Content uploaded by Jonathan Benchimol
Author content
All content in this area was uploaded by Jonathan Benchimol on Jul 30, 2023
Content may be subject to copyright.
Do Expert Experience and Characteristics
Affect Inflation Forecasts?
Jonathan Benchimol,†Makram El-Shagi‡and Yossi Saadon§
September 2022
Abstract
Each person’s characteristics may influence that person’s behaviors and
outcomes. This study builds and uses a new database to estimate experts’
performance and boldness based on their experience and characteristics. Our
study classifies experts providing inflation forecasts based on their education,
experience, gender, and environment. We provide alternative interpretations
of factors affecting experts’ inflation forecasting performance, boldness, and
pessimism by linking behavioral economics, the economics of education, and
forecasting literature. The study finds that an expert with previous experience
at a central bank appears to have a lower propensity for predicting deflation.
Keywords: Expert forecast, Behavioral economics, Survival analysis, Panel es-
timation, Global financial crisis.
JEL Codes: C53, E37, E70.
This paper does not necessarily reflect the views of the Bank of Israel. The authors
thank Philippe Andrade, Lahcen Bounader, Itamar Caspi, Paul Fontanier, Xavier Gabaix, Paul
Goldsmith-Pinkham, Ori Heffetz, Ben Schreiber, Michel Strawczynski, Matthias Weber, and Osnat
Zohar, participants at the 2021 Asian Meeting of the Econometric Society, the Bank of Israel and
Tsinghua University research seminars for their valuable comments.
†Research Department, Bank of Israel, Jerusalem, Israel. Email: jonathan.benchimol@boi.org.il
‡Center for Financial Development and Stability, Henan University, Kaifeng, China. Corre-
sponding author. Email: makram.el-shagi@cfds.henu.education
§Research Department, Bank of Israel, Jerusalem, Israel. Email: yosis@boi.org.il
1
1 Introduction
There has been increasing interest in professional economic forecasts in recent
decades, partly due to the increasing importance of expectations in modern work-
horse macroeconomic models. In a few of the most active fields, authors have com-
pared private-sector and central bank professional forecasters (Romer and Romer,
2000; El-Shagi et al., 2016), and professional forecasts to market-implied forecasts
(Adeney et al., 2017; Benchimol and El-Shagi, 2020). Others have assessed whether
forecasts (or forecast spreads) incorporate information regarding macroeconomic
uncertainty (Bachmann et al., 2013; Bloom, 2014; Rossi and Sekhposyan, 2015).
While there is ample evidence regarding the factors driving the overall (rel-
ative) performance of professional forecasts (Giacomini and Rossi, 2010; El-Shagi
et al., 2016), there is, to our knowledge, no assessment of the factors driving perfor-
mance at the individual level. Instead, studies looking at individual-level forecasts
have so far mostly looked at general time-series properties of individual forecasts
and their interaction with each other (Andrade and Le Bihan, 2013). The main
reason seems to be that there is no information regarding individual forecasters
linked to the available major forecast databases.
This paper overcomes this limitation. It combines a dataset of US CPI inflation
nowcasts and clearly identified forecasters with two new original databases with–
mostly web-sourced–detailed information about forecasters and institutions. Our
first database includes the key individual characteristics of experts, such as their
previous job experience (location, type, and duration), gender, educational attain-
ment (including the quality of their alma mater), and affiliation (type and place).
Our second database includes institutional characteristics, most importantly lo-
cation and type, which have been found to matter in the previous literature. We
merge these databases with CPI nowcast data extracted from the Bloomberg sur-
vey of professional forecasters, including the name and institution of each fore-
caster covered. To our knowledge, this is the first paper to exploit such intercon-
nected data.
The present study contributes to the literature in several ways. Most impor-
tantly, it can assess the role of the individual and institutional characteristics in
forecast performance to a new degree in the literature. At the same time, the novel
dataset allows unique insights into the behavioral aspects of forecasting. We look
at boldness in terms of a forecaster’s willingness to deviate from the consensus (or,
in other words, the herd), which can also be interpreted as overconfidence (Bor-
dalo et al., 2020). Similarly, we assess forecasters’ tendency to under or overreact,
i.e., a bias towards or away from the long-term mean. In an extension, we consider
pessimism, which we define as the tendency to predict deflation. While this list is
far from exhaustive regarding possible behavioral traits that might affect forecasts,
it provides insights into how factors beyond asymmetric information or disagree-
3
ment on the underlying data-generating process shape forecast dispersion. While
our study is only a first step in that direction, understanding how forecasts are
derived is essential to improve forecast pooling approaches and correctly weigh
different forecasts.
In addition to being the first to account in such detail for forecaster characteris-
tics, we move away from the previous literature by considering a range of nonlin-
earities and interactions between relevant indicators, particularly how forecaster
characteristics interact with growing experience.
Apart from its central interest, namely macroeconomic forecasts, our study
contributes to the exciting debate in cognitive sciences and behavioral economics
about the role of education—its level, field, and quality—in expert behaviors. We
examine experts’ underreactions and the influence of high inflation periods on ex-
perts’ forecasting performance and boldness. In addition, this paper contributes
to the debates in human resources about how experience within or between insti-
tution types matters in forecasting performance or boldness.
A deeper understanding of professional forecasts is crucial for three reasons.
First, economic agents, including policymakers such as central banks, strongly
rely on forecasts (Piotroski and Roulstone, 2004). Since expert forecasts are gen-
erally better than market-based forecasts (Adeney et al., 2017; Benchimol and El-
Shagi, 2020), the economic agents mostly rely on the average of expert forecasts1
(Genre et al., 2013; Budescu and Chen, 2015). Usually, the forecasts are aggregated
in very simplistic ways, such as the simple average, and policymakers rank them
without considering the experts’ characteristics (Alessi et al., 2014; Coibion et al.,
2020). Nevertheless, as pointed out by Giacomini and Rossi (2010), understanding
the (time-varying) conditional relative forecast performance of several forecasts
can help to select the appropriate one and generate better forecast combinations.
Forecast combinations are typically superior to individual forecasts because
they can overcome model uncertainty and pool private information from various
sources. Disentangling dispersion due to behavioral traits rather than actual dis-
agreement is thus essential when not pooling model-based forecasts but forecasts
provided by individual forecasters based on unknown models or experience. Sec-
1Although it is impossible to certify if the Fed decides according to expert forecasts, it is inter-
esting to know how often the Fed mentions expert forecasts. However, we can reasonably assume
the Fed has actively considered expert forecasts in their decision processes since the global finan-
cial crisis (GFC). The Federal Open Market Committee (FOMC) meeting minutes detail the record
of the committee’s policy-setting meetings and offer detailed insights regarding the FOMC’s stance
on monetary policy. They mention the word "forecaster" only six times over the pre-GFC decade
(88 meetings), with more than a fourfold increase over the post-GFC decade (25 times over 82 meet-
ings). The word "survey" does not appear in the interest rate announcements during the pre-GFC
decade, while it appears 29 times during the post-GFC decade. The difference in the Chairman’s
speeches is even more spectacular than the difference in interest rate announcements or monetary
policy committee minutes. The word "forecaster" appeared only eight times in 216 speeches by the
Chairman during the pre-GFC decade. However, it appeared 58 times in only 160 speeches over
the post-GFC decade–almost ten times more than during the pre-GFC decade.
4
ond, understanding professional forecasts help us to understand the behavioral
foundations of expectations. There is a small but growing literature exploring this
field. Contrary to our approach, the previous literature has focused on institu-
tional characteristics–particularly the location and type of institution–due to data
limitations. For example, regarding location, Bae et al. (2008) show that the earn-
ings forecasts of local financial experts are more precise–the local analyst advantage.
Berger et al. (2009) demonstrate that institutions based in Frankfurt (or with a sub-
sidiary in Frankfurt) are significantly better at predicting the ECB’s interest rate
decisions. With respect to the type of institution, Mitchell and Pearce (2007) find
evidence that predictions by some economists covered in the biannual Wall Street
Journal survey are consistently above the survey mean, while those of others are
consistently below, depending on the industry of the economists’ employers. In
this vein, economists with a public mission–e.g., academics, central bank, and gov-
ernment employees–demonstrate a tendency towards being pessimistic, whereas
bankers, in general, are overly optimistic about future stock market developments
(Veress and Kaiser, 2017).
Third, inflation expectations from learning-to-forecast experiments2are in line
with experts, households (Michigan), and industry (Livingston) survey forecasts
(Cornand and Hubert, 2020). Understanding the characteristics that influence
the outcomes of expert forecasts should contribute to identifying potential factors
driving the formation of expectations.
Our study produces a range of results that are novel to the literature. First, we
find that experts with central bank experience are less likely to predict deflation.
These experts are less pessimistic, but this is mitigated when pessimism turns out
to be justified. Second, we highlight the implications and nonlinearities of the
role of experience and traits in experts’ forecasting performance and boldness.
Third, we confirm that the influence of experts’ traits on forecasting performance
and boldness changed following the GFC. Fourth, we show that underperforming
experts are less likely to survive in our expert database, while boldness does not
significantly influence this survival rate.
The remainder of the paper is organized as follows. Section 2 outlines litera-
ture relevant to our theoretical background. Section 3 discusses our data. Section
4 outlines our empirical methodology and results, including characteristics-based
forecasting ability tests and expert characteristics related to pessimism, with their
interpretation. Section 5 presents the policy implications of our results and con-
cluding remarks. Appendix A describes a simple theoretical model of expectation
formation that demonstrates the importance of accounting for nonlinearities, and
Appendix B presents a proportional hazards model to examine the expert’s sur-
2Experiments are well-incentivized by remuneration. Subjects are asked to submit an inflation
forecast and are rewarded solely based on their forecasts ex-post accuracy (Marimon and Sunder,
1993).
5
vival.
2 Literature Review and Theoretical Background
With forecasts being both the result of forecasters’ labor and, to some degree, a
measure of their expectations, our research question is at the intersection of labor
economics and behavioral economics.
In labor economics and adjacent fields, especially economics of education (Min-
cer, 1974), there has been extensive discussion regarding what employee traits
improve (or limit) his or her labor productivity. This includes innate traits such
as gender and origin, as well as acquired permanent traits such as the field of
education and experience. It seems plausible that those factors that drive labor
productivity also drive forecasters’ performance and, thus, the quality of their
forecasts. However, as mentioned above, forecasts are not merely a product an-
alytically derived by the forecaster using his human capital, but also a reflection
of his expectations. As such, they are subject to a plethora of factors that psy-
chologically and rationally affect the forecaster, including attention to economic
variables (Gabaix, 2019, 2020), expert behavioral biases (Thomas, 1999; Davis and
Lleo, 2020), and asymmetric information (Keane and Runkle, 1990), among oth-
ers (Lim, 2001; Coibion and Gorodnichenko, 2015). For example, Malmendier and
Nagel (2011) find that people who experienced low stock market returns invest
less in the stock market, indicating how personal experience shapes optimism and
pessimism. Education not only equips a forecaster with the necessary knowledge
to perform his job but also imbues him with a specific world view. Specific life
experiences are much more (or less) likely depending on someone’s gender, ori-
gin, etc., making the role of those factors play in forecasts far less evident than the
traditionally estimated effects on productivity.
