaction network above, we get the following system of equations
for the evolution of 𝝃 𝑡 :
⎧𝑥𝜉1,𝑡+1 = (1 − 𝛿)𝑥𝜉1,𝑡 + 𝑦𝜉1,𝑡 ,
⎪ 𝜉 ( ) 𝜉 ( )
𝜉 𝜉 1 𝜉 1 𝜉
⎪𝑦1,𝑡+1 = 𝛼1 𝑥1,𝑡 + 𝛼2 + 𝐹 ′ (𝑦𝑡 ) 𝑦1,𝑡 − 𝜖 𝐹 ′ (𝑦𝑡 ) 𝑦1,𝑡 − 2 𝑦2,𝑡 − 2 𝑦3,𝑡 ,
𝑠 𝑠
⎪ 𝜉 𝜉 𝜉
⎪𝑥2,𝑡+1 = (1 − 𝛿)𝑥2,𝑡 + 𝑦2,𝑡 , ( )
⎨ 𝜉 𝜉 ( ′ 𝑠
) 𝜉 ′ 𝑠 𝜉 𝜉 (10)
⎪𝑦2,𝑡+1 = 𝛼1 𝑥2,𝑡 + 𝛼2 + 𝐹 (𝑦𝑡 ) 𝑦2,𝑡 − 𝜖 𝐹 (𝑦𝑡 ) 𝑦2,𝑡 − 𝑦1,𝑡 ,
⎪ 𝜉 𝜉 𝜉
⎪𝑥3,𝑡+1 = (1 − 𝛿)𝑥3,𝑡 + 𝑦3,𝑡 , ( )
⎪𝑦𝜉 𝜉 ( ′ 𝑠
) 𝜉 ′ 𝑠 𝜉 𝜉
⎩ 3,𝑡+1 = 𝛼1 𝑥3,𝑡 + 𝛼2 + 𝐹 (𝑦𝑡 ) 𝑦3,𝑡 − 𝜖 𝐹 (𝑦𝑡 ) 𝑦3,𝑡 − 𝑦1,𝑡 ,
where we have used the fact that the deviations 𝝃 𝑡 are small to Taylor-expand the function 𝐹 to first order. This formulation suggests
that the evolution of deviations of any single agent is given by the Jacobian of the dynamics corresponding to that agent, plus a
term of interaction with other agents. Extrapolating from this example, it is possible to see that the evolution of the whole vector
𝝃 𝑡 is generically given by
( )
𝝃 𝑡+1 = 𝑰𝑁 ⊗ 𝑱 (𝒔𝑡 ) − 𝜖 𝐹 ′ (𝒔𝑡 )𝑲 𝑳𝑲 ⊗ 𝑯 𝝃 𝑡 , (11)

where:

• 𝑰𝑁 is the 𝑁-dimensional identity matrix.

• 𝑱 (𝒔𝑡 ) is the 2-dimensional Jacobian of the dynamical system.
• ⊗ denotes the Kronecker product, i.e. a 2𝑁-dimensional matrix.
• 𝐹 ′ (𝒔𝑡 ) is the first derivative of 𝐹 , due to the fact that we are considering a first-order approximation around the synchronized
state. √
• 𝑲 is an 𝑁-dimensional square matrix with 1∕ 𝑘𝑖 on the main diagonal and zero everywhere else.
• 𝑳 is the Laplacian of the network. This is a key mathematical property of the network that is widely used in many applications.
𝐿𝑖𝑖 = 𝑘𝑖 and 𝐿𝑖𝑗 = −1 if 𝑖 and 𝑗 are connected. 𝑲 𝑳𝑲 is known as normalized Laplacian, and has similar properties to the
Laplacian.13
• 𝑯 is the 2-dimensional square matrix of connectivity in the dynamical system. Since here the only connectivity is through 𝑦,
it is 𝐻22 = 1 and zero everywhere else.

The problem with Eq. (11) is that the evolution of each component 𝑦𝜉𝑖,𝑡 depends on all other agents 𝑗 to which 𝑖 is connected. The
key trick of the master stability approach is to diagonalize 𝑲 𝑳𝑲 so as to decompose the deviations 𝝃 into orthogonal, uncoupled,
components. Following the terminology in the literature, we call these components eigenmodes. Let
( )
𝝃 𝑡 = 𝑸 ⊗ 𝑰2 𝜻 𝑡 , (12)

where 𝑸 is the matrix of eigenvectors of 𝑲 𝑳𝑲. Here, 𝜻 𝑡 can be interpreted as a projection of 𝝃 𝑡 in the eigenspace. Replacing Eq. (12)
in Eq. (11), we see that 𝜻 𝑡 evolves according to
( )
𝜻 𝑡+1 = 𝑰𝑁 ⊗ 𝑱 (𝒔𝑡 ) − 𝜖 𝐹 ′ (𝒔𝑡 )𝜦 ⊗ 𝑯 𝜻 𝑡 , (13)

13 Here we summarize a few properties of 𝑲 𝑳𝑲. Because the rows sum to zero, one eigenvalue is zero. Moreover, it is well known that the other eigenvalues

are positive and bounded between 0 and 2. The multiplicity of the 0 eigenvalue reflects the number of disconnected clusters in the network. We sort the
eigenvalues in increasing order, so that 𝜆1 = 0, 𝜆2 > 0 if the network is connected, 𝜆3 ≥ 𝜆2 , . . . , 𝜆𝑁 ≥ 𝜆𝑁−1 , 𝜆𝑁 ≤ 2. The eigenvalue 𝜆2 is knows as algebraic
connectivity and the corresponding eigenvector as Fiedler vector. The smaller 𝜆2 , the more the network has a modular structure, in which two or more clusters of
nodes have strong internal connectivity and weak external connectivity. In the case of two clusters, the components of the Fiedler vector are positive for nodes
in a cluster and negative for nodes in the other cluster. Fiedler vectors are commonly used for graph partitioning.

