Measuring the output gap using large datasets

Matteo Barigozzi Matteo Luciani

London School of Economics Federal Reserve Board
[email protected] [email protected]

October 23, 2019

Abstract

We decompose US aggregate output into potential output and the output gap by
means of a non-stationary dynamic factor model estimated on a large dataset of
macroeconomic indicators, combined with a non-parametric trend-cycle decomposi-
tion of the factors. We find that: (1) from the mid-90s to 2008 the US economy
operated above its potential; (2) as of 2018:Q4 the labor market was tighter than the
goods and services market; and (3) our output gap measure revises modestly in real-
time. Due to its purely data driven nature, our measure is a natural complementary
tool to the theoretical models used in policy institutions.

JEL classification: C32, C38, C55, E0.

Keywords: Output Gap; Non-stationary Approximate Dynamic Factor Model; Trend-
Cycle Decomposition.

∗
We thank for helpful comment: Stephanie Aaronson, Gianni Amisano, Andrew Figura, Manuel
Gonzales-Astudillo, James Mitchell, Filippo Pellegrino, John Roberts, and Andrea Stella.

Disclaimer: the views expressed in this paper are those of the authors and do not necessarily reflect the
views and policies of the Board of Governors or the Federal Reserve System.

Electronic copy available at: https://ssrn.com/abstract=3217816

1 Introduction
A fundamental issue in monetary policy and macroeconomics is to decompose aggregate
output into potential output and the output gap. Indeed, the former is a crucial input for
long-term projections; the latter is an important tool to measure slack in the economy, as
well as a potential gauge of future inflation. However, both these quantities are unobserv-
able; hence, a statistical or economic model must be used to estimate them from economic
data.
In this paper, we measure potential output and the output gap by fitting a non-
stationary dynamic factor model on a large dataset of US macroeconomic indicators. We
estimate the model by Maximum Likelihood via the Expectation-Maximization (EM) al-
gorithm, and we disentangle common trends from common cycles via the eigenanalysis of
the long-run covariance matrix of the latent common factors. Our measure of potential
output is that part of aggregate output that is explained by the common trends, while
our measure of the output gap is that part of aggregate output that is explained by the
common cycles.
Our approach differs from the one most commonly used in policy institutions—for
example, at the Congressional Budget Office (CBO)—which is the so-called “production
function approach” (Kiley, 2013). Differently from the CBO, which estimates the output
gap mainly based on theoretical macroeconomic models, our estimate is entirely data-
driven. Indeed, we do not make any structural macroeconomic assumption; instead, we
exploit the statistical properties of the data. This methodological difference makes our
measure a natural complementary tool to the theoretical models used in policy institutions.
In Figure 1, we provide a synthetic view of our main findings by showing our estimate
of the output gap together with the one produced by the CBO. As we can see, these two
measures look very similar in that the dating of the turning points perfectly coincides.
However, the CBO estimates that the output gap closed just in 2006, whereas our model
suggests a persistent overheating of the US economy from the mid-90s to 2008. This
difference comes from the fact that the two measures weight in a different way the signal
coming from inflation, with the CBO putting considerable weight on inflation, and our
model putting a smaller weight.
Our framework allows us also to estimate the unemployment rate gap (i.e., the part
of the unemployment rate that is driven by the common cycles). In general, we find
that the unemployment gap moves together with the output gap, a feature that in the
CBO framework is imposed via the Okun’s law. However, in 2018:Q4 our estimate of the

Electronic copy available at: https://ssrn.com/abstract=3217816

 Figure 1: Output gap
6

-2

-4
CBO
-6 BL

-8

-10

1970 1975 1980 1985 1990 1995 2000 2005 2010 2015

The blue line is the the output gap estimated by the CBO, while the red line is our estimate.
The shaded areas around our estimate are 68% and 84% confidence bands, respectively.

unemployment rate gap signals that the economy is operating above its potential, while our
estimate of the output gap suggests that the US economy was operating at its potential.
In other words, our model suggests that the labor market is tighter than the goods and
services market at the end of 2018.
Finally, we investigate the real-time properties of our model by conducting a quasi-
real-time evaluation. We find that our output gap measure revises modestly, and none of
the results presented in the previous paragraphs would have been overturned in real time.
In the last 30 years, many papers have suggested different ways to obtain a trend-cycle
decomposition of GDP, both based on univariate methods (e.g., Watson, 1986; Morley et al.,
2003; Kamber et al., 2018; Hamilton, 2018; Phillips and Shi, 2019), or on multivariate
techniques (e.g., Fleischman and Roberts, 2011; Aastveit and Trovik, 2014; Morley and
Wong, 2017; Jarociński and Lenza, 2018; Hasenzagl et al., 2018). In the latter group of
papers, two sets of them are particularly related to our approach.
The first set of papers includes those using non-stationary low-dimensional unobserved
component models to estimate the output gap (e.g., Fleischman and Roberts, 2011; Jaro-
ciński and Lenza, 2018; Hasenzagl et al., 2018). Compared with these papers, which typi-
cally consider estimates of the output gap using only highly aggregated variables such as
GDP, the unemployment rate, and price inflation, we estimate our model on a large dataset
comprising 103 variables, thus capturing information coming from a wider spectrum of the

Electronic copy available at: https://ssrn.com/abstract=3217816

economy. As a result, our output gap estimate is driven by all those forces driving the
co-movements in the US economy. Moreover, from a technical point of view, we use a large
number of variables because only by aggregating a large number of time series it is possible
to consistently disentangle macroeconomic fluctuations from idiosyncratic dynamics (Stock
and Watson, 2002; Doz et al., 2012; Barigozzi and Luciani, 2019a,b).
The second set of papers includes those using high-dimensional datasets for estimating
the output gap: (i) Aastveit and Trovik (2014) estimate a stationary factor model combined
with the Hodrick-Prescott (HP) filter; (ii) Morley and Wong (2017) estimate a stationary
Bayesian VAR combined with the Beveridge-Nelson (BN) decomposition. Differently from
these works, we do not need to transform the data to induce stationarity; rather, we
explicitly account for both stochastic and deterministic trends, which are a main feature
of macroeconomic data.
Our approach possesses three more appealing features. First, it models long-run output
growth as slowly drifting over time, because structural breaks in the long-run growth rate
would lead to biased estimates of the output gap (Ng, 2018). Second, it allows each
variable to load the common factors heterogeneously in time, because some indicators
lead (e.g., housing market indicators), and some others lag the business cycle (e.g., the
unemployment rate). Third, our decomposition does not impose a functional form for
trends and cycles, because their dynamics might be richer than just the standard random
walk vs. causal ARMA distinction as in the BN decomposition and the traditional state-
space model approach (Lippi and Reichlin, 1994).
The rest of this paper is structured as follows. In Section 2, we present our methodology,
and in Section 3, we describe the data used and the model setup. Then, in Section 4, we
discuss economically meaningful constraints that we apply to our model. Our empirical
analysis starts in Section 5, in which we discuss the characteristics of the co-movements in
the US economy. In Section 6, we present and discuss our estimates of potential output
and the output gap. In Section 7, we analyze the revision properties of our model through
a quasi-real-time evaluation. Finally, in Section 8, we compare our estimates with other
existing methods. Section 9 concludes.

Electronic copy available at: https://ssrn.com/abstract=3217816

2 Methodology
Let us assume to observe a vector of n time series y = {yit : i = 1, . . . , n, t = 1, . . . , T }
such that
yit = Dit + xit , (1)

where Dit is driven by a deterministic linear trend, and xit is the stochastic component
such that E[xit ] = 0. Letting xt = (x1t · · · xnt )0 , we assume xt ∼ I(1)—i.e., at least one of
its components has a unit root. In a high-dimensional setting, it is reasonable to assume
that fluctuations are the results of not only common trends and common cycles but also
idiosyncratic dynamics. Thus,

yit = Dit + Tit + Cit + ξit , (2)

where Tit ∼ I(1) is the component driven by the common trends, Cit ∼ I(0) is the com-
ponent driven by the common cycles, and ξit is the idiosyncratic component, which is
allowed to be either I(1) (in presence of idiosyncratic trends) or I(0) (e.g., measurement
errors). The components Tit and Cit are driven by factors common across all series, and
thus constitute the common component defined as χit = Tit + Cit . Hence, (2) is also written
as
yit = Dit + χit + ξit . (3)

The goal of this paper is to estimate (2). To do so, we first disentangle common
and idiosyncratic dynamics by fitting a non-stationary dynamic factor model on a large
dataset of US macroeconomic indicators. Then, we decompose the common component
χit into common trends and common cycles by applying a non-parametric trend-cycle
decomposition to the estimated latent common factors.

2.1 Non-stationary dynamic factor model

In a macroeconomic setting dynamic factor models are based on the idea that fluctuations
in each variable are caused by (i) a few common factors (ft ) capturing macroeconomic
dynamics—thus driving the common component (χit )—and (ii) an idiosyncratic component
(ξit ) capturing sectoral and local dynamics as well as measurement errors. Formally, (3) is

Electronic copy available at: https://ssrn.com/abstract=3217816

assumed to be such that
s
X
yit = Dit + λ0i` ft−` + ξit , (4)
`=0

Dit = bi + Dit−1 , Di0 = ai , (5)

p
X i.i.d.
ft = Ak ft−k + ut , ut ∼ N (0, Q), (6)
k=1
i.i.d.
ξit = ρi ξit−1 + eit , eit ∼ N (0, σi2 ), (7)

where ft = (f1t · · · fqt )0 are the q common latent factors capturing co-movements across
series and across time, λi` = (λi1` · · · λiq` )0 are the q factor loadings of series i at lag `,
s ≥ 0 and p ≥ 1 are finite integers, and ai and bi are finite real numbers.
In what follows, we summarize the main features of model (4)-(7), while we refer the
interested reader to Barigozzi and Luciani (2019b) for a formal treatment of the model.

2.1.1 Assumptions

Non-stationarity in the data is modeled by means of three main assumptions.

P
Assumption 1. There exists an integer d such that 0 < d < q and det(Iq − pk=1 Ak z k ) =
0 has (q − d) roots in z = 1, while the other d roots lie outside the unit circle. Moreover,
Q is positive definite.
Assumption 1 implies (i) that the vector of common factors ft is a cointegrated vector with
cointegration rank d; (ii) that while there are q sources of macroeconomic fluctuations, only
(q −d) of them are capable of permanently affecting the economy; and (iii) that ft is driven
by (q − d) common trends. As a consequence, the common component (χit ) has a non-
stationary part (Tit ) driven by the common trends, and a stationary (cyclical) part (Cit )
(see (2) and Section 2.2).
Assumption 2. Let nI be the number of variables among y1t , . . . , ynt for which ξit ∼ I(1).
Then, 0 < nI < n.
Assumption 2 implies that some (but not all) idiosyncratic components might be I(1) and
it is motivated by the fact that assuming that all idiosyncratic components are stationary
(i.e, nI = 0) would imply that any (q − d + 1)-dimensional vector would be cointegrated, an
assumption which in general is not supported by the data—see also the theory and empirical
results presented in Barigozzi et al. (2016, 2019). However, since some macroeconomic

Electronic copy available at: https://ssrn.com/abstract=3217816

variables are known to be cointegrated (e.g., income and consumption) it is reasonable
to assume nI < n. This intuition is confirmed on our dataset, as, on 103 variables that
we use for the empirical analysis, about half of them have a non-stationary idiosyncratic
component, i.e., nI = 44 (see Section 3 and the Appendix for details).
Assumption 3. Let nb be the number of variables among y1t , . . . , ynt for which bi 6= 0.
Then, 0 < nb < n.
Assumption 3 implies that there are no common deterministic linear trends. This is a
realistic assumption for a typical macroeconomic dataset. Indeed, for series belonging to
the real side of the economy, e.g., GDP, bi is likely to be strongly significant, whereas for
nominal series, e.g., inflation, it is likely to be the case that bi is not significantly different
from zero. This intuition is confirmed on our dataset, as, on the 103 variables that we use
for the empirical analysis, only about half of them display a significant linear trend, i.e.,
nb = 61 (see Section 3 and the Appendix for details).
The model is identified by means of the three following assumptions.
Assumption 4. Let Λ = (λ10 · · · λ1s · · · λn0 · · · λns )0 be the n×q(s+1) matrix of all factor
loadings, then there exists a q × q matrix H such that rk(H) ≥ q and limn→∞ n−1 Λ0 Λ = H.
Assumption 5. There exists two finite C1 , C2 > 0 independent of n such that C1 <
P P
Var( ni=1 wi eit ) < C2 for all wi such that ni=1 wi2 = 1.
Assumption 6. The common component χit is independent of the idiosyncratic compo-
nent ξjs , for all i, j, t, s.
Assumption 4 is equivalent to saying that the factors ft are pervasive, i.e., they have non-
negligible effects on most variables at least at one lag, while Assumption 5 implies that the
idiosyncratic innovations are weakly cross-sectionally correlated and therefore do not have
a pervasive effect; moreover, their covariance matrix is positive definite. These assumptions
are similar to the typical assumptions made in the factor model literature (e.g., Bai and
Ng, 2002, 2004, and Barigozzi et al., 2019). From Assumptions 4-6, it follows that in the
limit n → ∞ the common component can be recovered by aggregating the data, and this
justifies from a methodological point of view our choice of using large datasets.