What complicates matters is that the reported forecasts are not necessarily the
experts’ true expectations. The objective of the forecasters’ employer is typically
to obtain the best possible forecasts.3However, due to agency problems, the fore-
caster’s objective function does not necessarily mirror the objective function of the
employer.4While there might be intrinsic motivation to provide good forecasts,
3Their specific loss function can differ to some degree, depending on their use of the forecast.
For instance, risk-neutral investors might aim to minimize absolute forecast errors rather than
squared forecast errors that are more common in the literature. However, all those loss functions
that aim for unbiased forecasts with minimal errors are highly correlated, making the differences
mostly inconsequential.
4The outcomes achieved by any institution depend on its ability to take action today to
achieve its objectives tomorrow. Institutions use expert forecasts to shape their economic decisions
(Bernanke, 2007; Campbell and Sharpe, 2009). Consequently, expert forecasting accuracy generates
opportunity costs (Laster et al., 1999). Fulfilling these objectives depends on the performance of
the forecasts they use, including expert forecasts (See Footnote 1).
6
the main incentive for the forecaster to provide good forecasts is to maximize ex-
pected lifetime income.5
There has been discussion in the herding literature that failing "alone" is much
worse than failing as part of a group. Indeed, it seems likely that an employer
reads the failure of its experts as the experts’ fault when they alone fail, while
he might interpret them as inevitable if everybody fails. Consequently, forecasters
who deviate from the herd (following their own beliefs) have a lot to lose and little
to gain and might indeed be fired for their supposed incompetence. In this case,
their loss function–based on the underlying objective to maximize–would include
both the forecast loss and deviation from the herd of other forecasters.6
This is why we go beyond merely looking at forecast performance in this study
and consider boldness, which we define as the willingness to deviate from the
herd. Additionally, further highlighting the aforementioned psychological aspects
of forecasts, we consider pessimism.7
We extend the literature considerably by adding a battery of traits that might
affect either of those aspects of forecasts, including educational attainment (Bach-
elor, Master, and PhD), field (Economics, Finance, both, or others) and quality
(based on the Academic Ranking of World Universities, also known as Shanghai
Ranking). As the geography (i.e., location) and the institution type of both the ex-
pert and the institution (i.e., its affiliation) matter for forecasting (Batchelor, 2007;
Hong and Kacperczyk, 2010), we also contribute to the literature by extending the
expert’s experience characteristics with institution type (central bank, academic,
financial, or other) and location (of both the expert and the institutional affilia-
tion).
In accordance with Malmendier and Nagel (2016) and Malmendier et al. (2021),
which show that policymakers’ inflation experience affects their inflation fore-
casts, we test the influence of high inflation periods experienced by experts in
their forecasting performance and boldness. We also analyze the determinants of
experts’ over- and underreaction in predicting inflation along the lines of Barberis
et al. (1998) and Daniel et al. (1998).
3 Data
In this study, we combine three unique datasets: individual forecasts for the US
CPI (Section 3.1), web-sourced information, and characteristics of the forecaster’s
CV (Section 3.2), and web-sourced information about institutions where a fore-
caster worked (Section 3.3). These datasets cover the period from 1997:Q1 to
5More precisely, the present value of expected lifetime income.
6See Section 3.1, Eq. 1 and 2, for more details.
7For details on our definition, refer to Sections A and 4.3.
7
2017:Q4. We focus on inflation forecasts for both data availability in terms of quan-
tity of forecasters and forecasts and the sufficient variability of inflation compared
to interest rate forecasts. CPI inflation is also less prone to data revision than GDP
forecasts, and forecaster names are unavailable for most GDP forecasts.
3.1 Forecaster Behaviors
Most of the data, including individual point forecasts and the name and affilia-
tions of the experts, come from Bloomberg. Each expert can submit and update
their US inflation forecasts until the last day of the month. The publication of the
effective US CPI inflation during the corresponding month occurs about fifteen
days after this day. Since forgoing the chance to update despite new information
being available is irrational, we assume that the final forecasts are considered the
best possible forecasts by the submitting experts.
The expert forecast updates are accessible to other experts at any time, allowing
forecasters to react to each other. Therefore, any deviation from the herd can be
considered deliberate. This allows us to capture both the quality (Fig. 1) and
boldness (deviating from the herd) of forecasts.
Figure 1: Expert Forecast Performance
-1,20
-0,90
-0,60
-0,30
0,00
0,30
0,60
0,90
%
Notes: The blue line is the median of expert forecast errors, and the gray area represents the differ-
ence between the maximum and minimum forecast errors for each period.
8
First, we look at performance, measured through squared forecast errors as the
most commonly used loss function, i.e.,
Li,t=πtEi
tτ[πt]2, (1)
where iis a specific expert. We use tτto denote a point in time before tbut
clearly after t1, since forecasters can update until the last moment.
Second, we assess boldness, which we define as deviating from the “herd” of
other forecasters. In other words,
Bi,t=Ei
tτ[πt]Mit hEi
tτ[πt]i2, (2)
where Mit is the median operator over all expert forecasts iat time t.
It is worth noting that expert forecasts are published after they have been sub-
mitted, and that experts can still update until the very last minute until the dead-
line. Since forecasts are typically not submitted at the last second, we took the last
available forecast. This would allow forecasters to adjust to the perceived con-
sensus (median), causing perfect coincidence of forecasts by the time the data is
published. Therefore, we can assume that any deviation from this median is an in-
tentional deviation from the group, which corresponds to our model’s assumption
that each expert is isolated from others (and thus cannot observe other forecasts)
and is overconfident in their forecasts since they disregard information from other
forecasts.
There is, however, one caveat. Since we use nowcasts, where a detailed infor-
mation set is already available for all forecasters, the median forecast is typically
very close to the actual outcome. The correlation between Band Lis 0.573, making
it problematic to fully attribute variation in Las pure boldness.
These two measures will mainly be used as the dependent variable in our em-
pirical exercise. However, they will also serve as an explanatory variable in a sur-
vival analysis, where we assess the influence of those measures on the probability
of submitting another forecast.
3.2 Forecaster Characteristics
Our databases on expert characteristics are based on an automated collection,
which is manually augmented where needed. The primary source is LinkedIn.8To
rule out technical problems, we manually checked LinkedIn itself, where the auto-
mated search did not yield results, before looking for alternative sources for a CV,
8LinkedIn is a US-based employment-oriented Internet service founded in 2002. It is mainly
used for professional networking, with employers posting jobs and job seekers posting their CVs,
as well as people seeking to build their professional networks.
9
such as private and institutional websites. Partial database entries from LinkedIn
were completed in the same fashion.
The data we collected about experts encompass 151 experts and 164 institu-
tions that published at least one US CPI inflation forecast, with an average of
about 31 forecasts per expert. Since our panel specification includes forecaster
fixed-effects, we drop all forecasters who provided less than three forecasts dur-
ing our sample (1997:Q1–2017:Q4), leaving us with 118 experts who published an
average of about 41 US CPI inflation forecasts per expert.
For each forecaster, we collect data about the type and duration of previous job
experience and education. The experience type is separated into (a) academic ex-
perience, (b) central bank experience, and (c) experience in financial institutions.
The expert’s experience duration corresponds to the time since the first experience
in forecasting according to the expert’s CV (resume-based). Our model incorpo-
rates forecasting experience (expert experience according to the CV) as a variable
expressed in months and experience type (a, b, and c) as dummy variables.
Each forecaster has about 23 years of experience in forecasting on average. We
also gather data on crucial expert characteristics such as their location (local or for-
eign), education (level, type, and quality), gender (male or female), and affiliation
(type and location). About 41% of the experts are Americans, 12% are females, and
77% have at least one degree (BA, MA, or PhD) in Economics. Experts have an av-
erage of 218 months of experience in financial institutions and only ten months
of experience in academia as well as central banks. Their highest degree (Bache-
lor’s, Master’s, or PhD) and the corresponding field, where we only distinguish
between Economics, Finance, and other fields, are also identified. For about 8%,
their highest degree is a BA, 59% have an MA, and 28% hold a Ph.D. The rest
have no identified highest degree. Our database also includes the Shanghai Rank-
ing classification for the respective university, following both the Economics and
Finance rankings for the highest degree.9About 39% of the US CPI inflation ex-
perts graduated (BA, MA, or PhD) from a top-ranked university according to the
general ranking.
Lastly, we collect additional information, such as citizenship and age. The date
of birth is generally not reported on LinkedIn. As such, we scrapped the CVs
available on the Internet and Bloomberg to gather the date of birth, which by con-
struction led to less available but more precise data. The measure of age used in
the paper is based on the available date of birth and allows us to assume that ex-
perts, who faced high inflation periods during their youth or adolescence, were
born before 1973.
Table 1 presents the summary statistics of the dataset used in the estimations
9We use the Shanghai Ranking’s Global Ranking, Academic Subjects, 2017. In addition, we
decompose each ranking (Economics and Finance) into four levels: first tier, second tier, third tier,
and not ranked.
10
presented below.
Table 1: Summary Statistics
Average Std. Dev. Average
% of the sample
Forecasts per forecaster 43.4 36.6 Local forecaster 44.7
Forecast error 0.14 0.12 Local institution 30.6
Forecast boldness 0.08 0.09 Financial institution 92.9
Experience: Central bank (m) 10 31 Experience: Central bank 16.5
Experience: Academia (m) 10 26 Experience: Academia 20.0
Experience: Total (m) 185 143 Experience: High Inflation 54.1
Gender 12.9
Education: MA 58.8
Education: PhD 28.2
Education: Economics 71.8
Education: Finance 14.1
Ranking: Economics 28.2
Ranking: Finance 29.4
Notes: (m) indicates the number of months. This table presents a summary of the data at our
disposal.
3.3 Institutional Information
The data about experts’ affiliations describe the institution type and its primary
location (local or foreign headquarters). We classified all forecast providers into
several types: retail bank, investment bank, private bank, insurance company,
economic and financial analysis firm, fund, investment management firm, bro-
kerage, credit union, savings and loan firm, academia, central bank, and others.
Although we built an in-depth database separating several relevant hosting insti-
tution types, we rely only upon the simple (and more relevant) difference between
private financial, academic, and monetary institutions in our analysis.
4 Results and Interpretation
This section uses the data described in Section 3 to detect characteristics leading to
forecasting performance, boldness, and experts’ sentiments. We present a random
effects analysis to identify the role of inherent traits (Section 4.1), a panel estima-
tion to explore the role of experience (Section 4.2) and over- and under-reaction
(Section 4.4), a probit model to assess the expert’s pessimistic and optimistic be-
haviors (Section 4.3), and forecasting ability tests (Section 4.5) to identify the fore-
casting performance of characteristics-based groups of experts.
We report two sets of results in each of the following two subsections. The first
set is based on demeaned variables. Thus, we can interpret the effect of experi-
11
ence as effects at the mean, although it should be noted that the mean forecaster
cannot exist (as the dummies take values between 0 and 1 for this hypothetical
forecaster). This is particularly helpful in immediately gauging the average ef-
fect of experience in a typical case. Essentially, this is a more straightforward way
to interpret the significance of individual interaction terms, whereas one would
need to utilize joint significance tests to overcome the multicollinearity issues that
otherwise arise with interactions.