14
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

Fig. 8. Lyapunov exponents 𝜇1 and 𝜇2 as a function of the effective coupling 𝐾, both for a parameter combination that leads to a limit cycle and for a parameter
combination that leads to chaotic dynamics.

where 𝜦 is the matrix with the eigenvalues of 𝑲 𝑳𝑲 on the main diagonal and zero everywhere else. Each of the 𝑁 eigenmodes of
𝜻 𝑡 can then be written as
( )
𝜻 𝑖,𝑡+1 = 𝑱 (𝒔𝑡 ) − 𝐹 ′ (𝒔𝑡 )𝜖 𝜆𝑖 𝑯 𝜻 𝑖,𝑡 . (14)

In this basis, the evolution of each eigenmode 𝑖 only depends upon itself. We stress that 𝑖 now is an eigenvector (corresponding
to the eigenvalue 𝜆𝑖 of the normalized Laplacian of the network) and not an agent. As will be clarified in the examples below, the
evolution of each eigenmode corresponds to(higher-order ) properties of the network. Of course one can retrieve the dynamics of the
agents by applying the transformation 𝝃 𝑡 = 𝑸 ⊗ 𝑰2 𝜻 𝑡 .
The eigenmode 𝑖 = 1 corresponds to the eigenvalue 𝜆𝑖 = 0, and describes dynamics parallel to the synchronization manifold
(i.e., ‘‘in the same direction as the dynamics’’). It captures the phase shift due to a shock hitting the system at a given time 𝑡, as will
be clarified below. The eigenmodes 𝑖 > 1 correspond to dynamics orthogonal to the synchronization manifold. If these dynamics
always converge to zero after the shock, the synchronized state is stable.
Whether the orthogonal dynamics converge to zero is not obvious from Eq. (14). To know if they do, one must numerically
compute the Lyapunov exponents of the system. Letting 𝐾 = 𝜖 𝜆𝑖 denote an effective coupling for eigenmode 𝑖, one calculates
Lyapunov exponents 𝜇 in Eq. (14) for all values of 𝐾 that are typically obtained. (In our case, because 𝜖 ∈ [0, 1] and 𝜆𝑖 ∈ [0, 2], it is
of interest to only consider the interval 𝐾 ∈ [0, 2].) When Lyapunov exponents are negative, the corresponding eigenmodes die out
over time.
The Lyapunov exponents are shown in Fig. 8, for the baseline parameterizations of the model that lead to limit cycles and chaotic
dynamics (Appendix A.2). There are two Lyapunov exponents because the dynamical system is 2-dimensional. Under limit cycles,
both Lyapunov exponents 𝜇1 and 𝜇2 are always negative, except for 𝐾 = 0 in which case 𝜇1 = 0. This means that all eigenmodes
𝑖 > 1 die out, making the synchronized state stable. In other words, if countries followed deterministic limit cycles and were only
occasionally hit by idiosyncratic shocks, their business cycles would perfectly synchronize some time after each shock. Under chaotic
dynamics, there is a region under which 𝜇1 > 0, meaning that, even without exogenous shocks, under some conditions countries’
business cycles would not synchronize.

B.2. Examples

To clarify the formalism introduced in the previous section, we now consider two examples with the abstract framework for
business cycles and a simple interaction network. For simplicity, in all cases we focus on limit cycles rather than chaotic dynamics,
but the following analysis is analogous in the chaotic case. Moreover, we study the effect of single idiosyncratic shocks hitting the
agents in the network at a certain time 𝜏, and study its propagation.

B.2.1. Two connected agents

We start from the simplest possible system to study the interplay between dynamics and shocks in most detail, disregarding the
network structure. We consider two connected agents, agent 1 and agent 2, giving weight 1 − 𝜖 to their internal dynamics and 𝜖 to
the dynamics of the other agent. We simulate the two agents until they reach the synchronized state (which they always do given
that we consider limit cycle dynamics), and then hit agent 1 with a positive shock. We first show the numerical simulations, and
then make sense of the results using the formalism developed in Appendix B.1.
Fig. 9 shows the dynamics in the 60 time steps following the positive shock to agent 1. Panel A compares the original (unshocked)
to the perturbed dynamics, for both agents 1 and 2. It is clear that in the perturbed dynamics the two agents quickly reach
synchronization again, but their business cycles permanently lead those of the original dynamics by a few time steps. Next, panel
B shows how good the approximation of Eq. (11) is. In this panel, the perturbed dynamics of agent 1 (same as previously shown in
panel A) is compared to the linearized approximation 𝑦𝑠𝑡 + 𝑦𝜉𝑖,𝑡 . We see that the exact and approximate dynamics are not identical,
especially at the peaks and troughs, but overall the approximation is pretty faithful.
Having established this, we focus on the dynamics of the shocks in the original basis and in the eigenbasis. In the original basis
(panel C), the terms 𝑦𝜉1,𝑡 and 𝑦𝜉2,𝑡 represent the deviations of agent 1 and agent 2 from the synchronized state. We see that, after

15
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

Fig. 9. Example dynamics: two connected agents are synchronized in a limit cycle, and at time 𝜏 agent 1 gets hit by a positive shock. We show the resulting
dynamics at time 𝑡 > 𝜏. A: Original (unshocked) and perturbed dynamics of agent 1 (blue) and agent 2 (red). B: Perturbed dynamics of agent 1 (as in panel A)
compared with the linearized approximation 𝑦𝑠𝑡 + 𝑦𝜉𝑖,𝑡 . C: Deviations from the synchronized state in the original basis, for both agent 1 (𝑦𝜉1,𝑡 ) and agent 2 (𝑦𝜉2,𝑡 ).
D: Deviations from the synchronized state in the eigenbasis, for both eigenmode 1 (𝑦𝜁1,𝑡 ) and eigenmode 2 (𝑦𝜁2,𝑡 ).