2.1.2 Estimation

The model specified in (4)-(7) has a state-space form, and we estimate it by Quasi-
Maximum Likelihood, implemented through the Expectation Maximization (EM) algo-
rithm, where in the E-step the factors ft are estimated with the Kalman Smoother. This

Electronic copy available at: https://ssrn.com/abstract=3217816

approach in the stationary setting was originally proposed by Shumway and Stoffer (1982)
and Watson and Engle (1983), and further developed and studied by Doz et al. (2012),
Bai and Li (2016), and Barigozzi and Luciani (2019a) when considering large datasets.1
The case of non-stationary data has been considered in fewer works—see, e.g., Quah and
Sargent (1993) and Seong et al. (2013) for applications. The theoretical properties of the
estimators we use here are studied in Barigozzi and Luciani (2019b), where consistency of
the estimated loadings and factors is proved as n, T → ∞.
We opt for Quasi-Maximum Likelihood estimation rather than Principal Component
(PC) analysis for two main reasons: first, in this way (i) we can distinguish between sta-
tionary and non-stationary idiosyncratic components, thus explicitly enforcing Assumption
2; (ii) we can estimate the model without previously de-trending data, thus accounting for
Assumption 3; (iii) we are able to impose economically meaningful restrictions to the model
(see Section 4); (iv) we are able to allow for time variation in key parameters (see Section
4). Second, by modeling the lagged factors ft−1 , . . . , ft−s as additional latent states, we
can estimate the model with dynamic loadings, thus allowing heterogeneity in the way each
variable loads the common factors.2 Details on implementation of the EM are in Barigozzi
and Luciani (2019b).
To properly initialize the EM algorithm we use the approach proposed in Barigozzi
et al. (2019, Lemma 3 and Proposition 2), which is based on PC analysis of the model in
first differences (see also Bai and Ng, 2004). This approach is crucial since it allows us
to use PCs without incurring in spurious effects due to the presence of idiosyncratic unit
roots (Onatski and Wang, 2019) or deterministic linear trends (Ng, 2018). Moreover, the
model in first differences, together with Assumptions 4 and 5, allows us also to determine
q, (q −d), and s. Indeed, it can be shown that the covariance matrix of the differenced data
vector ∆yt has at most q(s + 1) eigenvalues diverging with n, all the others being bounded
for all n. Second, it is possible to show that the q largest eigenvalues of the spectral density
of ∆yt diverge with n at all frequencies but at zero-frequency where—due to the presence of
common trends—only (q − d) eigenvalues diverge, all the others eigenvalues being bounded
for all n. These three conditions allow us to recover the number of common factors q and
of common trends (q − d), and of lags s—see Hallin and Liška (2007), D’Agostino and
Giannone (2012), Barigozzi et al. (2019), and Barigozzi and Luciani (2019b) for further
1
Recent applications of this approach include Reis and Watson (2010), Bańbura and Modugno (2014),
Juvenal and Petrella (2015), Luciani (2015), and Coroneo et al. (2016).
2
Other papers that have estimated a factor model with dynamic loadings are Luciani and Ricci (2014);
D’Agostino et al. (2016); Antolin-Diaz et al. (2017). All these adopt a Bayesian approach.

Electronic copy available at: https://ssrn.com/abstract=3217816

details, and Sections 3 and 5 for implementation.
Finally, note that, similarly to Doz et al. (2012) and Barigozzi and Luciani (2019a), we
estimate the model without accounting explicitly for the dependence across idiosyncratic
shocks. That is in (7), we estimate only the diagonal terms of the covariance matrix of
the vector of idiosyncratic shocks et = (e1t · · · ent )0 , while setting to zero all off-diagonal
entries. Furthermore, when ξit is stationary, we impose ρi = 0.
Although by imposing such a simplified structure we are in fact estimating a mis-
specified model, Doz et al. (2012) and Barigozzi and Luciani (2019a,b) show that this
approach does not affect consistency of the estimates, provided that the stationary id-
iosyncratic components are weakly serially and cross-sectionally correlated.

2.2 Trend-cycle decomposition

Assumption 1 implies that the vector ft is cointegrated with d cointegration relations,
therefore it admits a factor representation:

ft = Ψτt + γt , (8)

where τt = (τ1t . . . τ(q−d)t )0 is the vector of (q − d) common trends, such that τjt ∼ I(1) for
all j and Ψ is q × (q − d), while γt is a q-dimensional stationary vector—see e.g., Escribano
and Peña (1994).
We characterize the common trends and cycles by the following assumption.
Assumption 7. There exists a q × d matrix A such that rk(A) = d and the common
cycles are given by ct = A0 γt . There exist a q × (q − d) matrix B such that rk(B) = (q − d)
and the common trends are given by τt = B0 ft .
Assumption 7 implies that the common cycles vector has the same dimension as the di-
mension of the stationary component of the factor vector, i.e., the number of cointegration
relations. This is a common assumption in many trend-cycle decompositions (e.g., Kasa,
1992; Gonzalo and Granger, 1995) and it is in line with the idea of considering cycles as
deviations from long-run equilibria (see the definition of output gap given by CBO and
discussed in Section 6). The requirement that common trends are linear combinations of
the data is also standard (e.g., Gonzalo and Granger, 1995, Proposition 2).
Moreover, from (8) and Assumption 7, we have

ct = A0 γt = A0 (Iq − ΨB0 )ft = A0 ft − A0 Ψτt , (9)

Electronic copy available at: https://ssrn.com/abstract=3217816

which shows that, since, by construction, ct is stationary, we must have A0 Ψ = 0d×(q−d) .
The vector ct = (c1 . . . cd )0 then has components cjt which are indeed common cycles for
ft (Vahid and Engle, 1993).
Clearly, different choices of A and B lead to different trend-cycle decompositions. The
literature has focussed on two main identification issues. The first is about the way common
trends are defined, in particular, whether they are restricted to be pure random walks (e.g.,
Stock and Watson, 1988) or instead are allowed to contain transitory elements as well (e.g.,
Lippi and Reichlin, 1994). The second is how to allocate the common shocks between the
permanent and transitory components: for example they can be independent (e.g., Harvey,
1985; Kasa, 1992) or generate independent fluctuations only in the long run (e.g., Gonzalo
and Granger, 1995). Here, we refrain from such distinctions and we instead adopt an
agnostic, non-parametric, approach.
We add the following identifying assumption for the common trends.
Assumption 8. The trend loadings are such that rk(Ψ0 Ψ) = (q − d).
Assumption 8 implies that the trends have a pervasive effect on all factors but not neces-
sarily on all observed series. Indeed, some components of the data yt might be stationary,
in which case we could require that the loadings polynomial of the non-stationary compo-
nent of the factors in (4) contain the term (1 − L). We note here that, in practice, such
constraint seems to be naturally satisfied when estimating the parameters in the M-step
on our dataset.3
Finally, we identify the common cycles as follows.
Assumption 9. There exists a q × d matrix φ such that rk(φ0 φ) = d and γt = φct .
As a consequence of Assumption 9, we must have A0 φ = Id , which implies A = φ(φ0 φ)−1 .
An immediate implication of this assumption is that common trends and cycles are uncor-
related, i.e., E[τit cjt ] = 0 for all i, j, t. This restriction is the price to pay for allowing us to
define the non-parametric trend-cycle decomposition below, however note that we do not
rule out lagged dependences, i.e., in general E[τit cjs ] 6= 0 for t 6= s.
Now, if Ψ is given, then a natural choice for A and B is as follows. First, from (9) we
can choose A = Ψ⊥ , where Ψ⊥ is q ×d and such that rk(Ψ0⊥ Ψ⊥ ) = d and Ψ0⊥ Ψ = 0d×(q−d) .
Second, thanks to Assumption 9 we can retrieve the common trends by linear projection,
3
This is, for example, the case when considering our benchmark specification with one trend loaded
with one lag: the estimated loadings of the trend for PCE inflation—but also of core PCE inflation and of
the unemployment rate—have the form λ bτ (L) = λ bτ bτ bτ bτ
PCE 0PCE + λ1PCE L with λ0PCE ' −λ1PCE (see the first
two columns in Table 3 in Section 6).

10

Electronic copy available at: https://ssrn.com/abstract=3217816

i.e., by choosing B = (Ψ0 Ψ)−1 Ψ. We then have the following identity

ft = Ψ(Ψ0 Ψ)−1 Ψ0 ft + Ψ⊥ (Ψ0⊥ Ψ⊥ )−1 Ψ0⊥ ft = Ψτt + Ψ⊥ (Ψ0⊥ Ψ⊥ )−1 ct , (10)

and therefore, letting λτi 0 (L) = λi 0 (L)Ψ and λci 0 (L) = λi 0 (L)Ψ⊥ (Ψ0⊥ Ψ⊥ )−1 , and by com-
bining (1), (4) and (10), we have the trend-cycle decomposition of the data:

yit = Dit + λτi 0 (L)τt + λci 0 (L)ct + ξit = Dit + Tit + Cit + ξit . (11)

This decomposition does not impose any parametric model for either the common trends
(e.g., random walk) or the common cycles (e.g., AR(2)). Furthermore, this decomposition
does not assume any specific relation between the permanent and transitory shocks. In
other words, there can be macroeconomic shocks that affect both the common trends and
the common cycles.
To estimate (10) and (11) we need to face two problems: (i) the factors ft are not
observed, and (ii) the loadings Ψ are unknown. In practice, we do have a consistent
estimator of the factors from Section 2.1.2, and, based on the factor representation of
the common factors (8), we can estimate the loadings by means of PC analysis. More
b = T −2 PT fbt fb0 and denote
precisely, for given consistent estimates of the factors fbt , let S t=1 t
b b
by (Ψ Ψ⊥ ) the q × q matrix with columns given by the normalized eigenvectors of S, b
ordered according to the decreasing value of the corresponding eigenvalues, and such that
b is q ×(q −d), Ψ
Ψ b ⊥ is q ×d. Then, due to orthonormality of the eigenvectors, the estimated
common trends and common cycles are given by the projections τbt = Ψ b 0 fbt ,
b 0 fbt , and cbt = Ψ
⊥
respectively. As shown in Peña and Poncela (2006) and Zhang et al. (2018), respectively,
these projections recover consistently the space spanned by the common trends and the
cointegration vectors, which under Assumption 7 coincides with the space spanned by the
common cycles.

2.3 Quantifying uncertainty

In order to quantify uncertainty around trends and cycles in equation (11) we use a boot-
strap procedure. Specifically, to generate the data we simulate the states in our model
by using the simulation smoother of Durbin and Koopman (2002, algorithm 2), while we
simulate the stationary residual of the model by using a stationary block bootstrap proce-

11

Electronic copy available at: https://ssrn.com/abstract=3217816

dure (Politis and Romano, 1994), with average length of the block equal to four quarters.4
Details on the algorithm used for constructing confidence bands are given in the comple-
mentary appendix.

3 Data and model setup

Our analysis is carried out on a large macroeconomic dataset comprising n = 103 quar-
terly series from 1960:Q1 to 2018:Q4 describing the US economy. Specifically, our dataset
includes national account statistics, industrial production indexes, different price indexes,
various labor market indicators including indicators from both the household survey and
the establishment survey, as well as labor cost and compensation indexes, monetary aggre-
gates, credit and loans indicators, housing market indicators, interest rates, the oil price,
and the S&P500 index.
Broadly speaking, we take log-transforms of all variables in levels that are not already
expressed in percentage points, then all the variables that are I(1) are not transformed,
while all the variables that are I(2) are differenced once. The complete list of variables
and transformations is reported in the Appendix.
Before estimating the model, we need to tackle three preliminary issues: first, we need
to choose which variables need to have a deterministic trend. Second, we need to choose
which idiosyncratic components to model as a random walk. Moreover, third, we need to
choose the number of lags s in the factor loadings and the number of lags p in the VAR
for the common factors.
In order to choose to which variable we should add a linear trend, we test for significance
of the sample mean of ∆yit . Then, if the sample mean of ∆yit is statistically different from
zero, we estimate bi together with all the other parameters; otherwise, we set bi = 0 (see
the Appendix for more details).
In order to choose which idiosyncratic components to model as a random walk, we
run the test proposed by Bai and Ng (2004) for the null hypothesis of a unit root. Then,
if we fail to reject the null, we set ρi = 1, while if we reject, we set ρi = 0. While we
refer the reader to the Appendix for detailed results of the test, here we make just three
4
The “states” in our model are the common factors, the I(1) idiosyncratic components, and the de-
terministic components of GDP, GDI, and the unemployment rate described in 4; while the “stationary
residual” of the model is composed of the I(0) idiosyncratic components, and the residual from the mea-
surement equation for those variable that have an I(1) idiosyncratic components, which is typically very
small.