However, interpreting the results of the interactions is less straightforward
since a dummy value of 0 no longer reflects that the dummy variable is not true.
Therefore, we reestimate the model, but while we still demean experience (our
only continuous variables), we do not demean the dummies for location, educa-
tion, gender, etc. Thus, the baseline for this second estimation is a young, male,
foreign forecaster, not from a top university, who holds neither a Master’s nor a
PhD, and has a background that is neither finance nor economics, has no experi-
ence in academia or at a central bank, did not go through high inflation periods,
and works at a non-US nonfinancial institution. Consequently, this makes the in-
terpretation of the dummy variable interactions much more straightforward.
Note that the results are mathematically equivalent, and this provides different
angles to look at our results and to ease interpretation.
4.1 Random Effects
Many of the forecaster traits we collect in our dataset do not vary over their fore-
cast history and, therefore, cannot be included in the fixed effects panel benchmark
models (see Section 4.2). In this preliminary analysis, we use a random-effects
estimator–where non-time-varying variables can be used–to have a look at those
variables separately. We keep the model as parsimonious as possible and only in-
clude time-varying data in the fixed effects model, where we can fully account for
unobserved heterogeneity.
To account for the fact that a large part of forecast errors comes from unpre-
dictable shocks to inflation, affecting all forecasts simultaneously, we include time
fixed effects.
In addition, we account for both forecaster-specific and institution-specific char-
acteristics, thereby guaranteeing that some institutions with easier access to fore-
casters with specific traits do not merely drive the effect of forecaster characteris-
tics (e.g., forecasters with a PhD or from the best institutions).
This yields the estimation equation
yit =β0+β1ln xit +ΓXit +ΛZit +ui+vt+εit, (3)
where Xand Zdenote forecaster- and institution-specific characteristics, respec-
12
tively, and Γand Λthe corresponding vectors of coefficients. ui,vt, and εit repre-
sent the forecaster random-effect, time fixed-effect, and idiosyncratic component
of the error term, respectively. β0and β1are the constant and experience regres-
sion coefficients, respectively.
We split our sample (1997:Q1-2017:Q4) into two subsamples: the pre-
GFC(1997:Q1-2008:Q1) and post-GFC (2008:Q1-2017:Q4). Table 2 presents the ran-
dom effects estimations for the US CPI inflation expert forecasts over the full sam-
ple and the pre- and post-GFC samples, with the correction for time-effects out-
lined in Eq. 3, for the variables presented in Section 3.
Table 2, given in this subsection, is based on demeaned variables. Results based
on the alternative estimation where all dummy variables are set to 0 and only
experience is demeaned are reported in Table 9 in the appendix.
Table 2: Panel Estimates with Individual Random Effects - Nonlinear Model
Performance Boldness
Full sample Pre-GFC Post-GFC Full sample Pre-GFC Post-GFC
Experience -0.0003 -0.0156** 0.0017 -0.0018** -0.0182*** -0.0004
Age >60 years old 0.0016 0 -0.0015 0.0042 0 0.0022
Local forecaster (LF) 0.0037 -0.0145 0.0137 0.0068 -0.0097 0.007
Local institution (LI) -0.0114 -0.0014 -0.0241** -0.008* -0.0113 -0.01**
Financial institution (FI) -0.0067* -0.0662*** -0.004 -0.0016 -0.0479*** -0.001
Experience: Central bank (CB) -0.0044 -0.0167 -0.0036 -0.0016 -0.0203 -0.0001
Experience: Academia (AC) -0.0117*** -0.057*** -0.0082** -0.0069*** -0.0512*** -0.0048**
Gender (G) 0.0014 0.0686*** 0.0039 0.0023 0.0733*** 0.002
Education: MA (MA) -0.0047 0 -0.0128** -0.0032 0 -0.0069**
Education: PhD (PhD) 0.0037 0.0267 0.0031 -0.0023 0.0332 -0.0025
Education: Economics (EC) 0.0147 0.0172 0.0139* 0.0062 0.013 0.0009
Education: Finance (EF) 0.0092 0.0355 0.0156* -0.0068 0.0479 -0.0049
Ranking: Economics (RE) 0.0024 -0.0062 0.0018 0.0020** -0.0077 0.0018**
Ranking: Finance (RF) -0.0007 0.0097 -0.0003 -0.0003 0.0101 -0.0001
Constant 0.0372 0.0329 0.0781*** 0.0028 -0.0316 0.0242**
Two-way interactions
MA RE -0.0174*** 0 -0.0152** -0.0093*** 0 -0.0079***
MA RF 0.0139** 0 0.0104** 0.0081** 0 0.0058**
PhD RE -0.0119* 0.0196 -0.0109* -0.009** 0.0214 -0.0089***
PhD RF 0.0075 -0.0298* 0.0077 0.0038 -0.0271* 0.0039
MA EC -0.0098 -0.0221 0 -0.015 0.0119 0
FI LI 0.0276*** 0.131*** 0.0224*** 0.0094** 0.0955*** 0.0061*
GEF -0.0434*** -0.184*** 0 -0.0218** -0.173*** 0
MA LI 0.0497*** 0.138*** 0.0677*** 0.0337*** 0.103** 0.0306***
MA LF -0.0508*** -0.169*** -0.0769*** -0.036*** -0.165*** -0.0372***
Observations 2829 324 2505 2832 324 2508
R20.422 0.402 0.418 0.167 0.192 0.152
Notes: ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. Coefficients
lower than 0.00005 are reported as 0. Given the scaling of our variables coefficients in this order
of magnitude do not have an economically meaningful effect.
Random effects estimates presented in Table 2 show that performance and
boldness are influenced differently. We find coefficients of determination around
13
0.42 for performance and around 0.16 for boldness. This difference can be ex-
plained by performance variance, which is greater than boldness variance, and
our focus on performance for the modeling strategy benchmark since boldness
and performance results are aligned in the literature. Given the generally high
precision of nowcasts and the correspondingly relatively low variance of our mea-
sures, this level of explanatory power is fairly high, particularly for performance.
As far as forecasting performance is concerned, working at a financial institu-
tion (FI) improves the expert’s forecasting ability more before the GFC than after.
However, experts at local financial institutions (FI LI) slightly mitigate this re-
sult as this interaction is more significant than FI or LI effects alone. Interestingly,
graduating from a top university (Finance ranking) appears to improve expert
forecasts before the GFC. However, this is not the case after the GFC or over the
full sample. This is mitigated by experts who graduated from a top university
(Finance ranking), who achieve lower performance with their forecasts.
Before the GFC, male experts appeared to provide better forecasts than fe-
males, but this is not true after the GFC or over the full sample. The fact that male
experts were more numerically dominant prior to the GFC, while the percentage
of female experts has increased since the GFC, mitigates this result. Our results
may also reflect that women benefit less than men from connections in job perfor-
mance, herding behavior, and subjective evaluation by others (Fang and Huang,
2017).
The situation is different regarding the expert’s boldness (herding behavior).
First, working at a financial institution does not influence the expert’s boldness.
Experience in a central bank increases the expert’s boldness, but it does not influ-
ence the expert’s forecasting performance. In addition, graduating in Economics
or Finance decreases boldness and increases the performance of expert forecasts
after the GFC, while having a PhD from a top-ranked institution in Economics
increases the expert’s boldness before the GFC. Interestingly, the increase in the
boldness of female experts with a degree in Finance was greater than the general
decrease in the boldness of females before the GFC. At the same time, the robust-
ness of this result is still questioned due to the paucity of female experts before the
GFC. After the GFC, graduating with a Master’s degree from a top university in
Finance significantly reduces the expert’s performance and boldness compared to
experts who graduated with a Master’s degree in Economics, which contributes
significantly to increased forecasting performances and boldness.
In line with Clarke and Subramanian (2006), performance and boldness results
in Table 2 are often similar in terms of significance and sign,10 confirming that sig-
nificant underperformers are more likely to issue bolder forecasts and vice-versa.
Like financial analysts who also tend to exhibit herding behavior, which some-
10This is also the case with our next results (Table 3).
14
times compromises accuracy; our results suggest that social forces (ranking, in-
stitution type, and location), education (type and level), and experience (type and
duration) influence an expert’s rational economic logic and cognitive biases–an in-
terpretation close to Christoffersen and Stæhr (2019) that is presented in our next
results (Section 4.2). It should be noted that the highly precise nature of nowcasts
partly drives the correlation between accuracy and boldness.
Interestingly, the difference between pre- and post-GFC may reflect the change
in the expert’s attention or biases induced by the crisis shock on characteristics’ ef-
fects (Andrade and Le Bihan, 2013; Christoffersen and Stæhr, 2019). Although
these results consider a time-fixed-effect, results considering individual fixed-
effects like experience are presented in Section 4.2.
A Master’s degree in Finance led to greater boldness but more herding behav-
iors after the GFC and over the full sample than in the pre-GFC period. As a result
of cognitive biases and an intuitive reaction to uncertainty and financial instabil-
ity, experts with lower risk tolerance may herd more (Christoffersen and Stæhr,
2019). Also, being a local forecaster with a Master’s degree seems to improve
forecasting performances and decrease boldness, which complements Clarke and
Subramanian (2006).
Better education quality in Economics11 improves experts’ forecasting perfor-
mance among those with a Master’s or a PhD degree (MA RE and PhD RE).
This was not the case for experts who graduated with a Master’s degree from a
non-top-ranked university. However, better education quality in Economics (RE)
may accentuate experts’ herding behaviors, which is not necessarily the case for
experts who graduated from a top-ranked university in Finance (RF).
Working at a local institution and having a Master ’s degree (MA LI) in-
creased forecasting performance after the GFC. This result is confirmed for local
experts having a Master’s degree over the full sample and pre-GFC period (MA
LF). Our findings that mix geography and education relate to several strands
of the literature. Our results demonstrate that the likelihood of herding increases
with the expert’s forecasting experience and is influenced by institution (Clement
and Tse, 2005). Our results also show that experience (Hong et al., 2000; Mikhail
et al., 2003) and education (De Franco and Zhou, 2009) influence social interac-
tions, cognitive biases, and intuitive reaction to uncertainty, an interpretation par-
tially shared with Christoffersen and Stæhr (2019).
Female experts underperform men before the GFC, but this result is not signif-
icant after the GFC, which may confirm a labor market entry selection bias. How-
ever, female experts with education in Finance outperform men (G EF) even
after the GFC. Female experts in a market segment in which their concentration is
11Measured with the ShanghaiRanking’s Global Ranking of Academic Subjects in Economics,
see Section 3.
15
lower than in others appear to have better-than-average skills due to self-selection
(Kumar, 2010).
While differences in views may persist through time, differences in informa-
tion sets only cannot explain such differences in opinion. Patton and Timmermann
(2010) show they stem from heterogeneity in priors or models and that differences
in opinion move countercyclically. Although this heterogeneity is most robust
during recessions, our results bring another layer to this conclusion. The GFC not
only changed differences in opinion but also modified the influence of experts’
characteristics on their forecasting performance and boldness.
