ArticlePDF Available

IMPROVING ECONOMIC PREDICTION: A NEW METHOD FOR MEASURING ECONOMIC CONFIDENCE AND ITS IMPACT ON THE EVOLUTION OF THE US ECONOMY 1

Authors:

Abstract and Figures

A number of recent contributions have tried to add to the understanding and forecasting of the macro economy by analysing news and narratives. In this contribution we report on a new approach to the content analysis of very large text databases. It draws on a new social-psychological theory of decision-making under uncertainty to focus content analysis on the presence or absence of specific groups of emotional words in texts. The words identified are ordinary English emotional words. The method identifies, very rapidly, a Relative Sentiment Shift series which measures changes through time in the relations between two core emotional groups, excitement (about gain) and anxiety (about loss). Results are reported using text from the Thomson Reuters News Archive for articles published in the US. We find that shifts in the new emotion series Granger cause changes in US GDP and Gross Domestic Fixed Capital Formation. The RSS series also adds significant explanatory power to the Survey of Professional Forecasters' consensus forecasts and is useful both in identifying phase shifts in the direction of the economy and in " nowcasting " .
Content may be subject to copyright.
1
IMPROVING ECONOMIC PREDICTION: A NEW METHOD FOR
MEASURING ECONOMIC CONFIDENCE AND ITS IMPACT ON THE
EVOLUTION OF THE US ECONOMY
1
.
David Tuckett, Paul Ormerod, Rickard Nyman and Robert Elliot Smith
(University College London, Centre for the Study of Decision-Making Uncertainty)
ABSTRACT
A number of recent contributions have tried to add to the understanding and forecasting of the
macro economy by analysing news and narratives. In this contribution we report on a new
approach to the content analysis of very large text databases. It draws on a new social-
psychological theory of decision-making under uncertainty to focus content analysis on the
presence or absence of specific groups of emotional words in texts. The words identified are
ordinary English emotional words. The method identifies, very rapidly, a Relative Sentiment
Shift series which measures changes through time in the relations between two core emotional
groups, excitement (about gain) and anxiety (about loss). Results are reported using text from
the Thomson Reuters News Archive for articles published in the US. We find that shifts in the
new emotion series Granger cause changes in US GDP and Gross Domestic Fixed Capital
Formation. The RSS series also adds significant explanatory power to the Survey of
Professional Forecasters’ consensus forecasts and is useful both in identifying phase shifts in
the direction of the economy and in “nowcasting”.
1
We are very grateful to George Akerlof, Stephen Hansen, Sujit Kapadia, David Laibson, David Romer and
Dennis Snower for help and advice with this paper. Viktor Manuel Strauss has made available his experimental
work. Philip Treleaven and David Vinson were also very helpful. David Tuckett wishes to acknowledge support
from the Institute of New Economic Thinking (grants no. IN01100025 and IN1300051) and the Eric
Simenhauer Foundation of the Institute of Psychoanalysis (London). We also wish to thank Chrystia Freeland,
Richard Brown and Maciej Pomalecki of Thomson Reuters for arranging access to the Reuters News archive.
2
JEL classifications: D81, D83, E32, E37
Key Words: uncertainty; algorithmic text analysis; forecasts; Granger causality
3
I. INTRODUCTION
The largely unexpected arrival of the global economic crisis and the largely unpredicted
slowness of the recovery from the Great Recession have created a search to find additional
methods to detect how an economy is evolving (Haldane, 2014). One avenue emerging is to
use the analysis of news media or other digital sources to derive information about future
expectations and behaviour (Ramey and Shapiro, 1999; Romer and Romer, 2010; Dominguez
and Shapiro, 2013; Baker, 2013, 2014; Choi and Varian, 2012; Haddow et al 2013).
The ability to draw valid conclusions from news and text analysis depends both on the
methods adopted to extract trends and on the analytic solutions chosen to deal with problems
of reverse causality. In this contribution we report on a highly specific approach to extracting
information about emotion shifts in very large text databases which uses a new word dictionary.
Each text analysed is date and time stamped and never revised. The text analysis comprises a
set of algorithms to detect sentiment shifts underlying confidence and is designed to analyse
the words in any text archive. The specific words that measure sentiment are pre-selected
ordinary English emotional words appearing in the text. They will be shown to have no direct
economic content. The analysis, therefore, creates a series which is orthogonal to any economic
news in the text.
The method is a conventional dictionary approach using conventional algorithms. It is
new because it operationalises a new social psychological theory of decision-making,
conviction narrative theory (CNT). CNT focuses on how agents recruit their emotions so that
they are confident enough to act under conditions of uncertainty. The theory directs the word
search towards just to two very specific groups of emotion thought to encourage or inhibit
action in conditions of uncertainty. Specifically, the emotion groups explored are those
4
associated with excitement about gain (evoking approach) and anxiety about loss (evoking
avoidance).
The method can extract from any text database a particular sentiment shift time series
(Relative Sentiment Shift, RSS) based on the relative strength of the two emotion groups just
mentioned, from now on referred to simply as excitement and anxiety. It can be applied to any
unstructured text archive containing documents labelled with time and date. The purpose is to
measure shifts in emotional conviction in individual narratives which, when aggregated, can
be conceived as indicators of overall shifts in confidence.
In this paper, we construct an RSS series of the US economy from the Thomson-Reuters
newsfeed archive, described below. We proceed as follows. First, we set out details of our
Directed Algorithmic Text Analysis (DATA) method for creating an RSS time series for the
United States economy and then we undertake analyses to demonstrate the following points:
I. Granger causality tests between RSS and the quarterly growth rates of real GDP
(DLGDP), and real gross private domestic investment (DLK) show clearly that RSS
Granger causes both DLGDP and DLK, but there is no causality from either DLGDP
or DLK to RSS.
II. The relationship between RSS and DLGDP is not determined spuriously through a
relationship with contemporaneous variables. In a linear regression of DLGDP on a
set of forward looking variables based on asset prices, the lagged RSS series adds
significant explanatory power.
III. The relationship between RSS and DLGDP/DLK is not explained by economic news
within the news database. We look at this in two ways. First we confirm the words
used to construct the RSS series are ordinary English emotion words, not changes in
economic conditions. Second, we look at a number of ways of capturing news about
5
changing economic conditions in the database, including the economic policy
uncertainty index (Baker et al. 2013, 2014), and find they are Granger caused by the
RSS with no causality in the reverse direction.
IV. The one period ahead forecasts of DLGDP of the Survey of Professional Forecasters
(SPF) are known to be unbiased predictors (Stark 2010, for example). We confirm
this result, but also show that RSS adds significant explanatory power to a regression
of DLGDP on the consensus SPF forecasts.
V. In a linear regression of DLGDP on a set of forward looking variables based on asset
prices, the contemporaneous and lagged values of the RSS series add significant
explanatory power so that RSS may be useful for ‘nowcasting’.
VI. Finally, the RSS measure appears to gives early warning of the 2008/09 economic
recession. It shows a marked increase in the relative strength of anxiety, indicating a
move into a low growth phase in the business cycle, well in advance of not only the
consensus economic forecasts, but of indicators such as the Anxiety Index reported by
the Federal Reserve Bank of Philadelphia.
In section II of the paper, we offer a brief discussion of RSS in the context of uncertainty
and the formation of expectations. Section III describes the data and the method of constructing
the RSS series, and compares and contrasts it with text based measures of sentiment and
uncertainty already in the economics literature. Sections IV-IX set out the empirical results.
Section X is a concluding discussion.
II. RELATIVE SENTIMENT SHIFT AND UNCERTAINTY
As noted earlier, RSS is constructed to be orthogonal to economic data. It is not, therefore,
intrinsically subject to any obvious problems of reverse causality. The words analysed within
6
the DATA algorithm are not “economic words”, such as reports of increased uncertainty or of
a crisis. They are ordinary emotion words in use in everyday English.
As background to consideration of the methodology and findings, it is useful to consider
RSS in the context of expectation formation in economic theory. The dominant paradigm, that
of rational expectations, requires considerable knowledge on the part of agents of the ‘true’
model which describes the operation of the economy. Agents either are already in possession
of the relevant model, or discover it through some form of Bayesian learning. However, in
many situations, especially in macroeconomics, there is unresolved uncertainty about the
model itself. For example, in the context of macroeconomic models of the US economy, a
major survey by Ramey (2011) shows that even in this rather narrow methodological context,
the size of the fiscal multiplier, a basic concept in this area, varies between 0.8 and 1.5
according to whichever model one selects. Looking back to the policy debates in the immediate
aftermath of the collapse of Lehman Brothers in September 2008, prominent economists,
including Nobel Laureates, could be found on both sides of the argument as to whether or not
to allow banks and other financial institutions to fail. It is hard to imagine that these groups of
protagonists had the same model of the economy in mind.
More generally, within the statistics literature, there is a widespread understanding that
model uncertainty is often an inherent feature of reality. It may simply not be possible to decide
on the ‘true’ model. Chatfield (1995) is a widely cited paper on this topic. In an economic
context, Onatski and Williams (2003), in a survey for the European Central Bank of sources of
uncertainty, concluded that “The most damaging source of uncertainty for a policy maker is
found to be the pure model uncertainty, that is the uncertainty associated with the specification
of the reference model”. Gilboa et al. (2008) note that “the standard expected utility model,
along with Bayesian extensions of that model, restricts attention to beliefs modelled by a single
probability measure, even in cases where no rational way exists to derive such well-defined
7
beliefs”. The formal argument that in situations in which the external environment is changing
rapidly, it is often not possible for agents to learn the ‘true’ model even by Bayesian process
goes back at least as far as Alchian (1950).
In short, in situations in which there is uncertainty about the true model which describes
the system, it may not be possible for agents to form rational expectations. As a result, agents
are uncertain about the probability distribution of potential outcomes. Conviction Narrative
Theory, which motivates the direction of our search for words eliciting the emotions of
‘excitement’ about gain and ‘anxiety’ about loss, is a brain and social science based theory of
decision-making under this type of uncertainty. It was developed to describe and explain how
agents develop the confidence to take decisions in such circumstances, rather than simply being
paralysed into inactivity.
III METHOD AND DATA
III.A Background
We offer a new approach to text and sentiment analysis using algorithms directed by a specific
theory of conviction. Machine learning algorithms are by now a well-established approach to
content analysis within a document set. They offer the potential to carry out rigorous, very
rapid and extensive analysis of large volumes of textual material in order to derive knowledge
(for example, Pang & Lee, 2005; Sebastiani, 2002; Turney, 2002, Tetlock et. al., 2008, Soo,
2013). Because analysis is of documents fixed in “real time” it requires no revision. So, in
contrast to economic indicators like GDP, which are estimated and revised, results are what
they are.
Choi and Varian (2012) have used Google search data to predict a range of economic
indicators. That approach depends on the argument that changes in search behaviour (for
example for job vacancies) are stable indicators of developments in the economy (for example,
8
alterations in employment rates). This assumption was undermined in efforts to use Google
data to predict influenza outbreaks (Lazer et al, 2014, Bentley et al. 2014).
Other approaches look at news content which may be expected to influence agent
expectations and behaviour. Ramey and Shapiro (1999) used news reports to identify changes
in government spending likely to impact the economy. Romer and Romer (2010) analysed text
sources (including presidential speeches) to measure the likely effects of tax changes.
Dominguez and Shapiro (2013) analysed newspaper and media sources to detect narrative
shifts that could account for the slowness of the economic recovery. Soo (2013) went beyond
simple news content to quantify the positive and negative tone of housing news (adapting the
Harvard IV-4 word lists for Increase and Rise) in local newspaper articles about the US housing
market, trying to isolate the roles of sentiment and fundamentals. In the UK the Bank of
England, taking the view that perceived levels of uncertainty influence agent behaviour, has
recently started to publish an uncertainty index based on a range of public information
(Haddow et al, 2013). For the US Baker et al (2013, 2014) have described a method to construct
an uncertainty index, based on analysing the presence of uncertainty words in news media. The
underlying idea in most of these studies is that changes in information about the economy may
alter expectations and so behaviour in it. Only Soo, using a dictionary approach and also
looking at the co-occurrence of emotion words and economic news about what she called
“fundamentals”, has tried to isolate the impact of emotion. Her work was limited to housing
price rises during the recent bubble.