agent 1 is hit by a positive shock at 𝜏 and deviates positively from the synchronized state, also agent 2 deviates positively from
the synchronized state after a few time steps. Then, after a transient, the deviations from the synchronized state become identical
across the two agents and start following a limit cycle. In the eigenbasis (panel D), the deviations 𝑦𝜁1,𝑡 and 𝑦𝜁2,𝑡 correspond to the first
and second eigenmode. While the first eigenmode starts following a limit cycle similar to the one in panel C, the second eigenmode
converges to zero after a negative fluctuation which then becomes slightly positive.
To make sense of these results analytically, we now consider the master stability formalism. For this basic network, the
eigenvalues of 𝑲 𝑳𝑲 are 𝜆1 = 0 and 𝜆2 = 2. Looking at Fig. 8 (left panel), the first eigenmode, corresponding to 𝜆1 = 0 and so 𝐾 = 0,
has largest Lyapunov exponent 𝜇1 = 0. Thus, from Eq. (14), it is 𝜻 1,𝑡+1 = 𝑱 (𝒔𝑡 )𝜻 1,𝑡 : the dynamics of the first eigenmode are always
described by the Jacobian of the synchronized state (black line in Fig. 9D). By contrast, the second eigenmode, corresponding to
𝜆2 = 2 and so 𝐾 = 0.6,14 , has largest Lyapunov exponent 𝜇1 ≈ −0.2. This means that it exponentially converges to zero, approximately
by 20% every time step (gray line in Fig. 9D). √ √
The eigenvectors corresponding to the two eigenmodes are 𝒗1 = (1, 1)∕ 2 and 𝒗2 = (−1, 1)∕ 2. Thus, at time 𝑡 the deviations
from the synchronized state in the original basis and in the eigenbasis are related by the following equations:
( ) ( )
𝑦𝜉1,𝑡 = √1 𝑦𝜁1,𝑡 − 𝑦𝜁2,𝑡 , 𝑦𝜁1,𝑡 = √1 𝑦𝜉1,𝑡 + 𝑦𝜉2,𝑡 ,
2 ( ) 2 ( ) (15)
𝑦𝜉2,𝑡 = √1 𝑦𝜁1,𝑡 + 𝑦𝜁2,𝑡 , 𝑦𝜁2,𝑡 = √1 −𝑦𝜉1,𝑡 + 𝑦𝜉2,𝑡 .
2 2

The second column in Eq. (15) helps to interpret what the eigenmodes represent. The first eigenmode is proportional to the sum
of the deviations of the two agents. As it never converges to zero because 𝑦𝜉1,𝑡 and 𝑦𝜉2,𝑡 keep fluctuating (Fig. 9C), it represents the
permanent phase shift in the dynamics due to the shock, also described in the literature as a perturbation that is ‘‘parallel to the
synchronization manifold’’. The second eigenmode, which is ‘‘transverse to the synchronization manifold’’, is proportional to the
difference of the deviations of the two agents from the synchronized state. As it eventually converges to zero, this suggests that
the two agents always reach synchronization √ again after the shock. This can also be seen from the first column in Eq. (15): when
𝑦𝜁2,𝑡 = 0, it is always 𝑦𝜉1,𝑡 = 𝑦𝜉2,𝑡 = 𝑦𝜁1,𝑡 ∕ 2.
We can use this example to study analytically how other shocks would impact the dynamics. That is, we consider generic 𝑢1 and
𝑢2 hitting the agents at time 𝜏, so that 𝑦𝜉1,𝜏 = 𝑢1 and 𝑦𝜉2,𝜏 = 𝑢2 . From the second column of (15) we can see that if 𝑢1 = 𝑢2 = 𝑢, 𝑦𝜁2,𝜏 = 0,
and so for all following time steps 𝑡 ≥ 𝜏, 𝑦𝜁2,𝑡 = 0: The eigenmode transverse to the synchronization manifold never takes positive
values. In other words, because the shock is common, √ it does not cause any relative difference in phase between the two agents. On
the contrary, the first eigenmode is maximal, 𝑦𝜁1,𝜏 = 2𝑢: the effect of the shock is to only shift the phase, by the maximum amount.
If 𝑢 > 0, the phase is shifted forward; if 𝑢 < 0, it is shifted backwards. Consider now the reverse case 𝑢1 = −𝑢2 = 𝑢. Now 𝑦𝜁1,𝜏 = 0 and

14 In this example, we take 𝜖 = 0.3.

16
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

Fig. 10. Example dynamics in a network with six agents and two cliques of three nodes each. The top part of the figure shows the network, the adjacency matrix
𝑨 and the normalized Laplacian 𝑲 𝑳𝑲. It also shows the eigenvalues 𝝀, the eigenvectors 𝑸 of the normalized Laplacian (each column of 𝑸 is the eigenvector
corresponding to the eigenvalue shown above) and the inverse matrix of eigenvectors 𝑸−1 . The bottom part of the figure shows the evolution of the deviations
in the original basis and in the eigenbasis.

√
𝑦𝜁2,𝜏 = 2𝑢. In this case, the transverse eigenmode is maximal, while the parallel eigenmode is null. So, after a transient in which
the phases are different, they synchronize again, with no consequence of the shock in terms of permanent phase shift (this is the
only case which does not imply a permanent phase shift).