4.2 Forecaster Fixed Effects
In our third set of analyses, we assess the effect of experience on the different
output measures. Since experience is the only time-varying trait we consider, this
essentially boils down to a simple univariate panel regression with time (t) and
forecaster (i) specific effects, such as
yit =β0ln xit +Γln xitXit +Λln xit Zit +ui+vt+εit, (4)
where yit is one of our two loss functions discussed in Section 3.1 (performance
and boldness) and xit is our experience duration measure presented in Section
3.2. β1is the regression coefficient on experience. ui,vt, and εit represent the
forecaster fixed-effect, time fixed-effect, and idiosyncratic component of the error
term, respectively.
There are, however, potential endogeneity issues with this specification since
the ability to gain experience (or in other words “to keep your job”) might depend
on forecast performance. We assess this possibility in detail as outlined in Appen-
dix B. The evidence for the existence of such an effect is mixed, and if it does exist,
it seems to be only moderately sized. We correct for it in a robustness test by in-
cluding a dummy for flagging the last five forecasts submitted by any forecaster.
While much simpler, this follows the spirit of selection estimators12 (Heckman,
1979). The resulting equation is given as
yit =β0ln xit +β11lt2last5(i)+Γln xit Xit +Λln xitZit +ui+vt+εit , (5)
where last5(i)is the set of the last five periods in which forecaster isubmits a
forecast to our dataset, Xand Zdenote forecaster- and institution-specific charac-
teristics, respectively, and Γand Λare the corresponding vectors of coefficients.
β0and β1are regression coefficients.
12In a full-fledged selection model, one would include a transform of the “inclusion” probability
rather than a determinant of the latter. The key problem is that we cannot truly estimate inclusion
probabilities here, but merely relative risks of being removed from the sample.
16
Table 3 presents the panel estimations for the US CPI inflation expert forecasts
for the model that controls for individual specific and time fixed-effects. All rows
except the first two – reporting the effects of our only two time-varying indicators,
namely log experience and the age dummy – are interactions with (log) experi-
ence. As the indicators are constant, the coefficients should be interpreted as a
response to a change in experience conditional on the dummy (or combination of
dummies in the case of three-way interactions).
In both Tables 3 and 10 (Appendix C), some of the coefficients seem relatively
large given the low disagreement between nowcasts and the correspondingly low
variance of performance and boldness. However, a unit change in log experience
is substantial. For the mean forecaster, one additional year of experience corre-
sponds to an increase of 0.1 in log experience.
For both subsamples and the full sample, we find that, on average, experience
decreases the squared forecast error and deviation from the consensus forecast.
This is very much in line with the previous literature. Experts may underreact
less to prior CPI information as experience increases, suggesting one reason why
experts, like analysts, become more accurate with experience (Mikhail et al., 2003).
This is similar to analysts’ firm experience, which is strongly and positively associ-
ated with analysts’ forecast boldness (Clarke and Subramanian, 2006; Huang et al.,
2017). At the mean, a one-year change in experience would lead to a reduction of
approximately 0.02 in the squared forecast error (i.e., one-tenth of the coefficient
value).
In the full sample and for most of the post-GFC period, domestic forecasters
(LF) benefit less from experience than foreign forecasters at the mean, which com-
plements Clarke and Subramanian (2006). This is intuitive since domestic fore-
casters can be expected to start with a better understanding of the local economy.
At the same time, experience is, on average, more helpful at local institutions (LI),
which offer better exposure to local data, networking, and news. Generally, people
with higher levels of education benefit from experience, which might either reflect
that they start with less "practical knowledge" or (hopefully) that their training en-
ables them to utilize information better to improve their performance over time.
People with experience in academia benefit less from experience, while a degree
in Economics or Finance seems to help in that respect.
In Table 3, the results on expert performance (left panel) are close to the bold-
ness results (right panel) in terms of sign and significance, except for several
interesting instances. Experience seems to embolden local forecasters over the
full sample and after the GFC, which also complements Clarke and Subramanian
(2006). However, a notable difference between performance and boldness results
concerning experts who faced high inflation periods shows that their boldness
is discouraged by experience without significantly influencing their performance,
17
Table 3: Panel Estimates with Individual Fixed Effects - Nonlinear Model
Performance Boldness
Full sample Pre-GFC Post-GFC Full sample Pre-GFC Post-GFC
Experience -0.176*** -2.29 -0.271*** -0.137*** -1.027 -0.127***
Age >60 years old -0.03*** 0 -0.032*** -0.009 0 -0.005
last50.002 -0.042 -0.005 0.001 -0.011 -0.004
Two-way interactions with experience
Local forecaster (LF) 3.428*** -1.059 0.281* 1.512*** -0.27 0.188***
Local institution (LI) -4.409*** -1.634 0.105 -2.009*** -1.309 -0.101
Financial institution (FI) 1.409*** -0.279 0.085 0.662*** 0.381 0.09
Experience: Central bank (CB) 0.816*** 0.365 0.814*** 0.326*** 0.061 0.334***
Experience: Academia (AC) 2.351*** -0.006 -1.359*** 0.944*** -0.157 -0.6***
Experience: High Inflation (HI) -0.018 -1.949** 0.034 -0.046** -1.121 -0.006
Gender (G) -0.145 -0.625 -0.007 -0.067 -0.76 0.063
Education: MA (MA) -2.681*** -0.438 -3.147*** -1.327*** -0.195 -1.462***
Education: PhD (PhD) -3.3*** 0 -3.29*** -1.552*** 0 -1.443***
Education: Economics (EC) 0.738*** -1.473* -0.859*** 0.36*** -0.815 -0.331**
Education: Finance (EF) -3.467*** 4.805 0.462 -1.582*** 7.649*** 0.193
Ranking: Economics (RE) 1.098*** -0.203 0.467*** 0.504*** -0.131 0.221***
Ranking: Finance (RF) -0.041*** 0.366 -0.028 0.004 -0.144 0.008
Three-way interactions with experience
FI RE -0.016 1.19 -0.003 0.013 0.46 0.017
MA RE -7.103*** -0.33 -6.857*** -3.449*** -0.296 -3.151***
PhD RE -7.117*** 0 -6.927*** -3.462*** 0 -3.201***
EC RE -3.573*** -0.727 0.262 0.544*** -0.466 0.067
FI RF -0.062 -1.275 -0.013 -0.027 -0.593 -0.028
GRF -0.162*** 0.427 -0.041 -0.111*** 0.492 0.02
MA RF -0.043** 0.083 -0.093** -0.031*** 0.089 -0.054**
EF RF 0.007 -3.304 -0.098 0.154*** -6.123*** -0.034
AC LI -15.052*** 2.069 7.791*** -6.39*** 1.104 3.248***
MA LI 21.214*** -4.6 20.684*** 10.349*** -3.378 9.555***
PhD LI 21.361*** 0 20.852*** 10.459*** 0 9.598***
FI LF -6.89*** 0 -6.908*** -3.278*** 0 -3.156***
GLF 6.441*** 4.387 6.725*** 2.776*** 2.463 2.956***
AC LF 14.892*** 0 -7.965*** 6.347*** 0 -3.207***
MA LF 0.023 2.564 0 0.06** 2.541 0
EC LF 7.453*** 0 -0.48* 3.392*** 0 -0.014
GFI -15.893*** 0 -15.794*** -7.647*** 0 -7.248***
EC FI -0.279 0 7.323*** 0.167 0 3.21***
GCB 7.149*** 0 7.017*** 3.322*** 0 3.019***
EC AC 3.749*** 00 1.641*** 00
EC MA -3.367*** 00 -1.357*** 00
Observations 2829 324 2505 2832 324 2508
R20.052 0.231 0.041 0.057 0.146 0.035
Notes: ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. Coefficients
lower than 0.00005 are reported as 0. Given the scaling of our variables coefficients in this order
of magnitude do not have an economically meaningful effect.
18
which mitigates Clarke and Subramanian (2006), who linked performance to bold-
ness in terms of sign and significance. High inflation, when the expert was young,
may contribute to a specific fear of remaining far from the consensus. Experi-
ence, therefore, encourages experts to herd. As a result of cognitive biases and an
intuitive reaction to uncertainty and financial instability, experts with lower risk
tolerance may herd more (Christoffersen and Stæhr, 2019).
We look at the full set of interactions to fully understand the results, which
is more accessible when looking at Table 10 (Appendix C). While the coefficient
on experience is positive in the model, where dummies are not demeaned, the
vast majority of (single) interactions is negative. Experience has a detrimental but
insignificant effect on our baseline forecaster, which is not true for the typical fore-
caster (who has good education and typically a background in an Economics or
Finance-related field). In the full sample and post-GFC, ceteris paribus, most devi-
ations from the benchmark case (i.e., the single interactions) are negative, making
experience more helpful. Two notable exceptions are gender (where women seem
to benefit less from experience than men or indeed suffer from experience) and fi-
nancial institutions where forecasters also seem to be deteriorating with growing
experience.
The simple (two-way) interactions and the three-way interactions have to be
considered jointly. For example, we find that the FI LF interaction compensates
for the negative effect of financial institutions. That is, local forecasters at financial
institutions benefit no less from experience than their counterparts in other insti-
tutions that provide forecasts. Similarly, we find a highly negative coefficient for
the G FI interactions, i.e., while women at nonfinancial institutions benefit less
from experience, women at financial institutions benefit more. That said, our gen-
der results should be taken with the caveat that there are relatively few women.
In a highly nonlinear model such as ours, this might imply that the results for a fe-
male local forecaster in a financial institution (for example) are driven by just one
or two women. Nevertheless, the large coefficient values make the combination of
the three aforementioned dummies an excellent example for interpretation.
Table 4 shows the marginal effects for all combinations of these three dummies,
where all other indicators match the baseline case. It is evidently impossible to
look at more than 8000 combinations in the same detail.
However, this way of looking at things sheds more light on seemingly extreme
results. For example, a hugely negative coefficient of local institutions is offset
mainly by interactions with higher degrees, i.e., the typical highly educated fore-
caster at those institutions is not that good.
For the modal forecaster in terms of characteristics (except experience), i.e., a
male foreign forecaster working at a foreign financial institution, who holds an
MA in Economics from a non-ranked university and has some experience with
19
Table 4: Financial Institutions, Local Forecasters, and Gender
FI LF G Marginal effect of experience
0 0 0 5.551***
0 0 1 9.034***
0 1 0 -0.521
1 0 0 7.298***
0 1 1 6.441***
1 0 1 -15.893***
1 1 0 -6.89***
1 1 1 5.02***
Notes: All other variables except inflation are set to 0, i.e., we are looking at a forecaster from a
foreign institution, without a higher degree, who studied neither Economics nor Finance and did
not go to an institution that excels at those issues. The forecaster has no experience with high
inflation regimes, academia, or central banking.
high inflation (but none in academia or central banking), the total benefit of expe-
rience is indistinguishable from zero. For this forecaster, the benefit of experience
would be much higher when having a degree in Finance instead of Economics, es-
pecially if this degree is from a top institution in Finance, or if he was female. He
would benefit much less when having a background in academia. Again, keep in
mind that a lower or even detrimental impact of experience does not necessarily
imply bad forecasting, but could also reflect the depreciation of relevant knowl-
edge that makes the initial forecasts particularly good.