Other papers in the economics literature have used computerized textual analyses. For
example, Gentzkow and Shapiro (2010) develop a method to compare phrase frequencies in
newspapers with phrase frequencies in the 2005 Congressional Record in order to identify
whether the newspapers’ language is more similar to that of a congressional Republican or a
congressional Democrat. Jensen et al (2012) also look at political polarization. Their method,
9
based on Gentzkow and Shapiro’s approach, was to construct “trigrams” (phrases) to compare
the Congressional Record and the Google Ngrams corpus. Other recent examples include work
by Hansen et al (2014) who used machine learning linguistic techniques to analyse the FOMC
transcripts and Lucca and Trebbi (2009) who developed what they described as a new
automated, objective and intuitive scoring technique to measure the content of central bank
communication about future interest rate decisions based on information from the Internet and
news sources.
In finance a research tradition of analysing texts using a dictionary approach to search
for words with positive and negative emotional content
2
has been developing (for example,
Tetlock 2007). However, Loughram and McDonald (2011) have shown that for the analysis of
10k and similar documents general positive and negative valence word lists performed less
well than those they developed specifically for the analysis of financial documents. Although,
their published word lists for different emotional valence contain many thousands of words
3
there are not yet any reports of their use in an economics context.
III.B The RSS emotion word list
We construct our Relative Sentiment Shift (RSS) index using a particular type of algorithmic
text analysis allied to a dictionary approach. The algorithm does not use natural language
processing and is not set to learn from the text. Rather, it is directed to search just for distinct
groups of emotional words. In consequence it can work unmodified on any text database.
As noted above, the RSS aim is to capture emotional shifts in narratives circulating
within an economy with the aim of capturing shifts in conviction, conceived as made up from
a balance of approach and avoidance stimulating emotions.
2
The technical term in the psychology literature is ‘valence’
3
http://www3.nd.edu/~mcdonald/Word_Lists.html
10
The word list which the algorithm uses to measure the extent of conviction is directed
in advance by our theory of narrative conviction (CNT) and remains unchanged. It is
orthogonal to the events that may be being described. It is made up of ordinary everyday
emotion words belonging to only two limited and pre-defined lists of words. Each list is about
150 words, one representing excitement about gain (approach) and the other one representing
anxiety about loss (avoid). Twenty randomly drawn example words can be found in Table 1.
Further details of the word list are in the supplemental material.
Table I
Randomly Drawn Selection of Words indicating excitement (about gain) and anxiety (about loss)
Anxiety Anxiety Excitement Excitement
Jitter Erodes Excited Perfect
Threatening Uneasy Incredible Win
Distrusted Distressed Ideal Amazes
Jeopardized Unease Attract Energizing
Jitters Disquieted Tremendous Gush
Hurdles Perils Satisfactorily Wonderful
Fears Traumas Brilliant Attracts
Feared Alarm Meritorious Enthusiastically
Traumatic Distrusting Superbly Exceptionally
Fail Doubtable Satisfied Encouraged
The idea that approach and avoidance words were creating the conviction to act was
derived from a cross disciplinary analysis of interview transcripts with fund managers (Tuckett,
2011). The theory that emerged draws on a growing body of work in social and brain science
which underlines the facilitating role of emotions in decision-making, particularly when
decisions “matter” and their outcomes are uncertain (Chong and Tuckett, 2014; Tuckett, Smith
and Nyman, 2014; Tuckett and Nikolic, 2015). Based on this idea, the specific words to be
included in the algorithm were first selected by one of us (DT) from very much longer lists of
categories of related emotion words in the Harvard IV-4 list
4
. The selection was intended to
capture words judged to belong in either the theoretically important emotion group of
4
The Harvard IV-4 dictionary is available at www.wjh.hamecat.htmrvard.edu/~inquirer/ho
11
excitement about gain (evoking approach) or anxiety about loss (evoking avoidance). The list
so drawn successfully identified groups in different states of mind in an earlier analysis of the
Enron email database (Tuckett, Smith and Nyman, 2014).
To check further whether words in the two lists evoked the relevant emotions Strauss
(2013) set up an online experiment using Burke and James’s (2006) recommendations for
giving instructions. 125 words (68 for anxiety, and 57 for excitement) were presented to
subjects derived from the word list but using only the lemmas of those words i.e. the not
inflected base forms as one finds them as entries in dictionaries. Previous work has
established that emotional values generalize to inflected forms, which makes it unnecessary
to rate all forms separately (Warriner et al, 2013). In line with Loughran and Mcdonald’s
result that words may be understood slightly different in an economics/finance context
respondents with a finance background were selected and further steps were taken to define
excitant for subjects as “a state of anticipation of pleasure and triumph, as about future gains”
and anxiety as “a state of uneasiness and apprehension, as about future uncertainties or
losses”. Each word was presented separately to subjects on the centre of the screen and the
order of the words was randomized for each subject.
Those who took part were randomly assigned so that they either rated words offered
them for anxiety or words for excitement on a virtual scale. The results show (after standard
procedures for removal of outliers) that a word rated highly by one participant on a particular
scale was usually rated highly by other participants too. Additionally, there was a strong
negative correlation between the mean ratings for anxiety and excitement
5
, indicating that
participants could clearly distinguish between anxiety and excitement words (Strauss 2013).
6
5
r(262) = -0.809, p < .0001
6
See supplementary material (Section 2).
12
A crucial point to emphasise again is that the two word lists used to construct RSS do
not contain any specific economic terms such as ‘crisis’, ‘boom’ or even ‘boost”. Rather they
are what we might term regular English words, words in everyday use in a wide variety of
contexts which convey the emotions of either ‘excitement’ or ‘anxiety’. Emotion
measurement focusing on the use of words is justified by robust findings in social psychology
which suggest that individuals, immersed in their culture and language, acquire automatic
associations between words or symbols and their (emotional) meaning. For instance,
emotional word perception goes along with affective reactions before the meaning of the
word is more deliberatively processed (Murphy and Zajonc, 1993; Siegle, Ingram and Matt,
2002; Hofmann et al, 2009). Additionally, emotionally-charged words have been shown to be
attention grabbing, to facilitate reaction times and accuracy in decision tasks, and to be more
easily recognised and remembered compared with neutral words (Ferré, 2003; Anderson,
2005; Sutton et al, 2007; Monnier and Syssau, 2008; Kousta, Vinson and Vigliocco, 2009).
Emotional words also cause distinct neural activity (Kensinger and Schacter, 2006; Herbert et
al., 2009; Anderson and Phelps, 2001).
III.C The RSS Statistic.
Given the word list, we extract RSS as a summary statistic of the two emotional traits from text
data by counting the two types of words within articles over a quarter.
Specifically, for the summary statistic of a collection of texts T we count the number
of occurrences of excitement words and anxiety words and then scale these numbers by the
total text size in number of articles. To arrive at a single statistic, reflecting the underlying
theory of conviction narratives, we subtract the anxiety statistic from the excitement statistic.
   

13
Data for all three data sets are available daily but in this paper we report results on a
quarterly basis
7
. In other words, we compute the above statistic for each quarterly collection of
articles in order to generate a quarterly series. We make the obvious remark that an increase in
this relative emotion score is due to an increase in excitement and/or a decrease in anxiety.
Also, as evidenced by the definition of the measure we do not control for possible
negations of our words, e.g. ‘not anxious’, ‘not excited’ nor do we control for their typical
(expected) frequencies in ordinary usage. Nonetheless we did devise a test to explore the
influence on results if “not” is present finding it made no difference
8
. The typical frequencies
of the words in ordinary usage might usefully be adjusted for to make the overall statistic more
robust to potential outliers among the most frequent words in terms of their usage in a particular
database, but the expected frequencies themselves are not meaningful in terms of sentiment
shifts over time.
The simplicity of this method is intentional. It leads to clarity about what is measured
and allows us to bring a range of statistical techniques to bear on correlations. To maintain
clarity we have not yet taken advantage of sophisticated natural language processing
techniques. We have also not employed approaches (e.g. Hansen et al, 2014, Socher et al, 2014)
that either use supervised learning to detect if the algorithm can learn from a pre-categorized
database, or find latent structure in the data using unsupervised learning methods, and rely on
fitting to a pattern to make improved predictions. In future studies such techniques could be
used to try to optimise the power of RSS, as for example by employing NLP (in an innovative
manner) to help understand the emotional intent of words, or by using learning from data
sources to weight some words more heavily than others. However, such approaches must be
7
See supplementary material section 4.
8
Supplemental material section 6.
14
specifically developed to maintain the orthogonality of RSS to domain or data source specific
words, and avoid data fitting.
III.D Data
We have discussed how we analyse content above. Here, we describe the source documents we
selected to analyse.
There are a variety of text data sources with a macroeconomic and financial sector
focus. For our purposes any text database thought to contain relevant data with a bearing on
relevant economic narratives with documents labelled with time and date would suffice. In
other studies we have successfully used Email data, Market Commentary prepared by analysts
at the Bank of England, Broker reports held at the Bank, Newspaper or Internet content, and
Twitter (Tuckett, Smith and Nyman, 2014; Nyman et al, 2014). The analysis in this paper is
based on the Thomson-Reuters News Archive (RTRS).
The Reuters News archive consisted of articles published between 1996 and 2014.
Reuters provide extensive documentation and for this paper we select only articles with
RTRS (Reuters) as the attribution, English as the language, published in the New York or
Washington offices, and so defined as US focused. Articles with the tags SPO (sports), ODD
(human interest) or WEA (weather) within the ‘tags’ field were removed
9
.
IV GRANGER CAUSALITY OF GDP
We explore whether the RSS series constructed Granger causes changes in US GDP growth or
vice versa.
IV.A Methodology
9
Supplementary material section 3.
15
We use the methodology described in Toda and Yamamoto (1996). In outline, in investigating
Granger causality between any two series, this is as follows:
I. Check the order of integration of the two series using Augmented Dickey-Fuller (Said
and Dickey 1984; p-values are interpolated from Table 4.2, p. 103 of Banerjee et al.
1993) and the Kwiatowski-Phillips-Schmidt-Shin (1992) tests. Let m be the maximum
order of integration found.
II. Specify the VAR model using the data in levelled form, regardless of what was found
in step 1, to determine the number of lags to use with standard method. We use the
Akaike Information Criteria
III. Check the stability of the VAR (we use OLS-CUSUM plots, which are reported in the
Supplement).
IV. Test for autocorrelation of residuals. If autocorrelation is found, increase the number of
lags until it goes away. We use the multivariate Portmanteau- and Breusch-Godfrey
tests for serially correlated errors. Let p be the number of lags then used.
V. Add m extra lags of each variable to the VAR.
VI. Perform Wald tests with null being that the first p lags of the independent variable have
coefficients equal to 0. If this is rejected, we have evidence of Granger-causality from
the independent to dependent variable.
We used the statistical program R to carry out the analysis, and the various packages
used to carry out the above Toda-Yamamoto procedure are documented in the Supplement.
