Direct Evidence on Sticky Information from the Revision Behavior
of Professional Forecasters
Karlyn Mitchell
Department of Business Management
Poole College of Management
North Carolina State University
Raleigh, NC 27695-7229
[email protected]
Douglas K. Pearce*
Department of Economics
Poole College of Management
North Carolina State University
Raleigh, NC 27695-8110
[email protected]
*
Corresponding author
April 1, 2017
Direct Evidence on Sticky Information from the Revision Behavior
of Professional Forecasters
Karlyn Mitchell
Department of Business Management
Poole College of Management
North Carolina State University
Douglas K. Pearce
Department of Economics
Poole College of Management
North Carolina State University
ABSTRACT
We provide evidence on the sticky-information model of Mankiw and Reis (2002) by
examining how often individual professional forecasters revise their forecasts. We draw
interest rate and unemployment rate forecasts from the monthly Wall Street Journal surveys.
We find evidence that forecasters frequently leave forecasts unchanged but revise more often
the larger the changes in the information set; additionally, the information sensitivity of
revision frequencies increased after 2007. We also find that, on average, forecasters in our
sample revise more frequently than found in previous research but that revised forecasts are
not consistently more accurate.
Keywords: Expectations, Sticky Information, Forecasts, Survey Data
JEL Codes: C53, D83, D84, E37, E47
1. Introduction
How economic agents process information to form expectations continues to be a central
issue in macroeconomics. Recent work proposes alternatives to the full information, rational
expectations model that presumes agents form expectations from complete information and
revise them when relevant new information appears.
Woodford (2003) relaxes the full
information assumption to develop a model in which agents extract signals from noisy
information (the noisy-information model). Sims (2003) considers limits to information
processing which lead rational agents to form expectations from incomplete information (the
rational inattention model). Reis (2006) and Mankiw and Reis (2002) posit significant costs of
acquiring and processing information that deter agents from updating their information sets and
revising their expectations every time new information arrives (the sticky-information model).
The sticky-information model has received empirical support from Mankiw, Reis, and
Wolfers (2003) and Coibion and Gorodnichenko (2015), who examine indirectly the frequency
with which professional forecasters revise their forecasts.
Mankiw, Reis, and Wolfers first
simulate inflation forecasts of agents who asynchronously collect information and revise their
forecasts using a sticky-information model. They then compare dispersion in the simulated
forecasts to dispersion in the actual forecasts of professional forecasters (consumers) and find
that the simulated series mirrors the actual series most closely when the agents revise their
inflation expectations about every 10 months (12.5 months).1 Coibion and Gorodnichenko
(2015) assume that professional forecasters make full information, rational expectations forecasts
but that costs prevent some from revising their forecasts every period. They estimate the
frequency of forecast revision by regressing the average forecast error for a specific horizon on
1
Mankiw, Reis, and Wolfers (2003) use the Livingston Survey for professional forecasts and the Michigan Survey
of Consumer Attitudes and Behavior for consumer expectations. Like Mankiw, Reis, and Wolfers, Carroll (2003)
also finds that households revise their expectations about once a year, based partially on professional forecasts.
1
the revision of the average forecast. They conclude that forecasters revise their inflation forecasts
once every 6 to7 months, on average.2
Later work investigates the sticky-information model using more direct methods.
Andrade and LeBihan (2013) measure the fraction of forecasters revising their forecasts each
quarter in the European Survey of Professional Forecasters. They find that, on average,
forecasters update their inflation forecasts about every 4 months, more frequently than found by
Mankiw, Reis, and Wolfers (2003) and Coibion and Gorodnichenko (2015).3 Pfajfar and Santoro
(2013) examine numbers of households revising their inflation forecasts in the Michigan Survey.
They find that households are more likely to revise their expectations the more newspaper
reports about inflation have appeared recently. Dovern et al. (2015) examine monthly GDP
forecasts by individual professional forecasters in thirty-six countries assembled by Consensus
Economics and find that forecasters revise their forecasts about every three months. These
findings challenge the sticky-information model (Coibion, 2015).
In this paper, we produce new evidence on the sticky-information model by studying
monthly forecasts of three economic variables made one to twelve months ahead by individual,
professional economists in the Wall Street Journal (WSJ) Economic Forecasting Survey from
2003 to 2014. Our evidence is new because forecasts from the WSJ survey have not, to our
knowledge, been used for this purpose. We begin by documenting properties of the economists’
forecasts and forecast revision behavior. Then, we estimate models of their revision behavior to
study how forecast horizon and economic changes affect the economists’ propensities to revise
2
Coibion and Gorodnichenko (2015) use the quarterly Survey of Professional Forecasters (SPF). Mertens and
Nason (2015) estimate a model similar to that of Coibion and Gorodnichenko (2015) on forecasts of the GDP
deflator from the SPF and find that forecasters reduced their revision frequency from about every 5 months in the
1970s to about every 7-8 months after 2000.
3
Andrade and LeBihan (2013) use the quarterly European Survey of Professional Forecasters. Armantier et al.
(2016) conduct an experiment on how households revise their inflation expectations and find that 42-47 percent do
not revise their expectations when given the opportunity.
2
forecasts. The dependent variable in our models is the fraction of forecasters not revising their
forecasts since the last survey because it implies a revision frequency measure comparable to that
of Coibion and Gorodnichenko (2015). We also examine whether the economists’ revision
behavior changed after the financial crisis. Then, we estimate the forecast revision model of
Coibion and Gorodnichenko (2015) on our data and compare its estimated forecast revision rates
with the revision rates we observe directly for the WSJ economists. Finally, we examine whether
economists who revise their predictions forecast more accurately.
Our use of the WSJ surveys has two main advantages. First, the surveys are monthly,
allowing forecasters to revise forecasts more frequently than possible with either the quarterly
(US and European) Surveys of Professional Forecasters or the semi-annual Livingston Survey.
Second, the surveys identify each forecaster by name, allowing us to associate forecasters with
their individual forecasts. Thus we can construct direct measures of forecast revision frequency
and conduct direct tests of whether forecasts are state-independent and constant through time, as
assumed by Coibion and Gorodnicheko (2015).
Our study uses the WSJ economists’ forecasts of three variables: the 10-year Treasury
bond rate, the fed funds rate, and the unemployment rate. We choose these variables because
they are rarely revised after being observed, avoiding ambiguity about whether forecasters
intended to forecast initially reported or revised values of a variable. We also choose them
because the WSJ survey asks for forecasts of the values of these variables on specific days or
months rather than asking for forecast averages over rolling horizons, like many other surveys.
Compared with forecast averages, single-date values yield cleaner measures of revision
frequency by avoiding revisions which correct for earlier forecast errors.4
4
For example, if we are forecasting the average annual inflation rate for 2014 and we make our new forecast in, say,
May 2014 after observing the actual monthly inflation rate for April 2014, we may change our forecast by replacing
3
To preview our results, we find some support for the sticky-information model in that
substantial numbers of forecasters do not revise their forecasts of the three variables we study at
every opportunity. The fraction of non-revisers varies with the variable forecasted. The fraction
is also state dependent: forecasters are more likely to revise their forecasts of a variable the
greater the change in that variable since the prior survey. This finding is significant because
indirect methods of studying revision frequency presume state-independent forecaster behavior.
We document how one such method misestimates the directly observed forecast revision
frequency in the WSJ survey. While we find that the WSJ forecasters do not revise their
forecasts at every opportunity, we also find that the average time between revisions is shorter
than reported by most previous studies, casting some doubt on how well the sticky-information
model can account for the persistence of macro-economic shocks. Additionally, we find that
recently revised forecasts are not consistently more accurate than unrevised forecasts.5
The rest of the paper is organized as follows. Section 2 details our data. Section 3
describes our tests of the state dependency of forecast revision behavior and reports our results.
Section 4 presents extensions of our basic model. Section 5 concludes our paper.
2. The Data
2.1. The Wall Street Journal Survey
We take our data from the Wall Street Journal Economic Forecasting Survey from March
2003 through December 2014. The forecasters include the chief economists from large
commercial banks and investment banks, heads of forecasting firms, and prominent business
our previous expectation of the April 2014 inflation rate with the actual value. We would be classified as revising
our forecast even if we did not change our expectations of monthly inflation for months from May to December.
5
Pfajfar and Santoro (2013) report a similar finding. The inability of professional forecasters to forecast more
accurately despite updating may support the noisy- information model. Dräger et al. (2016) find that professional
forecasters and consumers who form expectations consistent with economic theory forecast more accurately.
4
economists from industry. The economists submit forecasts of several economic variables in the
first or second week of each month and the WSJ publishes them on-line shortly thereafter.6
Economists’ names and employers appear along with their forecasts, unlike the Livingston
Survey, the (US and European) Surveys of Professional Forecasters and Consensus Economics.
This is important because it permits us to follow individual economists as they change their
employers and ensures that we record forecast revisions only for economists participating in
consecutive surveys.7 Over our sample period the number of economists in each survey ranges
from 45 to 60, averaging about 54. A total of 101 economists appear in our sample.
Features of the WSJ survey make it well-suited for investigating the sticky-information
model. The WSJ asks economists to predict the June 30 and December 31 values for the 10-year
Treasury bond rate and the fed funds rate and the June and December unemployment rate, inter
alia. (Before June 2007, the WSJ requested forecasts of the unemployment rate for May and
November). Since end dates are fixed, economists’ forecast horizons decline over time, from
twelve months to one. The economists can potentially access very recent information before
making their forecasts: current interest rate data are available almost contemporaneously and the
unemployment rate for a given month is announced on the first Friday of the next month,
generally before the new WSJ forecasts are due.8
6
Monthly surveys began in March 2003 but until 2008 no forecasts were collected at the start of January or July.