Three-way interactions show that experts having a Master’s degree or a PhD
from a top university according to the university rankings in Economics and Fi-
nance (MA RE, PhD RE, and MA RF), working at a financial institution and
being a local forecaster (FI LF), or having experience in academia and working
at a local institution (AC LI) generally benefit from experience over the full
and post-GFC periods. Experience is beneficial to the quality of education in Eco-
nomics (EC RE). For experts who graduated with a degree in Finance from a
top-ranked university, this is less true. However, experts who have a Master’s or
PhD in Economics from a top-ranked university can benefit from experience and
the latter seems to make them more confident, emboldening their forecasts.
Unlike local forecasters working at a financial institution (FI LF), working at
a financial and local institution (FI LI) does not influence the contribution of ex-
perts’ experience to performance or boldness. However, experience for experts at
a local institution while having experience in academia (AC LI) is beneficial and
encourages herding behaviors over the full sample, but experience seems detri-
mental and emboldens forecasts after the GFC. The total effect of experience on
experts at a local institution and a Master’s degree or PhD is beneficial and leads
20
to more herding behaviors, which is related to our previous findings for experts in
a local institution with a Master’s degree (Table 2). The total effect of experience
on local forecasters with a degree in Economics (LF EC) tends to be detrimental
during the full sample but remains neutral in terms of performance after the GFC.
Interestingly, experts with a degree in Economics benefit from experience differ-
ently if they hold a Master’s degree than those with experience in academia. The
findings regarding mixed geography and education relate to several strands of the
literature. Our results demonstrate that the likelihood of benefiting from experi-
ence in terms of boldness increases with the expert’s forecasting performance and
experience, and is influenced by the institution (Clement and Tse, 2005). Further-
more, the results complement the ones on experience (Hong et al., 2000; Mikhail
et al., 2003) and education (De Franco and Zhou, 2009) that influence social inter-
actions, cognitive biases, and intuitive reaction to uncertainty, an interpretation
partially shared with Christoffersen and Stæhr (2019).
Local female experts (G LF), female experts who graduated from a top-
ranked university in Finance (G RF), and female experts with central bank
experience (G CB) benefit more from experience than men with central bank
experience, confirming a labor market entry selection bias.
While differences in views may persist through time, differences in informa-
tion sets cannot explain such differences in opinion. Patton and Timmermann
(2010) show that they stem from heterogeneity in priors or models and that dif-
ferences in opinion move countercyclically. Although this heterogeneity is most
robust during recessions, our results bring another layer to this conclusion. The
GFC not only changed differences in opinion but also modified the influence of
experts’ characteristics on their forecasting performance and boldness.
The pre-GFC period significantly differs from the post-GFC one. However, our
results show that the GFC contributed significantly to changing the distribution
of the effects of characteristics and experience on forecasting performance and
boldness, a loosely documented phenomenon.13 Again, it should be noted that (a)
the results are not directly comparable due to different mean forecasters and (b)
the lack of significant results for the pre-GFC sample might be driven largely by
the low number of observations.
The results for boldness are generally similar to the results for performance.
Unsurprisingly, the average forecast is highly accurate at this time horizon. More
often than not, deviating from the herd is not beneficial.
Table 3 also reports the coefficients of determination for the fixed effects mod-
els. At first glance, these seem relatively low, falling in the range between 0.04
13There is a broad literature describing how economic conditions affect (expert) forecasts
(Adeney et al., 2017) or inflation perception. However, to our knowledge, there are few sources
documenting how economic conditions influence the changing effect of experts’ characteristics on
forecasting accuracy and boldness.
21
and 0.15. However, it must be kept in mind that the forecaster fixed effects–i.e.,
the actual characteristics–are not included in a within-R2. That is, the coefficient
of determination truly only captures the additional effect of experience and its in-
teractions. Also, as mentioned previously, nowcasts are highly precise since much
information is already available, limiting the influence of personal belief. Finding
that even for nowcasts, experience and characteristics have a statistically signifi-
cant effect, even though quantitatively small, is a strong result that indicates that
personal beliefs and abilities play an important role in understanding expectation
formation.
4.3 Pessimism
When looking at the behavioral side of forecasts, it seems evident that one of the
most relevant questions is whether forecasts–i.e., expectations–are optimistic or
pessimistic. While distinguishing optimism from pessimism is quite straightfor-
ward for business cycle forecasts, it is less so for inflation, where "good" and "bad"
are less clearly defined. One might look at the deviation from target inflation, but
it is hard to argue that 1.9% inflation is worse than 2.1%.
However, our sample includes two brief episodes when the US economy was
endangered by–and in some months experiencing–deflation. Unlike low infla-
tion, deflation is almost universally considered highly problematic in Economics,
allowing us to use those periods to assess the question of optimism vs. pessimism.
Figure 2 shows how in the early period of the GFC (late 2008 to early 2009) and
over most of 2015, forecasters disagreed on whether or not there would be defla-
tion.
For those subsamples we estimate a panel probit model explaining the proba-
blity that the forecast Eit (πit)would be below zero, taking the form:
p(Eit (πit)<0)=Φ(ψ0ln xit +XiΨ), (6)
where Φis the cumulative distribution function of the standard normal distribu-
tion and Ψis a vector of regression coefficients.
We split this sample into the subsamples when deflation was observed (i.e.,
when pessimism was justified, or–in other terms–the lack of deflation expectation
was overly optimistic), and when no deflation was observed (that is, when expect-
ing inflation can genuinely be seen as overly pessimistic).
Table 5 presents the probit panel regressions for a general situation (i.e., during
both inflationary and deflationary periods) and during only deflationary periods.
Table 5 shows that experts with more experience, a degree in Economics, or
central bank experience are less likely to predict deflation. However, this propen-
sity becomes less significant under deflation for experts with central bank expe-
22
Figure 2: Forecast Spread - Mixed Inflation and Deflation Expectations Periods
-3,0
-2,0
-1,0
0,0
1,0
2,0
3,0
4,0
5,0
6,0
7,0
%
Notes: The gray area represents the spread of forecasts (disagreement). The blue shaded back-
ground highlights the situations where both deflation and inflation were considered possible by
forecasters.
rience, which strongly influences private information. These experts are less pes-
simistic, but this is mitigated when pessimism turns out to be justified. These
results are not significant under deflation periods and may be explained by the
expert’s professional experience with few deflation periods and their education
curriculum, which may not contain extended attention to deflation or forecasting
deflation.
Positive numbers in Table 5 show that experts’ characteristics are more likely
to lead to expectations of deflation. Experts with central bank experience are op-
timistic, but people with experience in academia or high inflation are pessimistic.
Interestingly, this is mainly driven by actual deflation periods. Experts with cen-
tral bank experience and those with a degree in Economics do not (want to) see
deflation coming, but people with any high inflation experience or experience in
academia do.
Table 5 also presents the marginal effects, which translate the probit coefficients
into derivatives of the probability with respect to the explanatory variable. While
the marginal effects seem large, it has to be noted that we only look at periods
with disagreement on whether there will be deflation. Within this subsample, the
unconditional probability of predicting deflation exceeds 75%. Moreover, a unit
change of log experience at the mean (of this subsample) corresponds to almost 30
years of additional experience (with average log experience being 5.26).
At the mean, a unit change in experience reduces the probability of predicting
23
Table 5: Probit Estimates - Pessimism
All Deflation
Estimates Marginal Effects Estimates Marginal Effects
Experience -0.886*** -0.32 -0.933*** -0.37
Experience: Central bank (CB) -0.642** -0.23 -0.61* -0.24
Experience: Academia (AC) 0.613** 0.22 0.884*** 0.35
Education: Economics (EC) -0.858** -0.31 -0.697 -0.28
Ranking: Economics (RE) -2.083 -0.02 -1.444 -0.02
Experience: High Inflation (HI) 0.938** 0.34 1.135*** 0.45
Observations 359 257
Notes: ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. All stands for
both Inflation and Deflation periods, and Deflation stands for only deflation periods. Both samples
(All) are restricted to periods where there was disagreement, i.e., some experts predicted deflation
while others did not. The second sample (Deflation) is a subsample of All, where the pessimists
turned out to be correct.
deflation by 32 percentage points, corresponding to about 1.9 percentage points
for one year of experience (a log change of 0.06). This increases to 37 percentage
points (or 2.1 percentage points respectively) under deflation.
A degree in Economics reduces the probability of predicting deflation by 31
percentage points under normal times, while it is insignificant under deflation.
Central bank experience reduces the probability of predicting deflation by 23 per-
centage points under normal times, while this is less significant (but still signif-
icant at 10 percentage points) under deflation. Having academia or periods of
high inflation experience increases the probability of forecasting deflation by 22
percentage points and 34 percentage points, respectively. Under deflation, these
responses increase to 35 percentage points and 45 percentage points, respectively.
In line with our theoretical model (Appendix A), these results show that the
type and duration of an expert’s experience drive his or her sentiment. Their de-
gree of pessimism (optimism) may interact with their private signal processing.
According to Manzan (2011), heterogeneous sentiments may influence experts
in deciphering newly available information, involving a positive relationship be-
tween the interpretation of the mean signal and the prior sentiment (pessimistic
or optimistic). Hence, the effect of prior sentiment on an expert’s information
processing, discussed in Appendix A, may depend on the expert’s experience,
both duration and type (central bank or not).
We also complement the view stating that disagreement stems from hetero-
geneity in prior sentiments and moves countercyclically, with heterogeneity being
strongest during recessions, when forecasters appear to place greater weight on
their prior beliefs (Patton and Timmermann, 2010). Our results may confirm that
the weight of experts’ prior beliefs, and the beliefs themselves are influenced by
both experience and prior experience in a central bank, as well as the current state
24
of inflation (or deflation).
4.4 Over- and under-reaction
In this section, we provide evidence about the influence of cognitive factors on
experts’ forecast errors. We augment a standard Mincer and Zarnowitz (1969)
regression to allow for testing under- and over-reaction for different types of fore-
casters rather than just looking at the sample as a whole (Barberis et al., 1998;
Daniel et al., 1998).
Rather than merely estimating
Ei
t1[πt]πt=β0+β1πt+εit, (7)
where β0and β1are regression coefficients, and εit is the idiosyncratic component
of the error term. We estimate
Ei
t1[πt]πt=α0+ α1+
k
∑
n=1
αn+1xn,it!πt+ui+εt, (8)
where xn,it is the nth characteristic of forecaster iat time t. That is, through in-
teraction terms with πt, the effective coefficient β1from Equation 7 is allowed to
vary depending on forecaster characteristics. At the same time, we account for
unobserved heterogeneity through random effects14 (ui). We use the version of
the test that uses the forecast errors rather than forecasts on the left-hand side of
the equation; that is, we can test β1or α1+αnxn,it against 0 rather than against 1
to assess over- or under-reaction.
Table 6 presents the directional forecast error regression over the full sample.
Table 6 shows that experts underreact on average. However, experts with cen-
tral bank experience or more (accumulated) experience in forecasting tend to un-
derreact less than the average. Having central bank experience seems to reduce
underreaction behaviors more and more robustly than accumulating experience
in forecasting. Moreover, older experts (over 60 years old) underreact more than
the average. The underreaction of older experts may reflect a slight misinterpreta-
tion of private information and the decrease in forecasting performance with age
exhibited in Table 3.