IV.B Granger causality of GDP and its main components
RSS assesses shifts in the relationships between two clusters of expressed emotion words based
on the theory of conviction narratives and so in a conceptual sense entirely orthogonal to
16
“economy” words evoking emotion that might also exist in the texts. Figure 1 plots the
relationship between real quarterly growth in US Gross Domestic Product and the RSS series
Figure I
Quarterly percentage changes in US Real GDP and Relative Sentiment Shifts in Reuters US
based News Articles 1996-2014 (Source: Thomson Reuters News Archive)
Using the methods outlined steps I VI above, we consider Granger causality between
the RSS series and the quarterly change in real GDP, real gross private domestic investment,
and real personal consumer expenditures
10
(DLGDP, DLK and DLCE). The data is quarterly,
and the full sample period used is 1996Q1 through 2014Q3 in all tests reported on Granger
causality. The details of the tests carried out are described in the Supplement. Here, we report
step VI, namely the Wald tests of Granger causality
Table II
Wald test statistics of Granger-causality between the relative sentiment shift series (RSS) and
USDLGDP, DLK and DLCE
Direction Chi-Sq d.f. p-value
RSS -> DLGDP 3.5 1 0.06*
DLGDP -> RSS 0.037 1 0.85
RSS -> DLK 2.9 3 0.086*
DLK -> RSS 0.94 1 0.33
RSS -> DLCE 0.14 1 0.99
DLCE -> RSS 5.2 3 0.16
10
The data source is Table 1.1.6 of the NIPA tables at www.bea.gov
17
Note: *p<0.1; **p<0.05; ***p<0.01
The results in Table 2 show that there is evidence of Granger causality from the RSS
series to both DLGDP and DLK. The Wald statistics in each case are significant at a p-value
of < 0.10.
There is no evidence of Granger causality either way between RSS and DLCE. This is
consistent with the hypothesis of Hall (1978) regarding consumer spending. It is also consistent
with the psychological theory of decision making which underpins the RSS series. In contrast
to investment decisions, most consumption decisions relate to items which are routine, low
cost, have little implications for the future, and are easily reversed (e.g. if you buy a brand of
baked beans which it turns out you dislike, you can buy another brand next week). So the level
of uncertainty around the consequences of most consumption decisions is inherently much
lower than it is around investment. We would therefore not expect the RSS series to cause
consumer spending.
V ASSET PRICES
In terms of claiming causality from RSS to GDP, the question arises as to whether the result
is spurious; perhaps the RSS series is simply capturing the effects of forward-looking
indicators, such as asset prices. To explore this possibility we conducted further analyses
taking as the dependent variable the first difference of the natural log of real quarterly US
GDP (expenditure measure, denoted by DLGDP). We then considered the potential causal
role of forwarding looking indicators choosing the following series, using the quarterly
average values for each:
The Michigan Consumer Sentiment Index (MCI)
Changes in stock prices, measured by the Standard and Poor’s 500 (DSP)
The yield on 10 year US government bonds (BOND)
18
The yield on 3 month US Treasury bills (TB)
The TED spread (TED)
The VIX, the Chicago Board Options Exchange Market Volatility Index
To do this we initially estimated a completely general equation containing each of the
above variables at lags through 3 and the dependent variable at lags 1 through 3.(1)
       

    
As a first step, we eliminated all variables for which the null hypothesis that the
coefficient is zero could only be rejected at a p-value greater than 0.75. We then eliminated
further variables one at a time, in each step leaving out the one with the highest p-value.
The resulting regression is set out in Table III. The dependent variable is , the sample
period is 1997Q1 through 2014Q3.
Table III
Preferred linear regression of real US GDP on asset prices and lagged values of RSS
Dependent variable: DLGDP; Sample period: 1997Q1 through 2014Q3
Variable Estimated coefficient Standard error p-value
Constant 0.0075 0.0011 6.5e-09***
   0.0036 0.0011 0.002***
 0.0018 0.0006 0.007***
  -0.0103 0.0025 7.4e-05***
 -0.0045 0.0017 0.009***
 3.8e-05 9.1e-06 9.0e-05***
Note: *p<0.1; **p<0.05; ***p<0.01
Residual standard error: 0.0051; Adjusted R-squared 0.402. Test statistics and p-values: DW = 1.81 (0.16); Ramsey F (2, 63)
= 1.07 (0.35); BG(4) = 0.85 (0.93); KS = 0.098 (0.47). The p-values for the test statistics are the p-values at which the null
hypothesis is rejected. DW is the Durbin-Watson test for first order autocorrelation; Ramsey is the Ramsey RESET
specification test; BG is the Breusch-Godfrey test of residual autocorrelation from 1 through 4 lags; KS is the Kolmogorov-
Smirnov test for the normality of the residuals.
For the difference terms in RSS and TED, the null hypotheses that the relevant
coefficients in each case were equal but of opposite sign were not rejected. The equation is
well specified econometrically. Both the RSS series and series related to asset prices have
19
explanatory power in predicting short-term changes in real US GDP. When the above
equation is estimated leaving out the RSS terms, the adjusted R-squared falls to 0.278,
compared to the 0.402 of the equation above. So the incremental increase is 0.126 in
absolute terms, or 45 per cent of the R squared when the RSS terms are omitted. On this
analysis the RSS index does contain valuable new information.
VI ECONOMIC NEWS
As noted, above, the RSS series is constructed on psycholinguistic principles selecting English
language words which evoke emotions in two emotional clusters. But a large news database
like Reuters will include a number of more specific economic forecast words which evoke
emotion (for instance uncertainty, boom, etc.) or bits of economically relevant factual
information which, in principle, might be correlated with RSS or even causal of its shifts.
In this section we examine the question as to whether the relationship between RSS and
DLGDP is explained by economically relevant information within the news database rather
than by shifts in emotion. We looked at this possibility in two ways. First, we looked carefully
at the words used to construct the RSS series to confirm that what is measured are ordinary
English emotion words, not changes in economic conditions. Second, we looked for words
containing information about changing economic conditions and explored what relationship
they had to changing sentiment.
First, looking carefully at the two lists of words we used to construct RSS, selected as
emotional indicators via expert guidance validated in experiments, we find that all but six of
the words selected seem unequivocally free of potential economic information. The six possible
words which have both emotional and economic meaning are the words ‘uncertain’ and
‘uncertainty’ in the anxiety dictionary and the words ‘boost’, ‘boosted’, ‘boosts’, ‘exuberance’
20
and ‘exuberant’ in the excitement dictionary
11
. These words certainly indicate emotions but,
arguably, also convey information about economic conditions uncertainty, for instance, could
be a reason to delay investment, exuberance to join in. To test whether these words might be
influencing the results we re-examined the relationship between RSS and GDP leaving these
words out of the RSS measure. The new series correlated with the original RSS series at 0.99
in level form and 0.99 in difference form making no difference to the results reported above
12
.
The result appears to confirm that the RSS measure is an independent variable.
A second way we have attempted to examine the possibility that the correlation between
RSS and GDP is the outcome of changes in economic news rather than emotional shifts, has
been to examine the relationship between RSS and three relevant economic indicators:
I. The widely quoted Economic Policy Uncertainty Index (EPU) (Bloom,2014, Baker et
al, 2012) which captures discussions of uncertainty in major newspapers in the US.
II. A simple series (UNCERT) in which we counted the number of times each day the
words ‘uncertain’ and ‘uncertainty’ appear in the Reuters news archive for the US.
III. A further simple index (‘ECON’) we constructed to capture more specifically
economic terms, namely ‘boost’ and ‘boom’ (positive emotion words) and ‘crisis’,
recession’ and ‘fall’ (negative emotion words). A summary score indicates whether
economic news for that quarter is shifting in a positive or negative direction over
time.
We then conducted two sets of Granger causality tests. The first, we look at the
influence of each variable on GDP and, the second, to look at the causal relationships between
11
See Table 1. Also Table A1 in the Supplementary material contains a further 40 anxiety and 40 excitement
words selected at random from the total. The full list can be obtained from the authors on request.
12
Supplementary material page section 7.
21
the variables themselves and RSS. We report results with quarterly data, because this is the
periodicity of the national accounts.
Table IV
Wald test statistics of Granger-causality between the Economic Policy Uncertainty Index (EPU),
UNCERT, ECON and DLGDP
Direction Chi-Sq d.f. p-value
EPU -> DLGDP 1.2 1 0.27
DLGDP -> EPU 0.44 1 0.51
UNCERT -> DLGDP 3.3 1 0.07*
DLGDP -> UNCERT 0.28 1 0.6
ECON -> DLGDP 0.38 1 0.54
DLGDP -> ECON 1.3 1 0.25
Note: *p<0.1; **p<0.05; ***p<0.01
Table IV looks at the ability of these three new variables to predict changes in DLGDP
(as in Table II above) using the Wald test. There is no evidence of Granger causality of GDP
from either the EPU or the ECON variable. Interestingly, compared to the sophisticated EPU
measure, the simple UNCERT variable shows some evidence of Granger causing GDP.
Table V
Wald test statistics of Granger-causality between the relative sentiment shift series (RSS) and
UNCERT, ECON and EPU
Direction Chi-Sq d.f. p-value
RSS -> UNCERT 15.5 1 8.3e-05***
UNCERT -> RSS 1.6 1 0.21
RSS -> ECON 15.3 2 0.0005***
ECON -> RSS 4 2 0.13
RSS -> EPU 19.7 2 5.2e-05***
EPU -> RSS 1.5 2 0.48
Note: *p<0.1; **p<0.05; ***p<0.01
Table V reports the Granger causality tests (using the Wald test as before) we undertook
between the three new variables and RSS. It is clear that EPU, UNCERT and ECON are all
Granger caused by RSS but not vice versa.
22
VII PROFESSIONAL FORECASTS
The RSS series also appears to add credible information to improve the one-quarter
ahead forecasting record of real GDP growth in the United States.
The Survey of Professional Forecasters is the oldest quarterly survey of macroeconomic
forecasts in the United States. The survey began in 1968 and was conducted by the American
Statistical Association and the National Bureau of Economic Research. The Federal Reserve
Bank of Philadelphia took over the survey in 1990. Data on the consensus forecast for one-
quarter ahead real GDP growth is available at http://www.philadelphiafed.org/research-and-
data/real-time-center/survey-of-professional-forecasters/data-files/RGDP/
13
. A discussion of
the historical accuracy of the forecasts, for both GDP and other economic variables, is given in
Stark (op. cit.).
The consensus forecasts over time are unbiased. However, they are able to account for
only a relatively small fraction of the overall variance in quarterly real GDP growth. We
confirm this finding in the literature by regressing quarterly real GDP growth in quarter t on
the consensus forecast for quarter t made in quarter t-1 over the period 1996Q2 through
2014Q3.
Table VI
Regression of the actual values of quarterly growth in real US GDP (DLGDP) on the consensus
forecast made in the previous quarter (SPF) and the relative sentiment shift series (RSS)
Dependent variable:
---------------------------------------------------------------------------------------------------
DLGDP DLGDP
(1) (2)
Constant -0.430 0.325
(0.798) (0.829)
SPF 1.124*** 0.829***
13
Using data downloaded from the click through “Annualized Percent Change of Mean Responses”
23
(0.290) (0.271)
RSS 0.765**
(0.299)
Observations 74 74
R2 0.170 0.240
Adjusted R2 0.159 0.219
Residual Std. Error (df = 70) 2.469 2.379
F Statistic 14.77*** (df = 1; 72) 11.23*** (df = 2; 71)
Note: *p<0.1; **p<0.05; ***p<0.01
The RSS variable is statistically significant from zero, and the adjusted R squared
increases from 0.159 in the equation without RSS to 0.219 when it is included, an incremental
effect of 38 per cent.
The same results hold when the third vintage estimate of GDP is used rather than the
latest data used in Table VI
14
.
VIII NOWCASTING’ GDP GROWTH
A further way in which RSS seems to be of value is for nowcasting. First estimates for GDP in
any given quarter are available from national statistical offices some weeks after the quarter
ends. They are subsequently revised sometimes even several years later (Ormerod 1978,
Shrestha and Marini, 2013, Manski 2014). Estimates are often influenced by the state of the
business cycle at the moment they take place so that, for example, in the case of 2008, the move
into negative growth which took place during the year was not recognised in revised statistics
until about a year later (Shrestha and Marini, 2013). Efforts are underway to improve this
situation by “nowcasting” using new data sources such as Google searches (Scott and Varian,
14
See the supplementary material section 11.