The WSJ Economic Forecasting Survey web site is: http://projects.wsj.com/econforecast/#ind=gdp&r=20. Prior to
March 2003, the WSJ surveyed economists twice a year. For an analysis of the semi-annual forecasts, see Mitchell
and Pearce (2007).
7
Engelberg, Manski, and Williams (2011, footnote 9) note that id numbers in the Survey of Professional Forecasters
need not identify the same individuals over time. As a referee noted, following institutions might be preferable to
following individuals if forecasts are made by institution-specific models without forecaster adjustments.
8
Survey results posted at the WSJ survey web site contain some apparent errors. In instances where a forecaster’s
prediction is substantially different from the prediction for the same target date in the preceding and succeeding
surveys, we consider the prediction a probable transcription error. For example, one forecaster predicted that on
December 31, 2008 the 10-year bond rate would be 3.88 % in the September survey, 1.27 % in the October survey
and 3.68 % in the November survey. Appendix A lists the probable errors. We omit the questionable data points in
the results reported here, but including them has little effect.
5
Figure 1 plots the surveyed economists’ 4-months-ahead forecasts of the 10-year bond
rate, the fed funds rate, and the unemployment rate by target date.9 Horizontal bars denote actual
rates on the target dates. The plots show that the economists differ in their opinions, often
substantially, as is typical for forecast surveys.10 The sticky-information model explains differing
opinions as differences in the dates on which economists updated their forecasts, an explanation
which assumes economists make full-information, rational predictions whenever they update. In
general, differences in forecasts may reflect differential access to information, differences in
forecasting models, different loss functions, and/or differing prior beliefs (Manzan, 2011).
The economists’ 10-year Treasury bond rate forecasts exhibit roughly the same degree of
dispersion and accuracy throughout the 12-year sample period. The spread of forecasts in a
typical survey is about 150 basis points. In about half the surveys, the forecasts cluster mainly
above or mainly below the actual bond rate on the target date. The economists did not foresee
the plunge in the bond rate to 2.25% in December 2008 from around 4% during most of 2008.
The economists’ fed funds rate forecasts seem to reflect their reading of monetary policy.
Specifically, between March 2003 and mid-December 2008 the Federal Reserve set singlevalued funds rate targets which it moved in multiples of 25 basis points. In this same period the
economists’ forecasts are also in multiples of 25 basis points, with a spread of about 100 basis
points. The spread widened to 250 basis points in the March 2008 survey asking for the funds
rate on June 30, 2008, reflecting greater uncertainty. No one predicted the end-of-year drop in
the funds rate. In December 2008, the Fed simultaneously abandoned the single-valued target
for a target range of 0 to 25 basis points and issued forward guidance indicating it would hold the
9
The 4-month horizon is representative of middle range forecasts. For comparison, we show 10-months-ahead and
2-months-ahead forecasts in Appendix B. We do not show 12-months-ahead and 1-month-ahead forecasts because
they were not collected before 2008.
10
Mankiw and Reis (2003) use differences of opinion in forecast surveys to motivate their sticky-information
model. Dovern et al. (2015) presents a rigorous analysis of differences in opinion across three surveys of
professional forecasters.
6
funds rate within this range for a considerable time (Campbell et al., 2012). These changes may
have reduced the frequency of subsequent forecast revisions. Post-2008 most forecasts are inside
the target range, with a few forecasts above 50 basis points in multiples of 25 basis points.
Forecasts of the unemployment rate resemble forecasts of the bond rate in dispersion and
accuracy. The range of forecasts in a given survey is 1 to 1.5 percentage points. In about half the
surveys, the forecasts are roughly evenly divided between under- and over-predicting the actual
unemployment rate on the target date. Just as the economists failed to foresee declining interest
rates, they failed to foresee climbing unemployment in late 2008 and early 2009.
2.2. Forecast Revisions
We construct a direct measure of economists’ forecast revision behavior to investigate
information rigidity. We presume an economist revising a prior forecast updated his information
set before announcing the revision. Andrade and Le Bihan (2013, p. 973) observe that a
forecaster could update his information and not revise his forecast, thus forecast revision
frequency should be viewed as a lower bound on information updating frequency. Since actual
and predicted interest rates (unemployment rates) are reported to 2 decimal places (1 decimal
place), forecasters would revise their forecasts only if they predicted interest rate (unemployment
rate) changes of at least 1 basis point (10 basis points).
We compute our direct measure of forecast revision behavior as follows. We identify the
number of economists who supplied forecasts on both survey dates t-1 and t and then compute
the fraction of those economists who did not revise their forecasts, a fraction we call Nochanget;
we do this for every survey date. Nochanget is comparable to the proportion of forecasters not
updating, λ, that Coibion and Gorodnichenko (2015) estimate. Unlike λ, Nochanget is a direct
7
measure of the fraction of forecasters who do not update which requires no assumptions about
forecasting method or forecaster rationality.11
Figure 2 displays Nochanget for forecasts of the 10-year bond rate, fed funds rate, and
unemployment rate at each forecast horizon averaged across all surveys. For bond rate forecasts,
Nochanget averages about 0.35 for most horizons but appears lower for the one- and sevenmonth horizons. Economists in the WSJ survey make forecasts for these two horizons at the same
time, since the start of June (December) is one month before June 30 (December 31) and seven
months before December 31 (June 30), the dates for which they predict the bond rate. However,
a formal test that the Nochanget averages are equal across horizons does not reject this
hypothesis.12 Nochanget averages of about 0.35 imply that the economists revised their bond rate
forecasts roughly twice every three months. For unemployment rate forecasts, the economists’
forecast revision behavior resembles their behavior for bond rate forecasts: Nochanget ranged
between 0.30 and 0.45 with no apparent relationship between revision rate and forecast horizon,
implying economists revised their unemployment rate forecasts about twice every three months.
The surveyed economists revised their fed funds rate forecasts less frequently than their
bond rate or unemployment rate forecasts. For the fed funds rate forecasts, Nochanget averages
about 0.65, implying a revision rate of no more than once every three months. If instead of
predicting the actual fed funds rate the economists were predicting the fed funds rate target, this
revision rate suggests they expected a target change at about every other meeting of the Federal
Open Market Committee (FOMC), which meets roughly twice every three months. The
11
Changes in revision frequency could arise from changes in the panel of forecasters. While there is turnover in the
panel, about two-thirds of all revisions come from participants who responded to about eighty percent of the
surveys. See Engelberg Manski, and Williams (2011) for a discussion of how changes in the panel could affect the
usefulness of mean or consensus forecasts.
12
Hotelling T2 tests indicate that the average values of Nochanget are not significantly different for the three
variables, with F(10,1) values of 2.81, .64, and 10.01 for the bond rate, unemployment rate, and fed funds rate,
respectively. Coibion and Gorodnichenko (2015) report that their measure of information rigidity does not appear to
vary across forecast horizon.
8
economists may, in fact, have been updating their information sets more frequently than their fed
funds rate forecasts if new information was insufficient to predict a change in Fed policy.
The behavior of Nochanget both contrasts with and confirms findings of Coibion and
Gorodnichenko (2015). The average revision rates we observe for economists in the WSJ survey
forecasting the bond, fed funds and unemployment rates are greater than the revision rates they
estimate for economists in the Survey of Professional Forecasters forecasting the inflation rate
and other variables. These differences may reflect the difference between monthly and quarterly
surveys.13 Our finding that substantial proportions of forecasters forgo revising their forecasts at
every opportunity supports Coibion and Gorodnichenko’s interpretation of their results as
originating from costly revision rather than noisy information. However, the higher average
revision rates we observe also cast doubt on whether infrequently revised expectations can
account for the persistent effects of shocks at a quarterly frequency.
Figure 3 documents heterogeneity in the forecast revision behavior of the WSJ
economists with histograms of non-revision frequency. Specifically, for each economist having
at least 25 chances to revise a prior forecast of a variable we compute the percentage of forecasts
not revised; we do this for economists’ forecasts of the bond rate, the fed funds rate and the
unemployment rate. We then group the percentages into ten categories (0-10%, 11-20%, etc.)
and compute the percentage of economists in each category. (Economists who revised every
forecast (no forecasts) are in the 0-10% (91-100%) category.)
The economists show considerable heterogeneity in revising forecasts of all three
variables, but especially the bond rate. About one-third of economists revised their bond rate
13
Estimates reported in Figure1, Panel B of Coibion and Gorodnichenko (2015) imply forecast revision times for a
long-term interest rate (the AAA bond rate), a short-term interest rate (the three-month Treasury Bill rate), and the
unemployment rate of about once every 3.6 months, 4.5 months, and 5 months respectively. Andrade and LeBihan
(2013) report that forecasters in the European Survey of Professional Forecasters revise their forecasts about once
every four months. Of course, quarterly data restrict the minimum forecast revision time to once every 3 months.
Mankiw, Reis, and Wolfers (2003) report substantially less frequent revisions, once every 10 to 12 months.
9
forecasts frequently, leaving 20% or less of their forecasts unchanged (Panel A). Nearly half left
from 31% to 60% of their forecasts unchanged, while the remaining one-sixth left 61% to 80%
of their forecasts unchanged. The economists showed generally more reluctance to revise their
fed funds rate forecasts, likely reflecting the timing of FOMC meetings. Over the full sample
period, about sixty percent of economists did not revise between 51% and 70% of their forecasts
(Panel B). They behaved similarly over the 2003-2008 sub-period when the Fed used a singlevalue funds rate target (Panel C). The economists’ behavior is most homogeneous in revising
unemployment rate forecasts (Panel D): their unrevised forecast percentages span one less
category than their interest rate forecasts and the distribution of economists is fairly symmetric.14
The foregoing evidence on forecast revision frequency is consistent with the notion that
the costs of acquiring and processing information prevent forecasters from updating their
forecasts whenever new information becomes available. Heterogeneity in revision behavior
suggests that costs and/or benefits vary across forecasters. This evidence begs the question of
whether forecast revision rate is independent of the size of recent changes in the variable being
forecasted. Coibion and Gorodnichenko (2015) assume in their framework that the revision rate
is not state dependent, although they find evidence that more volatile periods exhibit less
information stickiness.15 We address this question in the next section.