B.2.2. Six agents with two cliques of three nodes each

To understand the effect of the network structure on shock propagation and synchronizing non-linear dynamics, we consider the
network in Fig. 10. This network is composed of two cliques of three nodes each, connected by a link between two of the nodes
of the cliques. These two nodes are colored lighter to highlight the stronger connectivity to the other clique. This network can be
thought of as a stylized model of six countries in two continents.
At time 𝜏, we apply the following shock vector:
( )
𝑦𝜉1,𝜏 , 𝑦𝜉2,𝜏 , 𝑦𝜉3,𝜏 , 𝑦𝜉4,𝜏 , 𝑦𝜉5,𝜏 , 𝑦𝜉6,𝜏 = (0.20, 0.12, 0.16, −0.12, −0.24, 0.00). (16)

This shock vector can be thought of as the combination of idiosyncratic shocks hitting all agents differently, combined with a positive
shock hitting the clique with agents 1,2 and 3, and a negative shock hitting the other clique. In the eigenbasis, this shock vector
corresponds to
( )
𝑦𝜁1,𝜏 , 𝑦𝜁2,𝜏 , 𝑦𝜁3,𝜏 , 𝑦𝜁4,𝜏 , 𝑦𝜁5,𝜏 , 𝑦𝜁6,𝜏 = (0.05, 0.35, 0.00, 0.06, −0.08, −0.17). (17)

It should be noted that the shock in the second eigenmode (0.35) is the largest one.
The evolution of the shocks is shown at the bottom of Fig. 10. In the panel on the left, we show the shocks in the original basis,
i.e. each line corresponds to an agent. It is clear that the dynamics vary across the two cliques, but they are very similar within
each of the two cliques. The cyan and orange lines, corresponding to the two ‘‘bridge’’ agents, are the ones that show the closest
dynamics to the ones in the other clique. Over time, all dynamics converge again to the synchronized state, with a permanent phase
shift (not visible in the figure).
The evolution of the shocks in the eigenbasis is shown in the right panel. The black line is the eigenmode which is parallel to
the synchronization manifold. The gray line corresponds to the second eigenmode, the one with largest initial value, and decays
very slowly. The green lines correspond to the third to sixth eigenmodes, and decay very quickly.
The interpretation for the eigenmodes is now clear. The second eigenmode corresponds to synchronization across cliques. Indeed,
because 𝑦𝜁2,𝑡 = 0.43𝑦𝜉1,𝑡 + 0.43𝑦𝜉2,𝑡 + 0.38𝑦𝜉3,𝑡 − 0.38𝑦𝜉4,𝑡 − 0.43𝑦𝜉5,𝑡 − 0.43𝑦𝜉6,𝑡 , it is maximal in absolute value when shocks are positive in
one clique and negative in the other one. This eigenmode takes time to decay, because 𝐾 = 𝜖 𝜆2 = 0.3 ⋅ 0.2 = 0.06 is relatively small
and so, as shown in Fig. 8, the largest Lyapunov exponent is not very negative (𝜇1 ≈ −0.02). This means that synchronization across

17
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

cliques takes time, as the eigenmode decays by approximately 2% every time step. The third to sixth eigenmodes correspond to
synchronization within cliques, and decay quicker (𝐾 = 𝜖 𝜆𝑖 , 𝑖 = 3, … , 6 ranges from 0.35 to 0.49, so the largest Lyapunov exponent
𝜇1 ranges between −0.14 and −0.18). This indicates that synchronization within cliques happens much quicker.
This phenomenon is general, because the second eigenvector of the Laplacian—the Fiedler vector—takes positive values in a
cluster of nodes, and negative values in the other one. If there were 𝑀 cliques, or more loosely 𝑀 clusters of highly connected
nodes weakly connected across one another, there would be 𝑀 − 1 small eigenvalues and then a large gap to the next eigenvalue.
The eigenmodes corresponding to the small eigenvalues would indicate, as in the case above, synchronization across clusters. The
remaining 𝑁 − 𝑀 eigenvalues would be much larger, and the associated eigenmodes would correspond to synchronization within
clusters.15

B.2.3. Weighted directed networks

Due to the way in which we built the networks in the examples above, the analysis was performed for unweighted, undirected
networks. However, nothing prevents us to apply it to the case of directed, weighted networks described by a generic connectivity
matrix 𝑾 (so that all rows are normalized to one): instead of calculating the eigenvalues of 𝑲 𝑳𝑲 one calculates those of 𝑰𝑁 − 𝑾 .
The only difference to the networks considered in the previous examples is that they have the convenient mathematical property
that all eigenvalues of the normalized Laplacian are real, positive and bounded between zero and two. When considering a generic
interaction network, the eigenvalues could instead be complex. Indeed, in the example considered in Section 4.2 we find that
the eigenvalues have a very small imaginary part. Applying master stability analysis in the case of complex eigenvalues is almost
identical to the case of real eigenvalues, except that one computes complex Floquet exponents (instead of real Lyapunov exponents)
and then verifies if the complex values of 𝐾 = 𝜖 𝜆𝑖 are inside the stable region in the complex plane (Pecora and Carroll, 1998).
Here, for simplicity we just consider the real part of the eigenvalues and eigenvectors, since the imaginary part is small relative to
the real part.