24
2012) but they are not necessarily reliable (Lazer, et al. 2014, Ormerod, Nyman & Bentley,
2014).
Our analysis suggests RSS could provide statistically reliable information to assist
nowcasting of this kind. For the US, we again begin with equation (1) above, except that we
add the contemporaneous value of all the series as well as their lagged values. All of them will
be known for certain at the end of the relevant quarter. We followed the same process, namely
eliminating all variables where the p-value was above 0.75, and then eliminating them one at a
time on a step by step basis. Results are in Table VII. The dependent variable is 
sample period is 1997Q1 through 2014Q3.
Table VII
Preferred linear regression of real US GDP on asset prices and contemporaneous and lagged values
of RSS. Dependent variable: DLGDP; Sample period: 1997Q1 through 2014Q3
Variable Estimated coefficient Standard error p-value
Constant 0.0074 0.0011 1.4e-08***
   0.0020 0.0010 0.061*
 0.0023 0.0006 0.0005***
  -0.0092 0.0025 0.0004***
 -0.0045 0.0017 0.0073***
 2.4e-05 1.0e-05 0.023**
 2.7e-05 9.2e-06 0.0047***
Note: *p<0.1; **p<0.05; ***p<0.01
Residual standard error: 0.0046; Adjusted R-squared 0.527. Test statistics and p-values: DW = 1.85 (0.18); Ramsey F (2, 62)
= 2.70 (0.075); BG(4) = 2.36 (0.67); KS = 0.075 (0.78). The p-values for the test statistics are the p-values at which the null
hypothesis is rejected. DW is the Durbin-Watson test for first order autocorrelation; Ramsey is the Ramsey RESET
specification test; BG is the Breusch-Godfrey test of residual autocorrelation from 1 through 4 lags; KS is the Kolmogorov-
Smirnov test for the normality of the residuals.
For the difference terms in RSS and TED, the null hypotheses that the relevant coefficients in
each case were equal but of opposite sign were not rejected. The equation is again well
specified econometrically. The value of RSS at time t is indeed statistically significant in the
regression, though its value at lag 2 has the equal but opposite sign. Compared to the
25
equation using only lagged values as explanatory factors, the adjusted R-squared increases
from 0.402 to 0.527.
IX TURNING POINTS AND PHASES
We examine the causality between the Anxious Index (ANX) reported by the Federal
Reserve Bank of Philadelphia
15
and RSS. ANX is based on a survey of professional
forecasters, which asks panellists to estimate the probability that real GDP will decline in the
quarter in which the survey is taken and in each of the following four quarters. The anxious
index is the probability of a decline in real GDP in the quarter after a survey is taken.
We consider two variants. We denote by ANXA the Anxious Indices as reported on
the Philadelphia Feb website. In other words, the probability of a recession in quarter t+1
estimated in quarter t. We denote by ANXB the ‘nowcast’. In other words, the probability of
a recession in quarter t estimated in the same quarter, t. The data is quarterly, and the full
sample period used is 1996Q1 through 2014Q3. Again, the details of the tests carried out in
steps 1 5 of the Granger causality procedure are described in the Supplement. We simply
report here step 6, namely the Wald tests of Granger causality
Table VIII
Wald test statistics of Granger-causality between the relative sentiment shift series (RSS) and ANX
and ANXA
Direction Chi-Sq d.f. p-value
RSS -> ANX 28.5 3 2.8e-06***
ANX -> RSS 0.85 3 0.84
RSS -> ANXA 20.3 2 3.9e-05***
ANXA -> RSS 2.1 2 0.34
Note: *p<0.1; **p<0.05; ***p<0.01
15
http://www.philadelphiafed.org/research-and-data/real-time-center/survey-of-professional-
forecasters/anxious-index/
26
There is very strong evidence of Granger causality going from RSS to both the Anxiety series,
and no evidence of causality in the reverse direction.
A further set of results suggest the RSS series can also be useful for determining when
turning points in the economy appear imminent, something which until now has proved very
difficult.
Figure II below shows the RSS time series (blue line) plotted with US GDP (black line).
The solid line in the centre marks the moment not forecasted at the time when GDP began
to shrink (2008Q1). The sharp downturn in RSS in mid-2007 is a clear turning point, so RSS
in this example is a leading indicator for what subsequently happens to GDP.
Figure II
RSS and USDLGDP
Over the period 2003Q2 2007Q2, the mean value of the RSS series (scaled) was 0.850 with
a standard deviation of 0.256. The value in 2007Q3 is 4.37 standard deviations (taking the
27
2003Q2-2007Q2 of the standard deviation) below this mean value. In 2007Q4 and 2008Q1 it
is, respectively, 3.54 and 4.42 standard deviations below. A clear shift is identified.
The Anxious Index discussed above, indicating the probability of a recession in the next
quarter, took values 2007Q3-2008Q1 of 13.14, 14.00 and 16.95 respectively, indicating only a
low probability. It was not until 2008Q2 that the value rose to 42.9. The mean annualised
growth rate of GDP in the Survey of Professional Forecasters made in 2008Q2 for 2008Q3 was
+2.17 per cent. It was not until 2008Q4 that the consensus prediction for growth in the next
quarter (2009Q1) became negative.
So the RSS series fell sharply below its levels in the mid-2003 to mid-2007 period in
the third quarter of 2007, a fall which was subsequently sustained. Consensus economic
forecasts were essentially of no value at all in predicting the recession, and the Anxious Index
only showed a substantial increase in the probability of recession in the next quarter in 2008Q2.
The only indicators which seem to have given a similar type of early warning are the Financial
Stress Indices produced by the various Federal Reserve banks. Putting the data on a quarterly
basis, the Cleveland Fed FSI, for example, takes a value of 0.85 standard deviations above its
2003Q2-2007Q mean in 2007Q3. An increase here means a rise in financial stress. In 2007Q4,
it was 2.58 standard deviations above the 2003Q2-2007Q2 mean, and in 2008Q1 3.61 above.
The St Louis Fed index was already 2.42 standard deviations above in 2007Q3, rising to 4.18
and 5.15 respectively in 2007Q4 and 2008Q1. However, Nyman et al. (2015) show that over
a longer period (1996-2014) the RSS series Granger causes both the Cleveland and St Louis
Financial Stress Indices.
We also find evidence that RSS may be of help in identifying different phases of GDP
growth. We use cluster analysis of the data on quarterly US GDP growth and the Reuters RSS
data over the period 1996Q1 through 2014Q3. We use the fuzzy clustering algorithm in the
28
command ‘fanny’ in the statistical package R (Kaufman and Rousseeuw (1990)). Dunn’s
partition coefficient (Dunn 1973) indicates that the data set is best described by just two
clusters. Full details of analysis are available in the Supplement. Each cluster is made up of
two sets of data which are essentially time-contiguous. One cluster comprises data from
1996Q1 through 2000Q3 and then from 2003Q2 through 2007Q2. The other contains the data
from 2000Q4 through 2003Q2 and from 2007Q3 through 2014Q2
16
. The mean values of both
real GDP growth and the RSS Reuters data are distinctly different in the two clusters. In the
former, GDP growth at an annualised rate averages 3.62 per cent and 1.22 per cent in the
second. The corresponding values for the normalised Reuters RSS data are +0.85 and -0.87
respectively.
In economic terms, the clusters identify relatively high and low growth regimes. The
value of the RSS data can be calculated on the last day of each quarter, and is never subject to
revision, since by that date all the articles in the quarter have been published. There is no
overlap at all in the normalised values of the RSS data in the two regimes identified by the
clustering. The maximum value in the low growth cluster is -0.031 and the minimum in the
high is +0.030. So a very good heuristic for initial identification of GDP growth for the current
quarter is given by the value of the RSS series.
X CONCLUSION
We have described a new method of content analysis which identifies, very rapidly, shifts in
the relations between two core emotional groups, excitement (about gain) and anxiety (about
loss), in different sets of documents primarily containing economic and financial news. As we
have emphasised the word lists which we use to measure these core emotion groups are
16
There are one or two minor exceptions, with, for example, the quarter in which LTCM collapsed (1998Q3)
being allocated to the second cluster.
29
essentially orthogonal to economic terminology. They contain no explicit economic terms, but
are made up of everyday English words, validated for their emotional content in a
psychological experiment.
In this paper we focused on GDP through analysis of the Reuters database. Other results
have demonstrated causality versus the ViX and the Michigan Consumer Sentiment Index. The
latter outperform those of the consensus forecasts and are much stronger using the more
textually rich series of Broker Reports or Bank of England Market commentary we had
available (Nyman and Ormerod, 2015; Nyman et al, 2015). The latter have more nuanced and
arguably more relevant content and suggest the power of RSS might be stronger with still more
suitable data. Other results are relevant to anticipating unforeseen financial risk (Tuckett et al,
2014; Nyman et al, 2015).
There is scope for much more work and the approach can be extensively developed
both by using it to analyse databases more explicitly tailored to financial and economic news
and through development of the methods themselves. As time goes by datasets will be longer
and cover more cycles. For the time being it is clear that using Directed Algorithmic Text
Analysis to capture Relative Sentiment Shifts does work as a predictor of a range of variables.
The reason it works, we suggest, is because it successfully captures the emotion within
narratives which people construct in order to make decisions under uncertainty. When this
emotion shifts the decisions they make shift and economic activity shifts too.
Until now economics has focused almost exclusively on a model of human behaviour
in which, on average, agents are able to assess the outcomes of their actions correctly. A
limitation of that approach is that it must break down insofar as information contexts exist
which are genuinely uncertain i.e. which leave scope for intelligent agents to interpret
information differently or to have different levels of commitment to the actions they consider
appropriate. Conviction narrative theory sets out how under these circumstances agents draw
30
on their cognitive and affective endowment to reach conviction when calculation alone would
be inadequate. The factors that influence conviction, in part, are likely to be developed through
social interaction. Agents, in short, are influenced by what they see others are doing and the
norms and practices that grow up around them.
A number of new empirically-backed theories developing in sociology, economics,
anthropology, psychology and neuroscience have suggested a broader model of human
behaviour. Within such frameworks narrative and emotion can be conceived to join with
cognitive and calculative skills to facilitate economic action (For example, Akerlof and Shiller,
2009; Bruner, 1990; Damasio, 1999; Lane and Maxfield, 2005; Mar and Oatley, 2008; Beckert,
2011, Barbelet, 2013; Pixley, 2009; Bandelj, 2009; Berezin, 2005, 2009; Tuckett, 2011).
Moreover, because narratives are developed through social interaction they create order or
equilibrium, albeit in a fragile way (Chong and Tuckett, 2014).
The methods described above show ideas of this type can be operationalised with
advantage. Focusing on specific action-enabling (excitement) or disabling (anxiety) emotions
in the narratives circulating in the economy helps us to understand its likely evolution. The
explanation is that such narratives are drawn on to support actions in the economy (making
expected outcomes less unbearably uncertain) and so drive the aggregate outcomes of those
actions, which subsequently emerge as GDP.
Two features of the prototype RSS measures developed to date can be stressed. First,
even when a very large textual database is interrogated, the analysis can be performed quickly,
enabling it to be used as a leading indicator for economic time series in a subsequent period.
Given the database, the time needed to calculate the RSS is, at most, a few hours. Second,
unlike many economic time series, the RSS is not subject to revisions. The Reuters newsfeed
text database in any given month, for example, contains articles that are not revisited and
revised; nor are additional articles added to it for that month.
31
Keynes (1936, 1937) argued that economic action is energised despite future
uncertainties by a combination of conventional judgment and animal spirits. The RSS measure
outlined is based on a modern social and brain science approach to the way action is taken
under uncertainty. It can be considered an operational indicator of animal spirits.
