Inflation at Risk

David López-Salido Francesca Loria

Federal Reserve Board

January 6, 2022

Abstract

We investigate how macroeconomic drivers influence the predictive inflation distribution and
establish two key findings. First, the recent muted response of the conditional mean of infla-
tion to economic conditions does not convey a complete picture of inflation dynamics. Indeed,
we find ample variability in the tails of the inflation outlook that remains even when focusing
on the most recent period of stable and low mean inflation. Second, we document that tight
financial conditions carry substantial downside inflation risks in the United States and in the
Euro Area, a feature overlooked by much of the literature but consistent with financial amplifi-
cation mechanisms. Finally, we show that evidence from financial market quotes, from survey
data and from a regime-switching model of inflation is consistent with our findings and use our
model to track inflation risks during the Covid-19 crisis.

JEL C LASSIFICATION: C21, C53, E31, E44.

K EYWORDS : Inflation Risks, Quantile Regression.

Special thanks to Ben Bernanke, Danilo Cascaldi-Garcia, Steve Cecchetti, David Cho, Jim Clouse, Olivier Coibion,
Deepa Datta, Giovanni Favara, Felix Galbis-Reig, Ed Herbst, Paul Lengermann, Antoine Lepetit, Ed Nelson, Claudia
Pacella, Andrea Prestipino, Giorgio Primiceri, Jeremy Rudd, Tatevik Sekhposyan and Srečko Zimic as well as seminar
participants of the Research & Statistics Lunch Seminar at the Federal Reserve Board, of the DG-E Research Seminar
at the European Central Bank, of the 2020 CEBRA Annual Meeting Session on “Inflation Dynamics and the Phillips
Curve”, the “Macro-at-Risk” Session at the ECB Working Group on Econometric Modelling, of the IMF Monetary and
Capital Markets (MCM) Policy Forum, the 27th International Conference of Computing in Economics and Finance
(CEF), the 2021 International Association for Applied Econometrics (IAAE) Annual Conference, the 2021 Meeting
of the Society for Economic Dynamics (SED), and of the International Conference on Economic Modeling and Data
Science (EcoMod2021) for their useful comments and suggestions to earlier versions of the paper. We also thank
Eugenio Cerutti, Michiel De Pooter, Caitlin Dutta, Joaquin Garcı́a-Cabo Herrero, Louisa Liles, Luke Lillehaugen,
Matteo Luciani, Paul Tran and Riccardo Trezzi for their help in collecting some of the data used in the paper.
Disclaimer: The views expressed in this paper are solely the responsibility of the authors and should not be interpreted
as reflecting the views of the Board of Governors of the Federal Reserve System or of anyone else associated with the
Federal Reserve System.
Francesca Loria (corresponding author). Federal Reserve Board, Research & Statistics Division. 20th Street and
Constitution Avenue N.W., Washington, D.C. 20551, USA. Email: [email protected].
David López-Salido. Federal Reserve Board, Monetary Affairs Division. 20th Street and Constitution Avenue N.W.,
Washington, D.C. 20551, USA. Email: [email protected].
“Monetary policy responded first in the summer of 2012 by acting to defuse the sovereign debt
crisis, which had evolved from a tail risk for inflation into a material threat to price stability.”

Mario Draghi, ECB President, Sintra, June 2019.1

Introduction

Since the upheavals of the global financial crisis, the emergence of downside risks to the inflation
outlook have increasingly become a source of macroeconomic concern. So far, most efforts have
been devoted to studying the factors underlying the muted response of the conditional mean of
inflation to economic and financial conditions. At the same time, much has been said on the in-
ability of labor market conditions to explain recent inflation outcomes. The Phillips curve linkages
seem to be breaking down. In this paper, we show that in the presence of tail risks, the conditional
inflation mean does not necessarily fully depict the inflation outlook, as reminded by President
Draghi’s quote.
Indeed, we show that the contrasting response of the inflation tails and median reveals a more
complete picture of the effects that real and financial shocks impinge on its outlook. Specifi-
cally, we find that there have been sizeable downside risks to the inflation outlook in the last 20
years, mainly accounted for by financial tightenings. This is consistent with the idea that, due to
amplification mechanisms, when financial conditions become tighter firms cut prices dispropor-
tionately more, on average. These concepts are related to recent research by Del Negro, Giannoni,
and Schorfheide (2015), Christiano, Eichenbaum, and Trabandt (2015), Christiano, Motto, and
Rostagno (2014) and Gilchrist, Schoenle, Sim, and Zakrajšek (2017) which shows that financial
conditions matter also for inflation dynamics. While these studies focus on the modal outlook, we
draw on their insights to study empirically the inflation tails.
Other studies already documented that looking at the entire predictive distribution of economic

1
Mario Draghi, “Twenty Years of the European Central Bank’s Monetary Policy,” speech delivered at the ECB
Forum on Central Banking in Sintra on June 18th, 2019 (available at https://www.bis.org/review/r190618c.htm).

1
growth can reveal additional insights into their dynamics. For instance, the deterioration in credit
market conditions led to a substantial decline in economic activity as well as a deterioration in the
odds of low growth and of high unemployment. Tight financial conditions moved the conditional
distribution of real GDP growth to the left (e.g., Adrian, Boyarchenko, and Giannone, 2019 and
Caldara, Cascaldi-Garcia, Cuba Borda, and Loria, 2020) – with its left tail being the most sensitive
to macroeconomic shocks (see Loria, Matthes, and Zhang, 2019) – and implied medium-term
upside risks to unemployment (see Kiley, 2018).
To summarize, in this paper we make three points. First, we offer evidence that some of the
macroeconomic factors covered under the “Phillips curve umbrella” – conventionally used to study
the conditional mean of inflation – are still at work in the tails. Indeed, we find ample variability
in the tails of the inflation outlook that remains even when focusing on the most recent period of
stable and low mean inflation. Second, we show that tight financial conditions carry substantial
downside risks to the inflation outlook, an aspect of inflation behavior overlooked by much of the
literature but consistent with financial amplification mechanisms. For instance, we replicate our
findings within the model by Gertler, Kiyotaki, and Prestipino (2019) - a nonlinear DSGE model
which features the possibility of a severe financial crisis. Third, we uncover that these two findings
are supported by evidence from inflation probabilities derived from financial contracts, predictive
densities from the SPF and a regime-switching model.
Our econometric strategy frames the effects of risk factors on inflation within an “augmented”
quantile Phillips curve model using data since the 1970s for the U.S. economy. That is, we extend
the standard regression analysis (e.g., Blanchard, Cerutti, and Summers, 2015) – designed to as-
certain the drivers of the conditional mean of inflation – to different inflation quantiles. This setup
allows to relate the risks to the inflation outlook to labor market slack, inflation inertia and inflation
expectations, as well as relative prices. More importantly, we extend the analysis to consider the
effect of financial conditions on the inflation distribution and on the odds of low inflation. To do
so, we construct the predictive distribution of the inflation outlook by fitting a flexible distribution
on the estimated inflation quantiles.

2
 We check whether probabilities of inflation falling within certain intervals, as priced by finan-
cial market quotes, are consistent with some of the conclusions about inflation risks derived from
our analysis. We show that these probabilities and the ones obtained from our quantile Phillips
curve model share a defining feature, namely that tight financial conditions are associated with
higher probabilities of low inflation and that this relationship weakens as one considers higher
inflation cutoffs. We also compare the predictive densities obtained from our quantile regression
model with those obtained from the SPF and find that they are remarkably similar.
To shed light on the interpretation and on the sources of identification of the inflation tails
coming from the quantile regression, we run a regime-switching version of our augmented Phillips
curve model. The regime-switching regression is consistent with the main findings from our quan-
tile regression approach. In particular, we show that the Phillips-curve linkages in the left, median
and upper inflation tails arising from the quantile regression are comparable to the Phillips-curve
linkages informed by regimes of low, moderate and high inflation. In terms of our main result,
we show that in regimes of low inflation, credit spreads disproportionately held down inflation as
compared to times of moderate and high inflation.
We consider the recent global financial crisis to illustrate how economic and financial head-
winds influenced the inflation outlook both in the United States and in the euro area. We show
how in the U.S., while average inflation experienced only a modest reduction despite the fall in
output triggered by the financial crisis, downside inflation risks moved considerably over this pe-
riod – mainly a reflection of soaring credit spreads during the financial meltdown. These patterns
have been less benign in the euro area, where the sovereign debt crisis triggered a more prolonged
increase in the odds of low inflation due to the more limited role of inflation expectations in coun-
teracting downside risks posed by the economic slowdown and financial distress.
Finally, we use our framework to characterize inflation risks in uncertain times. We find that
the inflation distributions implied by our model for the recent Covid-19 episode correctly display
an increase in downside inflation risks at the onset of the crisis and upside inflation risks in the
recovery.