14
We also investigated the role of employer type in forecast revision behavior. We defined ten employer types:
commercial banks, investment banks, investment-advising firms, forecasting and research firms, insurance
companies, other financial institutions (e.g., Fannie Mae), bond-rating firms, academia, professional associations,
and nonfinancial institutions. Using a subsample of economists who responded to at least 25 surveys we computed
the mean frequency of non-revision by employer type. Only economists at “other” financial institutions and bondrating firms have significantly different mean revision rates, revising their forecasts more frequently than economists
at other employer types. They represent only about 5 percent of the WSJ economists, however.
15
Coibion and Gorodnicheko (2015) report evidence that forecasters revise less frequently during the Great
Moderation. They note that “recessions, as periods of increased volatility, should be times when economic agents
update and process information faster than in expansions since the (relative) cost of ignoring macroeconomic shocks
in recession rises.” (page 2674)
10
3. Is the Degree of Information Stickiness State Dependent?
3.1. The Model
Empirically testing the state dependency of forecasters’ forecast revision processes
requires us to model changes in the information set for the economy. While we cannot measure
all incoming information forecasters might access, we can measure one seemingly important
piece of information: the amount of recent change in the variable a forecaster is predicting. In an
efficient Treasury bond market, for example, bond rate changes since the last survey should be a
good measure of new information which embeds itself in the current rate. Analogous arguments
can be made about changes in the funds rate and the unemployment rate. A practical advantage
of representing changes in the information set by recent changes in the variables forecasted is
that actual values of these variables are available to all economists at virtually no cost.
Some extreme examples illustrate the effect of information set changes on forecasts.
Specifically, after the Fed lowered the fed funds rate target by 125 basis points in January 2008
all economists in the February 2008 survey revised their fed funds rate predictions for June 30,
2008 and nearly all revised their predictions for December 31, 2008. Similarly, after seeing the
funds rate target fall by 100 basis points during October 2008 nearly all economists in the
November 2008 survey revised their fed funds rate forecasts for December 31, 2008.
We use the timing of the WSJ survey to define our change variables. While we observe
neither the exact date an economist submits a forecast nor the most recent value of the forecasted
variable he observed prior to submission, we do know that the WSJ assembles its surveys in the
first or second week of each month. This fact leads us to compute the change in the actual bond
rate, fed funds rate, and fed funds rate target from the last business day of the month before the
prior survey to the last business day of the month before the current survey. Analogously, we
11
compute the change in the unemployment rate as the difference in unemployment rates
announced at the starts of the prior and current months.16
Our forecast revision model relates the fraction of economists not revising their forecasts
of a variable (Nochanget) to the absolute change in that variable in the prior month (bond rate,
|∆it-1|; fed funds rate, |∆ffrt-1|; or unemployment rate, |∆Ut-1|) and to the forecast horizon. We
allow the horizon to have a nonlinear effect by including indicator variables for each horizon:
Nochanget = α + β |∆variablet-1| + Σj=sj=S γj Djt + et
(1)
where Djt is a zero-one indicator for forecast horizon of length j. 17 We expect larger values of
|∆it-1|, |∆ffrt-1|, and |∆Ut-1| to cause more economists to revise their forecasts, leading β to be
negative. The signs of the γj are unclear: Figure 2 shows that the unconditional means of
Nochanget may rise or fall as the target date grows more distant but differences in the
unconditional means by horizon are not statistically significant, as noted earlier.
The design of the WSJ survey leads us to estimate equation (1) for two different sets of
forecast horizons. At each survey, participants make shorter horizon (1- to 6-month-ahead) and
longer horizon (7- to 12-month-ahead) forecasts of each variable. For example, the March survey
reports bond rate, fed funds rate and unemployment rate forecasts the economists made at the
start of March for the ends of June and December, four and ten months ahead, respectively.18
New information arriving between the February and March surveys may affect economists’ June
and December forecasts. Given this survey design, we estimate equation (1) separately on data
16
The Bureau of Labor Statistics announces the unemployment rate on the first Friday of a month for the previous
month. Thus for example, we presume that economists submitting March 2010 unemployment rate forecasts for
June 2010 have observed the change in the unemployment rate from January 2010 to February 2010. We use the
announced unemployment rates in the real-time data set from the Federal Reserve Bank of Philadelphia (see
Croushore and Stark, 2001) to insure that survey participants had access to this information, since there are slight
adjustments subsequent to the initial unemployment rate announcements.
17
Since our dependent variable ranges from zero to one, OLS could give misleading results as it does not impose
this restriction. Consequently, we also estimated the models using the quasi-maximum-likelihood estimation method
of Papke and Wooldridge (1996). The results, which are very similar to the OLS results, appear in Appendix C.
18
Before June 2007 the WSJ survey reported economists’ unemployment rate forecasts for May and November.
12
for shorter- and longer-horizon forecasts. In estimates on shorter-horizon data, j = {2,3,4,5,6}
with j=1 being the omitted category; in estimates on longer-horizon data, j = {8,9,10,11,12} with
j=7 being the omitted category.
3.2. Model Estimates for the Full Sample Period
Table 1 reports estimates of equation (1) on data for the 2003-2014 sample period. Initial
estimates produced F-tests favoring constrained versions of equation (1); Table 1 reports these Ftests and estimates of the constrained models. (Unconstrained estimates are available upon
request.) Specifically, F-tests on the equation (1) estimate using shorter-horizon bond rate
forecasts imply that the coefficients of Dj, j={2,…,6}, are all non-zero but jointly equal; an
analogous statement applies to the equation (1) estimate using longer-horizon forecasts. Column
1.1 (1.2) reports the constrained model estimate on shorter-horizon (longer-horizon) forecasts.
The constraint is imposed by replacing Dj, j={2,…,6} with D1, an indicator for a one-month
horizon. (The constraint is imposed analogously in column 1.2). When |∆it-1| =0, the constrained
estimate implies that 38% (44%) of forecasters do not change their prior-month bond rate
forecasts with horizons of 2 to 6 months (8 to 12 months), and 23% (32%) do not change
forecasts with horizons of one month (seven months). When |∆it-1|  0, non-revisions decline
significantly: a two-standard-deviation change in the 10-year bond rate, about 38 basis points,
reduces the percentage of non-revisers by about 11 percentage points at all horizons (-.30 x .38).
Columns 1.3-1.6 report estimates of equation (1) on fed funds rate forecasts with no
horizon effects. (In unconstrained model estimates, F-tests cannot reject the hypothesis that all
γj=0, j={2,…,6} and j={8,…,12}.) We report model estimates for two alternative information
variables: |∆ffrt-1| and |∆ffrtargett-1|, the absolute change in the effective funds rate and the Fed’s
funds rate target, respectively. We use the latter on grounds that economists may consider target
changes as well as actual funds rate changes when forecasting. When |∆ffrt-1| = 0 or |∆ffrtargett-1|
13
= 0 about two-thirds of forecasters do not revise their shorter-run forecasts (columns 1.3 and 1.5)
and about sixty percent do not revise their longer-run forecasts (columns 1.4 and 1.6). Twentyfive-basis-point changes in the actual and target rates reduce Nochanget for shorter-horizon
forecasts by 7 and 13 percentage points, respectively, and reduce longer-horizon forecasts by 7
and 9 percentage points, respectively.19
The last three columns of Table 1 report estimates of equation (1) on unemployment rate
forecasts. (In unconstrained model estimates, F-tests cannot reject the hypothesis that all γj=0,
j={2,…,6}, but can reject the hypotheses that all γj=0 and all γj=, j={8,…,12}; further F-tests
show that γ8= γ9= 0 and γ10= γ11= γ12<0.) Column 1.7 reports an estimate of equation (1) on
shorter-horizon unemployment rate forecasts without horizon indicators. With no change in the
actual unemployment rate from the prior month, about 46% of forecasters leave their shorterhorizon forecasts unrevised. A two-standard-deviation change in the unemployment rate, about
24 basis points, reduces this fraction by about 16 percentage points. A similarly constrained
model estimated on longer-horizon forecasts yields nearly identical results (column 1.8). Adding
the horizon indicator D10+, defined as D10+D11+D12, reveals a small horizon effect (column 1.9).
Specifically with an unchanged unemployment rate, 50% of economists leave their
unemployment rate forecasts unrevised 7 to 9 months before the target date whereas only 42%
leave forecasts unrevised 10 to 12 months before the target date. A two-standard-deviation
change in the unemployment rate reduces both percentages by about 14 percentage points.20
19
When we include both rate changes in the same model, only the funds rate target change has a significant
coefficient in the model estimate using shorter-horizon forecasts while neither rate change has a significant
coefficient in the model estimate using longer-horizon forecasts.
20
As noted earlier, the WSJ economists make six forecasts at each survey – three variables and two horizons. To
study the possibility that an economist makes joint forecasts, we computed correlation coefficients between pairs of
Nochanget measures. All fifteen coefficients are non-negative with the highest correlations between Nochanget
measures of the same variable at shorter and longer horizons (about .85). Less correlated are Nochanget measures of
the bond rate and the funds rate (.35-.50). Coefficients are smaller for the remaining pairs of measures; half are
statistically insignificant. To accommodate possibly joint forecasts, we estimated the models reported in Table 1 by
seemingly unrelated regressions. Since SUR estimation requires balanced panels, we lose some observations. These
14
In summary, the evidence in Table 1 reveals three patterns. First, changes in the variables
economists forecast reduce the percentages of unrevised forecasts, consistent with state
dependency of forecast revisions. Second, recent changes in the variables economists forecast do
not push the percentages of unrevised forecasts to zero, consistent with the sticky information
model. Third, forecast horizon has little measurable effect on forecast revision frequency.