REFERENCES
Akerlof, G. and Shiller, R. (2009) Animal Spirits: How Human Psychology Drives the
Economy, and Why It Matters for Global Capitalism. Princeton: Princeton University Press.
Alchian, A. A. (1950). Uncertainty, Evolution and Economic Theory, Journal of Political
Economy, 58 (3), 211-221.
Anderson, A. K. (2005) Affective influences on the attentional dynamics supporting awareness.
Journal of Experimental Psychology: General, 134(2), 258.
Anderson, A. K. and Phelps, E. A. (2001) Lesions of the human amygdala impair enhanced
perception of emotionally salient events. Nature, 411(6835), 305-309.
Baker, S. R., Bloom, N., Davis, S. J. (2013). Measuring Economic Policy Uncertainty. Chicago
Booth Research Paper No 13-02. Available at SSRN: http://ssrn.com/abstract=2198490
Baker, Scott R., Bloom, N., Canes-Wrone, B., Davis, S. J. and Rodden, J. (2014) Why Has US
Policy Uncertainty Risen since 1960? American Economic Review, 104(5): 56-60. DOI:
10.1257/aer.104.5.56
Bandelj, N. (2009) Emotions in economic action and interaction. Theor. Soc. 38: 34736.
Barbalet, J. (2009) A Characterization of Trust, and its Consequences. Theory and Society, 38,
367-382.
Beckert, J. (2011) Imagined Futures: Fictionality in Economic Action. MPIfG Discussion
Paper 11/ 8.
Berezin, M. (2005) Emotions and the Economy. In Smelser, N. and Swedberg, R. (eds) The
Handbook of Economic Sociology (2nd edition). New York and Princeton: Russell Sage
Foundation and Princeton University Press, pp. 109-131.
_______ (2009) Exploring emotions and the economy: new contributions from sociological
theory. Theory and Society, 38, 335-346.
Bloom, N. (2014) Fluctuations in Uncertainty. Journal of Economic Perspectives, 28(2), 153-
176. DOI: 10.1257/jep.28.2.153
Bollen, Johan; Pepe, Alberto; Mao, Huina (2009) Modelling public mood and emotion: Twitter
sentiment and socio-economic phenomena. ARXIV eprint arXiv:0911.1583
Bruner, J. (1990) Acts of Meaning, Cambridge, Harvard University Press.
32
Chatfield, C. (1995) Model Uncertainty, Data Mining and Statistical Inference, Journal of the
Royal Statistical Society, 158(3), 419-466
Choi, H., and Varian, H. (2012). Predicting the Present with Google Trends. Economic Record
Special Issue: Selected Papers from the 40th Australian Conference of Economists. Volume
88, Issue Supplement s1 pp. 29 DOI: 10.1111/j.1475-4932.2012.00809.x
Chong, K. and Tuckett, D. (2014) Constructing Conviction through Action and Narrative: How
Money Managers Manage Uncertainty and the Consequences for Financial Market
Functioning. Socio-Economic Review. 1-26. doi:10.1093/ser/mwu020
Damasio, A. (1999) The Feeling of What Happens: Body and Emotion in the Making of
Consciousness, New York, Harcourt Brace.
Dominguez, K. M. and Shapiro, M. D. (2013) Forecasting the Recovery From the Great
Recession: Is This Time Different? American Economic Review, 103(3), 147-152.
Dunn, J.C. (1973) A fuzzy relative of the ISODATA process and its use in detecting compacted
well-separated clusters, Journal of Cybernetics 3(3): 32-57, 1973
Ferré, P. (2003) Effects of level of processing on memory for affectively valenced words.
Cognition and Emotion, 17(6), 859-880.
Gilboa, I., Postelthwaite A.W., Schmeidler, D. (2008), Probability and Uncertainty in
Economic Modelling, Journal of Economic Perspectives, 22(3), 173-88
Granger, C.W.J., and Newbold, P. (1974) Spurious regressions in economics, Journal of
Econometrics 4, 111-120
Haddow, A., Hare, C., Hooley, J. and Shakir, T. (2013) Macroeconomic uncertainty: what is
it, how can we measure it and why does it matter? Bank of England Quarterly Bulletin, Quarter
2, 100-110.
Haldane, Andrew (2014) The Corridor of Uncertainty. Speech at the Scarborough Business
Ambassadors’ dinner, Scarborough.18 June 2014.
http://www.bankofengland.co.uk/publications/Documents/speeches/2014/speech738.pdf
Hall, R.E (1978). Stochastic implications of the life-cycle permanent income hypothesis:
theory and evidence, Journal of Political Economy, 86(6), 971-987
Hansen, Stephen, McMahon, Michael & Andrea Prat (2014).Transparency and Deliberation
within the FOMC: A Computational Linguistics Approach. CEP Discussion Papers dp1276,
Centre for Economic Performance, LSE.
Hatzivassiloglou, V. and McKeown, K. R. (1998) Predicting the semantic orientation of
adjectives. ACL '98 Proceedings of the 35th Annual Meeting of the Association for
Computational Linguistics and Eighth Conference of the European Chapter of the Association
for Computational Linguistics archive, 174-181.
Hendry, D. and Mizon, G. E. (2010) On the Mathematical Basis of Inter-temporal
Optimization. Economics Series Working Papers 497, University of Oxford, Department of
Economics.
33
Herbert, C., Ethofer, T., Anders, S., Junghofer, M., Wildgruber, D., Grodd, W. and Kissler, J.
(2009) Amygdala activation during reading of emotional adjectivesan advantage for pleasant
content. Social Cognitive and Affective Neuroscience, 4(1), 35-49.
Hofmann, M. J., Kuchinke, L., Tamm, S., Võ, M. L. and Jacobs, A. M. (2009). Affective
processing within 1/10th of a second: High arousal is necessary for early facilitative processing
of negative but not positive words. Cognitive, Affective, & Behavioral Neuroscience, 9(4),
389-397.
Jensen, J., Naidu, S., Kaplan, E., and Wilse-Samon, L (2012) Political Polarization and the
Dynamics of Political Language: Evidence from 130 Years of Partisan Speech. Fall. Pages 1-
81. Brookings Papers on Economic Activity.
Kaufman, L and Rousseeuw, PJ (1990) Finding Groups in Data: An Introduction to Cluster
Analysis, Wiley, New York.
Kensinger, E. A. and Schacter, D. L. (2006). Processing emotional pictures and words: Effects
of valence and arousal. Cognitive, Affective, & Behavioral Neuroscience, 6(2), 110-126.
Keynes, J.M. (1936) The General Theory of Employment, Interest and Money. London,
Macmillan and Co.
Keynes, J.M. (1937) ‘The General Theory of Employment’, The Quarterly Journal of
Economics, 51(2), 209-223.
Kousta, S. T., Vinson, D. P. and Vigliocco, G. (2009). Emotion words, regardless of polarity,
have a processing advantage over neutral words. Cognition, 112(3), 473-481.
Lane, D. and Maxfield, R. (2005) Ontological Uncertainty and Evolution. Journal of
Evolutionary Economics, 15, 3-50.
Lazer, David, Ryan Kennedy, Gary King, and Alessandro Vespignani (2014). The Parable of
Google Flu: Traps in Big Data Analysis. Science 343, no. 14 March: 1203-1205. Copy at
http://j.mp/1ii4ETo
Lougrhan, Tim and McDonald, Bill (2011) When is a Liability not a Liability? Textual
Analysis, Dictionaries, and 10-Ks. Journal of Finance, vol. 66, February 2011, 35-65.
Lucca, David O. and Trebbi, Francesco (2009) Measuring Central Bank Communication: An
Automated Approach with Application to FOMC statements. NBER Working Papers 15367,
National Bureau of Economic Research, Inc.
Manski, C (2014) Communicating Uncertainty in Official Economic Statistics NBER Working
Papers 20098, National Bureau of Economic Research, Inc.
Mar, R. A. and Oatley, K. (2008) ‘The Function of Fiction is the Abstraction and Simulation
of Social Experience’, Perspectives on Psychological Science, 3, 173–192.
Monnier, C. and Syssau, A. (2008) Semantic contribution to verbal short-term memory: Are
pleasant words easier to remember than neutral words in serial recall and serial recognition?
Memory & Cognition, 36(1), 35-42.
34
Murphy, S. T. and Zajonc, R. B. (1993) Affect, cognition, and awareness: Affective priming
with optimal and suboptimal stimulus exposures. Journal of Personality and Social Psychology,
64(5), 723.
Nyman, R., Gregory, D., Kapadia, S., Ormerod, P., Tuckett, D and Smith, R (2015) News and
narratives in financial systems: exploiting big data for systemic risk assessment. ECB
Workshop on Big Data for Forecasting and statistics, First presented Frankfurt. April 7th/8th.
To appear in the Bank of England Working papers series.
Nyman, R., Ormerod, P. (2014) Big data, socio-psychological theory, algorithmic text analysis
and predicting the Michigan Consumer Sentiment Index’ paper presented at the European
Central Bank, April 2014, available online at arxiv.org/abs/1405.5695
Nyman, R., Ormerod, P. and Tuckett, D. (2015) The potential to improve financial stress
indices by incorporating information based on psychological theory and directed algorithmic
text analysis, Working Paper, Centre for the Study of Decision Making Uncertainty, University
College London
Onatski, A and Williams, N (2003) Modeling Model Uncertainty, Journal of the European
Economics Association, 1, 1087-1122
Ormerod, P., Nyman, R., Bentley, R.A, (2014) Nowcasting economic and social data: when
and why search engine data fails, an illustration using Google Flu Trends.
http://arxiv.org/abs/1408.0699
Ormerod, P (1978) The Effect of Revisions to Central Statistical Office Data for the UK
Economy, Oxford Bulletin of Economics and Statistics, 40,2, 165-171, 1978
Pixley, J. (2009) Time Orientation and Emotion-Rules in Finance. Theory and Society, 38,
353-400.
Ramey, V.A. (2011). Can Government Purchases Stimulate the Economy? Journal of
Economic Literature, 49(3), 673-85
Ramey, Valerie A. & Shapiro, Matthew D. (1999) "Costly Capital Reallocation and the Effects
of Government Spending," NBER Working Papers 6283, National Bureau of Economic
Research, Inc.
Romer, C. D. and Romer, D. H. (2010) The Macroeconomic Effects of Tax Changes: Estimates
Based on a New Measure of Fiscal Shocks. American Economic Review, 100, 763801.
http://www.aeaweb.org/articles.php?doi=10.1257/aer.100.3.763
Scott, S. and Varian, H. (2013) Bayesian Variable Selection for Nowcasting Economic Time
Series. NBER Working Paper No. 19567
Shrestha, M and Marini, M (2013) Quarterly GDP Revisions in G-20 Countries: Evidence from
the 2008 Financial Crisis. IMF Working Paper WP/13/60. March.
Siegle, G. J., Ingram, R. E. and Matt, G. E. (2002) Affective interference: An explanation for
negative attention biases in dysphoria? Cognitive Therapy and Research, 26(1), 73-87.
35
Socher, Richard., Alex Perelygin, Jean Wu, Jason Chuang, Christopher Manning, Andrew Ng
and Christopher Potts (2104) Recursive Deep Models for Semantic Compositionality Over a
Sentiment Treebank. Conference on Empirical Methods in Natural Language Processing
(EMNLP 2013). See http://nlp.stanford.edu/sentiment/index.html.
Soo, C. (2013) Quantifying Animal Spirits: News Media and Sentiment in the Housing Market.
What type of publication and from where?
https://www.stern.nyu.edu/sites/default/files/assets/documents/con_040497.pdf
Stets, J. (2010) Future Directions in the Sociology of Emotions. Emotion Review, 2(3), 265-
268.
Sutton, T. M., Altarriba, J., Gianico, J. L. and Basnight-Brown, D. M. (2007) The automatic
access of emotion: Emotional stroop effects in spanish-english bilingual speakers. Cognition
and Emotion, 21(5), 1077-1090.