The misinterpretation of genuine new private information affects forecasting
performance. Daniel et al. (1998) assume overconfidence about private informa-
tion involves stronger return predictability in firms with the greatest information
asymmetries,15 and suggest investigating whether the overconfidence of investors
and traders can be identified with specific characteristics. As in their model,
14A Hausman test prefers random effects to fixed effects in our specification.
15This also implies greater inefficiencies in the stock prices of small companies.
25
Table 6: Random Effects Estimates - Underreaction
Forecast Error
CPI -0.0356**
Experience (E) 0.0001*
Experience: Central bank (CB) 0.0079**
Age >60 years old -0.0007**
Constant 0.0093*
Observations 2904
Notes: ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. Unlike our
previous exercises, these results are obtained from undemeaned data, and the age variable used
here is the age of the expert rather than its dummy. A Hausman test justifies the use of a random
effects model.
the uninformed investors of our model could be interpreted as being contrarian-
strategy investors (whether institutions or individuals). Identifying the confidence
characteristics of different observable experts’ categories generate additional im-
plications.
4.5 Forecasting Ability
The characteristics identified in Section 4 should improve the predictive ability
of experts. We analyze time variation in the out-of-sample relative forecasting
performance to test this hypothesis. More precisely, we test for relative forecast
performance of characteristics-based expert groups in unstable environments, as
proposed by Giacomini and Rossi (2010). The null hypothesis is that the forecasts
under consideration perform equally well at every point in time. Exceeding the
critical value does not imply that one group of expert forecasts constantly outper-
forms the other but that there is a meaningful difference in the predictive ability
for a subsample. Fig. 3 presents these fluctuation tests for forecasts grouped by
characteristics.
The test rejects only for education level characteristics (MA and PhD, see Fig.
3) that is, experts with a PhD differ from the others significantly at least once,
and so do forecasters where the top degree is a Masters’s degree. Due to the low
number of experts holding neither an MA nor a PhD, those two tests capture al-
most the same information. Thus, it is unsurprising that the rejection is driven by
the same period in the mid-2000s, when PhD educated experts underperformed
the competition (mostly MAs), and MAs outperformed their competition (mostly
PhDs).
Interestingly, the test does not reject the remaining characteristics, confirming
that experts’ characteristics may have a significant role in their out-of-sample fore-
26
Figure 3: Forecasting Ability Tests
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
-2
0
2
Local forecaster
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
-2
0
2
Local institution
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
-2
0
2
Financial institution
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
-2
0
2
Experience: central bank
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
-5
0
5Education: MA
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
-5
0
5Education: Ph.D.
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
-2
0
2
Gender
Notes: The red dashed lines represent the critical values of the predictive ability test.
casting outcomes even in unstable environments. Our findings may have impli-
cations for policymakers. They may decide to select their inputs (multiple expert
forecasts) according to the expert’s characteristics to maximize their forecasting
precision and optimize their decisions.
Our results also provide a possible basis for an alternative to conventional fore-
cast combination methods in the literature. While the bias-adjusted combination
method is found to work well in practice (Capistrán and Timmermann, 2009b), we
demonstrate that a characteristics-based forecast combination is potentially more
desirable than equal-weighted or bias-adjusted forecast combination methods.
Inflation forecasts from the Survey of Professional Forecasters are biased, pre-
senting positive serial correlation in forecast errors, cross-sectional dispersion, and
predictability patterns depending on inflation variance. As we control for time
fixed-effects, we interpret experts’ shifts in forecasting performance not explained
by asymmetric loss and rational expectations (Capistrán and Timmermann, 2009a)
through their characteristics rather than inflation variance.
27
All in all, combining forecasts with respect to the expert’s characteristics gen-
erally improves out-of-sample forecasting performance.
5 Policy Implications and Conclusion
In line with our model, the characteristics we identify are shown to influence ex-
perts’ performance, boldness, forecasting ability, and sentiment (optimism or pes-
simism).
Investors and policymakers use forecasts to design or explain their decisions,
and sometimes, the efficiency of these decisions depends on forecasts (European
Central Bank, 2011, 2014; de Vincent-Humphreys et al., 2019). Identifying the char-
acteristics of the best experts could help firms and policymakers to achieve their
objectives efficiently. For instance, Carvalho and Nechio (2014) and Binder (2020)
show that households’ macroeconomic forecasts–about interest rates, inflation,
and unemployment–are not uniform across income and education levels. Fore-
casts also constitute an essential information channel leading investment portfolio
and spending decisions (Duca-Radu et al., 2021). The more accurate the forecast,
the less likely a surprise could occur, minimizing the required adjustment costs of
the investment portfolio and the corresponding market volatility when the data
become publicly available (Laster et al., 1999). We provide an in-depth picture
of the personal characteristics of professional forecasters that may affect informa-
tion rigidity, complementing Coibion and Gorodnichenko (2012, 2015). We also
confirm the influence of personal characteristics on forecasting performances of
consumers found by Duca-Radu et al. (2021) with professional forecasters.
Policy institutions extensively use expert forecasts for both decision-making
and forecasting purposes (Piotroski and Roulstone, 2004; Adeney et al., 2017).
While policymakers generally aggregate these forecasts in simplistic ways and
rank them without considering the expert characteristics (Alessi et al., 2014;
Coibion et al., 2020), the main takeaway from our results is that experts’ char-
acteristics drive forecasting outcomes, boldness, and sentiment.
Consequently, policymakers may use our results to group forecasters with re-
spect to some of their characteristics (Section 4.5) to increase the reliability of their
inflation forecasts compared to simple averaging, thus improving their policy
decision making processes. This should also hamper the spillover effect of pes-
simistic or optimistic behaviors on inflation forecasts, which is somehow frequent
during specific periods such as deflation, if policymakers group expert forecasts
according to experience (Section 4.3).
The current state of inflation (or deflation) and of the economy (after or before
a crisis like the GFC) influence experts’ behaviors and beliefs, and thus the trans-
mission is channeled from their characteristics to their forecasting performance
28
and boldness as demonstrated here.
Underperforming experts are more likely to no longer be part of our expert
database, i.e., they are less likely to be in charge of the inflation forecasts con-
tributed to the Bloomberg database, while boldness does not significantly influ-
ence the experts’ survival rate. Degrees in Finance or Economics do not offer
the same protection as career concerns or institutional labor market expectations,
while graduating from a top university decreases the expert’s survival rate.16
Comparing our results from the two subsamples reveals that the GFC changed
both financial institutions and the expert’s labor market. After the GFC, expert’s
experience, location, institution type, education field, or quality change their fore-
casting performance. Our panel estimations also show that before the GFC, only
the education field and quality mattered for forecasting performance and bold-
ness under fixed effects. More characteristics play a significant role under random
effects.
The expert’s location, institution location and type, experience type, and gen-
der affect the expert’s forecasting ability. The expert’s previous experience or pre-
vious experience in a central bank significantly influences the expert’s sentiment.
We interpret our results as evidence of the effect of characteristics on experts’
inflation forecast outcomes. One implication of our analysis is that experts’ char-
acteristics and experience matter for policymakers as long as expert forecasts are
considered in their decision-making process.
References
Adeney, R., Arsov, I., Evans, R., 2017. Inflation expectations in advanced
economies. RBA Bulletin March Quarter, Reserve Bank of Australia.
Alessi, L., Ghysels, E., Onorante, L., Peach, R., Potter, S., 2014. Central bank macro-
economic forecasting during the global financial crisis: the European Central
Bank and Federal Reserve Bank of New York experiences. Journal of Business &
Economic Statistics 32 (4), 483–500.
Andrade, P., Le Bihan, H., 2013. Inattentive professional forecasters. Journal of
Monetary Economics 60 (8), 967–982.
Bachmann, R., Elstner, S., Sims, E. R., 2013. Uncertainty and economic activity:
evidence from business survey data. American Economic Journal: Macroeco-
nomics 5 (2), 217–249.
16See MA RF and PhD RE in Table 8.
29
Bae, K.-H., Stulz, R. M., Tan, H., 2008. Do local analysts know more? A cross-
country study of the performance of local analysts and foreign analysts. Journal
of Financial Economics 88 (3), 581–606.
Barberis, N., Shleifer, A., Vishny, R., 1998. A model of investor sentiment. Journal
of Financial Economics 49 (3), 307–343.
Batchelor, R., 2007. Bias in macroeconomic forecasts. International Journal of Fore-
casting 23 (2), 189–203.
Benchimol, J., El-Shagi, M., 2020. Forecast performance in times of terrorism. Eco-
nomic Modelling 91, 386–402.
Berger, H., Ehrmann, M., Fratzscher, M., 2009. Forecasting ECB monetary policy:
Accuracy is a matter of geography. European Economic Review 53 (8), 1028–
1041.
Bernanke, B. S., 2007. Inflation expectations and inflation forecasting. Speech 306,
Board of Governors of the Federal Reserve System.
Binder, C., 2020. Coronavirus fears and macroeconomic expectations. Review of
Economics and Statistics 102 (4), 721–730.
Bloom, N., 2014. Fluctuations in uncertainty. Journal of Economic Perspectives
28 (2), 153–176.
Bordalo, P., Gennaioli, N., Ma, Y., Shleifer, A., 2020. Overreaction in macroeco-
nomic expectations. American Economic Review 110 (9), 2748–2782.
Breslow, N., Chatterjee, N., 1999. Design and analysis of two-phase studies with
binary outcome applied to Wilms tumour prognosis. Journal of the Royal Sta-
tistical Society Series C 48 (4), 457–468.
Budescu, D. V., Chen, E., 2015. Identifying expertise to extract the wisdom of
crowds. Management Science 61.
Campbell, S. D., Sharpe, S. A., 2009. Anchoring bias in consensus forecasts and its
effect on market prices. Journal of Financial and Quantitative Analysis 44 (2),
369–390.
Capistrán, C., Timmermann, A., 2009a. Disagreement and biases in inflation ex-
pectations. Journal of Money, Credit and Banking 41 (2-3), 365–396.
Capistrán, C., Timmermann, A., 2009b. Forecast combination with entry and exit
of experts. Journal of Business & Economic Statistics 27 (4), 428–440.
30
Carvalho, C., Nechio, F., 2014. Do people understand monetary policy? Journal of
Monetary Economics 66 (C), 108–123.
Christoffersen, J., Stæhr, S., 2019. Individual risk tolerance and herding behaviors
in financial forecasts. European Financial Management 25 (5), 1348–1377.
Clarke, J., Subramanian, A., 2006. Dynamic forecasting behavior by analysts: The-
ory and evidence. Journal of Financial Economics 80 (1), 81–113.
Clement, M. B., Tse, S. Y., 2005. Financial analyst caracteristics and herding behav-
ior in forecasting. Journal of Finance 60 (1), 307–341.
Coibion, O., Gorodnichenko, Y., 2012. What can survey forecasts tell us about in-
formation rigidities? Journal of Political Economy 120 (1), 116–159.