3.3. Did Forecaster Behavior Change After the Financial Crisis?
We put the hypothesis of state-dependent forecast revisions to a stronger test by
exploiting the presence in our sample period of both the end of the Great Moderation and the
2007-2009 financial crisis and its aftermath. Prior academic research shows that volatility in
many economic variables increased starting in 2007 (Clark, 2009; Stock and Watson, 2012). The
unanticipated bankruptcy of Lehman Brothers in September 2008 radically changed perceptions
about “too big to fail,” the reliability of government interventions into financial markets, and
financial market fragility. Andrade and LeBihan (2013) find greater dispersion in forecasts of
professional European economists after 2007 and Dovern (2013) reports higher probabilities that
international forecasters revised their forecasts during recessions.
With greater economic
uncertainty post-2008, we expect that forecasters revised their forecasts more frequently
following changes in the information set. We test this theory by comparing estimates of
constrained versions of equation (1) produced by forecasts from 2003-2007 and from 2008-2014.
Table 2 reports model estimates from bond rate and unemployment rate forecasts but not from
fed funds rate forecasts, since the Fed’s funds rate target remained unchanged after December
2008.21
estimates are reported in Appendix D, Tables D1 and D2. Breusch-Pagan tests indicate contemporaneously
correlated residuals. Nevertheless, the SUR estimates are qualitatively very similar to the OLS estimates reported in
Table 1.
21
Model estimates on data from before the December 2008 decision are very similar to those reported in Table 2 for
the whole period and are reported in Appendix D, Table D3.
15
Model estimates from bond rate forecasts show that post-2007, fewer economists revised
prior forecasts with an unchanged bond rate and more revised with a changed rate (Table 2,
Panel A). F-tests show significant differences in the model estimates for the two sub-periods.
For shorter-horizon forecasts pre-2008, an unchanged bond rate yields a Nochanget estimate of
about 33% for 2- to 6-month forecast horizons and about 7% for 1-month horizons; post-2007
the estimates are 40% and 25%. Pre-2008, a 38-basis-point (two-standard-deviation) bond rate
change from the prior month reduces Nochanget by about 5 percentage points; post-2007 the
reduction is about 14 percentage points. Longer-horizon forecasts show a similar pattern.
Model estimates from unemployment rate forecasts show that forecasters were more
sensitive to unemployment rate changes after 2007 (Table 2, Panel B). Pre-2008, an
unemployment rate change the month before a survey has no significant effect on Nochanget for
either shorter- or longer-horizon forecasts; post-2007, a 24-basis-point (two-standard-deviation)
change reduces Nochanget by about 15 percentage points for both forecast horizons. With no
change in the unemployment rate the month before a survey, the estimate of Nochanget is
roughly the same for both sub-periods and both forecast horizons: between 42% and 49%.22
4. Extensions
4.1. Estimates of the Coibion-Gorodnichenko Model
In this section we compare our direct method of testing the sticky-information model
using estimates of equation (1) with the indirect method developed by Coibion and
Gorodnichenko (2015), hereafter CG. CG develop a model to infer the average forecast revision
frequency for a sample of forecasters whose individual forecasts are unobserved. CG assume that
forecasters form full information, rational expectations predictions whenever they forecast but
22
The estimate of Nochanget is about 35% for forecast horizons of ten or more months in the 2003-2007 sub-period.
16
that frictions prevent continuous updating. CG also assume the probability of an individual
forecaster revising a forecast on a given date is (1-λ), making the average time between revisions
[1/(1-λ)]. CG derive a relationship between the average forecast error and the change in the
average forecast:
xt+h – Ft xt+h = [λ/(1- λ)] (Ft xt+h – Ft-1 xt+h) + vt+h,t
(2)
where xt+h is the actual value of the variable forecasted h periods ahead and Ft xt+h is the average
forecast at time t across all forecasters in the survey. CG note that an estimate of equation (2) on
aggregate data yields an estimate of [λ/(1-λ)], from which the average time between revisions
may be inferred. Support for the sticky-information model comes from evidence that λ>0 so that
[1/(1-λ)]>1, the average time between forecast revisions exceeds the time between surveys. Since
the probability of non-revision, λ, is analogous to our directly observed measure of non-revision,
Nochanget, we compare values of the two measures produced by forecasts from the WSJ survey.
Table 3 reports our results. Panel A shows estimates of equation (2) on shorter- and
longer-horizon forecasts of the bond rate, the funds rate and the unemployment rate. Statistically
insignificant estimates of β=λ/(1-λ) in the model estimates using bond rate forecasts and shorthorizon funds rate forecasts imply λ estimates of zero (columns 3.1a – 3.3a). Conversely,
statistically significant estimates of β in the estimates using longer-horizon funds rate forecasts
and unemployment rate forecasts imply λ estimates exceeding zero (columns 3.4a – 3.6a).
Panel B reports average numbers of months between forecast revisions produced by the
two methods. Values reported for the CG model are inferred from the β estimates in Panel A;
values reported for the Nochanget model are computed from the average values of Nochanget at
each forecast horizon, displayed as Figure 2.23 The direct and indirect methodologies produce
23
In the CG model, β = [λ/(1- λ)], λ = β/(1+β). The average time between revisions, 1/(1-λ), is 1+ β when the β
estimate is statistically significant, and one otherwise. Nochanget is conceptually similar to λ. The average number
of months between forecast revision is 1/(1- avg Nochange) where avg Nochange is the value of Nochange t for a
17
different estimates of forecast revision frequency. For the bond rate, the CG model estimates
imply monthly revision of forecasts whereas the Nochanget averages imply revisions closer to
every month and a half (columns 3.1b and 3.2b). For the fed funds rate, the CG model estimates
imply revisions once a month for shorter-horizon forecasts and once every two and two-thirds
months for longer-horizon forecasts; the Nochanget averages imply revisions closer to once
every three months for both forecast horizons. For the unemployment rate, the CG model
estimates imply revisions about every one and three-quarter months for shorter-horizon forecasts
and about every four months for longer-horizon forecasts; the Nochanget averages imply
revisions about every one and two-thirds months for both horizons. In summary, although the
indirect and direct estimates of average revision frequency are similar in magnitude, the former
can understate or overstate the observed degree of information rigidity.
4.2. Are Recently Revised Forecasts More Accurate?
The heterogeneity in forecast revision behavior documented in Figure 3 may reflect
differential rewards to forecast accuracy, leading us to investigate whether recently revised
forecasts are more accurate than unrevised forecasts. To test this hypothesis, we compute the
squared forecast error for each economist for every target date and horizon and then regress the
squared forecast errors on a binary indicator variable coded one if the economist’s forecast is
unchanged from the prior survey. Only forecasters who responded to both the current and
previous surveys are included. Each regression has a sample size of about 50, roughly the
average number of economists per survey in our 12-year sample period. Table 4 reports the
given variable 1 to 6 months or 7 to 12 months before the target date averaged over all of the surveys in our sample
period.
18
outcome of this experiment by reporting the percent of surveys in which revised forecasts are
significantly more or less accurate than unrevised forecasts by variable and forecast horizon.24
Revised forecasts are often more accurate than unrevised forecasts but not consistently
so. Differences in accuracy are greatest for the 10-year bond rate. At a one-month forecast
horizon, revised forecasts are significantly more accurate than unrevised forecasts in 64% of the
surveys and are significantly less accurate in 0% of the surveys; analogous metrics at a twomonth horizon are 48% and 4%. At the remaining horizons, revised forecasts are significantly
more (less) accurate than unrevised forecasts in at most 30% (5%) of the surveys. Declining
forecast accuracy at longer horizons is consistent with increasingly noisy information about the
target date. Differences in the accuracy of revised and unrevised unemployment rate forecasts
are smaller. At a one-month horizon, revised unemployment rate forecasts are significantly more
accurate than unrevised forecasts in just 28 % of the surveys and are significantly less accurate in
9% of the surveys. At the eleven other forecast horizons, revised forecasts are significantly more
(less) accurate than unrevised forecasts in at most 20% (13%) of the surveys. Revised fed funds
rate forecasts are significantly more accurate than unrevised forecasts only slightly more often
than they are less accurate.
For the majority of surveys, however, revised and unrevised
forecasts of the three variables are statistically indistinguishable in accuracy.
5. Conclusions
The sticky-information model predicts that forecasters will not revise their forecasts
when new information arrives if the costs exceed the benefits. This paper contributes to the
evidence on sticky information in several ways. First, we test the model using data from the WSJ
24
Revisers are significantly more (less) accurate than non-revisers in a survey if the coefficient estimate of the
indicator variable is statistically significant at the 10%-level or better and positive (negative). The coefficient
estimates appear in Appendix E.
19
Economic Forecasting Survey, which publishes the names and forecasts of professional
forecasters. From these data we can see precisely when forecasters revise their forecasts and
measure rates of forecast revision without making assumptions about forecast rationality, as
researchers must do when testing the sticky-information model using datasets without
individuals’ forecasts. Additionally, the WSJ Survey is monthly, permitting a higher frequency
investigation than prior research using quarterly, semi-annual or infrequent surveys. The paper
is, to our knowledge, the first to use the WSJ Survey to evaluate the sticky-information model.