3
Related Studies Our paper speaks to the literature which studies risks to the inflation outlook.
Andrade, Ghysels, and Idier (2012) introduced the nomenclature “inflation-at-risk” when con-
structing a measure of (left and right) tail risks to inflation using survey-based density forecasts.
They also showed that the magnitude and the asymmetry of inflation risks evolves over time and
that it is not only related to purely nominal factors but also informed by financial variables, among
others. Kilian and Manganelli (2007, 2008) derive inflation risk measures from the private sec-
tor agent’s preferences with respect to inflation. In a cross-section of countries, Cecchetti (2008)
computes t-distribution approximations to deviations of log GDP and log price level from their
trend and documents that asset price booms increase both growth and inflation risks. Manzan
and Zerom (2013) find that incorporating macroeconomic variables into quantile regressions im-
proves the accuracy of inflation density forecasts. Korobilis (2017) finds that predictive densities
from a quantile regression Bayesian model averaging (QR-BMA) model are superior to and bet-
ter calibrated than those of the traditional regression BMA model and that this methodology is
competitive with popular nonlinear specifications for U.S. inflation. Galvão and Owyang (2018)
find that financial conditions have stronger effect on inflation on periods of “financial stress” in
their factor-augmented smooth-transition vector autoregressive model (FASTVAR). Ghysels, Ia-
nia, and Striaukas (2018) construct measures of inflation risk using a Quantile Autoregressive
Distributed Lag Mixed-Frequency Data Sampling (QADL-MIDAS) regression model and find that
they contain useful information about future inflation realizations. Adams, Adrian, Boyarchenko,
and Giannone (2021) construct risks around consensus forecasts of inflation, among others.
We consider our contribution to this literature as augmenting a well-understood and micro-
founded time-series framework, that of a Phillips-curve, to study the risks to the entire distribution
of the inflation outlook, coming from both conventional inflation determinants as well as financial
conditions. Moreover, we offer a comprehensive review of how evidence on the role of financial
conditions for downside risks to the inflation outlook is supported by a comparison with euro-area
results as well as alternative frameworks to measure inflation-at-risk: survey evidence, financial
market quotes, a regime-switching model and a macroeconomic model with financial panics.

4
 Our approach differs from, and complements, studies that define inflation risks as the chance
of lost purchasing power resulting from negative inflation-adjusted returns. These studies evalu-
ate the inflation risk premium associated with the compensation required by investors for future
expected inflation or deflation – typically using information contained in financial market quotes
(see, e.g., Boons, Duarte, de Roon, and Szymanowska, 2020 among many others). An important
departure from our approach is that, in general, they lack an explicit link of these risks to specific
macroeconomic outcomes.
Our framework has been taken as a starting point in some recent papers, who confirm our
findings that financial conditions are important determinants of inflation risks, and especially of
downside risks. Two examples are Korobilis, Landau, Musso, and Phella (2021) in the context
of both our semi-structural Phillips curve model as well as other time series models with time-
varying parameters, and Banerjee, Contreras, Mehrotra, and Zampolli (2020) in a cross-section of
advanced and emerging economies using panel quantile regressions.

Outline In Section 1 we organize ideas by presenting our theoretical framework and empirical
strategy. As time-variation emerges in the characterization of the determinants of the inflation
distribution, we illustrate subsample results in Section 2 and use them to shed new light on modern
inflation linkages. In Section 3 we show supporting evidence for our main findings coming from
financial-markets-derived inflation probabilities, survey data and a regime-switching version of
our augmented Phillips curve model. We focus on the global financial crisis and compare the
United States and euro area inflation experiences in Section 4, while also exploring the role of
financial conditions in affecting the odds of low inflation during that period. In Section 5 we
show how our model tracks inflation risks during the Covid-19 crisis. Concluding remarks, future
research avenues and policy implications are offered in Section 6. A data guide, additional material
and robustness exercises are collected in the Appendix. We refer to this appendix material either
explicitly in the main text or in footnotes.

5
1 Characterizing Inflation-at-Risk

In many circumstances the study of the determinants of the conditional mean of inflation may be
sufficient to produce a good representation of the modal dynamics of inflation. In other cases,
however, studying the response of the tails of the predictive inflation distribution is essential for
providing a more complete picture. This is likely to be the case, for instance, in the presence
of large real or financial shocks, as it aids understanding the effects that these shocks have on
inflation. Because of these considerations, we extend the standard regression analysis – designed
to ascertain the drivers of the conditional mean of inflation – to the entire inflation distribution.
In this section we describe the econometric specification we use to link economic and financial
conditions with risks to the inflation outlook. We first describe conditional inflation quantiles
as a function of observed economic and financial variables (risk factors). Second, we use these
quantiles to approximate the inflation distribution. Variations in inflation risks are then measured
according to how much the tails of the inflation distribution vary with the evolution of economic
and financial factors. We refer to these “tail risks” as Inflation-at-Risk (IaR).
We frame the effects of different risk factors on inflation within an augmented quantile Phillips
curve model. This setup allows us to relate inflation risks to variations in the amount of slack in
the labor market, changes in inflation persistence, variations in inflation expectations, as well as
movements in relative prices (imported goods and/or oil). Our Phillips curve model is “augmented”
as it also incorporates financial conditions (approximated by credit spreads) as an additional factor
affecting not just the mean, but mainly the tails of the inflation distribution.

1.1 (Phillips-Curve) Quantile Regressions

Quantile regression models are a flexible tool for studying the determinants of IaR.2 Our inflation
measure of interest is the (annualized) average core CPI inflation rate over the next year (that is,
between quarter t + 1 and quarter t + 4, π̄t+1,t+4 ).3 We want to focus on quarterly inflation as this is

2
For an introduction to the quantile regression methodology, see Koenker (2005).
3
A similar approach is taken in Adrian, Boyarchenko, and Giannone (2019) for the average growth rate of GDP.

6
typically the main frequency of interest and as it allows to abstract from noisier movements in the
monthly series. We consider a linear model for the conditional inflation quantiles whose predicted
value
Q
bτ (π̄t+1,t+4 |xt ) = xt β̂τ , (1)

is a consistent linear estimator of the quantile function of π̄t+1,t+4 conditional on xt – where

τ ∈ (0, 1), xt is a 1 × k-dimensional vector of conditioning (risk) variables, and β̂τ is a k × 1-
dimensional vector of estimated quantile-specific parameters.4 Accordingly, a determinant xt may
exert non-linear effects on inflation dynamics if it affects differently the median and the tails.
Some observers might wonder about how having overlapping observations in our dependent
variable might affect our results. As discussed in Caldara et al. (2020), running a regression of
this type, is aking to a “direct” forecast, as opposed to an “iterated” forecast where the dependent
variable would be πt+1 and where one would iterate the one-step-ahead prediction to obtain multi-
horizon forecasts. As shown in that same paper, in simulation direct and iterated forecasts deliver
the same results, when they share the same linear (V)AR data-generating process. In empirical
data the direct model actually is comparable to if not better than an iterated model in terms of
coverage and correct calibration of the predictive density.5
Our model for conditional inflation quantiles extends the Phillips-curve model used in the liter-
ature. In particular, we closely follow Blanchard, Cerutti, and Summers (2015) which summarized
a vast empirical literature on inflation dynamics. Formally, the baseline quantile regression model
in (1) can be written as an augmented Phillips curve model:

∗
Q
bτ (π̄t+1,t+4 |xt ) = µ̂τ + (1 − λ̂τ )πt−1 + λ̂τ πtLT E + θ̂τ (ut − u∗t ) + γ̂τ (πtR − πt ) + δ̂τ Ft , (2)

where risk factors affecting the distribution of future inflation can be divided in different blocks.
4
Formally, the dependence between xt and a quantile τ ∈ (0, 1) of π̄t+1,t+4 is measured by the coefficient β̂τ :
T −h
τ · 1(π̄t+1,t+4 ≥xt β) |π̄t+1,t+4 − xt βτ | + (1 − τ ) · 1(π̄t+1,t+4 <xt β) |π̄t+1,t+4 − xt βτ | ,
X

β̂τ = arg min
βτ ∈Rk t=1
where 1(·) denotes the indicator function, taking the value one if the condition is satisfied.
5
This can be rationalized by the fact that the direct model is agnostic about the nature of the evolution of the
dependent variable over horizons and thus performs best when the DGP departs from a linear AR process.