Coibion, O., Gorodnichenko, Y., 2015. Information rigidity and the expectations
formation process: a simple framework and new facts. American Economic Re-
view 105 (8), 2644–2678.
Coibion, O., Gorodnichenko, Y., Kumar, S., Pedemonte, M., 2020. Inflation expec-
tations as a policy tool? Journal of International Economics 124, 103297.
Cornand, C., Hubert, P., 2020. On the external validity of experimental inflation
forecasts: A comparison with five categories of field expectations. Journal of
Economic Dynamics and Control 110, 103746.
Cox, D. R., 1972. Regression models and life-tables. Journal of the Royal Statistical
Society: Series B 34 (2), 187–202.
Daniel, K., Hirshleifer, D., Subrahmanyam, A., 1998. Investor psychology and se-
curity market under- and overreactions. Journal of Finance 53 (6), 1839–1885.
Davis, M., Lleo, S., 2020. Debiased expert forecasts in continuous-time asset allo-
cation. Journal of Banking & Finance 113 (C), 105759.
De Franco, G., Zhou, Y., 2009. The performance of analysts with a CFA desig-
nation: The role of human-capital and signaling theories. Accounting Review
84 (2), 383–404.
de Vincent-Humphreys, R., Dimitrova, I., Falck, E., Henkel, L., Meyler, A., 2019.
Twenty years of the ECB Survey of Professional Forecasters. Economic Bulletin
Articles 1.
Duca-Radu, I., Kenny, G., Reuter, A., 2021. Inflation expectations, consumption
and the lower bound: Micro evidence from a large multi-country survey. Jour-
nal of Monetary Economics 118 (C), 120–134.
31
El-Shagi, M., Giesen, S., Jung, A., 2016. Revisiting the relative forecast perfor-
mances of Fed staff and private forecasters: A dynamic approach. International
Journal of Forecasting 32 (2), 313–323.
European Central Bank, 2011. Monthly Bulletin. Executive Board 01/2011, Euro-
pean Central Bank.
European Central Bank, 2014. Monthly Bulletin. Executive Board 01/2014, Euro-
pean Central Bank.
Fang, L., Huang, S., 2017. Gender and connections among Wall Street analysts.
Review of Financial Studies 30 (9), 3305–3335.
Gabaix, X., 2019. Behavioral inattention. In: Bernheim, B. D., DellaVigna, S., Laib-
son, D. (Eds.), Handbook of Behavioral Economics - Foundations and Applica-
tions 2. Vol. 2 of Handbook of Behavioral Economics: Applications and Foun-
dations 1. Elsevier, pp. 261–343.
Gabaix, X., 2020. A behavioral New Keynesian model. American Economic Re-
view 110 (8), 2271–2327.
Genre, V., Kenny, G., Meyler, A., Timmermann, A., 2013. Combining expert fore-
casts: Can anything beat the simple average? International Journal of Forecast-
ing 29 (1), 108–121.
Giacomini, R., Rossi, B., 2010. Forecast comparisons in unstable environments.
Journal of Applied Econometrics 25 (4), 595–620.
Heckman, J. J., 1979. Sample selection bias as a specification error. Econometrica
47 (1), 153–161.
Hong, H., Kacperczyk, M., 2010. Competition and bias. Quarterly Journal of Eco-
nomics 125 (4), 1683–1725.
Hong, H., Kubik, J. D., Solomon, A., 2000. Security analysts’ career concerns and
herding of earnings forecasts. RAND Journal of Economics 31 (1), 121–144.
Huang, R., Krishnan, M., Shon, J., Zhou, P., 2017. Who herds? Who doesn’t? Es-
timates of analysts’ herding propensity in forecasting earnings. Contemporary
Accounting Research 34 (1), 374–399.
Jacob, J., Lys, T. Z., Neale, M. A., 1999. Expertise in forecasting performance of
security analysts. Journal of Accounting and Economics 28 (1), 51–82.
Kaplan, E. L., Meier, P., 1958. Nonparametric estimation from incomplete observa-
tions. Journal of the American Statistical Association 53 (282), 457–481.
32
Keane, M. P., Runkle, D. E., 1990. Testing the rationality of price forecasts: new
evidence from panel data. American Economic Review 80 (4), 714–735.
Kumar, A., 2010. Self-selection and the forecasting abilities of female equity ana-
lysts. Journal of Accounting Research 48 (2), 393–435.
Laster, D., Bennett, P., Geoum, I. S., 1999. Rational bias in macroeconomic fore-
casts. Quarterly Journal of Economics 114 (1), 293–318.
Lim, T., 2001. Rationality and analysts’ forecast bias. Journal of Finance 56 (1),
369–385.
Malloy, C. J., 2005. The geography of equity analysis. Journal of Finance 60 (2),
719–755.
Malmendier, U., Nagel, S., 2011. Depression babies: do macroeconomic experi-
ences affect risk taking? Quarterly Journal of Economics 126 (1), 373–416.
Malmendier, U., Nagel, S., 2016. Learning from inflation experiences. Quarterly
Journal of Economics 131 (1), 53–87.
Malmendier, U., Nagel, S., Yan, Z., 2021. The making of hawks and doves. Journal
of Monetary Economics 117 (C), 19–42.
Manzan, S., 2011. Differential interpretation in the survey of professional forecast-
ers. Journal of Money, Credit and Banking 43 (5), 993–1017.
Marimon, R., Sunder, S., 1993. Indeterminacy of equilibria in a hyperinflationary
world: experimental evidence. Econometrica 61 (5), 1073–1107.
Mikhail, M. B., Walther, B. R., Willis, R. H., 2003. The effect of experience on se-
curity analyst underreaction. Journal of Accounting and Economics 35 (1), 101–
116.
Mincer, J. A., 1974. Schooling, Experience, and Earnings. NBER Books. New York,
NY: National Bureau of Economic Research.
Mincer, J. A., Zarnowitz, V., 1969. The evaluation of economic forecasts. In: Eco-
nomic Forecasts and Expectations: Analysis of Forecasting Behavior and Perfor-
mance. NBER Chapters. National Bureau of Economic Research, pp. 3–46.
Mitchell, K., Pearce, D. K., 2007. Professional forecasts of interest rates and ex-
change rates: Evidence from the Wall Street Journal’s panel of economists. Jour-
nal of Macroeconomics 29 (4), 840–854.
33
Patton, A. J., Timmermann, A., 2010. Why do forecasters disagree? Lessons from
the term structure of cross-sectional dispersion. Journal of Monetary Economics
57 (7), 803–820.
Piotroski, J. D., Roulstone, B. T., 2004. The effects of analyst forecast properties
and country-level institutions on the cost of debt. Accounting Review 79 (4),
1119–1151.
Romer, C. D., Romer, D. H., 2000. Federal reserve information and the behavior of
interest rates. American Economic Review 90 (3), 429–457.
Rossi, B., Sekhposyan, T., 2015. Macroeconomic uncertainty indices based on now-
cast and forecast error distributions. American Economic Review 105 (5), 650–
655.
Scharfstein, D. S., Stein, J. C., 1990. Herd behavior and investment. American Eco-
nomic Review 80 (3), 465–479.
Thomas, L. B., 1999. Survey measures of expected U.S. inflation. Journal of Eco-
nomic Perspectives 13 (4), 125–144.
Veress, A., Kaiser, L., 2017. Forecasting quality of professionals: Does affiliation
matter? Quarterly Review of Economics and Finance 66 (C), 159–168.
Zwiebel, J., 1995. Block investment and partial benefits of corporate control. Re-
view of Economic Studies 62 (2), 161–185.
34
Appendix
A Model
In this section, we present a stylized model that motivates the inclusion of nonlin-
earities and interactions. We show that even using a straightforward and standard
form of learning–namely Bayesian learning–experience has nonlinear effects that
depend strongly on initial conditions. Our model belongs to the "noisy informa-
tion" model class. As in Coibion and Gorodnichenko (2012, 2015), agents continu-
ously monitor variables and consider the most updated information to formulate
their decisions. Different responses may occur in information acquisition when
fundamental shocks happen.
Consider a forecaster aiming to forecast inflation π, who is exposed to both
public and private signals. We assume both signals are drawn from a normal
distribution, they are isolated17 and mutually independent, and both are unbiased
but noisy. In other words:
su
t N πt,τ1
u, (9)
and
sr
t N πt,τ1
r, (10)
where su
tand sr
tare the public and private signals at time t, respectively.
We assume that the precision of the public signal, τu, is known and con-
stant. The precision of the private signal, τr, is unknown to the forecaster initially.
Rather, starting from a prior assumption, he learns about the quality (i.e., the pre-
cision) of his private signal over time through Bayesian learning.
For simplicity, we assume that the prior regarding the precision of the private
signal is Gamma distributed, where the initial prior has a mean of ˜
τ0and a vari-
ance of σ2
τ,0.18
Since the Gamma distribution is a conjugate prior to the precision of a normal
distribution with a known mean (in this case zero), the Bayesian updating yields
a new Gamma distribution with lower variance and a more accurate estimate of
the true mean every period.
More precisely, the mean of new distribution is given by
˜
τt=˜
τ2
t1/σ2
τ,0 +1/2
˜
τt1/σ2
τ,t1+1/2 sr
t1πt12, (11)
17Each expert cannot observe other forecasts when extrapolating this one-agent model to a
multiple-agent model, a simplifying assumption corresponding to the findings of Bordalo et al.
(2020).
18In the more common α,βparameterization, this corresponds to α0=˜
τ2
0/σ2
τ,0 and β0=˜
τ0/σ2
τ,0.
While this notation makes the updating equations more convoluted, it allows for deriving the
interest variable’s law of motion more straightforwardly.
35
and the variance by
σ2
t=˜
τ2
t1/σ2
τ,0 +1/2
˜
τt1/σ2
τ,t1+1/2 sr
t1πt122. (12)
The Bayesian point estimate for the average forecaster–i.e. the representative
forecaster that is repeatedly experiencing errors of a magnitude of exactly one
standard deviation–who starts forecasting in t=0 and does so every period at
time tis thus given by:
˜
τt=˜
τ2
0/σ2
τ,0 +(t1)/2
˜
τ0/σ2
τ,0 +t/2τr
. (13)
A forecaster who aims to maximize the expected precision of his forecast, will
then provide a weighted forecast
ft=τr
τr+˜
τt
su
t+˜
τt
τr+˜
τt
sr
t, (14)
implying a precision of
τt= τr
τr+˜
τt2
τ1
u+˜
τt
τr+˜
τt2
˜
τ1
t!1
. (15)
The dynamics implied by this model are quite straightforward:
1. Unless ˜
τ0=τu, the forecast performance will improve over time, since any
misconception–whether it is overly optimistic or pessimistic–regarding the
quality of the private signal will lead to a suboptimally weighted forecast.
2. Precision converges monotonically to τ=τu+τr.
We can now imagine a range of factors that can potentially drive forecast qual-
ity τrand the initial optimism (or pessimism) regarding one’s own forecast quality
˜
τ0.
Figure 4 shows some core scenarios–namely a good, a bad, and two aver-
age forecasters, with the latter two differing in the degree of optimism (or pes-
simism) regarding the quality of their private signal–linking expected forecast per-
formance to time, i.e., experience.