Second, we investigate the state dependency of forecast revision frequency by testing whether
frequency changes after an increase in the volatility of the variables forecasted. Third, we
compare direct estimates of forecast revision frequency with indirectly inferred estimates
produced using a technique from the literature. Finally, we examine whether forecast revision
improves forecast accuracy.
Our results both support the sticky-information model and cast doubt on the model’s
adequacy as an explanation for the persistence of macro-economic shocks. While we find that
many forecasters revise their forecasts only every other month or less frequently, we also find
that forecasters revise their estimates somewhat more frequently than other researchers have
found. Given that our measure of forecast revision frequency is likely a lower bound to the
frequency of information updating, our results suggest that frictions other than the costs of
acquiring and processing information likely play a role in the responses to economic shocks.
Forecasters in the WSJ Survey revise their forecasts of the fed funds rate less frequently than
forecasts of the 10-year U.S. Treasury bond rate or the unemployment rate, perhaps due to the
timing of FOMC meetings. Forecast horizon appears to exert little influence on the frequency of
forecast revision. Forecasters exhibit considerable heterogeneity in their revision frequencies,
consistent with substantial variation in the costs and benefits of revising across forecasters. We
20
find evidence that forecast behavior is state dependent, with forecasters revising their forecasts
more frequently in more volatile times. Our direct measures of revision frequency are similar in
magnitude to those estimated indirectly, but the latter can understate or overstate the observed
degree of information rigidity. Finally, we find only weak evidence that revising forecasts
improves forecast accuracy, particularly at longer horizons.
21
References
Andrade, P. and H. Le Bihan. “Inattentive Professional Forecasters.” Journal of Monetary
Economics, 60(8), 2013, 967-82.
Armantier, O., S. Nelson, G. Topa, W. Van der Klaauw and B. Zafar. “The Price is Right:
Updating Inflation Expectations in a Randomized Price Information Experiment.” Review of
Economics and Statistics, 98(3), 2016, 503-523.
Campbell, J.R., C.L. Evans, J.D. Fisher, A. Justiniano, C.W. Calomiris and M. Woodford.
“Macroeconomic Effects of Federal Reserve Forward Guidance” [with comments and
discussion]. Brookings Papers on Economic Activity, Spring 2012, 1–80.
Carroll, C. “Macroeconomic Expectations of Households and Professional Forecasters.”
Quarterly Journal of Economics, 118(1), 2003, 269-298.
Clark, T.E. “Is the Great Moderation Over? An Empirical Analysis.” Federal Reserve Bank of
Kansas City Economic Review, Fourth Quarter 2009, 5-42.
Coibion, O. “Comments on Dovern, Fritsche, Loungani and Tamarisa.” International Journal of
Forecasting, 31(1), 2015, 155-6.
Coibion, O. and Y. Gorodnichenko. “Information Rigidity and the Expectations Formation
Process: A Simple Framework and New Facts.” American Economic Review, 105(8), 2015,
2644-78.
Croushore, D. and T. Stark. "A Real-Time Data Set for Macroeconomists." Journal of
Econometrics, 105(1), 2001, 111-30.
Dovern, J. “When are GDP forecasts updated? Evidence from a large international panel.”
Economics Letters, 120(3), 2013, 521-524.
Dovern, J. “A Multivariate Analysis of Forecast Disgreement: Confronting Models of
Disagreement with Survey Data.” European Economic Review, 80(1), 2015, 16-35.
Dovern, J, U. Fritsche, P. Loungani and N. Tamirisa. “Information Rigidities: Comparing
Average and Individual Forecasts for a Large International Panel.” International Journal of
Forecasting, 31(1), 2015, 144-54.
Dräger, L., M.J. Lamla and D. Pfajfar. “Are Survey Expectations Theory-Consistent? The Role
of Central Bank Communication and News.” European Economic Review, 85(1), 2016, 84-111.
Engelberg, J., C.F. Manski and J. Williams. “Assessing the Temporal Variation of
Macroeconomic Forecasts by a Panel of Changing Composition.” Journal of Applied
Econometrics, 26(7), 2011, 1059-78.
22
Mankiw, N. and R. Reis. “Sticky Information Versus Sticky Prices: A Proposal to Replace the
New Keynesian Phillips Curve.” Quarterly Journal of Economics, 117(4), 2002, 1295-328.
Mankiw, N., R. Reis and J. Wolfers. “Disagreement about Inflation Expectations.” NBER
Macroeconomic Annual, 18, 2003, 209-48.
Manzan, S. “Differential Interpretation in the Survey of Professional Forecasters.” Journal of
Money, Credit and Banking, 43(5), 2011, 993-1017.
Mertens, E. and J.M. Nason. “Inflation and Professional Forecast Dynamic: An Evaluation of
Stickiness, Persistence, and Volatility.” Working Paper 06/2015, Centre for Applied
Macroeconomic Analysis, Crawford School of Public Policy, Australian National University,
2015.
Mitchell, K. and D. Pearce. “Professional Forecasts of Interest Rates and Exchange Rates:
Evidence from the Wall Street Journal's Panel of Economists.” Journal of Macroeconomics,
29(4), 2007, 840-54.
Papke, L.E. and J.M. Wooldridge. “Econometric Methods for Fractional Response Variables
with an Application to 401(k) Plan Participation Rates.” Journal of Applied Econometrics, 11(6),
1996, 619-32.
Patton, A.J. and A. Timmermann. “Why Do Forecasters Disagree? Lessons from the Term
Structure of Cross-Sectional Disagreement.” Journal of Monetary Economics, 57(7), 2010, 80320.
Pfajfar, D. and E. Santoro. “News on Inflation and the Epidemiology of Inflation Expectations.”
Journal of Money, Credit and Banking, 45(6), 2013, 1045-67.
Reis, R. “Inattentive Producers.” Review of Economic Studies, 73(3), 2006, 793-821.
Sims, C.A. “Implications of Rational Inattention.” Journal of Monetary Economics, 50(3), 2003,
665-90.
Stock, J.H. and M.W. Watson. “Disentangling the Channels of the 2007-2009 Recession.”
Brookings Papers on Economic Activity, Spring 2012, 81-135.
Woodford, M. “Imperfect Common Knowledge and the Effects of Monetary Policy.” In
Knowledge, Information, and Expectations in Modern Macroeconomics: In Honor of Edmund
Phelp, edited by P. Aghion, R. Frydman., J. Stiglitz, and M. Woodford, Princeton University
Press, 2003, 25-58.
23
Table 1
Forecast Revision and Recent Changes in Forecasted Variable
General Model: Nochanget = α + β |Δxt-1| + Σ γj Djt + et
Forecast Revisions of:
Horizon Length:
Column:
10-year Bond Rate
1-6
(1.1)
7 - 12
(1.2)
-.298**
(.059)
-.150**
(.028)
-.297**
(.055)
Fed Funds Rate
Effective
Target
1-6
7 - 12
1-6
7 - 12
(1.3)
(1.4)
(1.5)
(1.6)
Unemployment Rate
1-6
(1.7)
7 - 12
(1.8)
7 - 12
(1.9)
-.655**
(.089)
-.629**
(.083)
-.597**
(.080)
Explanatory Variable:
|Δxt-1|
D1
-.282**
(.053)
-.260**
(.044)
-.536**
(.047)
-.351**
(.071)
-.117**
(.023)
D7
-.079**
(.023)
D10+
Constant
F tests:
all γj=0
all γj=
R2
Sample size
.379**
(.016)
.442**
(.023)
.681**
(.017)
.614**
(.016)
.678**
(.016)
.605**
(.016)
4.65**
0.70
5.20**
0.36
.44
.43
.74
.36
.88
.87
.72
.51
.333
121
.279
121
.117
126
.088
125
.274
126
.125
125
.463**
(.018)
.471**
(.018)
1.79
2.20
2.61*
3.21*
.221
131
.272
119
.500**
(.019)
.337
119
The table reports OLS estimates of the forecast revision model shown for the 2003-2014 sample period. Nochange is the fraction of forecasters in the current WSJ
survey with forecasts unchanged from the prior survey. Forecasts are of the 10-year Treasury bond rate, the fed funds rate and the unemployment rate on a target
date (30 June or 31 December). Separate estimates are reported for surveys 1-6 months and 7-12 months before the target date. |Δxt-1| is the absolute change in x
from the last business day of the month before the prior survey to the last business day of the month before the current survey; x is the bond rate, the effective fed
funds rate, the target fed funds rate and the unemployment rate in columns (1.1)-(1.2), (1.3)-(1.4), (1.5)-(1.6), and (1.7)-(1.9), respectively. Dj = 1 if j is the number
of months until the forecast target date (30 June or 31 December) and 0 otherwise. D10+ =1 if the number of months until the forecast target date is 10 or more and
0 otherwise. Robust standard errors appear in parentheses. The F-tests are from unconstrained estimates of the models in which the full set of horizon indicators
appear (Dj, j={2,...,6} or j={8,…,12}). ** and * denote statistical significance at the 05. and .10 levels, respectively.