Appendix C. International trade data

In the international trade network interpretation of the model (Section 3.2), the interaction coefficients between countries
represent the share of a country’s output that is exported to another country. Importantly, the model also features ‘‘self-loops’’,
representing the share of a country’s output that remains within the same country. This internal demand, coming from domestic
firms and households, is generally much higher than international demand. This makes it possible to set the problem as one of
endogenous business cycles that originate within countries and synchronize internationally, as opposed to a problem in which
business cycles originate because of international linkages (as previously stressed, independent oscillations are a necessary ingredient
to apply synchronization theory). Further to the importance of internal demand, another key feature of the model is that countries
demand fixed proportions of imports from other countries. These two features of the model are crucial to determine which data
should be used to calibrate the interaction network.
To have a consistent framework that includes information on internal demand, exports and imports, we turn to international
input–output tables. Although we do not disaggregate our model at the industry-country level and do not distinguish between
intermediate and final products, we prefer to use international input–output data over computing internal demand from raw import
and export bilateral data and separately collected gross output data. Moreover, to test how good the assumption of fixed proportions
of imports is, we need international input–output data that span as many decades as possible.
For these reasons, the ideal dataset is a merge of the recently released Long-Run World Input-Output Database (LR-WIOD),
covering the period 1965-2000 (Woltjer et al., 2021), with the established 2013 release of the WIOD, spanning the years 1995-
2011 (Timmer et al., 2015). Although the creators of these databases warn against using them together at a detailed industrial
level because of differences in their construction (Woltjer et al., 2021), they say that ‘‘aggregate levels and trends are likely to
be comparable’’. Selecting the countries that are available in both datasets, we end up with 24 countries and a Rest of the World
aggregate.
Fig. 11 shows the dynamics of the share of internal demand for all the countries we consider. We show both data from LR-WIOD
and data from WIOD 2013, including data from both datasets in the period 1995–2000. This figure has three take-away messages.
First, the trends in the overlapping period do indeed look similar across the two datasets. This justifies our choice to merge the
two datasets and supports the hypothesis of Woltjer et al. (2021). To perform the merging, we use LR-WIOD up to 1994 and WIOD
2013 from 1995 on, after adjusting the WIOD 2013 data so that the levels in 1995 match with those of LR-WIOD (note however
that the levels were very close even without this transformation). Second, except for some irregular high-frequency movement,
the share of internal demand changes slowly, generally decreasing. Similarly, we find that the fraction of bilateral exports over
countries’ output changes slowly over time (not shown), compared to business cycle frequency. We think that these results support
our assumption of assuming fixed import shares as a first approximation. We will show that our results are robust to assuming
time-varying interaction coefficients (Appendix D.2.2). Third, the share of internal demand is almost always above 0.7, except for
Ireland in recent years. This supports our view that business cycles originate from strategic complementarities of demand within
countries, and then synchronize through international linkages, rather than originating from international demand.
We show an example of the interaction network 𝜖𝑖𝑗 in Fig. 12, for the most recent year available in our dataset.

15 There could also be a hierarchy of clusters. Imagine for example that there were two main clusters, and two sub-clusters within each cluster. Here, the Fiedler

vector would divide between the two main clusters, but then the following eigenmode (corresponding to a larger eigenvalue) would correspond to synchronization
between sub-clusters. There would be a time-scale separation so that synchronization would first occur within sub-clusters, then across sub-clusters within the
same main cluster, and finally across main clusters.

18
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

Fig. 11. Dependency on internal demand (self-loops), 1965–2011, for all countries in our sample. Solid lines are taken from LR-WIOD, whereas dashed lines
are taken from WIOD 2013.

Fig. 12. Interaction network in 2011. The position of the node is equal to the countries’ geographical centroid. Node size is proportional to countries’ gross
output, while link size is proportional to interaction coefficients 𝜖𝑖𝑗 (for visualization purposes, we applied nonlinear transformations to both node and link size).
RoW indicates the Rest of the World.

Appendix D. Empirical application

D.1. Additional result on pairwise correlations

Fig. 13 shows that, when the model is parameterized as a node, it not only produces very low correlations, but there is also no
clear pattern showing that pairs of countries with similar components of the Fiedler vector have higher pairwise correlation.

D.2. Robustness

We now test if the main result is robust to alternative specifications.

D.2.1. Shock persistence parameter

So far, we have used a value 𝜌 = 0.3 for the autocorrelation of idiosyncratic shocks. What happens when this persistence parameter
varies?

19
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

Fig. 13. Pairwise Pearson correlation coefficients between the dynamics in different countries when the model is parameterized as a node, averaged across 100
replications.

Fig. 14. Heatmaps showing the mean Pearson correlation coefficients across all country pairs and 20 simulation runs, for several values of the persistence
parameter 𝜌 and the standard deviation of idiosyncratic shocks 𝜎. The color bar is designed so that white corresponds to the mean of the empirical correlation
of employment and value added. We show three heatmaps, for each of the main parameterization of the model.

Fig. 14 shows that, as we increase the persistence parameter, correlation coefficients become lower in the limit cycle case. This
is not surprising, as higher persistence makes the shock processes lead the deterministic dynamics farther away from the values it
would take in the absence of shocks. However, for all values of 𝜌, there is a value of 𝜎 that makes the limit cycle model match the
empirical level of comovement. By contrast, in the focus and node cases there is not a big effect of the persistence parameter, and
in no case the model can produce correlations that are as high as in the data.

D.2.2. Time-varying network

From the description of the model in Section 3.2, we have been assuming that the interaction coefficients 𝜖𝑖𝑗 between country 𝑖
and country 𝑗 are fixed over time. This was in part supported by the discussion of the data in Section 4.2.1, where we argued that
the share of a country’s output that is exported to other countries varies slowly over time, at lower frequencies than typical business
cycle frequencies. Here, we relax the assumption of fixed interaction coefficients, and directly input to the model time-varying
interaction coefficients 𝜖𝑖𝑗 ,𝑡 , with 𝑡 = 1965, … , 2011, corresponding to the available data.
As we can see in Fig. 15, the results are virtually unchanged from Fig. 3, showing that indeed assuming fixed interaction
coefficients is a good approximation for this problem.