7
A full description of the data is provided in Appendix A.
∗
First, the variables πt−1 and πtLT E respectively represent average inflation over the previous
four quarters and a measure of long-term inflation expectations. Lagged average inflation captures
the role of “intrinsic persistence” or different forms of inertia in the price setting process that could
precipitate upward or downward drift in the aggregate inflation rate.6 In some models, this variable
proxies adaptive or non-rational expectations whereas in others it is used to capture backward-
looking or simple rule-of-thumb pricing rules. Long-term inflation expectations approximate the
importance of some firms setting prices in a rather forward-looking way. Which of these two
elements dominates the persistence observed in the distribution of aggregate inflation depends
on the size of the parameter λτ . To preserve the notion that inflation persistently deviates from
longer-run inflation expectations, we impose the homogeneity constraint in prices by constraining
the two coefficients to sum up to one. When λτ = 0, the model becomes an extension of the
accelerationist Phillips curve, where changes in inflation are a function of the unemployment gap.
We impose (1 − λτ ) + λτ = 1, 0 ≤ (1 − λτ ) ≤ 1 and 0 ≤ λτ ≤ 1 using the inequality constrained
quantile regression method developed by Koenker and Ng (2005).
The second risk factor is linked to variations in the amount of labor market slack – as measured
by the unemployment gap (ut −u∗t ), where ut is the civilian unemployment rate and u∗t is the natural
rate of unemployment. Most of the recent literature has concentrated on the stability over time of
the parameters λ and θ to explain the evolution of average inflation. This literature has focused, for
instance, on understanding the failure of average inflation to respond to unemployment – i.e., the
flattening of the Phillips curve – and on the increasingly dominant role of inflation expectations
in explaining inflation persistence – i.e., the well-anchoring of long-run inflation expectations.
In this paper we extend this analysis by exploring the effects of these variables on the tails of
the distribution of inflation. The importance of these effects is captured by the variation across
quantiles of the parameters λτ , (1 − λτ ) and θτ in expression (2).
The third risk factor in (2) is given by (πtR − πt ), which reflects variations in relative prices.
6
Wolters and Tillmann (2015) use a quantile regression model of core CPI and core PCE inflation which solely
conditions on past inflation to study how inflation persistence differs across quantiles.

8
We use the quarterly change in relative import prices (πtI − πt ). As in Blanchard, Cerutti, and
Summers (2015), this variable is usually included to capture the pass-through of both nominal
exchange rates and oil prices into core inflation measures and is perceived as having been a key
driver of the run-up of inflation in the late seventies and the eighties. Lately, this variable has been
used to approximate a wide range of risk factors, from changes in global commodity prices, taxes
and tariffs to other global influences on domestic inflation. Its effects on the inflation distribution
are captured by the cross-quantile variation in the parameters γτ in expression (2).
The last, but not least, risk factor that we consider is related to financial conditions. Accord-
ing to conventional wisdom, economic factors – labor market slack, inflation expectations, and
relative prices – have been considered as the major sources of variation in the conditional mean
of inflation. However, recent research by Del Negro, Giannoni, and Schorfheide (2015), Chris-
tiano, Eichenbaum, and Trabandt (2015), Christiano, Motto, and Rostagno (2014) and Gilchrist,
Schoenle, Sim, and Zakrajšek (2017) suggests that changes in firms’ financial conditions (proxied
by variations in credit spreads) also helps to explain inflation dynamics. After the financial stress
of the fall of 2008, these studies aim at explaining how the sharp contraction in economic activity
was accompanied by only a modest decline in (average) inflation. However, they mostly discuss
the role of financial frictions in amplifying the business cycle and creating adverse feedback loops,
while leaving its implications for inflation not fully developed. We thus allow for financial condi-
tions Ft in expression (2), to affect differently the conditional inflation quantiles. This allows a test
for the presence of differential effects of financial variables on the mean versus the tails of the infla-
tion distribution (e.g., through the variation in δτ ). Following these authors, and as recommended
by Gilchrist and Zakrajšek (2012), we approximate Ft by the credit spread, cst .

1.2 Quantile Function of Inflation

The estimated conditional quantiles are approximations to the so-called “quantile function”, that is,
Qτ (π̄t+1,t+4 |xt ) = Fπ̄−1
t+1,t+4
(τ |xt ), where F −1 (·) is the conditional inverse cumulative distribution
function (CDF) of average future inflation. We follow Adrian, Boyarchenko, and Giannone (2019)

9
in smoothing the quantile function using the skewed t−distribution proposed by Azzalini and
Capitanio (2003). This flexible distribution is characterized by four parameters and given by:
s !
2 κt + 1
f (π̄t+1,t+4 |xt , µt , σt , ηt , κt ) = × t (zt,t+4 ; κt ) × T ηt zt,t+4 2 ; κt + 1 , (3)
σt κt + zt,t+4

π̄t+1,t+4 (xt )−µt

where zt,t+4 = σt
and t and T respectively represent the density and cumulative dis-
tribution function of the student t-distribution. The constants µt ∈ R and σt ∈ R+ are location
and scale parameters, whereas the constants ηt ∈ R and κt ∈ Z+ control the skewness and the
kurtosis of the distribution, respectively. We compute these parameters at each point in time t to
minimize the squared distance between our estimated quantile function Q
bτ (π̄t+1,t+4 |xt ), obtained

from the quantile Phillips-curve model (2), and the quantile function of the skewed t−distribution
Fπ̄−1
t+1,t+4
(τ |xt , µt , σt , ηt , κt ) to match the 10th , 25th , 75th and 90th quantiles.

2 The Time-Varying Dynamics of Inflation-at-Risk

Two distinct subsamples emerge when characterizing the determinants of the inflation distribution
in the United States.7 The first period, running from 1973:Q1 to 1999:Q4, covers the OPEC shocks,
the subsequent Volcker disinflation and the early stages of the Great-Moderation. The presence of
large shocks to relative prices and the taming of inflation expectations induced large swings in the
upper quantile, while changes in unemployment and past inflation affected the median. These are
ubiquitous themes in the description of inflation dynamics.
The first period contrasts with the second subsample, from 2000:Q1 to 2019:Q1, characterized
by large movements in credit spreads, progressively well-anchored inflation expectations but sub-
dued inflation pressures. These patterns are studied in most of the literature discussing a favorite
whipping boy – the flatness of the Phillips curve. Well-anchored long-run inflation expectations
Yellen (2013), systematic monetary policy (e.g. Ball and Mazumder, 2019, McLeay and Tenreyro,

7
We study the role of economic and financial conditions for the risks to the United States inflation outlook using
our full sample in Appendix B.

10
2018) and mismeasurement of labor market slack (e.g., Stock and Watson, 2019) are the usual
suspects in explaining the observed muted response in average inflation, also referred to as the
“missing deflation/inflation puzzle” Williams (2010) (more on this in the next subsection). How-
ever, the financial crisis and the period in which monetary policy has been constrained by the zero
lower bound, have been followed by a period of underperformance of inflation relative to explicit
or implicit inflation targets. This period has ended with reductions, of different size, of long-term
inflation expectations. Some authors have pointed out that the risks of persistent below-target in-
flation are associated with the emergence of this phenomenon and claim that this set the seeds for
further downside risks to inflation. Through this section we will show that tight credit conditions
arising from financial crises also contributed to increasing odds of low inflation or even deflation,
pointing to a greater role of labor market recovery and well-anchored inflation expectations in sup-
porting average inflation. This point will be further investigated in the Section 4 using contrasting
evidence from the United States and the euro area.

2.1 Subsample Stability and the Missing Deflation/Inflation

To investigate how the importance of risk factors changed across the two subsamples, we report
their estimated quantile-specific slopes in Figure 1. Three results stand out. First, relative import
prices still pose threats to the upper inflation quantile, though to a lesser degree than prior to the
Great Moderation. Second, inflation inertia has completely lost its grip on inflation, crowning long-
run inflation expectations as the decisive inflation determinant among the variables in the modern
Phillips curve, as supported by the time-varying parameter model for mean inflation in Blanchard,
Cerutti, and Summers (2015).8 Third, long-run inflation expectations exert a symmetric effect on
the inflation distribution. In fact, in the recent period of well-anchored long-run inflation expecta-
tions the response of average inflation to labor market slack, financial conditions and relative price

8
Among others, we use long-term inflation expectations from Consensus Economics as it provides equivalent data
for the euro area, an interesting comparison to the U.S. experience (see Section 4). Results are similar if we use
long-term inflation expectations from the SPF or Michigan survey, which are respectively available from 1987Q1 and
1981Q1.

11
changes is dampened. However, it would be misleading to dismiss the role of these factors focus-
ing on the conditional mean only. Instead, credit conditions and, to a lesser extent, labor market
outcomes are key drivers of the asymmetry in the inflation distribution.

Figure 1: Quantile Regression Slopes Across Subsamples.