24
Table 2
Constancy of Forecaster Revision Behavior, 2003-2007 versus 2008-2014
Panel A: 10-year Bond Rate Forecast Revisions: Nochanget = β0 + β1 |∆it-1| + β2 Djt + et
Horizon Length:
Sample Period:
Column:
Explanatory Variable:
|∆it-1|
D1
1-6 months
2003-07
2008-14
(2.1a)
(2.2a)
-.140*
(.077)
-.256***
(.030)
-.361***
(.088)
-.146***
(.031)
D7
Constant
.328**
(.023)
F tests across time:
β0, β1, β2 =
R2
Sample size
.400**
(.021)
7-12 months
2003-07
2008-14
(2.3a)
(2.4a)
-.158**
(.064)
-.383***
(.088)
-.111***
(.035)
.375***
(.024)
-.103***
(.033)
.471***
(.021)
5.69***
.235
37
.389
84
3.75**
.193
45
.341
76
Panel B: Unemployment Rate Forecasts Revisions: Nochanget = β0 + β1 |∆Ut-1| + β2 Djt + et
Horizon Length:
Sample Period:
Column:
Explanatory Variable:
|∆Ut-1|
1-6 months
2003-07
2008-14
(2.1b)
(2.2b)
-.114
(.301)
-.639***
(.094)
.468***
(.035)
.428***
(.024)
D10+
Constant
F tests across time:
β0, β1, β2 =
R2
Sample size
7-12 months
2003-07
2008-14
(2.3b)
(2.4b)
.052
(.240)
-.136***
(.034)
.489***
(.028)
8.27***
.003
49
.272
82
-.617***
(.088)
-.047
(.030)
.470***
(.028)
4.74***
.261
41
.350
78
The table reports OLS estimates of the forecast revision models shown. Forecasts are of the 10-year
Treasury bond rate or the unemployment rate on 30 June or 31 December. Separate estimates are
reported for surveys 1-6 and 7-12 months before the target date. In Panel A, Nochange is the fraction of
forecasters in the current WSJ survey with bond rate forecasts unchanged from the prior survey. |∆i t-1| is
the absolute change in the bond rate from the last business day of the month before the prior survey to
the last business day of the month before the current survey. D1 =1 (D7 =1) if the number of months until
the forecast target date is 1 (7) and 0 otherwise. In Panel B, Nochange is the fraction of forecasters in
the current survey with unemployment rate forecasts unchanged from the prior survey. |∆U t-1| is the
absolute change in the unemployment rate from the last business day of the month before the prior
survey to the last business day of the month before the current survey. D10+ =1 if the number of months
until the forecast target date is 10 or more and 0 otherwise. In both panels robust standard errors appear
in parentheses. ***, ** and * denote statistical significance at the .01, .05 and .10 levels, respectively.
25
Table 3
The Coibion-Gorodnichenko Model
Panel A: CG Model estimates:
xt+h – Ft xt+h = β [Ft xt+h – Ft-1 xt+h] + εt+h
Average Forecast Errors of:
Horizon Length:
Column:
Explanatory Variable:
10-year Bond Rate
1-6
7-12
(3.1a)
(3.2a)
Fed Funds Rate
1-6
7-12
(3.3a)
(3.4a)
[Ft xt+h – Ft-1 xt+h]
.238
.216
.100
.286
.517
.340
F tests, horizon effects:
R2
Sample size
0.60
.007
130
0.57
.001
108
1.07
.031
135
1.650***
.356
1.35
.171
112
Unemployment Rate
1-6
7-12
(3.5a)
(3.6a)
.783**
.355
1.26
.080
135
3.081***
.559
0.60
.304
112
Panel B: Average Number of Months between Forecast Revisions: Indirect and Direct Estimates
Forecast Revisions of:
Horizon Length:
Column:
Average number of months
between forecast revisions:
10-year Bond Rate
1-6
7-12
(3.1b)
(3.2b)
Fed Funds Rate
1-6
7-12
(3.3b)
(3.4b)
Unemployment Rate
1-6
7-12
(3.5b)
(3.6b)
CG Model
1.00
1.00
1.00
2.65
1.78
4.08
Nochanget
1.41
1.54
3.35
2.75
1.61
1.62
Panel A reports estimates of the sticky-information model of Coibion and Gorodnichenko (2015) using the WSJ forecasts of
the 10-year Treasury bond rate, the fed funds rate and the unemployment rate on 30 June or 31 December. xt+h is the actual
value of the variable forecasted h periods ahead and Ft xt+h is the average forecast across all forecasters at time t. xt and Ft xt+h
refer to the bond rate, the fed funds rate and the unemployment rate in columns (3.1a)-(3.2a), (3.3a)-(3.4a), and (3.5a)-(3.6a),
respectively. Separate estimates are reported for surveys 1-6 months and 7-12 months before the target date. *** and **
denote statistical significance at the .01 and .05 levels, respectively. F tests are for unreported model estimates which include
dummy variables permitting different intercepts and slope coefficients by forecast horizon; the F tests are for the hypothesis
that these coefficients are jointly zero. Panel B compares the average number of months between forecast revisions from the
CG model estimates in Panel A and the direct measures of forecast revision plotted in Figure 2. In the CG model the average
number of months between forecast revisions is 1/(1-λ) = 1+β if β is statistically significant, and zero otherwise. Nochanget is
conceptually similar to λ. The average number of months between forecast revisions is 1/(1- avg Nochange) where avg
Nochange is the average value of Nochange for a given variable 1 to 6 months or 7 to 12 months before the target date
averaged over all of the surveys in our sample period.
26
Table 4
Forecast Accuracy of Revised versus Unrevised Forecasts, by Variable Forecasted and Forecast Horizon
Horizon, in months:
1
2
3
4
5
6
7
8
9
10
11
12
10-year bond rate:
Revisers more accurate
Revisers less accurate
Number of surveys (8/2003 – 12/2014)
64
0
14
48
4
23
30
9
23
30
0
23
26
4
23
21
0
14
30
4
22
14
5
22
18
5
22
23
0
22
14
0
22
9
0
11
Fed funds rate:
Revisers more accurate
Revisers less accurate
Number of surveys (1/2003 – 6/2008)
0
50
2
23
18
11
18
18
11
9
18
11
20
0
10
0
0
1
20
0
10
20
0
10
10
10
10
10
20
10
10
10
10
Unemployment rate:
Revisers more accurate
Revisers less accurate
Number of surveys (1/2003 – 12/2014)
28
9
22
13
0
23
13
4
24
8
13
24
20
7
15
9
5
22
9
4
23
13
9
23
0
9
23
13
9
23
14
0
14
Variable forecasted:
% of surveys:
15
8
13
This table summarizes statistically significant differences between the mean squared forecast errors of forecasters who did and did not revise their
forecasts from the previous survey. For each target date and forecast horizon we first compute the squared forecast error of every economist and
then regress the squared forecast errors on a binary indicator variable coded one if the economist’s forecast is unchanged from the prior survey. Each
regression has a sample size of about 50. (1-, 5-, 6-, 11- and 12-months-ahead forecasts of some variables are unavailable because the WSJ did not
consistently request them. For the fed funds rate, comparisons of forecast accuracy stop after mid-2008 when the Federal Reserve pegged the funds
rate target.) Revised forecasts are more (less) accurate if the estimated coefficient of the indicator variable is statistically significant at the 10% level
or better and positive (negative). The estimated coefficients are reported in Appendix E.
27
Figure 1 4-Months-Ahead Forecasts of the Bond Rate, Fed Funds Rate and Unemployment Rate
Panel A: 10-year bond rate
7
6
Percent
5
4
3
2
1
0
4-months-ahead forecasts
Actual 10-year bond rate
Panel B: Fed funds rate
7
6
Percent
5
4
3
2
1
0
4-months-ahead forecasts
Actual fed funds rate
28
Figure 1 -- continued
Panel C: Unemployment rate
10.5
9.5
Percent
8.5
7.5
6.5
5.5
4.5
3.5
4-months-ahead forecasts
Actual unemployment rate
29
Figure 2 Nochanget, the Fraction of Forecasters Not Revision Forecasts, by Forecast Horizon
0.80
0.70
0.60
Nochange
0.50
0.40
0.30
0.20
0.10
0.00
0
1
2
3
10-year bond rate
4
5
6
7
Forecast Horizon
Fed funds rate
8
9
10
11
12
Unemployment rate
30
Figure 3 Distribution of Forecasters by Percent of Unrevised Forecasts
Percent of forecasters
Panel A: 10-year bond rate forecasts, 2003 - 2014
40
35
30
25
20
15
10
5
0
Percent of unrevised forecasts
Panel C: Fed funds rate forecasts, 2003 - 2008
Percent of forecasters
40
35
30
25
20
15
10
5
0
40
35
30
25
20
15
10
5
0
Percent of unrevised forecasts
Percent of unrevised forecasts
Panel D: Unemployment rate forecasts, 2003 - 2014
Percent of forecasters
Percet of forecasters
Panel B: Fed funds rate forecasts, 2003 - 2014
40
35
30
25
20
15
10
5
0
Percent of unrevised forecasts
31
Appendix A
Questionable Entries in the WSJ Survey Data
1. 10-year Bond rate Forecasts
Survey
June 2008
Target Date
June 2008
Questionable data
Prakken & Varvares forecast is 1.65 with previous
forecast of 3.55
Correction: omitted
June 2008
Dec 2008
Prakken & Varvares forecast is 2.13 with previous
forecast of 4.1 and subsequent forecast of 4.2
Correction: omitted
Oct 2008
Dec 2008
Prakken & Varvares forecast is 1.27 with previous
forecast of 3.88 and subsequent forecast of 3.68
Correction: omitted
Nov 2008
Dec 2008
Sterne forecast is .9 with previous forecast of 3.70 and
subsequent forecast of 3.00
Correction: omitted
Oct 2008
June 2009
Prakken & Varvares forecast is 1.13 with previous
forecast of 4.35 and subsequent forecast of 3.36
Correction: omitted
Nov 2008
June 2009
Sterne forecast is 1.3 with previous forecast of 3.70 and
subsequent forecast of 3.50
Correction: omitted
Dec 2008
June 2009
Wilson forecast is 1.65 with previous forecast of 2.89 and
subsequent forecast of 2.80
Correction: omitted
April 2009
June&Dec2009
Wyss forecasts recorded as .028 and .03 with
previous being 2.9 and 3.1 and subsequent 3.2 and 3.5
Correction: changed to 2.8 and 3.0
July 2012
June 2013
Leamer/Shulman reported as 25 with before and after of 2.2
and 2.5
Correction: changed to 2.5
32
Appendix A -- continued
2. Fed funds rate forecasts
Survey
Feb 2003
Target Date
June 2003
Questionable data
Shilling forecast is .05 with previous forecast of .75 and
subsequent forecast of .5
Correction: omitted
Sept 2009
Dec 2009
Johnson forecasts are recorded as -.125 instead of .125
Correction: corrected to .125
June 2011
Dec 2011
June 2011
June 2012
Maki forecast is .0125 with previous forecast of .125 and
subsequent forecast of .125
Correction: corrected to .125
Maki forecast is .0125 with previous forecast of .125 and
subsequent forecast of .125
Correction: corrected to .125
July 2012
all
Several cases of forecasts recorded as .0125 with previous
and subsequent forecasts of .125
Forecasters for whom reported forecasts for Dec 2012 and
June 2013 are .0125 when for the survey before and after
the forecasts are .125 are Behravesh, Carey, Coronado,
Fiorini-Ramirez, Ethan Harris, Maury Harris, Maki,
Prakken/Varvanes, Resler, and Soss. These were changed
to .125. Daane also reported to forecast .0125 but this is
repeated in subsequent surveys.