D.2.3. Other parameter values

Next, we test if our results are also valid for parameters that are different from the baseline parameters indicated in Appendix A.2.
To do so, still using the 𝐹1 specification of the interaction function (Eq. (7)), we check for alternative values of strategic
complementarities at the steady state by randomly picking the 𝛽 parameters. We restrict the parameter values so that they do

20
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

Fig. 15. Mean Pearson correlation across the decision variables 𝑦 of the countries in our sample as a function of the standard deviation of idiosyncratic shocks,
in the model and in the data. The interpretation is the same as in Fig. 3.

Fig. 16. Left: Same as in Figs. 3 and 15, except that we only show results for many more parameter values than the baseline. We show the mean across 20
simulation runs (and not the error bands) and 10 combinations of the 𝛽 parameters for each of the cycle, focus and node cases. Right: same diagram as in
Fig. 6, for the same combinations of the 𝛽 parameters shown on the left.

not produce multiple equilibria and do not exit the stable region in Fig. 6 from below, as this would lead to a flip bifurcation and
a limit cycle of period 2, which is unlikely to be relevant in economic analysis (Beaudry et al., 2015).
As we show in Fig. 16 (left panel), the key results are robust to alternative parameter specifications, as only in the limit cycle
case the model can produce correlations that are as high as in the data. Depending on the specific parameterization, the general
level of comovement in the limit cycle case can be higher or lower than the baseline for a given strength of idiosyncratic shocks.
This is because, depending on the parameters, limit cycle fluctuations have different amplitude, and so can be more or less perturbed
by idiosyncratic shocks with a given standard deviation. It is also possible to note that we do not see a pattern for the focus case
that is similar to Fig. 3, as in this case the maximal correlation is 0.04, much smaller than 0.15 as in Section 5.2. This is because,
in that case, we selected the parameters so that the focus was very close to the boundary of the unstable region, and so oscillations
driven by exogenous shocks would die out very slowly. This is not the case for the parameter combinations that we select here in
the focus case, which are well within the stable region (Fig. 16, right panel).

D.2.4. Logistic specification of the interaction function

Finally, we also vary the function 𝐹 mediating the effect of interactions, using the alternative logistic specification in Eq. (8).
In this case, there is only one free parameter, namely the slope at the steady state 𝛽, proportional to the strength of strategic
complementarities. Thus, we simply consider 10 equally spaced values for 𝛽 for each of the cycle, focus and node regimes.
Fig. 17 is very similar to Fig. 16, mainly suggesting that the specification of the function 𝐹 as a quartic is not particularly
restrictive. It is also interesting to note that, for the focus parameterization that is very close to the Hopf bifurcation, and for small

21
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

Fig. 17. Left: Same as Fig. 16, except that for each of the cycle, focus and node cases we select 10 values of the 𝛽 parameters, representing the slope of the
logistic function at the steady state.

levels of noise, comovement in the model can match the empirical comovement of employment, although it is still too small to
match the empirical comovement of GDP.

D.2.5. Bifurcation parameters unrelated to strategic complementarities

Throughout the paper, by varying the 𝛽 parameters, we have used the strength of strategic complementarities at the steady state
as a bifurcation parameter: as strategic complementarities increase, the steady state moves from being a node, to being a focus, to
eventually become unstable and at the center of a limit cycle or chaotic attractor.
Our choice of focusing on strategic complementarities as a bifurcation parameter leads to a possible criticism: what if it is the
strength of strategic complementarities, rather than the type of dynamics, that causes higher comovement?
To address this issue, we fix the parameterization of the limit cycle considered as a baseline, and consider two alternative set of
parameters leading to the steady state being a node and a focus. For these two alternative parameter sets, we keep the 𝛽 parameters
fixed (and so the regime of strategic complementarities), and vary instead 𝛼1 , 𝛼2 and 𝛿 to get different stability properties of the
steady state. Specifically, we use:

• Node: 𝛼1 = −0.11, 𝛼2 = 0.2, 𝛿 = 0.7

• Focus: 𝛼1 = −0.04, 𝛼2 = 0.2, 𝛿 = 0.1

Thus, every difference in the resulting comovement must be due to the different dynamical regimes, rather than to the strength of
strategic complementarities.
Fig. 18 shows that the results are very similar to the previous figures. Thus, it is the type of dynamics, not the strength of strategic
complementarities, that leads to different levels of comovement.

D.2.6. Parameters heterogeneous across countries

Throughout the paper we assumed that all nodes/countries are described by the same parameters. This was mostly out of
analytical convenience, as the theory of complete synchronization developed in Appendix B is cleanest in this case (although it
can accommodate limited parameter heterogeneity, see Pikovsky et al. 2003). However, nothing prevents us from testing the effect
of country heterogeneity through simulation.
To do so, we select the 𝛼1 and 𝛼2 parameters, and allow them to vary across countries. Specifically, before we start the simulations
we draw a random number from a normal distribution with mean zero and standard deviation 10% of the baseline parameter values
(that is, those listed in Appendix A.2), and add these random numbers to the baseline parameters.
We report results in Fig. 19. This is almost indistinguishable from Fig. 3, suggesting that heterogeneity of parameters plays a
limited role, at least when it is moderate.

D.2.7. Common shocks

Trivially, if the standard deviation of common shocks is large compared to the standard deviation of idiosyncratic shocks, all
dynamical regimes (cycles, focus and node) would yield very similar level of comovement. We do not think that this situation is
interesting, nor relevant: really global events such as the Covid-19 pandemic are rare, and so country-specific shocks should be more
prevalent.

22
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

Fig. 18. Same as figures above, but in this case the parameters that lead to the node and focus preserve strategic complementarities, as detailed in Appendix D.2.5.