0
0.1
-0.2 0.08

-0.4 0.06
0.04
-0.6
0.02
-0.8 0
1973-1999 2000-2019 1973-1999 2000-2019

1
0.6
0.8
0.4 0.6

0.4
0.2
0.2

0 0
1973-1999 2000-2019 1973-1999 2000-2019

-0.5
0
-1

-1.5 -0.2

-2
-0.4
1973-1999 2000-2019

N OTE: The figure shows the estimated slopes of the quantile regression of average four-quarter-ahead core CPI
inflation in the U.S., defined in expression (2). Two different subsamples are considered: (i) 1973-1999 and (iii)
2000-2019. The bars illustrate the coefficients for the 10th quantile (blue), median (red) and 90th quantile (yellow).

The Role of Credit Spreads There is substantial subsample instability governing the link be-
tween the inflation tails and variations in credit spreads. The sub-period 1973-1999 is characterized
by relatively small variations in credit spreads in a period of high and volatile inflation induced in
part by systematic increases in energy prices – reason why the actual contribution of credit spreads
to the inflation outlook in that period is smaller. From 2000 onward, low variability of inflation

12
around 2 percent has been a notable aspect of the stability of the macroeconomic landscape that
has coexisted with substantial variation in credit spreads, a phenomenon amplified by the global
financial crisis.
Focusing on the most recent subsample, a novel result is that the link between the inflation
outlook and financial conditions is asymmetric, in that an increase in credit spreads is associated
with a larger increase in downside risks than in upside risks. (We performed ANOVA tests on the
equality of the slope coefficients across quantiles and can reject the null hypothesis of equality
between the slopes on credit spreads for the 10th and 50th quantile at 1% significance level).
This result resonates with the findings by Korobilis, Landau, Musso, and Phella (2021) in
the context of both time-varying parameter and stochastic volatility versions of both our semi-
structural Phillips curve model and competitive time series models, by Cecchetti (2008) and Baner-
jee, Contreras, Mehrotra, and Zampolli (2020) in a cross-section of countries and is reminiscent of
the results for the GDP growth outlook in Adrian, Boyarchenko, and Giannone (2019). Moreover,
this result can be generated in nonlinear models with amplification mechanisms (e.g., financial
accelerator models). Specifically, it captures the notion that when financial conditions become
tighter firms cut prices disproportionately more, on average.9 In Appendix C, we show that our
findings can for instance be replicated by applying our quantile regression framework to simulated
data generated from the model by Gertler, Kiyotaki, and Prestipino (2019) - a fully micro-founded
and nonlinear DSGE model which features the possibility of a severe financial crisis.
In Appendix D, we show that these results hold also with different measures of financial con-
ditions and of inflation. Most importantly, the same patterns emerge when controlling for SPF
forecasts, thus proving that the credit spread is informative for inflation beyond its forward-looking
component and, thus, its information content for future economic activity.

9
Gilchrist, Schoenle, Sim, and Zakrajšek (2017) explain why when firms experience a large drop in their liquidity
that pushes them to their constraints, such as in the Great Recession, one might see less deflation than otherwise pre-
dicted from a model with homogenous firms that does not take into account liquidity constraints. Liquidity-constrained
firms restrain from cutting prices below marginal costs to support their cash-flow. However, this does not mean that
tighter financial conditions result in “higher” inflation. The predominant effect is still the one on the extensive mar-
gin, whereby the not liquidity-constrained firms lower their prices disproportionately more than during normal times.
During regimes of low inflation, tight credit conditions thus impinge a more negative effect than usual.

13
A Modern Phillips Curve The time-varying sensitivity of inflation to inflation expectations (and
inflation inertia), as well as its ability to explain the “missing deflation/inflation” puzzle, had al-
ready been explored and established by Blanchard, Cerutti, and Summers (2015). In Appendix E
we extend their findings to the entire distribution of the inflation outlook by constructing “coun-
terfactual” inflation distributions and inflation-at-risk probabilities from the perspective of these
two different subsamples. Our results confirm that the conventional wisdom from the pre-2000
economy would have suggested much larger deflation risks during the Great Recession and much
lower upside risks to the inflation outlook during the recovery than the modern Phillips-curve re-
lationship. We take this as evidence that the model estimated starting from the 2000’s does indeed
better characterize the inflation dynamics in recent times.

Predictive Ability of Credit Spreads Next, we show that for the sample starting in the 2000s,
the augmented Phillips curve model outperforms the standard model that ignores credit spreads
in terms of predictive ability. In Section B.1 of Appendix B we find that the augmented model is
also superior in terms of coverage by applying the Rossi and Sekhposyan (2019) test for correct
calibration of the predictive density .
The reliability of the predictive distribution can be assessed by measuring the accuracy of the
model’s density forecasts through its predictive scores. These are computed by evaluating the
model’s predictive distribution at the realized value of the time series. A higher predictive score
indicates more accurate predictions, as the model assigns a higher probability to outcomes that are
closer to the realized value. We compute the predictive scores in an out-of-sample exercise where
the predictive distributions are calculated using an expanding window.
Figure 2 plots the out-of-sample predictive scores for the modern Phillips curve model esti-
mated starting in 2000:Q1 (see Section 2). The augmented Phillips curve model has a higher pre-
dictive ability, on average, than the standard Phillips curve model that ignores credit spreads. We
take this as evidence of the importance of considering financial conditions in the modern Phillips
curve in characterizing the inflation outlook.

14
 Figure 2: Out-of-Sample Predictive Scores. Model Estimated from 2000:Q1.

N OTE: The figure illustrates the out-of-sample predictive scores obtained from the density constructed by fitting a
skewed-t distribution on the conditional quantiles from the quantile regression model (2) that starts in 2000:Q1. The
quantiles are constructed via an expanding rolling windows estimation starting in 2006:Q4. In particular, for the first
out-of-sample evaluation, we use data for average core CPI inflation over the next year until 2005:Q4 to estimate the
model. Then, we use data for the explanatory variables in 2006:Q4 to construct the out-of-sample predictive scores.

3 Evidence from Financial Markets and Regime-Switching Model

In this section we provide further evidence of our main finding that higher credit spreads are asso-
ciated with higher downside risks to the inflation outlook. First, we look at inflation probabilities
derived from inflation options and show that credit spreads are most negatively associated with
low inflation probabilities and that their relationship weakens as one considers higher inflation cut-
offs. We then consider predictive densities of various inflation measures derived from the Survey
of Professional Forecasters and find that they are remarkably similar to those obtained from our
quantile regression model. Finally, we estimate a monthly regime-switching version of our aug-
mented Phillips curve model over the last 20 years of data and show how the main findings of our
quantile regression model are directly comparable to and confirmed by this alternative approach.

3.1 Evidence from Financial Markets

Two defining features of our quantile Phillips curve model are that tight financial conditions carry
substantial downside risks to inflation and that these risks diminish as one moves to the left to
the upper tail of the inflation distribution. To test whether this relationship also holds in financial

15
markets, we run an OLS regression of options-implied inflation probabilities of one-year-ahead
CPI inflation derived from inflation caps and floors contracts (as described by Kitsul and Wright,
2013) on the credit spread. In the left panel of Figure 3 we present the estimated coefficients of
that regression. The slopes are rescaled so as to facilitate the comparison with those coming from
our quantile Phillips curve model (the right panel reproduces the bottom-right box of Figure 8). In
particular, the coefficient for the probability of one-year-ahead inflation being below 1% is rescaled
to match the slope estimated on the lowest inflation quantile which arises from the quantile Phillips
curve model. Further, the coefficient for the probability of inflation being below 1% is transformed
from positive to negative – as a positive correlation between the credit spread and this probability
is equivalent to a negative relationship between the credit spread and the lowest inflation quantile.

Figure 3: Credit Spread Slopes, Quantile Regression vs. Financial Markets.

Financial Markets Quantile Regression

N OTE: The left panel reports the slopes of separate regressions of inflation probabilities on the credit spread (at
monthly frequency), along with their 95% confidence interval. The coefficient for the probability of future inflation
being below 1% is rescaled to match the slope estimated on the lowest inflation quantile which arises from the quantile
Phillips curve model (right panel, taken from Figure 8). The coefficients are transformed from positive to negative for
the probability of inflation being below 1% – as a positive correlation between the credit spread and this probability is
equivalent to a negative relationship between the credit spread and the lowest inflation quantile.

Despite the vast disparities in the construction of the tails of the inflation distribution, the
estimated slopes are very similar to each other. Most importantly, as the inflation probability
cutoffs increase, their relationship with credit spreads weakens. We thus confirm our findings from
the quantile regression that credit spreads exert their biggest downward pressure on the left tail of
the distribution and thus are an important factor behind its asymmetry.10
10
This result is robust to the inclusion of the regressors used to purify the inflation probabilities from the oil price
effects (see discussion on these effects in Appendix F).