Strauss, V. M. (2013) Emotional Values of Words in Finance: Anxiety about Losses and
Excitement about Gains. M.Sc. thesis in Social Cognition, University College London.
Tetlock, Paul C., Maytal Saar-Tsechansky, and Sofus Macskassy. (2008). “More than Words:
Quantifying Language to Measure Firms’ Fundamentals.” The Journal of Finance, 63(3): pp.
14371467.
Tuckett, D (2011) ). Minding the Markets: An Emotional Finance View of Financial Instability.
Palgrave Macmillan.
Tuckett, D (2014) Uncertainty, Conflict and Divided States: Some Psychological Foundations
for Macroprudential Policy. Bank of England interdisciplinary workshop on the role of
uncertainty in central bank policy. April 2nd.
http://www.bankofengland.co.uk/research/Documents/pdf/tuckett_0414.pdf
____________Smith, R. E. and Nyman, R. (2014a) Tracking Phantastic Objects: A Computer
Algorithmic Investigation of Narrative Evolution in Unstructured Data Sources. Social
Networks. 38 (1) 121 - 133. 10.1016/j.socnet.2014.03.001
Tuckett, D.A and Nikolic, M., (2015) The Role of Conviction in Decision-Making Under
Deep Uncertainty. (Forthcoming)
Warriner, A. B., Kuperman, V., & Brysbaert, M. (2013). Norms of valence, arousal, and
dominance for 13,915 english lemmas. Behavior Research Methods, 1-17.
36
Supplementary Material
1. Wordlists
Table A1 contains a random sample of 40 anxiety words and 40 excitement words.
Note that when the same word is spelled differently in the US and the UK we have included
both variations in the list. Table A1
Randomly Drawn Selection of Words indicating excitement (about gain) and anxiety (about loss)
Anxiety Anxiety Excitement Excitement
Jitter Terrors Excited Excels
Threatening Worries Incredible Impressively
Distrusted Panics Ideal Encouraging
Jeopardized Eroding Attract Impress
Jitters Terrifying Tremendous Favoured
Hurdles Doubt Satisfactorily Enjoy
Fears Traumatised Brilliant Pleasures
Feared Panic Meritorious Positive
Traumatic Imperils Superbly Unique
Fail Mistrusts Satisfied Impressed
Erodes Failings Perfect Enhances
Uneasy Nervousness Win Delighted
Distressed Conflicted Amazes Energise
Unease Reject Energizing Spectacular
Disquieted Doubting Gush Enjoyed
Perils Fearing Wonderful Enthusiastic
Traumas Dreads Attracts Inspiration
Alarm Distrust Enthusiastically Galvanized
Distrusting Disquiet Exceptionally Amaze
Doubtable Questioned Encouraged Excelling
2. Experimental word list validation
Strauss selected forty participants (14 females and 26 males with a mean age of 37). They
were recruited via email from a list provided by a financial newsletter. Most participants were
native speakers. The ones that were not (n = 15) had spent an average of 11.87 years (SD =
8.2, range = 4 34) in an English speaking country. Many participants had obtained degrees
in accountancy, finance, economics or business administration, or were students in one of
these areas. Some of them had obtained PhDs in finance or economics. Others were
37
professional investors or money managers. 125 words (68 for anxiety, and 57 for excitement)
were derived from the word list using only the lemmas of those words i.e. the not inflected
base forms as one finds them as entries in dictionaries. For instance, the database included
“conflict”, “conflicted”, “conflicting” and “conflicts” of which only “conflict” was kept.
Previous work has established that emotional values generalize to inflected forms, which
makes it unnecessary to rate all forms separately (Warriner et al., 2013). These word were
interspersed with 100 the most frequent words found in the Reuters News Archive. The study
was conducted online using Qualtrics, a software designed for online data collection (on
Internet based research, see Birnbaum, 2004) using Burke and James’s (2006)
recommendations for giving instructions about the purpose of the study the identity of the
researchers, privacy and data confidentiality and how to proceed. Participants were explicitly
instructed to rate the words in a financial context and were presented with an excerpt of a
financial article with an exciting or anxious tone (depending on condition) and excitement
was defined for them as “a state of anticipation of pleasure and triumph, as about future
gains” and anxiety as “a state of uneasiness and apprehension, as about future uncertainties or
losses”. Additionally, some business/finance related words were randomly included to help
participants to stay in a “financial mind-set” by adding context to the ratings. Those who took
part were randomly assigned to one of two conditions: they either rated (1) words offered
them for anxiety or (2) for excitement on a virtual scale from 0 (no anxiety/excitement at all)
to 10 (strongest anxiety/excitement possible). Hence each word was rated 20 times in each
condition. During the word rating procedure, each word was presented separately on the
centre of the screen, and participants could proceed to the next word at their own pace by
clicking on a button. The order of the words was randomized for each subject to avoid order-
of-presentation effects. It took participants an average of 35 minutes to complete the survey.
Outliers were tested and excluded in a standard way by comparing each subject’s ratings to
38
the mean ratings of all participants, to create a personal correlation coefficient (Ferr et al.,
2012). As a cut-off value a correlation of r < 0.2 was chosen, four participants (20%) in the
excitement condition, and two (10%) in the anxiety condition had to be excluded from the
dataset. After excluding these subjects the means and standard deviations for both scales were
re-calculated. The average correlation of a subject’s ratings with the mean ratings of each
word was very similar in both conditions: M = 0.76 for excitement (SD = 0.145, range =
0.303 0.911) and M = 0.75 for anxiety (SD = 0.132, range = 0.372 0.894). The results
show that a word rated highly by one participant on a particular scale was usually rated
highly by other participants too. Additionally, there was a strong negative correlation
between the mean ratings for anxiety and excitement, r(262) = -0.809, p < .0001. This
indicates that participants could clearly distinguish between anxiety and excitement words.
Taken together and as expected, the study shows that the ratings across participants were very
stable, and that two distinct categories of words (excitement and anxiety) were emerging out
of the scores. (Strauss 2013).
3. Article Selection
Given the Reuters database we select the relevant documents to analyse by the use of ‘regular
expressions’ – an algorithmic concept of matching logical patterns of characters with text.
The selection of US articles can be described formally as follows.
Definition 1: Filtering
Let D be the set of all database ‘objects’ having key, value pairs (example keys are ‘title’,
‘text’, ‘date’, ‘tags’ etc.), ‘’ be logical conjunction and ‘’ be set-inclusion. Let T denote the
collection of ‘text’ fields (e.g., article body texts) within the set of database objects D. Thus,
    .
39
Given T and a text pattern r we denote the subset of T consisting of those texts containing
(matching) the pattern r by,
        .
We treat a piece of text as an ordered set of characters so that we may formally use the subset
operator . We extend the above notation by conditioning the texts on multiple properties more
generally.
Definition 2: Conditioning
For any collection C of potential values (or a given value) for a given property i (e.g. the ‘date’
property, the ‘tags’ property etc.) we define the conditional collection of texts  by
      .
The extension to two collections and is straight forward:
             .
If C is a single value as opposed to a set, the use of relational symbols different from set
inclusion ‘is clear from the particular context. These definitions allow us to formally state
how we filter for, e.g., English articles written by Reuters in the US the text collection used
in the following sections. Finally, the set of non-sports and weather articles we consider US
focused can be written as follows,
      
      
where ‘|’ is the symbol of logical disjunction for a regular expression.
40
4. Article Aggregation
The articles in the Reuters database are available to us at a daily frequency. For the purpose
of this study we have aggregated the articles quarterly, but any frequency could be used (e.g.,
daily, weekly or monthly). In other words, the emotion statistic introduced in the main text,


is computed over all articles published in a given quarter. This is computed over all quarters
for which articles were available to us, 1996Q1 through 2014Q3. The next section explains
how words are counted over all articles in a given quarter to produce the above statistic.
5. Article Tokenization
In order to count the frequency of the words in our emotion dictionaries we carry out a simple
tokenization strategy. We split each article into a ‘bag-of-words’ (i.e., an unordered set of
words) using the following procedure:
1. Convert the full article into lowercase letters only (to match our lists of lowercase
emotion words)
2. Remove each occurrence of the quotation marks ’ and `
3. Replace each non alphabetic character by a single space character
4. Split the text into words whenever we encounter a sequence of at least one whitespace
character (including newlines, tabs and spaces)
5. Remove any remaining whitespace before or after the resulting words
Technically, we achieve steps 2-5 by replacing each match of the regular expression ‘['`]’ by
the empty string ‘’, and then replacing each match of the regular expression ‘[^a-zA-Z]+’ by
the space character ‘ ‘, and finally splitting the text at each match of the regular expression
\s+’. From the set of remaining words we then count how many matches there are with the
two emotion word dictionaries. The relevant anxiety and excitement word counts are
aggregated over the articles in the given period, which in this case is a given quarter. We can
41
then compute the relative sentiment score as described in the main text. This can be coded in
the Scala programming language as follows (this is not the most efficient procedure that we
actually use, but it gives the same result and is arguably easier to understand):
val anxiety : HashSet[String] = … // this is our set of anxiety words
val excitement : HashSet[String] = … // this is our set of excitement words
val articles : List[String] = … // this is our list of articles for a given period
val excitementCount : Double = articles.map { article =>
val tokens = article. toLowerCase.replaceAll("""['`]""", "").replaceAll("""[^a-zA-
Z]+""", " ").split ("""\s+""")
tokens.filter(excitement.contains(_)).length
} .sum
val anxietyCount : Double = articles.map { article =>
val tokens = article. toLowerCase.replaceAll("""['`]""", "").replaceAll("""[^a-zA-
Z]+""", " ").split ("""\s+""")
tokens.filter(anxiety.contains(_)).length
} .sum
val sentimentScore : Double = (excitementScore anxietyScore)/articles.length
6. Emotion Word Negation
We test for the potential impact of negation on the movements in the RSS series for the UK
and the US. We apply the simple method of negation detection reported in Loughran and
McDonald (2011). We proceed by excluding any emotion word found in the text that is
preceded, within three words, by either of the words ‘no’, ‘not’, ‘none’, ‘neither’, ‘never’ or
‘nobody’. In other words we do not consider the word at all if this is the case, as opposed to
changing the meaning of the word and treating it as belonging to the ‘opposite’ category.
The negation-modified series for the UK and the US remain correlated with the
original series as highly as 0.999 in level form and 0.999 in difference form. Although this
might initially seem counterintuitive, on second thought it is in fact trivial to understand that
negation will only effect the movement of the series if there is a systematic bias in its use for
a given period. In other words, if for a period of time a given word is more likely to be
negated than at other times. Negation could have an effect on the overall level of the series, if
for example excitement words are more likely negated than anxiety words. However, since
42
we are concerned with the movements of the series over time, as opposed to the actual levels,
this does not concern us.
7. Orthogonality of wordlist
We test the hypothesis that RSS is orthogonal to the economy, and as such to fundamental
news. We exclude all words in our excitement and anxiety lists that could potentially have
economic meaning independently of emotional connotation. The words ‘uncertain’ and
‘uncertainty’ from the anxiety dictionary and the words ‘boost’, ‘boosted’, ‘boosts’,
‘exuberance’ and ‘exuberant’ from the excitement dictionary were excluded on this basis and
a new RSS series produced using the remaining words. The new series remains correlated
with the original RSS series at 0.99 in level form and 0.99 in difference form. We can
conclude that the RSS measure is likely not effected by the presence of economic terms.
8. Granger causality R Packages
The packages in R used in the Toda-Yamamoto procedure to investigate Granger causality are
as follows:
tseries we use the two functions adf.test and kpss.test (the Augmented Dickey-Fuller
test and Kwiatkowski-Phillips-Schmidt-Shin test respectively) to check if series are
stationary or contain unit roots. The adf.test function allows you to define the alternative
hypothesis by the alternative argument. We use the default option of “stationary”.