12

Electronic copy available at: https://ssrn.com/abstract=3217816

comments about the most important variables. First, the test rejects the hypothesis of an
idiosyncratic unit root for GDP, GDI, the Federal funds rate, CPI, core CPI, PCE, and core
PCE inflation. Second, we impose stationary idiosyncratic components for both the 1-year
and 10-year Treasury bill yield, so that they share the same trend with the Federal funds
rate. Third, the test indicates that the idiosyncratic component of the unemployment rate
is I(1), and we treat this in a particular way that we discuss at the end of the next section.
Then, to choose the number of lags s in the factor loadings, based on the discussion
in Section 2.1.2, we proceed as follows: given a choice of q, we can choose s such that the
share of variance explained by the first r = q(s + 1) PCs of the covariance matrix of ∆yt
coincides with the share of variance explained by the first q PCs of the spectral density
matrix of ∆yt (averaged over all frequencies)—see also D’Agostino and Giannone (2012).
By looking at Table 1, we can see that r ' 2q meaning that s ' 1.

Table 1: Percentage of explained variance

1 2 3 4 5 6 7 8 9 10
q 33.3 46.0 53.4 58.9 63.5 67.2 70.3 73.1 75.5 77.7
r 22.9 34.0 42.3 48.1 52.1 55.6 58.4 60.7 62.9 64.9
This table reports the percentage of total variance explained by the q largest eigenvalues of the
spectral density matrix of ∆yt and by the r largest eigenvalues of the covariance matrix of ∆yt .

Finally, we chose the number of lags p in the VAR for the common factors as the
minimum number of lags such that the residuals of the VAR exhibit no autocorrelation.

4 Constraints and time variation

As we explained in Section 2.1.2, two advantages of our estimation approach are that it
allows us to impose economically meaningful constraints and that it allows time variation
in key parameters. Here we exploit both these features by imposing one constraint and by
allowing two key parameters to be time-varying. Specifically, in order to correctly estimate
the output and the unemployment gap, we carefully specify the long-run dynamics of
both GDP and the unemployment rate by introducing two additional states in the model.
Moreover, we add a constraint derived from the National Account concept of “output.”
Time-varying parameter 1: The mean of GDP growth drifts slowly over time.
The goal here is to capture the secular decline in long-run US growth that has been doc-
umented by several authors (see Antolin-Diaz et al., 2017, for example, and reference
therein). The literature has pointed to several factors that can explain this secular slow-

13

Electronic copy available at: https://ssrn.com/abstract=3217816

down citing, among others, slower productivity growth and demographic factors, such as
the decline of population growth and the slower pace of increase in educational attainment
(Gordon, 2018). Indeed, as pointed out by Ng (2018) a linear trend cannot adapt to those
changes, thus leaving too much predictable variations unexplained, which would, in turn,
inflate our output gap at the end of our sample.
In practice, we introduce the additional common latent state for GDP, which evolves
over time according to the following local linear trend:

DGDP,t = bGDP,t−1 + DGDP,t−1

i.i.d.
bGDP,t = bGDP,t−1 + ηt , ηt ∼ N (0, ση2 ).

Time-varying parameter 2: The mean of the unemployment rate drifts gradually over
time.
The idea here is that several idiosyncratic labor market and demographic characteristics
might generate an idiosyncratic trend in the unemployment rate (see Walsh, 1998; Daly
et al., 2011, among others). For example, labor supply factors such as the entry and
exit from the labor market of the baby boom generation, the increased participation rate
among women, and changes in youth labor force share. Or labor demand factors such as, for
example, possible skill or location mismatch between job openings and the characteristics of
the unemployed. Finally, the emergency and extended benefits, which are a standard policy
response to elevated cyclical unemployment that expire once the labor market recovers,
might have transitory effects on the natural rate of unemployment.5
To capture this idiosyncratic trend, we introduce and additional latent state for the
unemployment rate, which evolves over time according to the following local level model:

i.i.d.
DUR,t = DUR,t−1 + t , t ∼ N (0, σ2 ).

Notice that this modeling strategy is consistent with the unit root test performed in the
previous section.
Constraint: In the long run GDP and GDI are driven exclusively, and by the same
amount, by factors common to the whole economy and by the secular trend, while the
5
In the US standard unemployment benefits lasts for 26 weeks. After 26 weeks, there are two programs
that allow to extend further the unemployment benefits (the Emergency Unemployment Compensation and
Extended Benefits), which extend unemployment benefits eligibility up to 99 weeks. These two measures
were introduced during the great recession, the previous maximum eligibility was 65 weeks.

14

Electronic copy available at: https://ssrn.com/abstract=3217816

discrepancy between GDP and GDI is only temporary and driven by stationary idiosyncratic
factors and measurement errors.
Here the rationale is that GDP and GDI are both measures of US aggregate output—
GDP tracks all expenditures on final goods and services produced, while GDI tracks all
income received by those who produced the output. Therefore, the difference between GDP
and GDI is exclusively the result of measurement error—using the NIPA table definition
“statistical discrepancy.” Based on this rationale, in recent years, there has been interest
in combining GDP and GDI to come up with a better estimate of aggregate output (e.g.,
Aruoba et al., 2016; Council of Economic Advisers, 2015) to which the Council of Economic
Advisers refers to as Gross Domestic Output (GDO).
In practice, we implement this assumption by constraining the factor loadings of GDP
and GDI to be equal at all lags (λGDP (L) = λGDI (L)), which implies that the common
components of GDP and GDI coincide (χGDP,t = χGDI,t ),6 and by imposing DGDP,t =
DGDI,t . Finally, we define GDOt := DGDP,t + χGDP,t = DGDO,t + χGDO,t .

5 Co-movements in the US economy

A large number of papers estimate the output gap by imposing one trend and one cycle
(e.g., Fleischman and Roberts, 2011; Jarociński and Lenza, 2018; Hasenzagl et al., 2018).
What if we estimate a model with two common factors (one trend-one cycle), i.e., with
q = 2?
The left plot in Figure 2 shows the spectral density of the first difference of the common
trend and of the common cycle when the model is estimated with two common factors.
As we can see, the estimated cycle captures frequencies with period shorter than 2 years,
and its spectral density has approximately the same shape of the inflation rate (see the
right plot in Figure 3). In some sense, this result is not surprising, as several papers have
documented that the first two factors of the US economy appear to be one “real” and one
“nominal” factor (e.g., Stock and Watson, 2016). What our decomposition is adding to
what the literature have documented is that these two common factors can be seen as a
common “real” trend (the shape of its spectral density is similar to that of GDP, left plot
Figure 3), and as a “nominal” cycle.
6
This restriction is indeed corroborated by the data, as even if we do not impose it, the estimated
χGDP,t and χGDI,t are nearly identical. In numbers, the mean absolute difference of (∆xGDP,t − ∆xGDI,t )
is 0.38 with a standard deviation of 0.49, while the mean absolute difference of (∆χGDP,t − ∆χGDI,t ) is
halved to 0.19 with a standard deviation of 0.27.

15

Electronic copy available at: https://ssrn.com/abstract=3217816

 Figure 2: Spectral densities
Common trend and common cycles
q=2 q=3 q=4
5.0
τt τt τt
4.5 c1t 5.0 c1t 7.0 c1t
c2t c2t
4.0 c3t
6.0
4.0
3.5
5.0
3.0
3.0
2.5 4.0

2.0 3.0
2.0
1.5
2.0
1.0 1.0
1.0
0.5

0.0 0.0 0.0

10Y 5Y 2Y 1Y 10Y 5Y 2Y 1Y 10Y 5Y 2Y 1Y

The left plot reports the spectral densities of the common trend ∆b τt (blue line), and the common cycle ∆b ct (orange line)
estimated when q = 2. The middle plot reports the spectral densities of the common trend (blue line), and of the two common
cycles (orange and yellow lines) estimated when q = 3. The right plot reports the spectral densities of the common trend
(blue line), and of the three common cycles (orange, yellow, and purple lines) estimated when q = 4. On the horizontal axis
of the plots in the left column we report periods τ measured in years such that the corresponding frequencies are given by
θ = 2π
4τ
.

Figure 3: Spectral densities

GDP Unemployment rate PCE Inflation
2.5
2.0
3.5
1.8

2.0 3.0 1.6

2.5 1.4

1.5 1.2
2.0
1.0

1.0 1.5 0.8

0.6
1.0
0.5 0.4
0.5
0.2

0.0 0.0 0.0

10Y 5Y 2Y 1Y 10Y 5Y 2Y 1Y 10Y 5Y 2Y 1Y

On the horizontal axis of the plots in the left column we report periods τ measured in years such that the corresponding
frequencies are given by θ = 2π
4τ
.

To summarize, it is clear, that with a cycle dominated by high frequencies, it is im-

possible to get any meaningful estimate of the output gap or the unemployment rate gap
(see also the robustness analysis in the complementary appendix). Therefore, we conclude
that the representation “one trend-one cycle” is rejected by the data, and thus we look for
a richer parametrization. Indeed, the test by Onatski (2009) and the information criterion
by Hallin and Liška (2007) suggest a value of q between three and four, while the infor-
mation criterion by Barigozzi et al. (2019), clearly indicates (q − d) = 1. Therefore, the
above statistical procedures suggest the presence of one common trend, which is in line
with many theoretical models assuming a common productivity trend as the sole driver of
long-run dynamics (e.g., Del Negro et al., 2007), and either two or three common cycles.
The middle plot in Figure 2 shows the spectral density of the first difference of the
common trend and of the two common cycles obtained when the model is estimated with
three common factors. As we can see, the additional cycle captures mainly frequencies with

16

Electronic copy available at: https://ssrn.com/abstract=3217816

period of about 5 years, and its spectral density has a shape quite similar to that of the
unemployment rate. In other words, under this parametrization we have a representation
of the economy with one common “real” cycle, and one common “nominal” cycle.
The spectral densities of the estimated cycles shows that the model with three common
factors is a serious candidate. Is there any rationale to include a fourth factor?
To answer this question Table 2 shows the in-sample mean squared error of the model
for different number of factors, which is the variance of the idiosyncratic component—note
that the idiosyncratic component of all the variables in the table are stationary. As we can
see, most of the variance of GDP is accounted for by the common component even with
two factors, and, moreover, the MSE of GDP decreases just modestly if additional factors
are included in the model.7 By contrast, the MSE of the other variables is quite sensitive
to the number of factors. Starting with inflation as measured by the quarter-on-quarter
percent change in the PCE price index, the model with three factors has an MSE that is
more than 90% lower than the one with two common factors. Moving to the unemployment
rate, the model with three common factors is not able to explain a good share of the total
variation, whereas once the fourth factor is included the MSE is more that 90% lower than
the MSE of the model with three factors. The same is true for other job market indicators
shown in Table 2, such as the labor force participation rate (LFPR), the employment to
population ratio (EPR), and private payroll employment (EMP). For all these variables,
the MSE of the model with four common factors is more than 80% lower than the model
with three common factors.
Table 2: In sample Mean squared error
GDP PCE PCEx FFR UR LFPR EPR EMP
q =2 0.279 0.223 0.191 7.181 0.931 2.491 1.079 5.578
q =3 0.290 0.012 0.037 2.122 0.712 1.641 2.034 5.382
q =4 0.235 0.008 0.017 2.037 0.017 0.118 0.056 0.922
q =5 0.232 0.015 0.028 0.634 0.013 0.062 0.026 0.744
This table shows the in-sample MSE of the model. “GDP” stands for Gross Domestic Product, “PCE” stands for Personal
Consumption Expenditure price index, “PCEx” is PCE excluding food and energy (aka core PCE prices), “FFR” is Fed
Funds Rate, “UR” stands for Unemployment Rate, “LFPR” stands for Labor Force Participation Rate, “EPR” stands for
employment to Population Ratio, “EMP” is total private payroll employment

Why do we need a third cycle to explain the unemployment rate? A possible expla-
nation is that as documented by Barnichon (2010), among others, prior to the mid 1980s,
productivity was procyclical; afterwards, it was acyclical or countercyclical. As a result,
Note that Table 2 shows the variance of the level of the estimated idiosyncratic component of log GDP.
7

Log GDP in 2018:Q4 is 983.98, so a variance of 0.28 is negligible.