Fig. 19. Same as figures above, but in this case the parameters 𝛼1 and 𝛼2 are heterogeneous across countries. See Appendix D.2.6 for more details.

Therefore, we set the standard deviation of idiosyncratic shocks 𝜎𝑢 to 0.08, a value at which the limit cycle regime matches the
empirical level of comovement. Next, we vary the standard deviation of common shocks 𝜎𝑣 from 0 to 0.04, following the assumption
that common shocks can be at most half as large as country-specific shocks.
Fig. 20 shows the results. We find that for only the largest values of 𝜎𝑣 the node regime can match the level of comovement
of employment, while it cannot match the level of comovement of GDP for any level of common shocks. Instead, the focus regime
matches both the level of comovement of employment and GDP for sufficiently large values of 𝜎𝑣 in this range.
These results confirm that common shocks are a powerful method to obtain high comovement. However, we emphasize that this
is a great simplification, as in our view modeling common shocks as a common AR(1) process is a very strong assumption—properly
modeling common shocks requires using a more structural model.

Data availability

Replication materials: [Link]

23
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

Fig. 20. Same as figures above, but in this case we vary the standard deviation of common shocks, for fixed standard deviation of idiosyncratic shocks. See
Appendix D.2.7 for more details.

References

Arkolakis, C., Ramanarayanan, A., 2009. Vertical specialization and international business cycle synchronization. Scand. J. Econ. 111, 655–680.
Asano, Y.M., Kolb, J.J., Heitzig, J., Farmer, J.D., 2021. Emergent inequality and business cycles in a simple behavioral macroeconomic model. Proc. Natl. Acad.
Sci. 118, e2025721118.
Backus, D.K., Kehoe, P.J., Kydland, .F.E., 1992. International real business cycles. J. Polit. Econ. 100, 745–775.
Barnett, W.A., Chen, P., 1988. The aggregation-theoretic monetary aggregates are chaotic and have strange attractors. In: Barnett, W.A., Berndt, E.R., White, H.
(Eds.), Dynamic Econometric Modeling. Cambridge University Press, pp. 199–246.
Barnett, W.A., Serletis, A., 2000. Martingales, nonlinearity, and chaos. J. Econ. Dyn. Control 24, 703–724.
Baxter, M., King, R.G., 1999. Measuring business cycles: approximate band-pass filters for economic time series. Rev. Econ. Stat. 81, 575–593.
Baxter, M., Kouparitsas, M.A., 2005. Determinants of business cycle comovement: a robust analysis. J. Monetary Econ. 52, 113–157.
Beaudry, P., Galizia, D., Portier, F., 2015. Reviving the Limit Cycle View of Macroeconomic Fluctuations. Technical report, National Bureau of Economic Research.
Beaudry, P., Galizia, D., Portier, F., 2020. Putting the cycle back into business cycle analysis. Amer. Econ. Rev. 110, 1–47.
Boldrin, M., Woodford, M., 1990. Equilibrium models displaying endogenous fluctuations and chaos: a survey. J. Monetary Econ. 25, 189–222.
Bonart, J., Bouchaud, J.-P., Landier, A., Thesmar, D., 2014. Instabilities in large economies: aggregate volatility without idiosyncratic shocks. J. Stat. Mech.
Theory Exp. 2014, P10040.
Brenner, T., Weidlich, W., Witt, U., 2002. International co-movements of business cycles in a phase-locking model. Metroeconomica 53, 113–138.
Brock, W.A., 1986. Distinguishing random and deterministic systems: Abridged version. J. Econom. Theory 40, 168–195.
Bullard, J., 1994. Learning equilibria. J. Econom. Theory 64, 468–485.
Bulow, J.I., Geanakoplos, J.D., Klemperer, P.D., 1985. Multimarket oligopoly: Strategic substitutes and complements. J. Polit. Econ. 93, 488–511.
Calderon, C., Chong, A., Stein, E., 2007. Trade intensity and business cycle synchronization: Are developing countries any different? J. Int. Econ. 71, 2–21.
Cesa-Bianchi, A., Imbs, J., Saleheen, J., 2019. Finance and synchronization. J. Int. Econ. 116, 74–87.
Christiano, L.J., Fitzgerald, T.J., 2003. The band pass filter. Internat. Econom. Rev. 44, 435–465.
Clark, T.E., Wincoop, [Link]., 2001. Borders and business cycles. J. Int. Econ. 55, 59–85.
Cooper, R., John, A., 1988. Coordinating coordination failures in Keynesian models. Q. J. Econ. 44, 1–463.
De Grauwe, P., 2011. Animal spirits and monetary policy. Econom. Theory 47, 423–457.
De Haan, J., Inklaar, R., Jong-A-Pin, R., 2008. Will business cycles in the euro area converge? A critical survey of empirical research. J. Econ. Surv. 22, 234–273.
Di Giovanni, J., Levchenko, A.A., Mejean, I., 2018. The micro origins of international business-cycle comovement. Amer. Econ. Rev. 108, 82–108.
Fazzari, S., Ferri, P., Greenberg, E., 2008. Cash flow, investment, and Keynes–Minsky cycles. J. Econ. Behav. Organ. 65, 555–572.
Feenstra, R.C., Inklaar, R., Timmer, M.P., 2015. The next generation of the penn world table. Am. Econ. Rev. 105, 3150–3182.
Feldkircher, M., Huber, F., 2016. The international transmission of US shocks—Evidence from Bayesian global vector autoregressions. Eur. Econ. Rev. 81, 167–188.
Foley, D.K., 1992. A contribution to the theory of business cycles. Q. J. Econ. 107, 1071–1088.
Frankel, J.A., Rose, A.K., 1998. The endogenity of the optimum currency area criteria. Econom. J. 108, 1009–1025.
Goodwin, R.M., 1951. The nonlinear accelerator and the persistence of business cycles. Econometrica 1–17.
Goodwin, R.M., 1967. A Growth Cycle. Cambridge University Press.
Gualdi, S., Bouchaud, J.-P., Cencetti, G., Tarzia, M., Zamponi, F., 2015. Endogenous crisis waves: stochastic model with synchronized collective behavior. Phys.
Rev. Lett. 114, 088701.
Guerini, M., Luu, D.T., Napoletano, M., 2023. Synchronization patterns in the European Union. Appl. Econ. 55, 2038–2059.
Haxholdt, C., Kampmann, C., Mosekilde, E., Sterman, J.D., 1995. Mode-locking and entrainment of endogenous economic cycles. Syst. Dyn. Rev. 11, 177–198.
Hicks, J.R., 1950. A Contribution to the Theory of the Trade Cycle. Oxford University Press.
Hommes, C.H., Manzan, S., 2006. Comments on testing for nonlinear structure and chaos in economic time series. J. Macroecon. 28, 169–174.
Hsu, C.-C., Wu, J.-Y., Yau, R., 2011. Foreign direct investment and business cycle co-movements: The panel data evidence. J. Macroecon. 33, 770–783.
Hwang, S.H., Kim, Y.J., 2021. International output synchronization at different frequencies. Econ. Model. 104, 105627.
Imbs, J., 2004. Trade, finance, specialization, and synchronization. Rev. Econ. Stat. 86, 723–734.
Imbs, J., 2006. The real effects of financial integration. J. Int. Econ. 68, 296–324.
Inklaar, R., Jong-A-Pin, R., De Haan, J., 2008. Trade and business cycle synchronization in OECD countries—A re-examination. Eur. Econ. Rev. 52, 646–666.
Johnson, R.C., 2014. Trade in intermediate inputs and business cycle comovement. Am. Econ. J.: Macroecon. 6, 39–83.
Judd, K.L., 1985. On the performance of patents. Econometrica 567–585.
Kaldor, N., 1940. A model of the trade cycle. Econom. J. 78–92.