16
 Finally, in Appendix F we plot the time series of these financial-market-based probabilities
together with the credit spread (see Figure F-3). The graphs confirm the progressive weakening
in the relationship between the credit spread and inflation probabilities as one moves further up
in the inflation distribution until it completely breaks down once the upper tail is reached. We
also show that inflation probabilities coming from financial markets and from our quantile Phillips
curve model point in the same direction.

3.2 Evidence from Surveys of Professional Forecasters

We now compare the predictive densities coming from our quantile Phillips curve model with those
coming from the Surveys of Professional Forecasters (SPF). In Figure 4 we do this for both core
CPI (left column) and core PCE inflation (right column), as both inflation measures are available
in the SPF and applicable to our statistical model. We display the predictive densities of average
one-year ahead inflation for several dates before and after the Great Recession; and always for the
last quarter in a year, so that the horizon considered by the SPF and by the quantile regression
model are aligned.11 The SPF forecast densities (red dashed) are computed by applying a kernel
smoothing function on the point estimates across survey respondents.12 The quantile regression
densities (blue solid) are computed by fitting a skewed-t distribution on the estimated quantiles, as
described in Section 1.2. Finally, we report the realized value with a green solid vertical line.
The two densities are similar over time and when they do not align, the quantile Phillips curve
model seems to be closer to the realized value. In 2007-Q4, for instance, the SPF densities are
more optimistic than our statistical model, which is informed by the increase in credit spreads.

11
Ganics, Rossi, and Sekhposyan (2020) discuss the importance of predictive densities based on fixed-horizon
density forecasts.
12
Del Negro, Bassetti, and Casarin (2020) construct subjective SPF forecast distributions using non-parametric
Bayesian methods.

17
 2 2.2

Core CPI
PDF
1 2
Figure 4: Predictive Densities from Quantile Regression vs. from SPF Forecasts.
1.8
Core CPI (Left) and Core PCE Inflation (Right).
0
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

2 1.52 2
2.2
1

Core PCE
Core CPI
Core CPI
1.8
PDF

PDF
11
PDF

1 2
1.6
0.5
1.8 0.5 1.4
0 0
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5 0.1-1.50.2-1 0.3
-0.5 0.4
0 1 1.5
0.5 0.6 0.7 20.8 2.50.9 3 3.5 0.1 0.2 0.3 0.4
0 0
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 2
1.5 0.8
1

Core PCE
0.6

Core CPI
0.8 1 2
PDF

Core CPIPDF
0.4 1
0.6 0.5
0.5
PDF

0.2
1.5
0.4 0
0 0 0
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5 0.1-1.50.2-1 0.3
-0.5 0.4
0 0.5 0.6
1 1.5
0.7 20.8 2.50.9 3 3.5 0.1 0.2 0.3 0.4
0.2
1
2
0 0.8 2 2
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Core PCE
Core CPI
0.6
PDF

PDF
1.5 1.5
0.4
1 21
0.2
1
Core CPI

0 0 1
PDF

-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5 0.1-1.50.2-1 0.3

1.5 -0.5 0.4
0 0.5 0.6
1 1.5
0.7 20.8 2.50.9 3 3.5 0.1 0.2 0.3 0.4
0.5
1 2 2
1

Core PCE
1
Core CPI

0
PDF

PDF

-1.5 0.5-1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5 1.5 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.5
0.5

2.2
0 1 0 1
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5 0.1-1.50.2-1 0.3
-0.5 0.4
0 0.5 0.6 0.7 20.8 2.50.9 3
1 1.5 3.5 0.1 0.2 0.3 0.4
PDF Core CPI

1 2
PDF

1.5
2.2 1.8
2
0.5

Core PCE
Core CPI

1 2
1.61 1.8
PDF

0 0.5 1.8
0.5 0.1 1.6
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
1.6
1.4
0 0
2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5 0.1-1.50.2-1 0.3
2.4 -0.5 0.4
0 0.5 0.6
1 1.5
0.7 20.8 2.50.9 3 3.5 0.1 0.2 0.3 0.4
Core CPI

2 2.4 2.2
2.22
PDF

Core PCE

1
Core CPI

2
PDF

PDF

2.2
1
21 1.8
2
0 1.6
-1.5 -1
0 -0.5 0 0.5 1 1.5 2 2.5 3 3.5 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5 0.1-1.50.2-1 0.3
-0.5 0.4
0 0.5 0.6
1 1.5
0.7 20.8 2.50.9 3 3.5 0.1 0.2 0.3 0.4

N OTE: The figure shows the estimated skewed t-Student densities from the quantile Phillips curve model (blue solid)
of average four-quarter-ahead core PCI inflation (left panels) and core PCE inflation (right panels). The figure also
reports the realized value (green vertical line) and the density of SPF forecasts (red dashed) of the respective year-
on-year inflation measures, computed by applying a kernel smoothing function on the point estimates across survey
respondents.
18
3.3 Evidence from Regime-Switching Regression

We analyze how a nonlinear model such as a regime-switching regression compares to quantile

regression estimates, as Caldara, Cascaldi-Garcia, Cuba Borda, and Loria (2020) in the growth-
at-risk context. We do so to add further evidence that the relationships we established in our
main analysis are not an artifact of the quantile regression but a genuine feature of the data that
alternative nonlinear models would identify as well.
Since in the regime-switching model we estimate more parameters than in the quantile regres-
sion (in particular, the volatility of the error term and the regimes’ transition probabilities), we are
going to move to monthly frequency to allow for more observations and better identification of the
regimes. To this end, we first introduce a monthly version of our quantile regression model and
show that the results mirror the ones from the quarterly specification. Next, we compare the find-
ings from the regime-switching model to the ones from quantile regression.13 As we will show,
the asymmetric effect of credit spreads on the inflation distribution and the symmetric effect of
inflation expectations is a robust finding across these approaches.

Monthly Quantile Regression We explore the monthly version of the quantile regression model
described in (2) by interpolating the inflation expectations data. The sample runs from from Jan-
uary 2000 to April 2019, the last available date for core CPI inflation over the next 12 months.
Formally, we estimate the following quantile Phillips curve

bτ (π̄t+1,t+12 |xt ) = µτ (1 − λ̂τ )π ∗ + λ̂τ π LT E + θ̂τ (ut − u∗ ) + γ̂τ (π O − πt ) + δ̂τ cst ,

Q (4)
t−1 t t t

where π̄t+1,t+12 is the (annualized) average core CPI inflation rate between month t + 1 and month
t + 12, and where as in the quarterly model several risk factors affect the quantiles. Absent the
import price index at monthly frequency, we use oil price inflation πtO in the relative price term.
As shown in Figure 5, the results resembles those found at quarterly frequency for the most recent
sample (see Figure 1).
13
In our baseline specification we stick to quarterly frequency as it allows to abstract from more volatile and
temporary movements in inflation and it does not require to interpolate data.

19
 Figure 5: Slopes of Inflation Determinants Across Quantiles, Monthly Model.

N OTE: Estimated 10th , 50th and 90th quantile regression slopes from monthly quantile regression model (4).

Monthly Regime-Switching Regression The Markov-switching regression model is given by

∗
π̄t+1,t+12 = µ(st ) + (1 − λ(st )) πt−1 + λ(st )πtLT E + θ(st )(ut − u∗t ) + γ(st )(πtO − πt ) + δ(st )cst + σ(st )εt , (5)

where both the coefficients Θ(st ) ≡ {µ(st ), λ(st ), θ(st ), γ(st ), δ(st )} and the standard deviation
σ(st ) vary depending on an unobserved regime variable st ∈ {1, 2, 3} which indicates the regime
prevailing at time t. The latent variable st is governed by a discrete time, discrete state Markov
stochastic process, defined by the transition probabilities:
3
X
pij = P r (st+1 = j|st = i) , pij = 1, ∀i, j ∈ {1, 2, 3} (6)
j=1

Estimation is done via Bayesian methods, details on priors and model fit are in Appendix G.

Inflation Regimes The estimated regimes broadly correspond to states where inflation is low,
moderate or high. This becomes clear when comparing realized average inflation over the next
year against the estimated regime probabilities, as we do in Figure 6.

20
 Figure 6: Regime Probabilities from Markov-Switching Regression.

N OTE: Estimated regime probabilities from regime-switching regression (5) featuring three regimes and one Markov
chain for both coefficients and volatility.