17

Electronic copy available at: https://ssrn.com/abstract=3217816

the cyclical relationship between unemployment and output changed. This change cannot
be captured by the model with just one “real” cycle; hence the need for an additional cycle.
Figure 4, which shows the estimated common trend and the three estimated cycles,
supports this assumption: the third and fourth cycle are pretty positively correlated until
the early 1990s and then negative correlated after that. This timing roughly matches the
change in the cyclical behavior of productivity described in Barnichon (2010).
To conclude, the results shown in this Section show how the model with three common
factors is a possible candidate, but at the same time it has weak explanatory power for
the unemployment rate.8 Therefore, since the unemployment rate is a crucial indicator of
the cyclical position of the economy, our benchmark specification includes four common
factors, q = 4. As we can see from Figure 4, under this parametrization, the first cycle
looks like an inflation cycle, the second cycle seems to be related with GDP, and the third
cycle seems to be related with the unemployment rate; hence, we have a representation of
the economy with two common “real” cycles and one common “nominal” cycle (see also the
right plot in Figure 2).

Figure 4: Trend and Cycles

τ
bt c
b1t c
b2t c
b3t
3 3 3 3

2 2 2 2

1 1 1 1

0 0 0 0
Levels

-1 -1 -1 -1

-2 τt -2 -2 -2

-3 -3 c1t -3 c2t -3 c3t

-4 -4 -4 -4

-5 -5 -5 -5

1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015

4 4 4
c2t 4
First differences

2 2 2 2

0 0 0 0

-2 -2 -2 -2

-4 -4 -4 -4
τt c1t c3t

-6 -6 -6 -6

-8 -8 -8 -8
1975 1980 1985 1990 1995 2000 2005 2010 2015 1975 1980 1985 1990 1995 2000 2005 2010 2015 1975 1980 1985 1990 1995 2000 2005 2010 2015 1975 1980 1985 1990 1995 2000 2005 2010 2015

This figure reports the estimated common trend and the estimated common cycles when q = 4. The upper row shows levels
of the estimated quantities, while the lower row shows first differences.

In the complementary appendix we show that the model with three common factors fails to provide a
8

sensible estimate of the cyclical component of the unemployment rate.

18

Electronic copy available at: https://ssrn.com/abstract=3217816

6 Potential output and the output gap
Based on our dynamic factor model and our trend-cycle decomposition, we define potential
output as the trending component of GDO:
s
X
d
P bGDO,t +
Ot = D b0
λ b τt−` . (12)
GDO,` Φb
`=0

In other words, potential output in our framework is the sum of the local linear trend,
which captures the secular decline in long-run US output growth, and the part of GDO
which is explained by the “real” common trend. Similarly, we define the output gap as the
cyclical component of GDO:
s
X
dt = b0
λ b bt−` (13)
OG GDO,` Φ⊥ c
`=0

that is, the output gap is that part of GDO that is accounted for by the common cycles.
In what follows, we compare our estimates with the one published by the Congressional
Budget Office (CBO). The CBO estimates potential output and the output gap by using
the so-called “production function approach” (Kiley, 2013) according to which potential
output is that level of output consistent with current technologies and normal utilization
of capital and labor, and the output gap is the deviation of output from potential output.
To be more precise, the CBO model is based upon a textbook Solow growth model, with
a neoclassical production function, and estimates trends in the components of GDP based
on a variant a variant of the Okun’s law. In other words, in the CBO framework, actual
output is above its potential (the output gap is positive), when the unemployment rate is
below the natural rate of unemployment, which is in turn defined as the non-accelerating
inflation rate of unemployment (NAIRU), i.e., that level of unemployment consistent with
a stable inflation—for further details see Congressional Budget Office (2001). Notice that
also for the CBO the output gap is assumed to revert to zero in the long run as it imposes
in its forecast that in 10 years the output gap will be zero—see, e.g., Congressional Budget
Office (2004).
Figure 5 shows our measure of potential output (red line), where the shaded areas
around our estimate are 68% and 84% confidence bands, respectively, together with the
estimate produced by the CBO (blue line), and the log-level of GDP (black line). As we
can see, the two trend estimates start to diverge at the end of the ’90s, with the CBO

19

Electronic copy available at: https://ssrn.com/abstract=3217816

 Figure 5: Potential Output
Levels Four-quarter percent change
8
980

6 GDP
960
CBO
BL
940 4
GDP
CBO
BL
920 2

900
0

880
-2

860
-4

1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 1975 1980 1985 1990 1995 2000 2005 2010 2015

The black line is log GDP, the blue line is the estimate of potential output computed by the CBO, while the red line is our
estimate. The shaded areas around our estimate are 68% and 84% confidence bands, respectively.

estimating higher potential growth (i.e., a steeper increase in the trend) than our model
(4.2% vs. 3.7% in 2000:Q1). However, at the beginning of the 2000s, both trend estimates
slowdown, and they reunite at the end of the sample.
Our results contrast with those of Coibion et al. (2018), who estimated that potential
output fell during the financial crisis, but then since 2010 has started to grow at approxi-
mately the same pace as before the crisis. However, as pointed out by Ng (2018) the path
of potential output estimated by Coibion et al. (2018) is heavily influenced by their choice
to include a deterministic linear trend in the model for GDP, which is responsible for the
large output gap they estimate at the end of their sample in 2017:Q4 (see also Bullard,
2012).

Figure 6: Output gap

Levels Four-quarter percent change
6 8

4 6
CBO
BL
2 4

0 2

-2 0

-4 -2
CBO
-6 BL -4

-8 -6

-10 -8

1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 1975 1980 1985 1990 1995 2000 2005 2010 2015

The blue line in the left is the level of the output gap estimated by the CBO, while the red line is our estimate. The blue line
in the right is the four-quarter percent change of the output gap estimated by the CBO, while the red line is our estimate.
The shaded areas around our estimate are 68% and 84% confidence bands, respectively.

20

Electronic copy available at: https://ssrn.com/abstract=3217816

 The left plot of Figure 6 shows our estimate of the level of the output gap, together
with the one produced by the CBO. Overall, our estimate is remarkably similar to that of
the CBO. In particular, the peaks and troughs of the two measures align, thus indicating
that the dating of the turning points perfectly coincides. However, we also see two puzzling
results: 1) the Great Recession was just slightly more severe than the 1990-1991 recession,
and 2) the output gap was positive throughout the 2001 recession.
The right plot in Figure 6, which shows the four-quarter percent change of our output
gap estimate, allows us to clarify the differences mentioned above. Our measure and
the CBO measure interpret the 2008 recessions broadly in the same way, as according to
both estimates, the output gap fell 5.5% in the four quarters through 2009:Q2. Moreover,
according to our measure the 1991 recession was less severe than the 2008 recession (-2.7%
in the four quarters through 1991:Q1). Last, our model estimates that during the 2001
recession, the output gap fell 2.2% in the four quarters through 2001:Q4.
Finally, despite the above discussion, an important difference between the two esti-
mates is visible during the period 1997-2008. The two output gap measures started to
diverge somewhat significantly at the end of 1997, which approximately dovetails with the
divergence in the estimate of potential output. Then, starting in 2000:Q3 the two measures
roughly co-move; however, according to the CBO the level of the output gap was negative
between 2002:Q1 and 2005:Q4, whereas according to our measure in that same period the
output gap was positive. In other words, while the CBO estimates that the US economy
reached its potential just a couple of years before the financial crisis, our model signals a
persistent overheating of the economy from the mid-90s to 2008.
What does explain this divergence of the two estimates? We believe that it has to
do with how the two considered measures extract signal from inflation data. Indeed,
the CBO measure by definition takes signal from inflation data through the NAIRU. By
contrast, our measure extracts different signals from the different dimensions of the US
economy (including inflation), and it constructs the output gap measure by letting the
data weighting these signals.
Table 3 shows the estimated loadings of the three common cycles. As we can see, the
first cycle, which we labeled “nominal” cycle in Section 5, is heavily loaded by inflation,
while it is weakly loaded by GDP. Now, from 1995 onward inflation has been stable and
low—on average core PCE price inflation between 1995:Q1 and 2007:Q4 was approximately
1.8%. Therefore, the CBO estimates that there was slack in the economy (i.e., a negative
output gap), while our measure indicates that the economy was operating above potential.

21

Electronic copy available at: https://ssrn.com/abstract=3217816

 Table 3: Factor loadings
bτ
λ bτ
λ b c1
λ b c1
λ b c2
λ b c2
λ b c3
λ b c3
λ
0 1 0 1 0 1 0 1
GDP 0.55 0.39 -0.25 -0.20 -0.95 -0.03 -1.53 -0.14
UR 0.34 -0.50 -0.07 0.80 0.28 -0.98 1.84 1.07
PCE 1.89 -1.82 3.94 0.26 -0.54 0.62 1.74 -1.43
PCEx 4.41 -4.30 3.71 1.90 -2.05 2.28 3.98 -3.23

The results in Figure 7, which shows estimates obtained with our model when we
exclude all price indicators from our dataset, confirm this interpretation. Two main results
are worth discussing: first, if we exclude all price indexes our estimate of the output gap is
even larger between 2002 and 2008, which shows how our model takes signal from inflation,
but not as much as the CBO. Second, even excluding all price indexes our estimate in the
last 10 years is mainly unchanged; the only difference here is in 2009 when the output gap
grows at a faster pace if we include all prices in the dataset. This make sense as in 2009
the 12-month percent change in core PCE prices rebounded from less than 1% to about
1.7%.
Figure 7: Output gap
No price inflation indexes
Levels Four-quarter percent change
6 8

4 6
BL
Excl. prices
2 4

0 2

-2 0

-4 -2

-6 -4

BL
-8 Excl. prices -6

-10 -8

1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 1975 1980 1985 1990 1995 2000 2005 2010 2015

In each plot the grey line is the estimate obtained when the model is estimated on a dataset excluding all price
indexes (ID 18, 19, 36–52). The red line is our benchmark estimate and the red shaded area are the 64% and 84%
confidence bands.

So far, we have proceeded as if the output gap is the sole indicator of the cyclical
position of the US economy. However, another widely used indicator of the cyclical position
of the economy is the unemployment rate gap, and indeed in the framework adopted by the
CBO, the output gap and the unemployment rate gap are linked via the Okun’s law. Using
our framework, we can construct an estimate of the unemployment rate gap as the cyclical
component of the unemployment rate, i.e., as that part of the unemployment rate driven by

22

Electronic copy available at: https://ssrn.com/abstract=3217816

the common cycles. Therefore, in our framework, the output gap and the unemployment
rate gap are not linked by a structural relationship, but rather are driven by the same
cycles.
The right plot of Figure 8 shows our estimate of the unemployment rate gap, together
with that of the CBO. Mirroring the discussion of the output gap, our estimate of the
unemployment rate gap is relatively similar to that of the CBO. The dating of the cycle
is nearly identical (i.e., peaks and through coincides), while the main difference emerges
during the 90s. Furthermore, and perhaps more importantly, by comparing the left plot
in Figure 6 and the right plot in Figure 8, we can clearly see how the output gap and the
unemployment rate gap estimated with our model are sending contrasting signals at the
end of the sample: the output gap is indicating that there is still some residual slack in the
economy (-0.3%), whilst the unemployment rate gap is suggesting that the US economy is
operating above potential (-0.9%).

Figure 8: Unemployment gap

Levels Four-quarter change
6 5

5
4
CBO CBO
4 BL BL
3
3

2
2

1 1

0
0

-1
-1
-2

-2
-3

-4 -3
1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 1975 1980 1985 1990 1995 2000 2005 2010 2015

The blue line is the unemployment rate gap estimated by the CBO, while the red line is our estimate. In both plots the
shaded areas around our estimate are 68% and 84% confidence bands, respectively.

Where does the 2018 decoupling of the output gap and the unemployment rate gap
come from? To answer this question, first we have to point out that in the CBO model
the relationship between the output gap and the unemployment rate gap is constrained by
the Okun’s law. In our model, instead, the output gap and the unemployment rate gap
are driven by a three-dimensional cycle process cbt , and there are no restrictions on the
impact of cbt on these measures. That said, as we can see from Figure 9, on average our
estimated output gap and unemployment rate gap are related approximately in the same
way as those estimated by the CBO; however, this relationship holds less tightly in our
model, and because of this it is possible to have a decoupling of the unemployment rate

23

Electronic copy available at: https://ssrn.com/abstract=3217816

gap and the output gap.

Figure 9: Okun’s law

BL CBO
5 5
R2 =0.755 R2 =0.794
4 4
β̂ =-0.579 β̂ =-0.603
3 3
Unemployment gap

Unemployment gap
2 2

1 1

0 0

-1 -1

-2 -2

-3 -3

-8 -6 -4 -2 0 2 4 6 -8 -6 -4 -2 0 2 4 6
Output gap Output gap

This figure shows the Okun’s law relationship with the output gap on the horizontal axis and the unemployment gap on
the vertical axis for the time period from 1962Q4 to 2018:Q4. In the left (right) plot the red (blue) dots correspond to
an output gap - unemployment gap pair at time t estimated by us (CBO). In both plots, the green line is the best fit
line. β̂ is the estimated coefficient of the linear regression of the output gap on the unemployment gap, and R2 is relative
goodness-of-fit coefficient.