3. Unemployment rate forecasts
Survey
Feb 2006
Target Date
May 2006
Questionable data
Swonk forecast of 3.4 with previous forecast of 5.0 and
subsequent forecast of 4.8
Correction: omitted
Feb 2006
Nov 2006
Swonk forecast of 3.6 with previous forecast of 5.0 and
subsequent forecast of 4.7
Correction: omitted
August 2006 Nov 2006
Duncan forecast of 2.8 with previous forecast of 4.8 and
subsequent forecast of 4.8
Correction: corrected to 4.8
33
Appendix A -- continued
3. Unemployment rate forecasts -- continued
Survey
May 2008
Target Date
June 2008
Questionable data
Sterne forecast is 2.9 with previous forecast of 5.2
Correction: omitted
May 2008
Dec 2008
Sterne forecast is 2.4 with previous forecast of 5.0 and
subsequent forecast of 5.1
Correction: omitted
Dec 2008
Dec 2008
Brinkman forecast of 8.3 with previous forecast of 6.9
Correction: left in since subsequent surveys are also high
Feb 2009
June 2009
Meil forecast is 5.8 with previous forecast of 8.3 and
subsequent forecast of 9.0
Correction: omitted
Nov 2013
Dec 2013
Handler forecast is 1.7 and probably should be 7.1
Correction: corrected to 7.1
34
Appendix B
2- and 10-Months-Ahead Forecasts
Panel A: 2-month-ahead bond rate forecasts
7
6
5
4
3
2
1
0
Panel B: 10-month-ahead bond rate forecasts
7
6
5
4
3
2
1
0
35
Appendix B -- continued
Panel C: 2-month-ahead Federal funds rate forecasts
7
6
5
4
3
2
1
0
Panel D: 10-month-ahead Federal funds rate forecasts
7
6
5
4
3
2
1
0
36
Appendix B -- continued
Panel E: 2-month-ahead unemployment rate forecasts
10.5
9.5
8.5
7.5
6.5
5.5
4.5
3.5
Panel F: 10-month-ahead unemployment rate forecasts
10.5
9.5
8.5
7.5
6.5
5.5
4.5
3.5
37
Appendix C
Papke-Wooldridge Estimates of Forecast Revision and Recent Changes in Forecasted Variable
General Model: Nochanget = α + β |Δxt-1| + Σ γj Djt + et
Forecast Revisions of:
Horizon Length:
Column:
10-year Bond Rate
1-6
(C.1)
7 - 12
(C.2)
-1.750***
(.365)
[-.358]
-.935***
(.222)
[-.160]
-1.494***
(.293)
[-.339]
Fed Funds Rate
Effective
Target
1-6
7 - 12
1-6
7 - 12
(C.3)
(C.4)
(C.5)
(C.6)
Unemployment Rate
1-6
(C.7)
7 - 12
(C.8)
7 - 12
(C.9)
-3.036***
(.469)
[-.714]
-2.872***
(.422)
[-.680]
-2.743***
(.404)
[-.650]
Explanatory Variable:
|Δxt-1|
D1
-1.243***
(.301)
[-.284]
-1.144***
(.263)
[-.278]
-2.546***
(.398)
[-.583]
-1.510***
(.386)
[-.367]
-.566***
(.118)
[-.120]
D7
-.344***
(.102)
[-.081]
D10+
Constant
F tests:
all Dj=0
all Dj=
Sample size
-.424***
(.076)
20.49**
3.94
121
-.212**
(.072)
24.55**
1.44
121
.758***
(.076)
.471***
(.069)
.751***
(.073)
.428**
(.066)
-.125
(.078)
2.29
1.84
126
3.88
1.54
125
4.93
3.97
126
3.78
2.19
125
9.38
9.20
131
-.095
(.075)
13.27**
13.14**
119
.029**
(.080)
119
The table reports Papke-Woolridge estimates of the forecast revision model shown for the 2003-2014 sample period. Nochange is the fraction of forecasters in the
current WSJ survey with forecasts unchanged from the prior survey. Forecasts are of the 10-year Treasury bond rate, the fed funds rate and the unemployment rate
on a target date (30 June or 31 December). Separate estimates are reported for surveys 1-6 months and 7-12 months before the target date. |Δxt-1| is the absolute
change in x from the last business day of the month before the prior survey to the last business day of the month before the current survey; x is the bond rate, the
effective fed funds rate, the target fed funds rate and the unemployment rate in columns (C.1)-(C.2), (C.3)-(C.4), (C.5)-(C.6), and (C.7)-(C.9), respectively. Dj = 1 if
j is the number of months until the forecast target date (30 June or 31 December) and 0 otherwise. D10+ =1 if the number of months until the forecast target date is
10 or more and 0 otherwise. Robust standard errors appear in parentheses. Numbers in brackets are the marginal effects of the variable, comparable to the OLS
estimates in Table 1. *** and **denote statistical significance at the .01 and .05 levels.
38
Appendix D
Table D1
SUR Estimates of Forecast Revision and Recent Changes in Forecasted Variable
General Model: Nochanget = α + β |Δxt-1| + Σ γj Djt + et
Forecast Revisions of:
Horizon Length:
Column:
10-year Bond Rate
Fed Funds Rate
Unemployment Rate
1-6
(D1.1)
7 - 12
(D1.2)
Effective
1-6
7 - 12
(D1.3)
(D1.4)
1-6
(D1.5)
7 - 12
(D1.6)
-.209**
(.054)
-.164**
(.032)
-.216**
(.056)
-.359**
(.067)
-.671**
(.119)
-.624**
(.096)
Explanatory Variable:
|Δxt-1|
D1
-.317**
(.063)
-.109**
(.033)
D7
-.035**
(.014)
D10+
Constant
R2
Sample size
.365**
(.016)
.421**
(.016)
.752**
(.018)
.684**
(.018)
.457**
(.020)
.487**
(.018)
.310
111
.232
111
.194
111
.163
111
.217
111
.267
111
The table reports SUR estimates of the forecast revision model shown for the 2003-2014 sample period. Nochange is the fraction of forecasters in the current
WSJ survey with forecasts unchanged from the prior survey. |Δxt-1| is the absolute change in x from the last business day of the month before the prior survey to
the last business day of the month before the current survey; x is the bond rate, the effective fed funds rate, and the unemployment rate in columns (D1.1)-(D1.2),
(D1.3)-(D1.4) and (D.5)-(D.6), respectively. Dj = 1 if j is the number of months until the forecast target date (30 June or 31 December) and 0 otherwise. D10+ =1
if the number of months until the forecast target date is 10 or more and 0 otherwise. ** denotes statistical significance at the .05 level. Breusch-pagan test χ2(15)
= 322.582 with a p value =0.000.
39
Appendix D -- continued
Table D2
SUR Estimates of Forecast Revision and Recent Changes in Forecasted Variable
General Model: Nochanget = α + β |Δxt-1| + Σ γj Djt + et
Forecast Revisions of:
10-year Bond Rate
Fed Funds Rate
Unemployment Rate
Target
Horizon Length:
Column:
1-6
(D2.1)
7 - 12
(D2.2)
1-6
(D2.3)
7 - 12
(D2.4)
1-6
(D2.5)
7 - 12
(D2.6)
-.211**
(.054)
-.164**
(.032)
-.215**
(.056)
-.497**
(.072)
-.349**
(.073)
-.663**
(.119)
-.602**
(.096)
Explanatory Variable:
|Δxt-1|
D1
-.111**
(.033)
D7
-.035**
(.015)
D10+
Constant
R2
Sample size
.366**
(.016)
.421**
(.016)
.738**
(.015)
.667
(.017)
.456**
(.020)
.484**
(.018)
.310
111
.232
111
.298
111
.174
111
.217
111
.267
111
The table reports SUR estimates of the forecast revision model shown for the 2003-2014 sample period. Nochange is the fraction of forecasters in the current
WSJ survey with forecasts unchanged from the prior survey. |Δxt-1| is the absolute change in x from the last business day of the month before the prior survey to
the last business day of the month before the current survey; x is the bond rate, the target fed funds rate and the unemployment rate in columns (D2.1)-(D2.2),
(D2.3)-(D2.4), and (D2.5)-(D2.6), respectively. Dj = 1 if j is the number of months until the forecast target date (30 June or 31 December) and 0 otherwise. D10+
=1 if the number of months until the forecast target date is 10 or more and 0 otherwise. ** denotes statistical significance at the .05 level. Breusch-pagan test
χ2(15) = 317.249 with a p value =0.000.