24
M. Pangallo Journal of Economic Behavior and Organization 229 (2025) 106827

Kalemli-Ozcan, S., Papaioannou, E., Peydro, J.-L., 2013. Financial regulation, financial globalization, and the synchronization of economic activity. J. Finance
68, 1179–1228.
Kleinert, J., Martin, J., Toubal, F., 2015. The few leading the many: Foreign affiliates and business cycle comovement. Am. Econ. J.: Macroecon. 7, 134–159.
Kose, M.A., Yi, K.-M., 2006. Can the standard international business cycle model explain the relation between trade and comovement? J. Int. Econ. 68, 267–295.
Krugman, P., 1996. The Self-Organizing Economy. Blackwell Publishers.
Kydland, F.E., Prescott, E.C., 1982. Time to build and aggregate fluctuations. Econometrica 50, 1345–1370.
Liao, W., Santacreu, A.M., 2015. The trade comovement puzzle and the margins of international trade. J. Int. Econ. 96, 266–288.
Matsuyama, K., 1999. Growing through cycles. Econometrica 67, 335–347.
Matsuyama, K., 2007. Credit traps and credit cycles. Amer. Econ. Rev. 97, 503–516.
Matsuyama, K., Sushko, I., Gardini, L., 2014. Globalization and Synchronization of Innovation Cycles. Centre For Macroeconomics.
Nikolaidi, M., Stockhammer, E., 2017. Minsky models: A structured survey. J. Econ. Surv. 31, 1304–1331.
Pangallo, M., 2019. Endogenous Fluctuations in Game Theory and Macroeconomics (Ph.D. Dissertation). University of Oxford.
Pecora, L.M., Carroll, T.L., 1990. Synchronization in chaotic systems. Phys. Rev. Lett. 64 (821).
Pecora, L.M., Carroll, T.L., 1998. Master stability functions for synchronized coupled systems. Phys. Rev. Lett. 80 (2109).
Pikovsky, A., Kurths, J., Rosenblum, M., Kurths, J., 2003. Synchronization: A Universal Concept in Nonlinear Sciences, vol. 12, Cambridge University Press.
Ravn, M.O., Uhlig, H., 2002. On adjusting the Hodrick–Prescott filter for the frequency of observations. Rev. Econ. Stat. 84, 371–376.
Scheinkman, J.A., LeBaron, B., 1989. Nonlinear dynamics and stock returns. J. Bus. 62, 311–337.
Selover, D.D., Jensen, R.V., 1999. Mode-locking and international business cycle transmission. J. Econom. Dynam. Control 23, 591–618.
Shintani, M., Linton, O., 2004. Nonparametric neural network estimation of Lyapunov exponents and a direct test for chaos. J. Econometrics 120, 1–33.
Shleifer, A., 1986. Implementation cycles. J. Polit. Econ. 94, 1163–1190.
Silverberg, G., Lehnert, D., 1993. Long waves and evolutionary chaos in a simple schumpeterian model of embodied technical change. Struct. Change Econ. Dyn.
4, 9–37.
Soares, M.J., et al., 2011. Business cycle synchronization and the euro: A wavelet analysis. J. Macroecon. 33, 477–489.
Strogatz, S., 2004. Sync: The Emerging Science of Spontaneous Order. Penguin UK.
Timmer, M.P., Dietzenbacher, E., Los, B., Stehrer, R., De Vries, G.J., 2015. An illustrated user guide to the world input–output database: the case of global
automotive production. Rev. Int. Econ. 23, 575–605.
Woltjer, P., Gouma, R., Timmer, M.P., 2021. Long-run world input-output database: version 1.1 sources and methods. GGDC Research Memorandum 190.

25