7 Quasi-real-time estimation of the output gap

Since the seminal work of Orphanides and van Norden (2002), who show that end-of-sample
revisions of GDP are of the same order of magnitude as the output gap, there has been
much debate on the reliability of such estimates in real time. There are two reasons why a
model-based estimate of the output gap revises in real-time: first, because the data upon
which the model is estimated get revised over time. Second, because as new observation
comes in, both the estimates of the model parameters and of the latent factors, might
change.
The goal of this section is to study the real-time properties of the output gap estimate.
However, because we were able to retrieve real-time data vintages for all 103 variables in
our dataset only from August 30, 2013, a fully real-time exercise would be based on a
too short sample to draw any reliable inference. Therefore, to investigate the real-time
properties of our model, we perform a quasi-real-time exercise, which allows us to extend
the period of analysis back to 1985.
Specifically, with “quasi-real-time exercise” we mean that by using the vintage of data
from March 29, 2019, we estimate the model on expanding windows, where the first window
ends in 1985:Q1 and the last one in 2018:Q4—this gives us four different estimates of the

24

Electronic copy available at: https://ssrn.com/abstract=3217816

output gap for each year for a total of 136 estimates. Note that, because the quasi-real-time
exercise uses already revised data, the output gap estimates obtained with this exercise
are free of the data revision problem; hence, any potential revision to our measures are
the sole result of parameter estimation, and filtering. In the complementary appendix, we
present the shorter fully real-time exercise where we also address the issue of data revision,
and we show that these play a negligible role.
Figure 10 shows the quasi -real-time estimates of the output gap (left plot) and of the
unemployment gap (right plot). In each plot, the blue line is the output gap (unemployment
gap) estimate obtained on the sample ending in 1985:Q1, while the red line is the estimate
obtained with the sample ending in 2018:Q4, which corresponds to the results shown in
Section 6. The black dots represent the output gap (unemployment gap) estimate for
quarter Q and year Y obtained with the sample ending at quarter Q and year Y. Finally,
for each of the 136 black dots, there is a corresponding yellow line, showing the output gap
(unemployment gap) estimate obtained using the associated sample.

Figure 10: Cyclical position of the economy

Quasi-real-time estimation
Output gap Unemployment gap
5
4
4

3 3
2
2
1

0
1
-1

-2 0

-3
-1
-4

-5 -2

-6
-3
-7

1985 1990 1995 2000 2005 2010 2015 1980 1985 1990 1995 2000 2005 2010 2015

In each plot the blue line is the estimate of the output gap obtained on the sample ending in 2000:Q1, the red line is the
estimate obtained on the benchmark sample ending in 2018:Q4, and the yellow lines are the estimate obtained on expanding
window from 2000:Q1 to 2018:Q4. Finally, the black dots represent the estimate of the output gap for quarter Q and year
Y obtained on the sample ending at quarter Q and year Y.

Our output gap measure (left plot of Figure 10) revises very little when considering
samples no more than five years shorter than our benchmark sample, while the size of
the revision becomes relevant only when considering samples that are ten or more years
shorter than our benchmark sample (see Table 4). However, these revisions are far from
being dramatic—the only exception to this statement being the early 1990s when the model
was pointing to a much tighter economy in quasi-real-time. In particular, our quasi -real-
time estimates still indicate a constant overheating of the economy from the mid-90s to

25

Electronic copy available at: https://ssrn.com/abstract=3217816

the great recession.

Table 4: Average absolute revision

Sample length OGt U Gt
5 years 0.12 0.38
10 years 0.79 0.61
15 years 1.20 0.70
20 years 1.04 0.64
all 1.33 0.85

Similar comments apply to the real-time properties of our unemployment gap estimate
(right plot of Figure 10). However, differently from our output gap estimates, the quasi-
real-time estimate of the unemployment gap points out to a level of slack at the beginning
of the 1990s, very close to the one estimated on the full sample.
Summing up, in general, the conclusions reached in Section 6 are still valid.

8 Comparison with alternative statistical models

In this section, we compare our output gap estimate with that obtained with four different
purely statistical models:
1. A multivariate BN decomposition derived from a Large Bayesian VAR estimated on
all the 103 variables in our dataset as in Morley and Wong (2017).9
2. A BN decomposition estimated as in Kamber et al. (2018), who develop a “BN filter”
with a low signal-to-noise ratio that is imposed by estimating a univariate AR model
with Bayesian methods.
3. The HP filter with a smoothing parameter set at 1600, the value used in the literature
for quarterly data.
4. The Hamilton (2018) filter, where the trend and the cycle are computed exactly as
in equation (20) of his paper, i.e., the trend is the 8-step ahead direct-forecast of the
quarterly GDP growth rate obtained by using the four most recent values of quarterly
GDP growth as of date t, and the cycle is the residual from such regression.
Figure 11 compares the estimate of the output gap obtained with our model with those
obtained with the four models just described. As we can see from the first column of Figure
To be more precise, to estimate the VAR we follow Morley and Wong (2017) by applying the same
9

data transformations, and by setting the shrinkage parameter to 0.04 (the value that Morley and Wong
(2017) find optimal for their “Large” model), and by including four lags in the VAR.

26

Electronic copy available at: https://ssrn.com/abstract=3217816

 Figure 11: Output gap estimates with statistical models
Multivariate BN BN Filter HP Filter Hamilton filter
6 6 6 6

4 4 4 4

2 2 2 2

0 0 0 0
OGt

-2 -2 -2 -2
BL
MW
-4 -4 -4 -4
BL BL
HP H
-6 -6 -6 -6
BL
-8 -8 KMW -8 -8

-10 -10 -10 -10

1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015

8 8 8

10
6 6 6
BL BL BL BL
MW KMW HP H
4 4 4

5
2 2 2
∆4 OGt

0 0 0
0
-2 -2 -2

-4 -4 -4
-5
-6 -6 -6

-8 -8 -8
-10
1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015

This figure compares the estimate of the output gap obtained with our model, and those obtained with a multivariate BN
decomposition derived from a Large Bayesian VAR estimated on all the 103 variables in our dataset as in Morley and Wong
(2017) (first column), the “BN filter” developed by Kamber et al. (2018) (second column), the HP filter (third column), and
the Hamilton filter (fourth column). The plots on the top row show estimates of the level of the output gap, while the plots
on the lower row show four-quarter percent change.

11, the multivariate BN decomposition of Morley and Wong (2017) yields an output gap
estimate often very similar to our estimate. In particular, similar to our model, it points to
a large output gap before the great financial crisis. However, in contrast with our estimate,
it points to a much tighter economy at the end of the sample. This latter result might be
likely linked to the fact that, as explained by Morley and Wong (2017), the unemployment
rate is the most important source of information for their model, and, as we have discussed
in Section 6, according to our model the labor market is pointing to a tighter economy than
the signal coming from the goods and services market. Moreover, as we can see from the
lower plot of the first column of Figure 11, the multivariate BN is slightly lagging compared
with our measure in particular during recessions. This might be the consequence of the
fact that Morley and Wong (2017) are using a VAR, which by nature uses only past values
to construct the output gap measure, whereas the factor model also uses contemporaneous
values.
Moving to the univariate models, as we can see from Figure 11, although the peaks
and the troughs of the output gap estimated by our measure are broadly consistent with
those estimated with univariate benchmarks, some significant differences emerge. First,
the output gap estimated with the BN filter exhibits very mild deviations from potential
output, to the point that the four-quarter percentage change of this measure decreased
only 2% during the financial crisis, thus implying that permanent factors were the main
drivers of the great recession.

27

Electronic copy available at: https://ssrn.com/abstract=3217816

 Second, the output gap estimated with the HP filter is more volatile than our estimate.
This is a consequence of the statistical definition of the HP filter, which leaves out all low
frequencies, and, as a result, it yields an output gap estimate that is relatively short-term
(Phillips and Jin, 2015). Moreover, recent literature has heavily criticized the use of the
HP filter because the common choice of the tuning parameter for quarterly data fails to
completely remove stochastic process thus introducing spurious dynamic relations (Phillips
and Jin, 2015; Hamilton, 2018).
Finally, among the univariate models, the Hamilton filter is the only one producing an
output gap with large fluctuations. However, as we can see from the plot of the four-quarter
percent change, the Hamilton filter fluctuates too much.

9 Conclusions
In this paper, we measure potential output and the output gap by fitting a non-stationary
dynamic factor model on a large dataset of US macroeconomic indicators via the EM algo-
rithm, and by disentangling common trends from common cycles via the eigenanalysis of
the long-run covariance matrix of the latent common factors. Our methodology allows us to
capture the signal coming from a large information set, to account for stochastic and deter-
ministic trends, to model long-run output growth as a slow-moving process, and to model
the impact of common factors on the different variables in a dynamically heterogeneous
way.
We compare our estimate of the output gap with the one produced by the Congressional
Budget Office (CBO), and we find that: first, the dating of the turning points by the two
measures perfectly coincides. Second, since the late ’90s, the two measures diverge: the
CBO estimates that the output gap closed just a couple of years before the great financial
crisis, whereas our model suggests a persistent overheating of the economy from the mid-
90s to 2008. Third, at the end of the sample in 2018:Q4 our estimate of the output gap
suggests that the US economy was operating at its potential, whereas our estimate of
the unemployment rate gap signals that the economy was operating above its potential.
Finally, by means of a quasi-real-time exercise, we show that our output gap measure
revises modestly. This last result support the intuition that, by pooling a large number of
variables, it is possible to get a robust estimate of the signal in the data while parsing out
the noise.
In conclusion, our model proved capable of extracting signals from different dimensions

28

Electronic copy available at: https://ssrn.com/abstract=3217816

of the US economy, and to construct the output gap measure by weighting these different
signals. In particular, our output gap measure takes strong signal from the labor market
while taking little signal from inflation, thus making it a complementary measure of the
cyclical position of the economy with respect to the theoretical models used in policy
institutions.

References
Aastveit, K. A. and Trovik, T. (2014). Estimating the output gap in real time: A factor
model approach. The Quarterly Review of Economics and Finance, 54:180–193.
Antolin-Diaz, J., Drechsel, T., and Petrella, I. (2017). Tracking the slowdown in long-run
GDP growth. The Review of Economics and Statistics, 99:343–356.
Aruoba, S. B., Diebold, F. X., Nalewaik, J., Schorfheide, F., and Song, D. (2016). Improv-
ing GDP measurement: A measurement-error perspective. Journal of Econometrics,
191:384–397.
Bai, J. and Li, K. (2016). Maximum likelihood estimation and inference for approximate
factor models of high dimension. Review of Economics and Statistics, 98:298–309.
Bai, J. and Ng, S. (2002). Determining the number of factors in approximate factor models.
Econometrica, 70:191–221.
Bai, J. and Ng, S. (2004). A PANIC attack on unit roots and cointegration. Econometrica,
72:1127–1177.
Bańbura, M. and Modugno, M. (2014). Maximum likelihood estimation of factor models
on datasets with arbitrary pattern of missing data. Journal of Applied Econometrics,
29:133–160.
Barigozzi, M., Lippi, M., and Luciani, M. (2016). Dynamic factor models, cointegration,
and error correction mechanisms. FEDS 2016-18, Board of Governors of the Federal
Reserve System.
Barigozzi, M., Lippi, M., and Luciani, M. (2019). Large-dimensional dynamic fac-
tor models: Estimation of impulse-response functions with i(1) cointegrated factors.
arXiv:1602.02398.
Barigozzi, M. and Luciani, M. (2019a). Quasi maximum likelihood estimation and inference