Relationship to Quantile Regression In Figure 7, we report the regime-specific fitted values

along with the estimated quantiles from the monthly quantile regression model (4), which are
remarkably similar. In particular, the low inflation regime corresponds to the 10th quantile, the
moderate inflation regime to the median and the high inflation regime to the 90th quantile.
Notice how over the last 20 years, periods of stable inflation have been accompanied by the
presence of sizeable downside risks arising from tight financial conditions. It is beyond the scope
of this paper to explore which structural factors influencing our inflation determinants kept risks in
check and thus stood behind the remarkable resistance of realized inflation from falling to negative
territory during the financial crisis; but one can reasonably speculate that swift monetary policy
communication and intervention into financial markets might have provided the needed cushion.

21
 Figure 7: Fitted Values: Regime-Switching vs. Quantile Regression.

N OTE: The figure reports regime-specific fitted values of average core CPI inflation over the next year along with the
conditional quantiles of that same variable estimated from the quantile regression. The model features one Markov
chain governing both coefficients and volatility.

What explains the similarities in the assessment of risks between the regime-switching and the
quantile regression models? Intuitively, this is due to the fact that the regime-switching model
identifies regimes of low, moderate and high inflation that roughly correspond to the subsample of
datapoints that most heavily inform the slopes on the 10th , 50th and 90th quantile in the quantile
regression. This becomes most evident by comparing the estimated coefficients from the regime-
switching regression (5) with the ones from the quantile regression (4), as we do in Table 1.
The regime-switching regression confirms our main findings from the quantile regression ap-
proach. Historically, credit spreads disproportionately held down inflation in regimes when infla-
tion was low, in line with Galvão and Owyang (2018) who also found that financial conditions have
stronger effect on inflation in periods of financial stress. On the contrary, inflation expectations
have a symmetric effect across regimes. Moreover, in regimes of high inflation, the unemployment
rate made less dent on inflation outcomes. Finally, we notice that the Markov-switching regression
identifies little or no changes in the volatility of inflation across regimes.
Thus, the inflation distribution arising from the quantile regression should be interpreted as
capturing the range of outcomes that inflation might take, on average, over the next year given the
historical relationships with its determinants during regimes of high, moderate and low inflation.

22
 Table 1: Estimated Coefficients Across Models

Regime-Switching Regression
Low Inflation Moderate Inflation High Inflation
µ(st ) 0.30 [ 0.25, 0.41] 0.37 [ 0.18, 0.33] 0.50 [ 0.39, 0.63]
λ(st ) 1.00 [ 0.99, 1.01] 1.00 [ 0.96, 1.02] 1.00 [ 0.85, 1.02]
θ(st ) -0.16 [-0.19,-0.13] -0.17 [-0.18,-0.15] -0.07 [-0.10,-0.05]
γ(st ) 0.00 [-0.01, 0.01] 0.00 [-0.02, 0.04] 0.00 [-0.02, 0.15]
δ(st ) -0.33 [-0.38,-0.28] -0.15 [-0.18,-0.13] -0.15 [-0.18,-0.12]
σ(st ) 0.18 [ 0.16, 0.21] 0.13 [ 0.11, 0.15] 0.14 [ 0.11, 0.19]
Quantile Regression
th
10 Quantile Median 90th Quantile
µτ 0.24 [-0.57, 1.05] 0.12 [-0.34, 0.58] 0.52 [-0.03, 1.08]
λτ 1.00 [ 0.63, 1.37] 1.00 [ 0.80, 1.20] 1.00 [ 0.77, 1.23]
θτ -0.16 [-0.20,-0.13] -0.16 [-0.18,-0.13] -0.05 [-0.08,-0.02]
γτ 0.00 [ 0.00, 0.00] 0.00 [ 0.00, 0.00] 0.00 [ 0.00, 0.00]
δτ -0.33 [-0.39,-0.27] -0.12 [-0.15,-0.08] -0.14 [-0.20,-0.09]

N OTE: Estimated coefficients from regime-switching regression (5) and from quantile regression (4). Both models
are estimated at monthly frequency from January 2000 to April 2019. Values in brackets for regime-switching model
indicate lower (5%) and upper (95%) bound of coefficients derived from their posterior distribution calculated via
MCMC using 20,000 replications, a burn-in of 1,000 replications and a thinning parameter of 10 (thus for a total
of 210,000 effective replications). Values in brackets for quantile regression model indicate lower (16%) and upper
(84%) bound of coefficients computed via “blocks-of-blocks” bootstrap (see Appendix H) using 10,000 replications.

4 The Great Recession: United States vs. Euro Area

We now analyze the effect of inflation drivers on the evolution of the inflation distribution during
the last 20 years of data, comparing the United States experience with that of the euro area. For
the eurozone, the analysis focuses on euro-area-wide core HICP inflation – measured by headline
HICP inflation excluding energy and unprocessed food.14 As for the U.S., the quantile regression
model (2) uses the sample period available for the euro area, that is, from 1999:Q1 to 2017:Q4.15
14
Busetti, Caivano, and Rodano (2015) use dynamic quantile regressions to forecast the conditional distribution
of euro-area inflation. Tagliabracci (2018) estimates this distribution conditioning on Eurocoin. Busetti, Caivano,
Monache, and Pacella (2019) investigate the role of domestic and global determinants of euro area core inflation by
estimating a Phillips curve via expectile regression.
15
The last date for which average inflation over the next four quarters is available is thus 2016:Q4. The data is
described in great detail in Appendix A.

23
Quantile Phillips Curve Figure 8 displays the estimated slopes of the quantile regression model
(2) for the euro area (left column) and the United States (right column). The information is orga-
nized as follows. Boxes in each row correspond to the covariates of the model. The black squares
report the point estimates of the 10th , 50th , and 90th quantile-specific slopes. The length of the verti-
cal lines around the point estimates corresponds to the 68 percent confidence intervals constructed
by “blocks-of-blocks” bootstrap (see Appendix H). The OLS point estimates and their 95 percent
confidence intervals are given by the horizontal red lines.
The unemployment gap generates fairly similar responses of median inflation in the euro area
and the U.S. but important differences emerge when looking at the tails of inflation. In the euro area
the upper tail is the most sensitive segment of the inflation distribution to unemployment, while the
lower tail responds little. Thus, the relative odds of high inflation risks arising from a substantially
tight labor market outweigh the downside risks of low inflation associated with substantial labor
market slack. This pattern is reversed in the U.S., though the degree of asymmetry and the role of
unemployment in general is much more muted.
During this period, changes in the relative price of imported goods played a minor role in the
U.S., and a slightly larger role in the euro area. Still, there are some interesting differences across
economies. In the eurozone, the median and the lower tail of the distribution seem more responsive
than the upper tail of the distribution of inflation. For the U.S., these results are consistent with the
previous section, in which we pointed to a greatly reduced relevance of relative prices as inflation
determinants starting with the Great Moderation.
Longer-term inflation expectations and inflation inertia influence differently the overall infla-
tion distribution in the two countries. While in the U.S. inflation expectations dominate all parts of
the inflation distribution, in the euro area they only play a major role in explaining the upper tail of
inflation as odds of low inflation are also driven by past inflation. In other words, in the eurozone,
unmoored reductions in inflation expectations result in more persistent increases in downside risks
as their negative effect is propagated over time through a higher inflation inertia.

24
 Figure 8: Quantile Regression Slopes and Confidence Intervals.

Euro Area Core HICP United States Core CPI

N OTE: The figure displays the slope coefficients of the quantile regression of average four-quarter-ahead euro area
core HICP (left) and United States Core CPI inflation (right) defined in (2). The black squares correspond to the
point estimates whereas the vertical lines to the 68% confidence intervals computed via “blocks-of-blocks” bootstrap
using 10,000 replications (see Appendix H) for the 10th quantile (blue), median (red) and 90th quantile (yellow). The
estimation period is 1999:Q1 to 2017:Q4. The OLS estimates and their 95% confidence intervals are respectively
represented by the solid and dashed red lines.
25
 The last row in Figure 8 presents the role of credit spreads across inflation quantiles. In the
euro area higher credit spreads (i.e., tighter credit conditions) shift the inflation distribution to the
left as they have a fairly symmetric negative effect across inflation quantiles. This contrasts with
the U.S. in which most of the reduction in inflation following high spreads is concentrated in the
lowest tail, while the effects on the upper tail are not very significant.16 Most importantly, in the
U.S. financial conditions are the only significant source of asymmetry in the inflation distribution.
We performed ANOVA tests on the equality of the slope coefficients across quantiles. The
test rejected the null hypothesis of equality between the slopes on the 10th and 50th quantile at 1%
significance level for average past inflation in the Euro Area and for credit spreads in the United
States, thus confirming the importance of inflation inertia for the euro area and of credit conditions
for the U.S. in characterizing downside risks to and left-skewness in the inflation distribution of
the respective countries.17

Inflation Quantiles Figure 9 highlights key aspects of the evolution of the inflation outlook by
displaying the time series of the median, the 10th and the 90th inflation quantiles. The top panel
shows the evolution for the euro area while the bottom panel focuses on the United States.
In the eurozone, looking at the lower tail, it appears that downside inflation risks have been
important since the inception of the euro. Strikingly, the inflation distribution tends to tilt to the
upside around the three recessionary episodes, while also widening up significantly. By the end
of the 2001-2003 recession, downside risks to inflation started to trend down (i.e., the lower tail
fell) and after a faint recovery subsequently failed to rebound to the pre-contraction level. This
phenomenon is observed during the global financial crisis of 2008-2009 and repeated around the
2011-2013 recession, when downside risks increased further without recovering since then.