40
Appendix D -- continued
Table D3
Estimates of Forecast Revision and Recent Changes in Forecasted Variable Prior to the Zero Lower Bound
General Model: Nochanget = α + β |Δxt-1| + Σ γj Djt + et
Forecast Revisions of:
Horizon Length:
Column:
Fed Funds Rate
Effective
Target
1-6
7 – 12
1-6
7 - 12
(D3.1)
(D3.2)
(D3.3)
(D3.4)
Explanatory Variable:
|Δxt-1|
-.223**
(.065)
-.168**
(.078)
-.503**
(.070)
-.287**
(.091)
Constant
.668**
(.034)
.564**
(.029)
.668**
(.034)
.566**
(.028)
.094
53
.058
53
.271
53
.141
53
R2
Sample size
The table reports OLS estimates of the forecast revision model shown for the 2003-2007 sample period. Nochange is the fraction of forecasters in the current
WSJ survey with forecasts unchanged from the prior survey. Forecasts are of the fed funds rate on a target date (30 June or 31 December). Separate estimates are
reported for surveys 1-6 months and 7-12 months before the target date. |Δxt-1| is the absolute change in x from the last business day of the month before the prior
survey to the last business day of the month before the current survey; x is the effective fed funds rate and the target fed funds rate in columns (D3.1)-(D3.2) and
(D3.3)-(D3.4). Robust standard errors appear in parentheses. ** denotes statistical significance at the .05 level, respectively.
41
Appendix E
Difference between Average Squared Forecast Errors of Non-Revisers and Revisers, by Horizon
Panel A: 10-year Bond Rate Forecasts
Horizon, in months:
Target Date:
Dec 2003
June 2004
Dec 2004
June 2005
Dec 2005
June 2006
Dec 2006
June 2007
Dec 2007
June 2008
Dec 2008
June 2009
Dec 2009
June 2010
Dec 2010
June 2011
Dec 2011
June 2012
Dec 2012
June 2013
Dec 2013
June 2014
Dec 2014
1
2
3
4
5
.043**
1.308***
.022
-.023
.187**
.271***
.051
-.017
.170***
.019**
.229**
.231*
.103**
.032
.025
.015
-.015
.278***
.028
.028**
.029**
.025
.148**
.035
-.340**
.463***
.043
.285
.233**
.121***
.009
.291***
.076
.024
.055
.224**
.275**
.052
-.061*
.116*
.055
.013
.051**
.015
.041
-.060
.006
.659***
.047
.106
.016
.232*
.031
.255***
.074
.169
-.113**
.122
.184**
.249**
-.092
-.129
.112
.106
.085*
.001
.018
.084
-.208
.006
-.227
.294*
.097
.044
-.080
.042
.892***
.223*
.200*
.022
.036
.236**
.435**
.009
-.011
.246**
.316**
.131*
-.002
-.045
.010
-.002
-.144
-.803**
.250
.045
.034
.018
.087
.459
-.065
.217*
.141**
-.003
.420
.472**
6
7
8
9
10
11
12
-.054
-.242
.776***
-.024
.039
.111
-.175
.260
.196**
-.048
.085
.247**
.365
.284
.012
.015
.377**
.336***
.097
.088
.084
-.904**
.077
-.699
-.402
.272
-.002
.361**
.090
.994**
.092
1.471***
.001
-.006
.531
.706***
.097
-.079
.138
.182**
.059
-.069
.115
-.121
-.063
-.391
.051
.132
.213
-.184
-.032
.245
-.155
-.454
.270**
-.202*
.388
.689***
.120
.251**
.447**
-.022
.145
.063
.041
-.005
-.070
-.367
-.001
.203
.054
.065
.214**
-.026
.493
-.010
.244**
-.208**
.327
.280
-.010
.209
.704***
.155
.200*
.061
.116
-.183
.082
.382
.194
.103
.058
.273
.265
-.243
.899
.579*
.226**
-.089
.238
.451*
.231
.131
.373
.216
-.114
.242***
-.033
-.222
-.199
.390
-.388
.103
.517*
.453
.215
.102
1.339
-.154
.064
.027
.0001
1.024**
-.200
---‡
.143
-.088
.125
-1.606
-.411
.103
-.110
.113
.601**
.457
This panel reports differences in mean squared forecast errors of forecasters who did and did not revise their forecasts of the 10-year government bond rate from
the previous survey for the target date shown, by months until the target date. Positive differences indicate larger mean squared errors for non-revisers. The
WSJ survey did not request 1-, 6- and 12-months-ahead forecasts of the bond rate until June 2008. ***, ** and * indicate differences significantly different
from zero at the .01, .05, and .10 levels. ‡ No forecast revision data are available for this date.
42
Appendix E-- continued
Panel B: Fed Funds Rate Forecasts
Horizon, in months:
Target Date:
June 2003
Dec 2003
June 2004
Dec 2004
June 2005
Dec 2005
June 2006
Dec 2006
June 2007
Dec 2007
June 2008
1
2
-.035
-.011*
-.039*
.008
.038**
.027**
.001
-.002
.024**
-.010
-.023**
.004
.003
3
-.076*
.004
.026**
.032**
-.020**
.019
.016
-.001
-.015
.018
-.046
4
-.039
.011
-.006
.016
-.006
-.087*
.058*
.012
-.150**
.101
.049
5
-.064
.017
-.009
-.068
-.036
.093**
.040
.015
-.124
.239**
---†
6
7
8
9
10
11
.265
.014
-.044
.066*
.035
-.034
.080
.040**
.209
.210
.975
.201
-.136
.214**
.075
-.021
.109
-.024
.132
.051
.356*
.276
-.058
.226
-.199**
.181**
.153
-.025
.127
-.319
-.322
.549**
.006
.035
-.304**
.099
-.376*
-.043
.240
.019
1.210
.292
.110
.204
-.209
-.210**
.524**
.197
.151
-.185
.474
12
This panel reports differences in the mean squared forecast errors of forecasters who did and did not revise their forecasts of the fed funds rate from the
previous survey for the target date shown, by months until the target date. Positive differences indicate larger mean squared errors for non-revisers. The WSJ
survey did not request 1-, 6- and 12-months-ahead forecasts of the fed funds rate until June 2008. ***, ** and * indicate differences significantly different
from zero at the .01, .05, and .10 levels.
† All forecasters revised their forecasts.
43
Appendix E-- continued
Panel C: Unemployment Rate Forecasts
Horizon, in months:
1
2
3
4
5
6
7
8
9
10
11
12
Target Date
May 2003
-.003
.020
.060*
-.031
Nov 2003
-.007
-.015
-.012
-.185
.288
-.106
-.043
.262
.269
May 2004
-.007**
.010
.007
.004
.003
-.013
.082
-.129
.031
Nov 2004
-.001
-.011
-.010
.012
-.017
.024
.045*
.024
.006
May 2005
-.004
.009
.003
.012
.002
-.011
.011
.007
-.043
Nov 2005
.001
-.004
-.005
-.003
.016
-.009
.007
.026
.015
May 2006
.008**
.011
-.022
.047
.019
.023
-.051*
-.060*
.028
Nov 2006
.024**
.015
.001
-.096
-.030
.005
-.004
-.069
.104*
**
May 2007
.004
.033
-.020
.037
.021
.060
.100
-.070
-.142*
Nov 2007
.011
-.006
.040
.048
Dec 2007
.004
-.003
.004
-.014
-.035
June 2008
-.204
-.042*
.085*
-.050
.072
.038
.029
-.010
.056
.239*
**
Dec 2008
.003
---†
.016
.425
.106
.083
-.199
-.391
.017
.655
-.079
.251
June 2009
.234**
-.179
.417
-.355
-.671
.935
12.987**
.060
.915
-.331
1.178
Dec 2009
.040*
-.158
.014
-.034
-.017
.110
-.202
.483
-.444
3.430**
-.666
.072
*
**
June 2010
-.033
-.038
-.011
-.087
.038
.085
-.195
.166
.010
.247
-.017
.599*
*
Dec 2010
-.039
.008
.053
-.043
-.021
-.026
.018
.067
.124
.025
.062
.042
June 2011
-.010
.048
.044
.030
.123*
.074
.033
-.115*
-.046
.061
-.003
.080
Dec 2011
.160**
-.050
.062
-.165** -.074
-.075
.097**
.039
.030
.075
.176**
-.037
June 2012
.008*
-.014
.006
.016
.023
.070
.052
-.089
-.052
-.312*
.064
.053
**
Dec 2012
-.003
.006
.018
-.015
-.010
-.092
.028
.047
.002
.094**
.081
.141
June 2013
-.003
.007
.021*
.011
.001
.039*
.027
.127**
.071
.075
-.069
-.128**
Dec 2013
.066**
-.022
-.053
-.068**
.101*
-.005
-.147**
-.029
-.110
.119
-.160
.091
**
**
June 2014
.013
-.001
-.056
.028
.115
.243
.128*
-1.044
-.181**
-.116
.001
-.009
Dec 2014
-.006
.026**
-.015
-.028
-.068***
.069
-.027
-.031
-.029
.054
.010
.494**
This panel reports differences in the mean squared forecast errors of forecasters who did and did not revise their forecasts of the unemployment rate from the
previous survey for the target date shown, by months until the target date. Positive differences indicate larger mean squared errors for non-revisers. The WSJ
survey switched from requesting unemployment rate forecasts for May and November to June and December starting in June 2007. The WSJ survey did not
consistently request 1-, 6- and 12-months-ahead forecasts of the unemployment rate until after June 2009. The large outlier at the 8-month horizon for the June
2009 target date is because all but one forecaster revised their forecasts following a 40-basis-point drop in the unemployment rate from the prior month.
***, ** and * indicate differences significantly different from zero at the .01, .05, and .10 levels.
† All forecasters revised their forecasts.
44