29

Electronic copy available at: https://ssrn.com/abstract=3217816

 of large approximate dynamic factor models via the em algorithm. arXiv:1910.03821.
Barigozzi, M. and Luciani, M. (2019b). Quasi maximum likelihood estimation of non-
stationary large approximate dynamic factor models. arXiv:1910.09841.
Barnichon, R. (2010). Productivity and unemployment over the business cycle. Journal of
Monetary Economics, 57.
Bullard, J. (2012). Global output gaps: Wave of the future? Presentation by President
James Bullard at the Monetary Policy in a Global Setting: China and the United States,
Beijing, China, March 28, 2012.
Coibion, O., Gorodnichenko, Y., and Ulate, M. (2018). The cyclical sensitivity in estimates
of potential output. Brookings Papers on Economic Activity. forthcoming.
Congressional Budget Office (2001). CBO’s method for estimating potential output: An
update.
Congressional Budget Office (2004). A summary of alternative methods for estimating
potential GDP. CBO Background Paper.
Coroneo, L., Giannone, D., and Modugno, M. (2016). Unspanned macroeconomic factors
in the yield curve. Journal of Business and Economic Statistics, 34:472–485.
Council of Economic Advisers (2015). A better measure of economic growth: Gross Do-
mestic Output (GDO). CEA Issue Brief.
D’Agostino, A. and Giannone, D. (2012). Comparing alternative predictors based on large-
panel factor models. Oxford Bulletin of Economics and Statistics, 74:306–326.
D’Agostino, A., Giannone, D., Lenza, M., and Modugno, M. (2016). Nowcasting business
cycles: A Bayesian approach to dynamic heterogeneous factor models. In Koopman, S.
and Hillebrand, E., editors, Dynamic Factor Models, volume 35 of Advances in Econo-
metrics, pages 569–594. Emerald Publishing Ltd.
Daly, M. C., Hobijn, B., and Valletta, R. G. (2011). The recent evolution of the natural
rate of unemployment. Working Paper Series 2011-05, Federal Reserve Bank of San
Francisco.
Del Negro, M., Schorfheide, F., Smets, F., and Wouters, R. (2007). On the fit of New
Keynesian models. Journal of Business and Economic Statistics, 25:123–143.
Doz, C., Giannone, D., and Reichlin, L. (2012). A quasi maximum likelihood approach

30

Electronic copy available at: https://ssrn.com/abstract=3217816

 for large approximate dynamic factor models. The Review of Economics and Statistics,
94(4):1014–1024.
Durbin, J. and Koopman, S. J. (2002). A simple and efficient simulation smoother for state
space time series analysis. Biometrika, 89:603–615.
Escribano, A. and Peña, D. (1994). Cointegration and common factors. Journal of Time
Series Analysis, 15:577–586.
Fleischman, C. A. and Roberts, J. M. (2011). From many series, one cycle: improved
estimates of the business cycle from a multivariate unobserved components model. FEDS
2011-046, Board of Governors of the Federal Reserve System.
Gonzalo, J. and Granger, C. (1995). Estimation of common long-memory components in
cointegrated systems. Journal of Business and Economic Statistics, 13:27–35.
Gordon, R. J. (2018). Declining american economic growth despite ongoing innovation.
Explorations in Economic History, 69:1–12.
Hallin, M. and Liška, R. (2007). Determining the number of factors in the general dynamic
factor model. Journal of the American Statistical Association, 102:603–617.
Hamilton, J. D. (2018). Why you should never use the hodrick-prescott filter. The Review
of Economics and Statistics, 100:831–843.
Harvey, A. C. (1985). Trends and cycles in macroeconomic time series. Journal of Business
& Economic Statistics, 3:216–227.
Hasenzagl, T., Pellegrino, F., Reichlin, L., and Ricco, G. (2018). A model of the Fed’s
view on inflation. CEPR Discussion papers DP12564.
Jarociński, M. and Lenza, M. (2018). An inflation-predicting measure of the output gap
in the euro area. Journal of Money Credit and Banking, 50:1189–1224.
Juvenal, L. and Petrella, I. (2015). Speculation in the oil market. Journal of Applied
Econometrics, 30:1099–1255.
Kamber, G., Morley, J., and Wong, B. (2018). Intuitive and reliable estimates of the output
gap from a Beveridge-Nelson filter. Review of Economics and Statistics, 100:550–566.
Kasa, K. (1992). Common stochastic trends in international stock markets. Journal of
Monetary Economics, 29:95–124.
Kiley, M. T. (2013). Output gaps. Journal of Macroeconomics, 37(C):1–18.

31

Electronic copy available at: https://ssrn.com/abstract=3217816

Lippi, M. and Reichlin, L. (1994). Common and uncommon trends and cycles. European
Economic Review, 38:624–635.
Luciani, M. (2015). Monetary policy and the housing market: A structural factor analysis.
Journal of Applied Econometrics, 30:199–218.
Luciani, M. and Ricci, L. (2014). Nowcasting Norway. International Journal of Central
Banking, 10:215–248.
Morley, J., Nelson, C. R., and Zivot, E. (2003). Why are unobserved component and
Beveridge-Nelson trend-cycle decompositions of GDP so different. The Review of Eco-
nomics and Statistics, 85:235–243.
Morley, J. and Wong, B. (2017). Estimating and accounting for the output gap with large
Bayesian vector autoregressions. CAMA Working Papers 2017-46, Centre for Applied
Macroeconomic Analysis.
Ng, S. (2018). Comments on the cyclical sensitivity in estimates of potential output.
Brookings Papers on Economic Activity. forthcoming.
Onatski, A. (2009). Testing hypotheses about the number of factors in large factor models.
Econometrica, 77:1447–1479.
Onatski, A. and Wang, C. (2019). Spurious factor analysis. manuscript, Cambridge Uni-
versity.
Orphanides, A. and van Norden, S. (2002). The unreliability of output-gap estimates in
real time. The Review of Economics and Statistics, 84:569–583.
Peña, D. and Poncela, P. (2006). Nonstationary dynamic factor analysis. Journal of
Statistical Planning and Inference, 136:1237–1257.
Phillips, P. C. and Jin, S. (2015). Business cycles, trend elimination, and the hp filter.
Discussion Paper 2005, Cowles Foundation.
Phillips, P. C. and Shi, Z. (2019). Boosting the hodrick-prescott filter. Technical Report
1905.00175, arXiv.
Politis, D. N. and Romano, J. P. (1994). The stationary bootstrap. Journal of the American
Statistical Association, 89:1303–1313.
Quah, D. and Sargent, T. J. (1993). A dynamic index model for large cross sections. In
Business cycles, indicators and forecasting. University of Chicago Press.

32

Electronic copy available at: https://ssrn.com/abstract=3217816

Reis, R. and Watson, M. W. (2010). Relative goods’ prices, pure inflation, and the Phillips
correlation. American Economic Journal Macroeconomics, 2:128–157.
Seong, B., Ahn, S. K., and Zadrozny, P. A. (2013). Estimation of vector error correction
models with mixed-frequency data. Journal of Time Series Analysis, 34:194–205.
Shumway, R. H. and Stoffer, D. S. (1982). An approach to time series smoothing and
forecasting using the EM algorithm. Journal of Time Series Analysis, 3:253–264.
Stock, J. H. and Watson, M. W. (1988). Testing for common trends. Journal of the
American Statistical Association, 83:1097–1107.
Stock, J. H. and Watson, M. W. (2002). Forecasting using principal components from a
large number of predictors. Journal of the American Statistical Association, 97:1167–
1179.
Stock, J. H. and Watson, M. W. (2016). Dynamic factor models, factor-augmented vector
autoregressions, and structural vector autoregressions in macroeconomics. In Taylor,
J. B. and Uhlig, H., editors, Handbook of Macroeconomics, volume 2, pages 415–525.
Elsevier.
Vahid, F. and Engle, R. F. (1993). Common trends and common cycles. Journal of Applied
Econometrics, 8:341–360.
Walsh, C. E. (1998). The natural rate, nairu, and monetary policy. Economic Letter
1998-28, Federal Reserve Bank of San Francisco.
Watson, M. W. (1986). Univariate detrending methods with stochastic trends. Journal of
Monetary Economics, 18:49–75.
Watson, M. W. and Engle, R. F. (1983). Alternative algorithms for the estimation of dy-
namic factor, mimic and varying coefficients regression models. Journal of Econometrics,
23:385–400.
Zhang, R., Robinson, P., and Yao, Q. (2018). Identifying cointegration by eigenanalysis.
Journal of the American Statistical Association. forthcoming.

33

Electronic copy available at: https://ssrn.com/abstract=3217816

Appendix Data Description and Data Treatment
This Appendix present the dataset used in the analysis. All variables where downloaded
from Haver on March 29th 2019. This vintage of data incorporates the second estimate of
Gross Domestic Product in 2017:Q4, as well as the first estimate of Gross Domestic Income
in 2018:Q4.A1 None of the variables where adjusted for outliers but variables 57, 83, 87,
and 94. All variables are from the USECON database. All monthly series are transformed
into quarterly observation by simple averages.
In order to choose whether or not to add a linear trend to a variable we test for zero
mean of the first difference process. Let q
mi be the sample mean of ∆yit , γi (j) be the auto-
P
covariance of order j of ∆yit , and γ̄i = T1 Jj=−J γi (j), with J = 15, then if |mγ̄ii | < 1.96
we set bi = 0.

List of Abbreviations
Source:
BLS = US Department of Labor: Bureau of Labor Statistics
BEA = US Department of Commerce: Bureau of Economic Analysis
ISM = Institute for Supply Management
CB = US Department of Commerce: Census Bureau
FRB = Board of Governors of the Federal Reserve System
EIA = Energy Information Administration
WSJ = Wall Street Journal
CBO = Congressional Budget Office
FRBPHIL = Federal Reserve Bank of Philadelphia

F = Frequency T = Transformation SA ξ = Idiosyncratic

Q = Quarterly 0 = None 0 = no 0 = I(0)
M = Monthly 1 = log 1 = yes 1 = I(1)
D = Daily 2 = ∆ log

D = Deterministic Component U = Units

P
0=b ai = T1 Tt=1 ∆yit , bbi = 0 1000-P = Thousands of Persons
1 = OLS Detrending 1000-U = Thousands of Units
BoC = Billions of Chained
$-B = Dollars per Barrel
A1
Note that in a normal year at the end of March the BEA publishes the third estimate GDP in Q4
of the previous year. However, because of the The United States federal government shutdown, which
occurred from December 22, 2018 until January 25, 2019, on March 29th 2019 the BEA published just the
second estimate.

34

Electronic copy available at: https://ssrn.com/abstract=3217816

N Series ID Definition Unit F S SA T D ξ
1 GDPH Real Gross Domestic Product BoC 2012$ Q BEA 1 1 1 0
2 GDYH Real Gross Domestic Income BoC 2012$ Q BEA 1 1 1 0
3 FSH Real Final Sales of Domestic Product BoC 2012$ Q BEA 1 1 1 1
4 IH Real Gross Private Domestic Investment BoC 2012$ Q BEA 1 1 1 1
5 GSH Real State & Local∗ BoC 2012$ Q BEA 1 1 1 1
6 FRH Real Private Residential Fixed Investment BoC 2012$ Q BEA 1 1 1 0
7 FNH Real Private Nonresidential Fixed Investment BoC 2012$ Q BEA 1 1 1 1
8 MH Real Imports of Goods & Services BoC 2012$ Q BEA 1 1 1 0
9 GH Real Government∗ BoC 2012$ Q BEA 1 1 1 1
10 XH Real Exports of Goods & Services BoC 2012$ Q BEA 1 1 1 0
14 CH Real Personal Consumption Expenditures (PCE) BoC 2012$ Q BEA 1 1 1 0
11 CNH Real PCE: Nondurable Goods BoC 2012$ Q BEA 1 1 1 0
12 CSH Real PCE: Services BoC 2012$ Q BEA 1 1 1 0
13 CDH Real PCE: Durable Goods BoC 2012$ Q BEA 1 1 1 0
15 GFDIH Real National Defense Gross Investment BoC 2012$ Q BEA 1 1 1 0
16 GFNIH Real Federal Nondefense Gross Investment BoC 2012$ Q BEA 1 1 1 0
17 YPDH Real Disposable Personal Income BoC 2012$ Q BEA 1 1 1 0
18 JI Gross Private Domestic Investment? 2012=100 Q BEA 1 2 0 0
19 JGDP Gross Domestic Product? 2012=100 Q BEA 1 2 0 1
20 LXNFU Unit Labor Cost† 2012=100 Q BLS 1 1 1 1
21 LXNFR Real Compensation Per Hour† 2012=100 Q BLS 1 1 1 1
22 LXNFC Compensation Per Hour† 2012=100 Q BLS 1 1 1 1
23 LXNFH Hours of All Persons† 2012=100 Q BLS 1 1 1 0
24 LXNFA Output Per Hour of All Persons† 2012=100 Q BLS 1 1 1 0
25 LXMU Unit Labor Cost‡ 2012=100 Q BLS 1 1 1 1
26 LXMR Real Compensation Per Hour‡ 2012=100 Q BLS 1 1 1 1
27 LXMC Compensation Per Hour‡ 2012=100 Q BLS 1 1 1 1
28 LXMH Hours of All Persons‡ 2012=100 Q BLS 1 1 0 1
29 LXMA Output Per Hour of All Persons‡ 2012=100 Q BLS 1 1 1 1
30 IP Industrial Production (IP) Index 2012=100 M FRB 1 1 1 1
31 IP521 IP: Business Equipment 2012=100 M FRB 1 1 1 1
32 IP511 IP: Durable Consumer Goods 2012=100 M FRB 1 1 1 0
33 IP531 IP: Durable Materials 2012=100 M FRB 1 1 1 1
34 IP512 IP: Nondurable Consumer Goods 2012=100 M FRB 1 1 1 0
35 IP532 IP: nondurable Materials 2012=100 M FRB 1 1 1 0
∗ Consumption Expenditures & Gross Investment
? Chain-type Price Index
† Nonfarm Business Sector
‡ Manufacturing Sector