The function also allows you to manually specify the lag order k to calculate the test
statistic. We use the default option     
, where N is the length of the
series and trunc is a function built into R truncating the value towards zero
vars we use the function VARselect to compute the Akaike Information Criteria for
VAR(p) processes with p from 1 through 20. The function takes a number of arguments.
We make use of the “lag.max” argument, which we set to 15 and the “type” argument
43
which we set to “const”, indicating that information criteria should be computed for
lags from 1 through 15 and that a constant term should be included in the VAR,
respectively. We use the VAR function for estimating a VAR(p) process. Similarly to
VARselect we use the “p” argument specifying the number of lags to include and the
“type” argument specifying whether to include a constant term, or a trend or both. In
all cases we set this argument to “const”, indicating that only a constant term should be
included. We use the function serial.test to compute the multivariate Portmanteau- and
Breusch-Godfrey tests for serially correlated errors in a VAR(p) process. We use the
default number of lags for each test. In the case of the Portmanteau test we keep the
default value of the “lags.pt” argument at 16 and in the case of the Breusch-Godfrey
test we keep the default value of the “lags.bg” argument at 5. We set the “type”
parameter to either “PT.asymptotic” or “BG” to compute the Portmanteau- or Breusch-
Godfrey test respectively. We use the function stability to compute empirical
fluctuation processes according to the OLS-CUSUM method. We use the default values
of each argument, in particular the “type” argument which defaults to “OLS-CUSUM”
for the OLS-CUSUM method. The figures for the empirical fluctuation processes are
generated by the use of the built in plot function on the returned object from the call to
stability
aod we use the function wald.test to perform the Wald tests for granger causality.
We use three of the function’s available arguments. The argument “Terms” specifying
which terms of the model to include in the null hypothesis of the Wald test, given as a
vector of term indices. The argument “b” specifying a vector of the coefficients of the
model. The argument “Sigma” specifying the variance-covariance matrix of the model.
To specify the values of the latter two arguments, we use the coef method on the
relevant equation from the VAR to extract the relevant coefficients and the vcov method
44
on the relevant equation from the VAR to extract the relevant variance-covariance
matrix
Here is an illustration of the full procedure when testing for Granger causality between RSS
and DLGDP.
# create time series variables from vectors
rss <- ts(RSS, start=c(1996,1), freq=4)
dlgdp <- ts(DLGDP, start=c(1996,1), freq=4)
data <- cbind(dlgdp, rss)
colnames(data) <- c(“DLGDP”, “RSS”)
# test variables for order of integration
adf.test(rss)
kpss.test(rss)
adf.test(diff(rss))
kpss.test(diff(rss))
# similarly for dlgdp. Let m be the maximum order of integration found
# compute information criteria for VAR models of different lags
VARselect(data, lag=15, type=”const”)
# test residuals of chosen lag p
serial.test(VAR(data, p=p, type=”const”))
serial.test(VAR(data, p=p, type=”const”), type=”BG”)
plot(stability(VAR(data, p=p, type=”const”)))
# if OK proceed and create a new VAR model adding m lags.
var.model <- VAR(data, p=(p+m), type=”const”)
# perform a Wald test on the first p lags of the independent variable for both equations.
# E.g., if p =1 and we test Granger causality from RSS to DLGDP:
wald.test(b=coef(var.model$varresult[[1]]), Sigma=cvov(var.model$varresult[[1]]),
Terms=c(2))
# and to test the second equation, Granger causality from DLGDP to RSS:
wald.test(b=coef(var.model$varresult[[2]]), Sigma=cvov(var.model$varresult[[2]]),
Terms=c(1))
# note that the independent variables are ordered by the number of lags
45
9. Linear models, R packages
We use the lm function available in the stats package to fit linear models. The specifications of
the models are tested using functions from lmtest.
lmtest - we use the function bgtest to perform the Breusch-Godfrey test for higher order
serial correlation. We set the “order” argument equal to 4 and leave the values of other
arguments to their default setting. We use the function bptest to perform the Breusch-
Pagan test against heteroskedasticity. All arguments except for the formula are kept at
default values. We use the dwtest function to perform the Durbin-Watson test for
autocorrelation of disturbances. All arguments except for the formula are kept at default
values. We use the reset function to perform the RESET test for functional form. All
arguments except for the formula are kept at default values.
Stats we use the ks.test function from the stats package to perform the Kolmogorov-
Smirnov test for normality. The function is used to test the residuals of a model for
normality by calling ks.test(form$res, pnorm, sd=sd(form$res))
10. Fuzzy Clustering, R packages
We use the cluster package to perform the fuzzy clustering analysis. We use the fanny function
to compute the fuzzy clustering of a data matrix. We use the argument “k” specifying the
desired number of clusters. We can estimate the number of clusters using the Dunn coefficient
of the resulting clustering. We created the following function to determine a suitable number
of clusters
cluster <- function (data, min, max) {
index <- c()
dunn <- c()
for (k in min:max) {
46
index <- c(index, k)
cl <- fanny(data, k)
dunn <- c(dunn, cl$coeff[['normalized']])
print(c(k, cl$coeff[['normalized']]))
}
return (index[which.max(dunn)])
}
We run this with “min” set to 2 and “max” set to 10.
11. Supplementary Tables
Table A2
Regression of the third estimates of quarterly growth in real US GDP (DLGDP) on the consensus
forecast made in the previous quarter (SPF) and the relative sentiment shift series (RSS)
Dependent variable:
---------------------------
DLGDP
SPF 0.760***
(0.271)
RSS 0.793***
(0.266)
Constant 0.722
(0.737)
Observations 73
R2 0.278
Adjusted R2 0.258
Residual Std. Error 2.115 (df = 70)
F Statistic 13.485*** (df = 2; 70)
Note: *p<0.1; **p<0.05; ***p<0.01
Table A3
Augmented Dickey-Fuller and Kwiatkowski-Phillips-Schmidt-Shin tests for stationarity of RSS,
DLGDP, DLK and DLCE
47
Variable ADF p-value KPSS p-value
RSS Level -2.31 0.45 1.63 < 0.01***
RSS Diff -4.10 < 0.01*** 0.05 > 0.1
DLGDP Level -2.88 0.22 0.82 < 0.01***
DLGDP Diff -5.06 < 0.01*** 0.02 > 0.1
DLK Level -2.17 0.51 0.52 0.04**
DLK Diff -3.81 0.023** 0.05 > 0.1
DLCE Level -2.76 0.27 1.42 < 0.01***
DLCE Diff -4.08 0.01*** 0.03 > 0.1
Note: *p<0.1; **p<0.05; ***p<0.01
The Augmented Dickey-Fuller tests were carried out with lag order 4 and the Kwiatkowski-Phillips-
Schmidt-Shin tests with truncation lag parameter 1.
Table A4
Multivariate Portmanteau- and Breusch-Godfrey tests for serially correlated errors of VAR models
involving RSS, DLGDP, DLK and DLCE
VAR model Portmanteau d.f. Breusch-Godfrey d.f.
RSS/DLGDP 44.22 60 12.04 20
RSS/DLK 53.63 60 19.71 20
RSS/DLCE 42.33 52 18.97 20
Note: *p<0.1; **p<0.05; ***p<0.01
Table A5
Augmented Dickey-Fuller and Kwiatkowski-Phillips-Schmidt-Shin tests for stationarity of ANX
Variable ADF p-value KPSS p-value
ANX Level -2.94 0.19 0.29 > 0.1
ANX Diff -3.93 0.018** 0.034 > 0.1
Note: *p<0.1; **p<0.05; ***p<0.01
The Augmented Dickey-Fuller tests were carried out with lag order 4 and the Kwiatkowski-Phillips-
Schmidt-Shin tests with truncation lag parameter 1.
48
Table A6
Multivariate Portmanteau- and Breusch-Godfrey tests for serially correlated errors of VAR models
involving RSS, ANX and ANXA
VAR model Portmanteau d.f. Breusch-Godfrey d.f.
RSS/ANX 48.06 52 25.78 20
RSS/ANXA 48.11 56 19.07 20
Note: *p<0.1; **p<0.05; ***p<0.0
Table A7
Augmented Dickey-Fuller and Kwiatkowski-Phillips-Schmidt-Shin tests for stationarity of EPU
Variable ADF p-value KPSS p-value
EPU Level -1.76 1.68 0.67 < 0.01***
EPU Diff -3.6 0.08* 0.04 > 0.1
Note: *p<0.1; **p<0.05; ***p<0.01
The Augmented Dickey-Fuller tests were carried out with lag order 4 and the Kwiatkowski-Phillips-
Schmidt-Shin tests with truncation lag parameter 1.
Table A8
Multivariate Portmanteau- and Breusch-Godfrey tests for serially correlated errors of VAR models
involving RSS and EPU
VAR model Portmanteau d.f. Breusch-Godfrey d.f.
RSS/EPU 46.94 56 27.96 20
Note: *p<0.1; **p<0.05; ***p<0.01
Table A9
Augmented Dickey-Fuller and Kwiatkowski-Phillips-Schmidt-Shin tests for stationarity of UNCERT
and ECON
Variable ADF p-value KPSS p-value
UNCERT Level -2.09 0.54 0.8 < 0.01
UNCERT Diff -4.36 0.01 0.04 > 0.1
49
ECON Level -2.33 0.44 1.44 < 0.01
ECON Diff -4.19 0.01 0.03 > 0.1
Note: *p<0.1; **p<0.05; ***p<0.01
The Augmented Dickey-Fuller tests were carried out with lag order 4 and the Kwiatkowski-Phillips-
Schmidt-Shin tests with truncation lag parameter 1.
Table A10
Multivariate Portmanteau- and Breusch-Godfrey tests for serially correlated errors of VAR models
involving RSS, UNCERT and ECON
VAR model Portmanteau d.f. Breusch-Godfrey d.f.
RSS/UNCERT 62.18 60 15.89 20
RSS/ECON 52.21 56 19.22 20
Note: *p<0.1; **p<0.05; ***p<0.01
Table A11
Multivariate Portmanteau- and Breusch-Godfrey tests for serially correlated errors of VAR models
involving EPU, UNCERT, ECON and DLGPD
VAR model Portmanteau d.f. Breusch-Godfrey d.f.
EPU/DLGDP 60.95 60 21.23 20
UNCERT/DLGDP 66.93 60 26.22 20
ECON/DLGDP 51.64 60 18.75 20
UNCERT/EPU 50.84 56 23.09 20
Note: *p<0.1; **p<0.05; ***p<0.01
Table A12
Akaike Information Criteria (AIC) for VAR models with RSS and the corresponding variables of the
table. The AIC score for the selected number of lags to use in the corresponding VAR is highlighted
LAGS DLGDP DLK DLCE ANX ANXA EPU UNCERT ECON
1 -11.236 2.565 -12.199 3.346 3.363 4.575 -2.190 -2.211
50
2 -11.199 2.598 -12.142 3.382 3.169 4.672 -2.082 -2.229
3 -11.081 2.702 -12.176 3.257 3.227 4.594 -2.032 -2.150
4 -10.977 2.758 -12.095 3.276 3.272 4.660 -2.028 -2.087
5 -10.905 2.840 -12.034 3.294 3.382 4.680 -1.943 -2.031
6 -10.856 2.828 -12.138 3.336 3.318 4.709 -1.825 -1.968
7 -10.823 2.869 -12.077 3.389 3.386 4.658 -1.760 -1.966
8 -10.698 2.986 -12.000 3.451 3.516 4.729 -1.825 -1.944
9 -10.730 3.006 -11.918 3.413 3.422 4.691 -1.917 -1.983
10 -10.618 3.043 -11.796 3.445 3.428 4.719 -2.260 -1.909
11 -10.670 3.136 -11.660 3.514 3.517 4.718 -2.237 -1.787
12 -10.622 3.251 -11.625 3.590 3.531 4.809 -2.137 -1.691
13 -10.606 3.248 -11.535 3.545 3.565 4.737 -2.211 -1.735
14 -10.573 3.352 -11.558 3.636 3.654 4.792 -2.160 -1.695
15 -10.466 3.315 -11.610 3.654 3.668 4.900 -2.065 -1.745
Note: The number of lags with the lowest AIC score is not selected if the null hypothesis of no serial
correlation in the residuals of the corresponding VAR is rejected at the 10% level by the Portmanteau-
or Breusch-Godfrey test. The VAR with UNCERT and RSS gets smaller AIC scores for lags 10, 11
and 13, but residuals show no signs of serial correlation with only 1 lag, so this is selected
Table A13
Akaike Information Criteria (AIC) for VAR models involving UNCERT, DLGDP, EPU and ECON.