16
This result is also robust to limiting the sample to 1995:Q1-2007:Q4 (i.e., discarding the Great Recession ) as
well as to using the short-term instead of the long-term CBO NAIRU measure for u∗t (the former is higher after the
Great Recession). Further, in Appendix I we extend these results to two alternative measures of inflation, core PCE
and the Stock and Watson (2019) Cyclically Sensitive Inflation index. The effects are more symmetric in the case of
core PCE, while the CSI measure exhibits a similar asymmetry as core CPI although of somewhat larger magnitude.
17
The slopes for the 90th and 50th quantile are statistically different only for the U.S. unemployment rate.

26
 Figure 9: Time Evolution of Selected Conditional Inflation Quantiles.
Euro Area Core HICP
3

2.5

Core HICP Inflation

2

1.5

0.5
= 0.1
0 = 0.5
= 0.9
-0.5
'99 '00 '01 '02 '03 '04 '05 '06 '07 '08 '09 '10 '11 '12 '13 '14 '15 '16

United States Core CPI

3

2.5
Core CPI Inflation

1.5

0.5
= 0.1
0 = 0.5
= 0.9
-0.5
'99 '00 '01 '02 '03 '04 '05 '06 '07 '08 '09 '10 '11 '12 '13 '14 '15 '16

N OTE: The figure displays the time evolution of the conditional quantiles of euro area core HICP inflation (top) and
of Unites States core CPI inflation (bottom) estimated from the quantile regression defined in (2). Shaded bars indicate
NBER-dated recessions for the United States and OECD-based recession indicators for the euro area.

The estimated quantiles for the U.S. in the bottom panel of Figure 9 show some salient dif-
ferences with the eurozone. First, the downward tilts to the distribution associated with the two
recessions were primarily a result of a drop in the left tail, unlike in the euro area where the down-
ward push was more pronounced for the median and the upper tail. This was particularly acute
during the global financial crisis. However, the substantial increase in the odds of low inflation
was followed by a sustained recovery until the distribution became again tightly centered slightly
above 2 percent with the 10th quantile moving back to close to 2 percent. This contrasts the euro-
zone experience, in which the left tail failed to recover after the global financial crisis. In Appendix
J we complement these results by exploring which factors contributed to the recovery of the left
tail in the United States and to the failed recovery of the left tail in the euro area.

The Role of Credit Spreads We now turn our attention to the increasing role played by changes
in credit conditions in influencing the downside risks of inflation throughout the recovery. In Figure

27
10, we illustrate the time evolution of the 10th conditional inflation quantile of euro area core HICP
inflation (left) and of Unites States core CPI inflation (right), both for the baseline version of the
model (blue solid) and for a version where the effect of credit spreads is set to zero (black dashed).
The gap between the two lines captures the partial effect of credit spreads on the 10th quantile.
It is evident how tighter financial conditions exert a persistent downward pressure on downside
inflation risks and more strongly so, when credit spreads are high. It is remarkable how the large
spike in credit spreads observed in 2008 in the U.S. (bottom right panel of Figure 10) pushed down
the lower inflation quantile, which slowly moved back to about 2 percent by the end of 2016.

Figure 10: Partial Effect of Credit Spread on 10th Inflation Quantile.

10th Quantile
2 3.5
3
2.5
1.5
2
1.5
1
1
0.5
0
0.5 -0.5
'99 '01 '03 '05 '07 '09 '11 '13 '15 '99 '01 '03 '05 '07 '09 '11 '13 '15

Credit Spreads
3.5 8

3 7

2.5 6

2 5

1.5 4

1 3

0.5 2

0 1
'99 '01 '03 '05 '07 '09 '11 '13 '15 '99 '01 '03 '05 '07 '09 '11 '13 '15

Euro Area United States

N OTE: The figure displays the evolution of the 10th conditional inflation quantile of average one-year-ahead euro area
core HICP inflation (left) and of U.S. core CPI inflation (right) estimated from the quantile regressions model (2), in
its baseline version (blue solid) and in its version in which the effect of credit spreads is set to zero (black dash-dotted).
Shaded bars indicate NBER-dated recessions for the U.S. and OECD-based recession indicators for the euro area.

The eurozone is a slightly different story. Financial conditions, which played a more limited
role in the lower tail inflation dynamics, became increasingly important after the global financial

28
crisis. To see this, let us focus on the bottom left panel of Figure 10. This figure clearly shows
that the tightening in credit conditions occurred in two consecutive waves. The initial tightening
in financial conditions happened around 2008 and 2009 and marked a sharp reduction in the lower
quantile of the distribution that, even after some recovery in credit conditions during 2009, would
never rebound. The second wave, linked to the European sovereign debt crisis, exacerbated this
change. As of 2012, the deterioration in credit conditions lifted up substantially the odds of low
inflation. It is remarkable how, early in 2013, economic conditions pointed to a recovery in the
lower quantile. According to our model, however, this would have portrayed a misleading picture
reflecting the lack of consideration of the substantial downward pressures in place originated by
the still very tight credit conditions at that time. To see this more clearly, we translate the variation
in these quantiles into changes of the entire distribution of inflation, to which we turn next.

The Distribution of Inflation At a speech in London in July 2012, Mario Draghi – President of
the European Central Bank from November 2011 to October 2019 – announced the ECB’s com-
mitment of doing “whatever it takes” to preserve the euro. The eurozone was in the throes of crisis,
bond yields and credit spreads of weak euro-member governments were soaring, and financial mar-
kets doubted that European institutions could avert disaster. This is part of the historical context
reflected in Figure 11, which plots the estimated euro area core HICP predictive inflation distribu-
tions (left column) and their associated quantile functions (right column) across four selected dates
(for details on the construction of these objects please refer back to Section 1.2). In Appendix K
we also report the inflation probabilities associated with these distributions. We start at the dawn
of the global financial crisis (2007:Q4), then explore those periods in which downside risks were
most acute (2008:Q4 and 2011:Q4) and finally zoom in the end of our sample (2016:Q4). The blue
solid lines correspond to the baseline quantile model whereas the black dash-dotted lines parse out
the effect of credit spreads.
The results reaffirm that tight financial concerns played a crucial role in shifting to the left the
inflation distribution during the end of 2008 and remarkably so in the last quarter of 2011 – a few

29
months before Draghi’s speech. It is noteworthy how by the end of 2007, inflation-at-risk above
3 percent was virtually non-existent, while the left-tail pointed to some minor downside risks of
inflation running below 1 percent. Overall, credit conditions barely affected these conclusions.
The two waves in which the financial crisis was reflected in tight credit conditions translated into
a remarkable change of the inflation outlook. The distribution shifted to the left and concentrated
around a median inflation little below 1 percent, with the odds of low inflation – or even deflation
– soaring. By the end of 2016 the distribution of inflation had fatter tails, with the odds of high
inflation above those observed at the peak of the crisis, but with substantial downside risks still
remaining. The effects of credit conditions on inflation are also illustrated in the right column
of Figure 11, which shows that the inflation quantiles which condition on credit spreads were
significantly below those that solely rely on economic factors.
In Figure 12, we compare these results with the experience of the United States, for which
we consider similar dates.18 We focus on 2008:Q4 as this is when downside risks were most
pronounced in the U.S. following the sharp rise in credit spreads and dire economic conditions.
As in the case of the eurozone, the effect of financial variables on the inflation distribution is
striking during those episodes in which financial distress was most acute. In 2008:Q4, for example,
tighter credit conditions contributed to pushing the entire inflation distribution to the left, while
making it more dispersed and poking down substantially its left tail to the point of placing non-
zero probability of deflation occurring on average within the next year (as we will show below).
Looking at the right columns of Figures 11 and 12, one important difference emerges between
the euro area and the U.S. experience – a difference we had already encountered when analyzing
the quantile-specific slopes of credit spreads on average future inflation in Figure 8. While in
the eurozone higher credit spreads pushed down the inflation distribution symmetrically across
quantiles, in the United States its effects were mostly reflected in the left tail. As we show below,
the translation of these effects into the probability of low inflation (i.e., downside inflation-at-risk)
is more pronounced, the more the inflation distribution is right-skewed (i.e., the fatter its left tail).
18
This chart is similar to Figure 4 in Section 3.2, the only difference being that the densities have been computed
using the same sample as for the Euro Area.