35

Electronic copy available at: https://ssrn.com/abstract=3217816

N Series ID Definition Unit F S SA T D ξ
36 PCU CPI-U: All Items 82-84=100 M BLS 1 2 0 0
37 PCUSE CPI-U: Energy 82-84=100 M BLS 1 2 0 0
38 PCUSLFE CPI-U: All Items Less Food and Energy 82-84=100 M BLS 1 2 0 0
39 PCUFO CPI-U: Food 82-84=100 M BLS 1 2 0 0
40 JCBM PCE: Chain Price Index 2012=100 M BEA 1 2 0 0
41 JCEBM PCE: Energy Goods & Services-price index 2012=100 M BEA 1 2 0 0
42 JCNFOM PCE: Food & Beverages-price index∗ 2012=100 M BEA 1 2 0 0
43 JCXFEBM PCE less Food & Energy-price index 2012=100 M BEA 1 2 0 0
44 JCSBM PCE: Services-price index 2012=100 M BEA 1 2 0 0
45 JCDBM PCE: Durable Goods-price index 2012=100 M BEA 1 2 0 0
46 JCNBM PCE: Nondurable Goods-price index 2012=100 M BEA 1 2 0 0
47 PC1 PPI: Intermediate Demand Processed Goods 1982=100 M BLS 1 2 0 0
48 P05 PPI: Fuels and Related Products and Power 1982=100 M BLS 0 2 0 0
49 SP3000 PPI: Final Demand Personal Consumption Gds∗∗ 1982=100 M BLS 1 2 0 0
50 SP3000 PPI: Finished Goods 1982=100 M BLS 1 2 0 0
51 PIN PPI: Industrial Commodities 1982=100 M BLS 0 2 0 0
52 PA PPI: All Commodities 1982=100 M BLS 0 2 0 0
53 FMC Money Stock: Currency Bil. of $ M FRB 1 2 0 0
54 FM1 Money Stock: M1 Bil. of $ M FRB 1 2 0 1
55 FM2 Money Stock: M2 Bil. of $ M FRB 1 2 0 0
56 FABWC C & I Loans in Bank Credit† Bil. of $ M FRB 1 1 1 1
57 FABWQ Consumer Loans in Bank Credit† Bil. of $ M FRB 1 1 1 1
58 FAB Bank Credit† Bil. of $ M FRB 1 1 1 1
59 FABW Loans & Leases in Bank Credit† Bil. of $ M FRB 1 1 1 1
60 FABYO Other Securities in Bank Credit† Bil. of $ M FRB 1 1 1 1
61 FABWR Real Estate Loans in Bank Credit† Bil. of $ M FRB 1 1 1 0
62 FOT Consumer Credit Outstanding Bil. of $ M FRB 1 1 1 1
63 HSTMW Housing Starts: Midwest 1000-U M CB 1 1 0 0
64 HSTNE Housing Starts: Northeast 1000-U M CB 1 1 0 0
65 HSTS Housing Starts: South 1000-U M CB 1 1 0 0
66 HSTGW Housing Starts: West 1000-U M CB 1 1 0 0
67 HPT Building Permit? 1000-U M CB 1 1 0 0
68 FBPR Bank Prime Loan Rate Percent M FRB 0 0 0 0
69 FFED Federal Funds [effective] Rate Percent M FRB 0 0 0 0
70 FCM1 1-Year Treasury Bill Yield‡ Percent M FRB 0 0 0 0
71 FCM10 10-Year Treasury Note Yield‡ Percent M FRB 0 0 0 0
∗ Purchased for Off-Premises Consumption
† All Commercial Banks
? New Private Housing Units Authorized by
‡ at Constant Maturity
∗∗ [Finished Consumer Gds]

36

Electronic copy available at: https://ssrn.com/abstract=3217816

N Series ID Definition Unit F S SA T D ξ
72 LP Civilian Participation Rate: 16 yr + Percent M BLS 0 0 1 0
73 LQ Civilian Employment/Population Ratio: 16 yr + Percent M BLS 0 0 1 0
74 LE Civilian Employment: Sixteen Years & Over 1000-P M BLS 0 1 1 0
75 LR Civilian Unemployment Rate: 16 yr + Percent M BLS 0 0 0 0
76 LU0 Civilians Unemployed for Less Than 5 Weeks 1000-P M BLS 0 1 1 0
77 LU5 Civilians Unemployed for 5-14 Weeks 1000-P M BLS 0 1 1 1
78 LU15 Civilians Unemployed for 15-26 Weeks 1000-P M BLS 0 1 1 1
79 LUT27 Civilians Unemployed for 27 Weeks and Over 1000-P M BLS 0 1 1 1
80 LUAD Average [Mean] Duration of Unemployment Weeks M BLS 0 1 1 1
81 LANAGRA All Employees: Total Nonfarm 1000-P M BLS 0 1 1 1
82 LAPRIVA All Employees: Total Private Industries 1000-P M BLS 0 1 1 0
83 LANTRMA All Employees: Mining and Logging 1000-P M BLS 0 1 0 1
84 LACONSA All Employees: Construction 1000-P M BLS 0 1 1 1
85 LAMANUA All Employees: Manufacturing 1000-P M BLS 0 1 0 1
86 LATTULA All Employees: Trade, Transportation & Utilities 1000-P M BLS 0 1 1 1
87 LAINFOA All Employees: Information Services 1000-P M BLS 0 1 1 1
88 LAFIREA All Employees: Financial Activities 1000-P M BLS 0 1 1 1
89 LAPBSVA All Employees: Professional & Business Services 1000-P M BLS 0 1 1 1
90 LAEDUHA All Employees: Education & Health Services 1000-P M BLS 0 1 1 1
91 LALEIHA All Employees: Leisure & Hospitality 1000-P M BLS 0 1 1 1
92 LASRVOA All Employees: Other Services 1000-P M BLS 0 1 1 1
93 LAGOVTA All Employees: Government 1000-P M BLS 0 1 1 1
94 LAFGOVA All Employees: Federal Government 1000-P M BLS 0 1 0 1
95 LASGOVA All Employees: State Government 1000-P M BLS 0 1 1 0
96 LALGOVA All Employees: Local Government 1000-P M BLS 0 1 1 0
97 PETEXA West Texas Intermediate Spot Price FOB∗ $-B M EIA 0 2 0 0
98 NAPMNI ISM Mfg: New Orders Index Index M ISM 1 0 0 1
99 NAPMOI ISM Mfg: Production Index Index M ISM 1 0 0 1
100 NAPMEI ISM Mfg: Employment Index Index M ISM 1 0 0 1
101 NAPMVDI ISM Mfg: Supplier Deliveries Index Index M ISM 1 0 0 0
102 NAPMII ISM Mfg: Inventories Index Index M ISM 1 0 0 0
103 SP500 Standard & Poor’s 500 Stock Price Index 41-43=10 D WSJ 0 1 1 0
∗ Cushing, Oklahoma

Series ID Definition Unit F Source

GDPPOTHQ Real Potential Gross Domestic Product BoC 2012$ Q CBO
NAIRUQ Natural Rate of Unemployment percent Q CBO
GDPPLUS US GDPplus percent Q FRBPHIL

37

Electronic copy available at: https://ssrn.com/abstract=3217816

Measuring the output gap using large datasets
Complementary appendix

Matteo Barigozzi Matteo Luciani

London School of Economics Federal Reserve Board
[email protected] [email protected]

A Real-Time
In this section, we present the fully real-time exercise. We retrieved real-time data vintages for
our 103 variables dataset starting in August 30 2013, that is after the 2013 NIPA comprehensive
data revision, and ending on March 29 2019, i.e., the vintage of data that we use to produce the
results in the paper.a Overall, we estimate the output gap on a total of 52 vintages, where each
vintage corresponds approximately to the data available the first Friday after GDI is released.
Details on the construction of the real-time data vintages are in the next section.
Figure A1 shows real-time estimates of the output gap for both our measure and the CBO
measure.b By looking at the figure, two main conclusions can be drawn: first, our estimate of
the output gap revises very little—the average size of the revision for our estimate is 25 basis
points, with a standard deviation of 35 basis points—and less than the CBO estimate. And,
second, by comparing the estimate obtained on the first vintage of data (the blue line) with the
one obtained on the last vintage (red line), it is interesting to notice that both our model and
the CBO model revised in the same direction, i.e., both models interpreted incoming data as
signaling that there was less slack in the economy.
Figure A2 shows the real-time and quasi-real-time estimates (presented in the paper) of the
output gap together in the same graph. As we can see, the real-time estimates and the quasi-
real-time estimates are very similar—the average size of the revision computed on comparable
samples is 18 basis points in the quasi-real-time exercise and 25 basis points in the real-time
exercise, with standard deviations of 26 and 35 basis points, respectively. Therefore, since the
difference between the real-time estimates and the quasi-real-time estimates are solely the results
a
Each year the BEA at the end of July revises NIPA data up to five years back in time, the so-called “annual
revision”. Every five years, the BEA produces a broader revisions of NIPA data, the so-called “comprehensive
revision”, in which it can revise the whole history of every national account variable.
b
The CBO publishes estimates of potential output two times a year (January and August). The first CBO
release that we have in our real-time database is from January 2014, while the last one is January 2019. We were
able to use all the CBO releases from January 2014 to January 2019, but the one from August 2018, for a total of
10 vintages. We cannot use the CBO estimate from August 2018, because the CBO started publishing potential
output in Chained 2012$ only in January 2019, meaning that potential output published in August 2018 was
expressed in Chained 2009$, while GDP was expressed in Chained 2012$—the BEA started publishing National
Account Statistics in Chained 2012$ following the 2018 NIPA comprehensive data revision—thus making the
computation of the output gap very complicated.

Electronic copy available at: https://ssrn.com/abstract=3217816

 Measuring the output gap using large datasets: Complementary Appendix

Figure A1: Output gap

Real-time estimation
BL CBO
1 1

0 0

-1 -1

-2 -2

-3 -3

-4 -4

-5 -5

-6 -6

-7 -7

2011 2012 2013 2014 2015 2016 2017 2018 2011 2012 2013 2014 2015 2016 2017 2018

In each plot the blue line is the estimate of the output gap obtained with the vintage of data from August 30 2013, the
red line is the estimate obtained with the vintage of data from on March 29 2019„ and the yellow lines are the estimate
obtained with all the remaining 31 vintages. Finally, the 19 black dots represent the estimate of the output gap for quarter
Q and year Y obtained with the vintage of data ending at quarter Q and year Y corresponding to the first release of GDI
for quarter Q and year Y. Note that, for CBO the blue line represents the estimate available as of April 11 2014.

of data revisions, we can confirm the results in Orphanides and van Norden (2002) according
to which output gap revisions are mainly caused by parameter estimation, rather than data
revision.
Figure A2: Output gap
Real-time and quasi-real-time estimation
Levels Four-quarter percent change
0 3

2.5
-1
Fin
2 QRT
RT
-2
Fin 1.5
QRT
RT
-3 1

0.5
-4

-5
-0.5

-6 -1
2010 2011 2012 2013 2014 2015 2016 2017 2018 2010 2011 2012 2013 2014 2015 2016 2017 2018

In each plot the red line is the estimate of the output gap obtained on our benchmark sample (the “final” estimate),
the green line is the quasi-real-time estimate, and the yellow line is the real-time estimate, i.e., the green line is the line
connecting the black dots in Figure A1.

A.1 Real-time data vintages

The real-time data vintages are constructed as follows: vintages from February 24 2017 onwards
where saved by us every Friday, while vintages prior to February 24 2017 are constructed using
a real-time database maintained by the Division of Research and Statistics of the Board of
Governors of the Federal Reserve System. All the variables were retrieved from that database
but variable 66 “Housing Starts: West”, which was retrieved from the ALFRED database of the
Federal Reserve Bank of St. Louis. Furthermore, some missing observations for GDI and the oil
price were filled by using real-time vintages downloaded from ALFRED.

Page II of VII

Electronic copy available at: https://ssrn.com/abstract=3217816

 Measuring the output gap using large datasets: Complementary Appendix

For the first four vintages we were able to retrieve only 101 variables out of 103, as we were
not able to find vintages for two PPIs (variables 47 and 49).
Potential output and the output gap were estimated on about 9 data vintages per year
corresponding to the dates after the release of GDI. In a normal year, the first estimate of Q1
GDI is released at the end of May, and then is revised at the end of June and at the end of July;
The first estimate of Q2 GDI is released at the end of August, and then is revised September;
The first estimate of Q3 GDI is released at the end of November, and then is revised at the end
of December and at the end of January of the following year; The first estimate of Q4 GDI is
released at the end of March of the next year.
The dates in which GDI is released varies from year to year, therefore each vintage corre-
sponds to the data that were available the closest Friday to the day in which “Personal Income
and Outlays” are released. Of course, we considered only “Personal Income and Outlays” after
GDI is released.