In the left panel, we show the “raw numbers”, i.e., time (experience) tand
precision τt, obtained from substituting Eq. 13 into Eq. 15. In the right panel,
we show the data using the transformations that will be applied in the empiri-
cal approach, i.e., the variance (corresponding to the expected values of squared
forecast errors) and the natural logarithm of time (experience). The different tra-
jectories of performance in response to time–i.e., experience–are visible. That is,
36
Figure 4: Evolution of forecast precision and variance over time
1,0
1,5
2,0
2,5
3,0
12345678910
Precision
Experience
Good Bad Optimistic Pessimistic
0,3
0,4
0,5
0,6
0,7
0,8
12345678910
Variance
Experience (log)
Good Bad Optimistic Pessimistic
Notes: The good forecaster has a precision of two with a prior of one. The bad forecaster has a
precision of 0.5 with the same prior. The optimistic forecaster has a precision of one with a prior
of two, and the pessimistic forecaster has a precision of one with a prior of 0.5. All priors have a
variance of zero, and τu=1.
factors that drive forecast performance–through both actual quality and subopti-
mal weighting–also affect the performance’s response to experience (here shown
as time). In other words, it is necessary to include forecast experience and constant
factors driving forecast performance in a model, but also to consider possible in-
teractions.
The model predicts that both heterogeneity in prior expectations and under-
standing of news across forecasters may drive dispersion in forecasts. Heteroge-
neous interpretation of the incoming information might exacerbate the dispersion
of forecasts. Similar results are observed in Coibion and Gorodnichenko (2012,
2015). We assume this interpretation of incoming information is influenced by
experts’ characteristics such as experience and education, for example.
To some degree, it is always good to include your own information19 so fore-
casts do not converge–worse forecasters stay worse. The two experts who have the
same "true" quality of the private signal–the pessimist and the optimist–eventually
converge as they learn their signal’s quality.
B Survival Analysis
In our first set of analyses, we assess forecasters’ probability of providing a further
forecast in the future, i.e., to survive in the market. Our dataset does not allow us
to distinguish the reasons for possible removal from the dataset. Being fired, or at
least removed from this particular responsibility, are possibilities, but merely re-
tiring or even being promoted are equally possible. However, the main reason for
19Since the signal error is unrelated, even a high variance signal is meaningful.
37
us to conduct this analysis is that bad forecast performance may lead to exclusion
from the sample, thereby creating endogeneity issues as outlined above.
The approach we chose is a proportional hazards model pioneered by Cox
(1972). In this model, we estimate
hit =h0(t)eλ0f(Yit1)+XiΛ, (16)
where Yit1is the past history of yit ,h0(t)corresponds to the time-specific ef-
fect, Λis a vector of regression coefficients, and λ0is a regression coefficient. We
use three different transformations f(Yit1)describing slightly different possibil-
ities of how performance review might be conducted by the hiring institutions:
last-period performance or boldness20 (Raw), five-period backward-looking mov-
ing average (MA5), and weighted moving average where the most recent periods
have a higher weight (wMA5). First, we simply take the last forecast’s value, i.e.
Yit1=yi,t1. Second, we look at a moving average over the past five months, in-
dicating a slightly longer horizon rather than penalizing extremely bad forecasts
immediately, i.e. Yit1=1/5 ∑5
s=1yits. Finally, we look at a weighted moving
average, i.e., Yit1=∑5
s=1wsyits, where the weights decline exponentially by a
power of two.
Since we want to stay true to our general multivariate framework that includes
continuous time-varying variables, we abstain from traditional Kaplan and Meier
(1958) type survival plots, which allow assessing actual expected survival time
for distinct subgroups. Therefore, we cannot interpret our results in terms of ad-
ditional months of survival in the job, but merely interpret the sign and relative
magnitude of coefficients.
Before presenting our survival analysis, we present preliminary results in Ta-
ble 7 to capture our research question’s intuition. Tables 7 and 8 present three
transformations for each dependent variable. In these tables, positive significant
coefficients show that low forecasting performance or boldness increases the prob-
ability of removal (i.e., not being a forecaster in the next period).
Table 7 presents the survival estimates for US CPI experts without interactions
with few variables.
Table 7 presents the survival estimates for US CPI experts, without interactions,
related only to the expert’s location and education field (Economics). It shows
that these variables matter even in this simplified model. Although the expert’s
survival depends on his past performance, especially MA5 and wMA5, it also
depends on his education in Economics and on his current location.
A local inflation expert may survive longer than a foreign expert, which may
reflect employment protection or information advantage effects, in line with the
literature on financial analysts (Malloy, 2005). We also find that an expert with a
20See Eq. 1 and Eq. 2, respectively, for the definitions of performance and boldness.
38
Table 7: Survival Estimates - Linear Model
Performance
Raw MA5 wMA5
Dependent variable 2.435 7.359*** 4.699**
Local forecaster (LF) -0.478** -0.474** -0.473**
Education: Economics (EC) -0.725*** -0.724*** -0.718***
Notes: ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. Raw stands
for the last period dependent variable, MA5 stands for five-period moving average (backward-
looking) and wMA5 for weighted moving average where the most recent periods have a higher
weight.
degree in Economics may survive longer than those without an Economics degree,
a result interpreted below (Table 8).
Although the coefficient looks large in Tables 7 and 8, squared forecast errors
for inflation nowcasts are small. Thus, the total impact of forecast quality on sur-
vival probability is small, which mitigates endogeneity issues.
Table 7 shows that a low past forecasting performance (MA5 and wMA5) in-
creases the probability of being removed.21 Providing bolder forecasts decreases
expert survival but less significantly. According to these preliminary results, the
expert’s future seems to be determined by current and past performance.
Table 8 presents the survival estimates for US CPI experts with interactions.22
As in Table 7, Table 8 shows that having a degree in Economics immunizes the
expert from being removed as an economist. Interestingly, Table 8 shows that a
degree in Finance does not offer the same protection as having graduated in Eco-
nomics. Experts with a degree in Economics may provide more convincing expla-
nations to back their information processing and subsequent forecasts of the US
CPI, an economic variable par excellence, than experts with a degree in Finance,23
increasing their survival rate.24 Having graduated from a top-ranked university
in Economics or Finance25 seems to improve the expert’s survival rate, and having
experience in academia seems to impair the expert’s survival rate.
The expert’s survival rate seems to be determined more by current and past
21The presented results use the exact partial likelihood. Under the Breslow and Chatterjee (1999)
approximation, low past performance in forecasting inflation (MA5) increases the probability of
being removed.
22See Section 3 for more details about the variables.
23Whatever the expert’s outcomes, providing more convincing economic explanations about
their inflation forecasts helps.
24In a previous version of this paper, an application to Fed fund rates shows that experts with a
degree in Finance survive better than those with a degree in Economics. We attributed this result
to the fact that the nominal interest rate is both an economic and a financial variable and is thus
better explained by experts with an education in Finance or both Economics and Finance.
25See Section 3.2 for more details about the ranking variables we use.
39
Table 8: Survival Estimates - Nonlinear Model
Performance Boldness
Raw MA5 wMA5 Raw MA5 wMA5
Dependent variable 2.509 8.14*** 5.173** 3.912 11.612** 7.651*
Age >60 years old -0.089 -0.124 -0.103 -0.115 -0.137 -0.13
Local forecaster (LF) -0.674* -0.668* -0.668* -0.682* -0.69* -0.685*
Local institution (LI) 0.173 0.199 0.19 0.196 0.176 0.196
Financial institution (FI) -0.767* -0.711* -0.754* -0.779** -0.744* -0.765**
Experience: Central bank (CB) 0.4 0.425 0.408 0.398 0.4 0.399
Experience: Academia (AC) 0.759* 0.799** 0.791** 0.737* 0.739* 0.747*
Gender (G) 0.319 0.38 0.349 0.335 0.376 0.355
Education: MA (MA) -0.196 -0.276 -0.231 -0.218 -0.249 -0.236
Education: PhD (PhD) -0.588 -0.697 -0.641 -0.576 -0.596 -0.591
Education: Economics (EC) -1.126** -1.11** -1.086** -1.151*** -1.157*** -1.141***
Education: Finance (EF) -0.528 -0.534 -0.483 -0.565 -0.539 -0.534
Ranking: Economics (RE) -0.07 -0.076 -0.076 -0.086 -0.094 -0.093
Ranking: Finance (RF) 0.057 0.067 0.057 0.063 0.069 0.064
Interactions
MA RF -0.537* -0.562** -0.55* -0.54* -0.558** -0.548**
PhD RE -0.598** -0.602** -0.596** -0.58** -0.578** -0.571*
Notes: ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. Raw stands
for the last period dependent variable, MA5 stands for five-period moving average (backward-
looking) and wMA5 for weighted moving average where the most recent periods have a higher
weight.
performance, and to a lesser extent, by only current performance. Low average
performance (MA5) significantly decreases the expert’s survival rate, while herd-
ing influences the experts’ survival rate less significantly.26
Several theories of reputation and herd behavior indicate that agents’ perfor-
mance and boldness may vary with career concerns (Scharfstein and Stein, 1990;
Zwiebel, 1995). Our results suggest that increased reputational capital may in-
crease the labor market attractiveness of top-ranked profiles, thus leading experts
originating from or targeting academia to leave the profession more frequently
than other experts. This might also mean that these experts are fired or moved to
another occupation (in the same or another institution) more often because, for in-
stance, they are hired based on higher or different expectations than others, which
they may not meet.
Table 8 provides additional insights about experts’ survival. First, it shows that
having graduated with a Master’s degree from a top Finance institution improves
the expert’s survival. This result sharply contrasts with an education in Finance
that does not significantly influence the expert’s survival. Second, an MA from a
top institution in Finance and a PhD from a top institution in Economics increase
26In general, most experts follow others or are moderately bold. It is evident that, as long as the
consensus provides the best forecast on average, an expert always far from the consensus (bold)
will have a lower survival rate.
40
the survival probability of the expert. This increase in the expert’s survival rate is
slightly more significant for a PhD from a top institution in Economics than for a
Master’s degree from a top institution in Finance.
It is theoretically possible that the effect we attribute to a PhD degree (and
other expert-specific factors) is indeed driven by better institutions hiring based on
those factors (whether they matter or not). An expert with a PhD might increase
his chances of working at a better institution, with better information-gathering
functions, than an expert without a PhD. Financial institutions may have better
information-gathering functions, data, and private information access. Also, ex-
perts with a PhD may be attracted by these institutions, and these institutions may
prefer recruiting people with a PhD, making a case for firm effects that are both
an intermediary and a factor. Although we do not have enough companies with
more than one forecaster in the sample (otherwise, we would be able to control for
firm fixed effects), we believe that it is unlikely that the institution is an intermedi-
ary rather than a factor. Indeed, as long as we use Bloomberg forecasts, meaning
all our experts have access to at least a Bloomberg terminal, we can reasonably
assume that all the forecasters have access to almost the same information and
necessary equipment to build their forecasts. Hence, the effect of the PhD is more
prevalent than firm effects.
Tables 8 mitigates existing career-concern-motivated herding theories (Hong
et al., 2000). While our results show that experts are more likely to lose their jobs
after<