30
 Figure 11: Selected Time Episodes of Predictive Densities (Left) and Skewness (Right).
Euro Area Core HICP Inflation.

3
0.8

Core HICP
0.6 2.5
PDF

0.4
2
0.2

0 1.5
-0.5 0 0.5 1 1.5 2 2.5 3 3.5 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

3
1

0.8 2.5

Core HICP
0.6 2
PDF

0.4 1.5

0.2 1

0 0.5
-0.5 0 0.5 1 1.5 2 2.5 3 3.5 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

2.5
2

2
1.5
Core HICP
PDF

1.5
1

0.5 1

0 0.5
-0.5 0 0.5 1 1.5 2 2.5 3 3.5 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

2.5

0.8
2
Core HICP

0.6
PDF

1.5
0.4
1
0.2

0 0.5
-0.5 0 0.5 1 1.5 2 2.5 3 3.5 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

N OTE: The left panels show the estimated skewed t-Student densities of average four-quarter-ahead euro area core
HICP inflation, in its baseline version (blue solid) and in its version where the effect of credit spreads is set to zero
(black dash-dotted). The same panel also reports the realized value of average four-quarter-ahead United States core
PCE inflation (green vertical line) and the density of SPF forecasts of year-on-year core HICP inflation (red dotted)
computed using a kernel smoothing function. The SPF density is only available in 2016-Q4 as that is the first date
for which the SPF forecasts are available and the last data point in our sample. The right panels show the estimated
skewed t-inverse cumulative associated with the conditional densities in the left panels.
31
 Figure 12: Selected Time Episodes of Predictive Densities (Left) and Skewness (Right).
United States Core CPI Inflation.

2.6
3
2.4
2.5

Core CPI
2 2.2
PDF

1.5 2
1
1.8
0.5
0 1.6
-0.5 0 0.5 1 1.5 2 2.5 3 3.5 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

2.5

3 2

1.5

Core CPI
PDF

2 1

0.5
1
0

0 -0.5
-0.5 0 0.5 1 1.5 2 2.5 3 3.5 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

1.5 2.5

1 2
Core CPI
PDF

0.5 1.5

0 1
-0.5 0 0.5 1 1.5 2 2.5 3 3.5 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

2.6
3

2.5 2.4
Core CPI

2
PDF

1.5 2.2

1
2
0.5

0 1.8
-0.5 0 0.5 1 1.5 2 2.5 3 3.5 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

N OTE: The left panels show the estimated skewed t-Student densities of average four-quarter-ahead United States
core CPI inflation, in its baseline version (blue solid) and in its version where the effect of credit spreads is set
to zero (black dash-dotted). The same panel also reports the realized value of average four-quarter-ahead United
States core CPI inflation (green vertical line) and the density of SPF forecasts of year-on-year core CPI inflation (red
dotted) computed using a kernel smoothing function. The right panels show the estimated skewed t-inverse cumulative
associated with the conditional densities in the left panels.

32
5 Tracking Risks During the Covid-19 Crisis
Risks to the inflation outlook have been front and center at the peak and in the recovery from
the Covid-19 crisis. We show how our model, augmented by credit spreads, allows to identify
important changes in the inflation distribution during this historical episode.
In doing this, we consider a monthly version of our model in which we use for past inflation πt∗
the average inflation rate between t and t − 11 when forecasting one-year-ahead inflation π̄t+1,t+12
as of time t. This is to allow to a more “real-time” assessment of inflation risks that incorporates
information on the current period inflation rate.19 The model is estimated from January 2000 to
April 2020, using as a dependent variable one-year-ahead inflation up to April 2021.
Figure 13 shows results for average inflation over the next 12 months, with core PCE inflation
on the left and core CPI inflation on the right. Specifically, it shows the distributions since January
2020, with markers given at the median and at the 90th quantile. Both core PCE and core CPI
inflation distributions have shifted to the right since the beginning of this year. Indeed, the selected
months show the quick build-up of downside risks to inflation during the harsh initial months of
the global pandemic (January to May 2020, top panels) which were then followed by the recent
increase in upside risks to inflation of the recent months (bottom panels).

We now ask what would have been the predictive distributions over the next 12 months had
the financial conditions deterioration of March 2020 persisted in May 2021. Figure 14 presents
distributions of average inflation over the next 12 months in May 2021 for our baseline (blue solid
lines) and in a counterfactual scenario in which credit conditions are those prevailing in March
2020 (black dashed lines). The easing of financial conditions since the onset of the pandemic
has moved the distributions of both core PCE and core CPI inflation rates to the right reducing
downside risks of low inflation. Indeed, had the financial conditions not eased during the period
from their May 2020 state, the probability of inflation running at or above 2.5 percent would be
about 7 percentage points lower for both core PCE and core CPI inflation rates as of May 2021.

19
In our baseline model, we do not assume that the researcher knows inflation as of time t when forecasting future
inflation, due to lags in the releases of inflation data.

33
 Figure 13: Predictive Densities of One-Year-Ahead Inflation Measures During Covid-19.

Core PCE Core CPI

N OTE: Predictive densities of inflation over the next 12 months in selected episodes. Onset of the COVID-19 pan-
demic (top panel) and 2021 (bottom panel), for core PCE inflation (left) and core CPI inflation (right). Markers appear
at the median and at the 90th quantile.

Figure 14: Predictive Densities of One-Year-Ahead Inflation Measures in April 2021.

The Role of Credit Spreads.

Core PCE Core CPI

N OTE: Predictive densities of inflation over the next 12 months in May 2021. The baseline case is in solid blue lines
and the counterfactual case (in which credit spreads are at March 2020) levels in dotted black lines, for core PCE
inflation (left) and core CPI inflation (right). Markers are at the median and at the 90th quantile.

34
6 Conclusion

In this paper we show that the recent muted response of the conditional mean of inflation to eco-
nomic conditions does not necessarily convey an adequate representation of inflation dynamics.
Indeed, we find ample variability in the tail risks to inflation, even when focusing on the post-2000
period of stable and low mean inflation. In particular, we document that tight financial conditions
generated times of substantial and persistent downside risks to inflation. We also show that ev-
idence from financial market quotes, from SPF forecasts and from a regime-switching model of
inflation is consistent with the previous findings.
Our paper offers a new empirical perspective to macroeconomic modelers and to policymakers,
showing that changes in credit conditions are key to understand tail-risk dynamics of inflation. Our
results provide empirical guidance and suggest more efforts in modeling the linkages between the
entire inflation distribution and financial markets in the context of nonlinear models.
As a future research avenue, we encourage the exploration of whether credit spreads in specific
sectors make the inflation outlook particularly vulnerable and whether certain inflation components
are particularly sensitive to movements in financial markets.

35
References

Adams, P. A., T. Adrian, N. Boyarchenko, and D. Giannone (2021). Forecasting Macroeconomic

Risks. International Journal of Forecasting 37(3), 1173–1191.

Adrian, T., N. Boyarchenko, and D. Giannone (2019). Vulnerable Growth. American Economic
Review 109(4), 1263–89.

Andrade, P., E. Ghysels, and J. Idier (2012). Tails of Inflation Forecasts and Tales of Monetary
Policy. Working Papers, Banque de France.

Azzalini, A. and A. Capitanio (2003). Distributions Generated by Perturbation of Symmetry with

Emphasis on a Multivariate Skew t-distribution. Journal of the Royal Statistical Society: Series
B (Statistical Methodology) 65, 367–389.

Ball, L. and S. Mazumder (2019). A Phillips Curve with Anchored Expectations and Short-Term
Unemployment. Journal of Money, Credit and Banking 51(1), 111–137.

Banerjee, R. N., J. Contreras, A. Mehrotra, and F. Zampolli (2020). Inflation at Risk in Advanced
and Emerging Economies. BIS Working Papers 883, Bank for International Settlements.

Baumeister, C. and L. Kilian (2016). A General Approach to Recovering Market Expectations

from Futures Prices with an Application to Crude Oil.

Berkowitz, J., I. Biegean, and L. Kilian (1999). On the Finite-Sample Accuracy of Nonparametric
Resampling Algorithms for Economic Time Series.

Blanchard, O., E. Cerutti, and L. H. Summers (2015). Inflation and Activity: Two Explorations
and Their Monetary Policy Implications. Working Paper Series WP15-19, Peterson Institute for
International Economics.

Boons, M., F. Duarte, F. de Roon, and M. Szymanowska (2020). Time-Varying Inflation Risk and
Stock Returns. Journal of Financial Economics 136(2), 444–470.

36