B Forecasting
As we discussed in the Introduction of the paper, one of the reasons why economists are very
much interested in the estimation of the output gap is that, potentially, the output gap is a
gauge of future inflation. Moreover, because the economy cannot grow forever above/below its
potential growth rate, it is reasonable to expect that the output gap is also a potential gauge
of future growth. Therefore, to further test the reliability of our model and our output gap
estimate, in this Section, we test its out-of-sample behavior. To be more precise, we answer the
following questions: is our output gap estimate better than the one computed by the CBO in
forecasting GDP growth? Is it better in forecasting inflation?
In order to test the forecasting properties of our output gap measure, we run a quasi-real-
time forecasting exercise. For GDP, we forecast the annualized GDP growth rate between time
t and t + h ( h4 (GDPt+h − GDPt )) by using the following forecasting equation:

4
(GDPt+h − GDPt ) = α + β × OGt + t (I)
h

where OGt is the level of the output gap at time t. For inflation, we forecast the annualized
inflation rate between time t and t + h ( h4 (Pt+h − Pt ))by using the following Phillips curve type
forecasting equation:

X 5 X 5
4
(Pt+h − Pt ) = α + βi × OGt−i + γj πt−j + t (II)
h
i=0 j=0

where πt = Pt − Pt−1 (Kamber et al., 2018, among others, also used a similar specification of
the Phillips curve to forecast inflation).
The exercise works as follows: starting from 2000:Q1, (I) we estimate our model thus ob-
taining an estimate of OGt , and then (II) we produce forecasts of output (GDP) growth and
inflation up to h steps ahead by using (I) and (II);c then, we increase the sample size of one
c
Note that for the CBO estimate we use the final estimate, which means that we are using the final CBO

Page III of VII

Electronic copy available at: https://ssrn.com/abstract=3217816

 Measuring the output gap using large datasets: Complementary Appendix

quarter and repeat the procedure up to the end of our sample for a total o (76 − h) forecasts.
Table B1 shows the ratio of the Mean Squared Errors (MSE) obtained by forecasting output
growth and inflation with our output gap estimate, over the MSE obtained when the CBO
estimate is used, together with the p-values of a Diebold and Mariano (1995) type test of equal
predictive ability. A value below 1 implies that, on average, our model did better. As we can
see, the two output gap measures perform about the same in forecasting output growth and
inflation. None of the differences in forecasting performance are statistically significant, with
the only exception being that our output gap estimate performs better than the CBO estimate
in predicting GDP growth at longer horizons.

Table B1: Relative Mean Squared Errors

h 1 2 3 4 5 6 7 8
GDP 0.977 1.004 0.996 0.976 0.951 0.928 0.901 0.876
0.33 0.90 0.93 0.60 0.29 0.11 0.03 0.01
Inflation 0.965 1.040 1.012 1.039 1.116 1.156 1.171 1.175
0.65 0.69 0.91 0.76 0.45 0.37 0.37 0.39
This table shows Relative MSE, i.e., the ratio of the MSE obtained by forecasting output growth and inflation with our
output gap estimate, over the MSE obtained when the CBO estimate is used. A value below 1 implies that on average our
model did better. The forecasts are produced by using equation (I) and (II). Below each of the Relative MSE, we report
the p-value of a Diebold and Mariano (1995) type test of equal predictive ability.

To sum up, our output gap measure performs as good as, if not marginally better than, the
CBO measure in forecasting output growth, and as good as the CBO measure in forecasting
inflation.

C Robustness analysis
In this Appendix we show robustness results about model specification. Specifically, (1) we test
the robustness of our model to changes in the number of common factors q, and (2) to changes
in the number of lags in the factor loadings s and in the VAR for the common factors p. Finally,
(3) we investigate the effect of not allowing for local linear trend in GDO and a local level model
in the unemployment rate.
Figure C1 shows results when the model is estimated with q = 1, . . . , 5 common factors.
The results in Figure C1 confirms those discussed in the paper: in order to have a meaningful
estimate of the output gap it is necessary to include at least three factors in model. By contrast,
in order to have a meaningful estimate of the unemployment rate gap at least four factors should
be included. Finally, a model with five factors produces results in line with those obtained with
our benchmark specification, but simply produce slightly more volatile estimates.
Figure C2 compares results from our benchmark specification {q = 4, s = 1, p = 3}, with a
specification which allows for reacher dynamic {q = 4, s = 4, p = 4}. As we can see, adding lags
to our model has nearly no effects on our estimates, but to add a bit more volatility.
Figure C3 compares results from our benchmark specification, to those obtained without
allowing for a local linear trend in GDO and for a local level model in the unemployment rate.
As we can see, not allowing for a local linear trend in GDP produces a much higher estimate
estimate to forecast the final estimate of GDP growth and PCE price inflation, while we are using our quasi-
real-time estimate for predicting the final estimate of GDP growth and PCE price inflation.

Page IV of VII

Electronic copy available at: https://ssrn.com/abstract=3217816

 Measuring the output gap using large datasets: Complementary Appendix

Figure C1: Robustness analysis: Number of common factors

Output gap Unemployment rate gap
6
6
5
4
4
q=2
2 q=3
3
q=4
q=5
0
2
q=2
-2 q=3 1
q=4
q=5
-4 0

-6
-1

-2
-8
-3
-10
-4
1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015

In each plot the blue line is the estimate obtained when the model with two common factors is estimated. The yellow
line is the estimate obtained when the model with three common factors is estimated, while the red line is the estimate
obtained when the model with four common factors, which is our benchmark specifications, is estimated. The green
line is the estimate obtained when the model with five common factors is estimated.

Figure C2: Robustness analysis: Number of lags s and p

Output gap Unemployment rate gap
6
6
5
4 s=1, p=3
4 s=4, p=4
2
3

0
2

-2 1

-4 0

-6
-1

s=1, p=3 -2
-8 s=4, p=4
-3
-10
-4
1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015

In each plot the blue line is the estimate obtained when the model is estimated by allowing the common factors to affect
each variable with a maximum delay of five lags, and when the VAR for the common factors includes four lags. The
red line is the estimate obtained when the model is estimated by allowing the common factors to affect each variable
with a maximum delay of three lags, and when the VAR for the common factors includes three lags.

of potential growth after the financial crisis, and a lower output gap at the end of the sample.d
Moving to the unemployment rate, not allowing for a local level model produces an estimate of
the unemployment rate gap that it is essentially unchanged.

D Constructing confidence bands

In this Appendix, we explain in detail the algorithm used for constructing confidence bands
around our estimates. From the EM algorithm, we have an estimate of the states b
ft , ξbt , D
bGDO,t ,
bUR,t ; and an estimate of the parameters of the model that we store in the vector ϕ.
and D b Let
ζt be an n-dimensional vector containing the I(1) idiosyncratic components, i.e., ζit = ξit if
ξit ∼ I(1), and ζit = 0 otherwise; and let zt be the vector containing the stationary residual of
the model, i.e., zit = ξit if ξit ∼ I(0), and ζit = t otherwise, where t is the residual from the EM
algorithm. Finally, let P b f = Eϕb[(b ft − ft )(b
ft − ft )0 ], be the conditional variance of the common
t|T
d
Note that not all estimated higher potential is translated into a lower output gap, as actually most of the
higher potential is translated into lower idiosyncratic component at the end of the sample (not shown).

Page V of VII

Electronic copy available at: https://ssrn.com/abstract=3217816

 Measuring the output gap using large datasets: Complementary Appendix

Figure C3: Robustness analysis: No Time-Varying parameters

Output gap Unemployment rate gap
6
6
5
4 BL
4 no TV parameters
2
3

0
2

-2 1
BL
no TV parameters 0
-4

-6
-1

-2
-8
-3
-10
-4
1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015

In each plot the blue line is the estimate obtained when the model is estimated without a local linear trend for GDO
and a local level model for the unemployment rate. The red line is our benchmark estimate and the red shaded area
are the 64% and 84% confidence bands.

b ζ be the conditional variance of the non-stationary idiosyncratic

factors, and likewise let P t|T
components, and P b D be the conditional variance of the time-varying deterministic component
t|T
of GDO and the unemployment rate. Then, for each repetition b = 1, . . . , B.

1. Simulate the states by the simulation smoother (Durbin and Koopman, 2002):

(a) Common factors:

(b)
i. simulate e
f1 ∼ N (b bf )
f1 , P1|T
(b) b
ii. simulate ue t ∼ N (0q , Q)
(b) P (b) (b)
iii. for t = 2, . . . , T generate e bke
ft = k=1 pA et
ft−k + u
(b) I(1) idiosyncratic components: for each i
(b)
i. simulate ζei1 ∼ N (ζbi1 , Pbi,1|T
ζ
)
(b)
ii. simulate eeit ∼ N (0, 1)
(b) (b)
iii. for t = 2, . . . , T generate ζeit = ζeit−1 + eeit
(c) Time-Varying deterministic components:
◦ GDO
(b) e (b) e(b)
DGDO
i. simulate ebGDO,1 ∼ N (bbGDO,1 , Pb1|T ), and set D GDO,1 = bGDO,1
(b)
ii. simulate ηet ∼ N (0, 1)
(b) (b)
iii. for t = 2, . . . , T generate ebGDO,t = ebGDO,t−1 + ηet
e (b)
iv. for t = 2, . . . , T generate D e
GDO,t = DGDOt−1 + bGDO,t
◦ Unemployment rate
e (b) ∼ N (bbUR,1 , PbDUR )
i. simulate D UR,1 1|T
(b)
ii. simulate ηet ∼ N (0, 1)
e (b) = D
iii. for t = 2, . . . , T generate D eURt−1 + ηe(b)
UR,t t

2. Simulate the stationary residual of the model by the stationary bootstrap (Politis and
Romano, 1994).

Page VI of VII

Electronic copy available at: https://ssrn.com/abstract=3217816

 Measuring the output gap using large datasets: Complementary Appendix

3. Generate the data: for t = 1, . . . , T generate

(b) e (b) + Ps λ
(a) yeit = D b e(b) e(b) e(b) , for GDP, GDI and the Unemployment rate
it `=0 i` ft−` + ζit + z it
(b) bit + P s b e (b) e(b) (b)
(b) yeit = D eit , for all other variables
`=0 λi` ft−` + ζit + z

(b)
4. using y
et run the EM algorithm to get a new estimate of the parameters ϕ
b(b) , and the
(b) (b) (b)
states f̌t , ζ̌t , Ďt
(b) (b) (b) (b) (b) (b) b (b) =
5. center the estimated states by: b
ft = b
ft − e
ft + f̌t , ζbt = ζbt − ζet + ζ̌t , and Dt
bt − D
D e (b) + Ď(b) .
t t

(b)
6. Run the TC decomposition on the estimated factors b ft to get a new estimate of the
(b) (b) b(b) .
b(b) and ψ
common trend τbt , the common cycles cbt , and the parameters ψ ⊥

d (b)
b (b) Ps b (b) 0
b d b(b) .
7. Estimate, for example, potential output as P Ot = DGDO,t + `=0 λGDO,` Φ τ t−`

(b)
d
Repeating this procedure several times gives a distribution of, say, potential output: {P Ot , b =
1, . . . 1000}. In order to compute the (1 − α) confidence interval, at each quarter t we compute
(b)
the sample variance of {P d O −P d Ot }, which we denote as σ 2 , and then we construct the (1 − α)
t t
d
confidence interval is given by [P d
Ot + zα/2 σt , P Ot + z1−α/2 σt ], where zα/2 = −z1−α/2 is the
α/2 quantile of a standard normal.

References
Diebold, F. X. and Mariano, R. S. (1995). Comparing predictive accuracy. Journal of Business
and Economic Statistics, 13:253–263.

Durbin, J. and Koopman, S. J. (2002). A simple and efficient simulation smoother for state
space time series analysis. Biometrika, 89:603–615.

Kamber, G., Morley, J., and Wong, B. (2018). Intuitive and reliable estimates of the output gap
from a Beveridge-Nelson filter. Review of Economics and Statistics, 100:550–566.

Orphanides, A. and van Norden, S. (2002). The unreliability of output-gap estimates in real
time. The Review of Economics and Statistics, 84:569–583.

Politis, D. N. and Romano, J. P. (1994). The stationary bootstrap. Journal of the American
Statistical Association, 89:1303–1313.

Page VII of VII

Electronic copy available at: https://ssrn.com/abstract=3217816