The AIC score for the selected number of lags to use in the corresponding VAR is highlighted
LAGS UNCERT/DLGDP UNCERT/EPU EPU/DLGDP ECON/DLGDP
1 -10.559 4.887 -4.064 -11.134
2 -10.555 4.911 -4.055 -11.070
3 -10.512 4.983 -4.001 -11.024
4 -10.499 5.089 -3.921 -10.936
5 -10.501 4.986 -3.820 -10.854
6 -10.438 5.048 -3.875 -10.809
7 -10.401 5.138 -3.875 -10.727
8 -10.281 5.246 -3.766 -10.644
9 -10.251 5.253 -3.675 -10.629
10 -10.355 5.166 -3.779 -10.576
11 -10.345 5.237 -3.972 -10.725
12 -10.263 5.287 -3.940 -10.600
13 -10.311 5.300 -3.855 -10.592
51
14 -10.360 5.366 -4.008 -10.535
15 -10.324 5.373 -3.894 -10.464
Note: The number of lags with the lowest AIC score is not selected if the null hypothesis of no serial
correlation in the residuals of the corresponding VAR is rejected at the 10% level by the Portmanteau-
or Breusch-Godfrey test
Table A14
Regression of the actual values of quarterly growth in real US GDP (DLGDP) on the consensus
forecast made in the previous quarter (SPF)
Dependent variable:
---------------------------
DLGDP
SPF 1.123***
(0.292)
Constant -0.430
(0.804)
Observations 74
R2 0.170
Adjusted R2 0.159
Residual Std. Error 2.469 (df = 72)
F Statistic 14.770*** (df = 1; 72)
Note: *p<0.1; **p<0.05; ***p<0.01
Table A15
Quarterly time series from 1996Q1 through 2014Q3 extracted from RTRS; RSS, UNCERT and
ECON
RSS UNCERT ECON RSS Cont. UNCERT Cont. ECON Cont.
1.192 -1.198 0.847 0.770 -0.543 0.899
1.344 -1.264 0.946 0.748 -0.791 0.862
0.979 -1.118 0.895 0.557 -0.832 0.906
1.560 -1.310 0.948 0.479 -0.363 0.928
1.042 -1.050 0.910 0.704 -0.706 0.834
1.299 -0.831 0.904 1.346 -1.210 0.926
1.442 -1.129 0.846 0.770 -1.147 0.629
0.863 -0.834 0.479 0.618 -1.260 0.869
0.839 -0.474 -0.109 -0.267 -0.805 0.411
52
0.659 -0.777 0.298 -0.054 -0.437 0.073
-0.421 0.546 -0.292 -0.280 -0.634 -1.109
0.728 -0.320 -0.037 -0.090 -0.741 -0.299
0.578 -0.683 0.444 -1.670 0.973 -0.838
0.878 -0.703 0.359 -1.516 0.957 -3.949
0.843 -0.674 0.552 -1.109 0.340 -2.562
0.820 -0.659 0.732 -0.594 -0.012 -1.742
1.080 -0.434 0.655 0.133 -0.361 -1.184
0.644 -0.172 0.803 0.232 -0.603 -0.747
0.540 -0.395 0.614 -0.272 0.645 -0.651
-0.199 1.771 0.518 -1.459 0.803 -1.666
-0.062 0.152 -0.198 -1.013 1.377 -0.761
0.031 0.306 0.106 -0.198 0.769 -0.368
-1.281 1.110 -0.282 -1.021 0.339 -0.701
-0.568 1.691 -0.390 -1.533 0.557 -0.459
-0.468 0.624 -0.246 -2.578 2.063 -2.279
-0.897 0.874 0.312 -2.151 1.803 -2.820
-1.611 1.198 0.101 -1.096 0.187 -0.824
-0.871 1.540 0.491 -1.731 1.433 -1.525
-1.199 2.084 0.397 -1.229 1.404 -0.763
0.515 0.356 0.660 -1.118 3.018 -0.643
1.092 -0.894 0.814 -0.798 0.973 -0.175
1.128 -0.923 0.878 -0.721 -0.072 0.283
1.130 -0.950 0.897 -0.367 0.376 -0.019
0.966 -0.790 0.979 -0.791 1.341 -0.055
0.714 -0.609 0.807 -0.834 -0.051 -0.488
1.038 -0.178 0.831 -0.031 -0.596 -0.076
1.123 -1.023 0.914 -0.087 -0.352 -0.250
0.760 -0.699 0.949
Table A16
Dunn coefficients for different number of fuzzy clusters
Clustering dataset is DLGDP and RSS; Sample period: 1996Q1 through 2014Q3
Number of clusters Dunn coefficient
2 0.5104659
3 0.4738599
4 0.4371471
53
5 0.4457091
6 0.4071845
7 0.4115111
8 0.4210204
9 0.4098577
10 0.4204649
Table A17
Degrees of membership of the two fuzzy clusters
Clustering dataset is DLGDP and RSS; Sample period: 1996Q1 through 2014Q3
Quarter Cluster 1 Cluster 2
1996 Q1 0.88685490 0.11314510
1996 Q2 0.84685511 0.15314489
1996 Q3 0.93422573 0.06577427
1996 Q4 0.79521342 0.20478658
1997 Q1 0.92415453 0.07584547
1997 Q2 0.85845158 0.14154842
1997 Q3 0.82210916 0.17789084
1997 Q4 0.94757708 0.05242292
1998 Q1 0.94879506 0.05120494
1998 Q2 0.91754430 0.08245570
1998 Q3 0.22704999 0.77295001
1998 Q4 0.93834001 0.06165999
1999 Q1 0.88287754 0.11712246
1999 Q2 0.94620064 0.05379936
1999 Q3 0.94868900 0.05131100
1999 Q4 0.94839083 0.05160917
2000 Q1 0.91630083 0.08369917
2000 Q2 0.91148600 0.08851400
2000 Q3 0.86272802 0.13727198
2000 Q4 0.35651414 0.64348586
2001 Q1 0.45175847 0.54824153
2001 Q2 0.51985715 0.48014285
2001 Q3 0.11001407 0.88998593
2001 Q4 0.16258030 0.83741970
2002 Q1 0.20461811 0.79538189
2002 Q2 0.08320127 0.91679873
54
2002 Q3 0.16768603 0.83231397
2002 Q4 0.08499398 0.91500602
2003 Q1 0.09640159 0.90359841
2003 Q2 0.84865231 0.15134769
2003 Q3 0.91373523 0.08626477
2003 Q4 0.90507415 0.09492585
2004 Q1 0.90461570 0.09538430
2004 Q2 0.93596331 0.06403669
2004 Q3 0.93517246 0.06482754
2004 Q4 0.92473046 0.07526954
2005 Q1 0.90640869 0.09359131
2005 Q2 0.94464309 0.05535691
2005 Q3 0.94593337 0.05406663
2005 Q4 0.94281427 0.05718573
2006 Q1 0.87206226 0.12793774
2006 Q2 0.82664593 0.17335407
2006 Q3 0.93196732 0.06803268
2006 Q4 0.84637565 0.15362435
2007 Q1 0.94571317 0.05428683
2007 Q2 0.90103351 0.09896649
2007 Q3 0.31212842 0.68787158
2007 Q4 0.45811162 0.54188838
2008 Q1 0.30452200 0.69547800
2008 Q2 0.43181222 0.56818778
2008 Q3 0.17874308 0.82125692
2008 Q4 0.15047175 0.84952825
2009 Q1 0.08499362 0.91500638
2009 Q2 0.15296599 0.84703401
2009 Q3 0.59369138 0.40630862
2009 Q4 0.66358518 0.33641482
2010 Q1 0.30920546 0.69079454
2010 Q2 0.14065715 0.85934285
2010 Q3 0.08015810 0.91984190
2010 Q4 0.35717008 0.64282992
2011 Q1 0.08024467 0.91975533
2011 Q2 0.15322801 0.84677199
2011 Q3 0.29851775 0.70148225
55
2011 Q4 0.25426107 0.74573893
2012 Q1 0.08375086 0.91624914
2012 Q2 0.19026221 0.80973779
2012 Q3 0.10104620 0.89895380
2012 Q4 0.08569688 0.91430312
2013 Q1 0.09476221 0.90523779
2013 Q2 0.11363785 0.88636215
2013 Q3 0.25525868 0.74474132
2013 Q4 0.09619544 0.90380456
2014 Q1 0.08919172 0.91080828
2014 Q2 0.47496301 0.52503699
2014 Q3 0.43436688 0.56563312
12. Supplementary Figures
Figure A1
OLS-CUSUM plot for the selected VAR with RSS and DLGDP
56
Figure A2
OLS-CUSUM plot for the selected VAR with RSS and DLK
57
58
Figure A3
OLS-CUSUM plot for the selected VAR with RSS and DLCE
59
Figure A3
OLS-CUSUM plot for the selected VAR with RSS and ANX
60
Figure A4
OLS-CUSUM plot for the selected VAR with RSS and ANXA
61
Figure A5
OLS-CUSUM plot for the selected VAR with RSS and EPU
62
Figure A6
OLS-CUSUM plot for the selected VAR with RSS and UNCERT
63
Figure A7
OLS-CUSUM plot for the selected VAR with RSS and ECON
64
Figure A8
OLS-CUSUM plot for the selected VAR with EPU and DLGDP
65
Figure A9
OLS-CUSUM plot for the selected VAR with UNCERT and DLGDP
66
Figure A10
OLS-CUSUM plot for the selected VAR with ECON and DLGDP
67
Figure A11
OLS-CUSUM plot for the selected VAR with UNCERT and EPU
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
We use the digitized Congressional Record and the Google Ngrams corpus to study the polarization of political discourse and the diffusion of political language since 1873. We statistically identify highly partisan phrases from the Congressional Record and then use these to impute partisanship and political polarization to the Google Books corpus between 1873 and 2000. We find that although political discourse expressed in books did become more polarized in the late 1990s, polarization remained low relative to the late 19th and much of the 20th century. We also find that polarization of discourse in books predicts legislative gridlock, but polarization of congressional language does not. Using a dynamic panel data set of phrases, we find that polarized phrases increase in frequency in Google Books before their use increases in congressional speech. Our evidence is consistent with an autonomous effect of elite discourse on congressional speech and legislative gridlock, but this effect is not large enough to drive the recent increase in congressional polarization.
Article
Full-text available
Financial assets are abstract entities. Their value depends on beliefs which are inherently social and, we argue, emotional. Recent events have revealed profound uncertainty at the heart of financial markets, the manifest existence of emotion and the way confidence is crucial to orderly market functioning. Using findings from two interview studies, supported by ethnographic observation, we elaborate on the irreducible cognitive and emotional conflicts which face actors engaged in financial markets and threaten their daily operations. We introduce the term conviction narrative to analyse how they manage these conflicts on a day-to-day basis, and with what collective consequences. Our thesis is that expertise and conviction in financial markets have constantly to be created and renewed through a combination of psychological and social action with the implication at the macro level that while financial markets can be orderly they are so in an intrinsically fragile way.
Article
Full-text available
Obtaining an accurate picture of the current state of the economy is particularly important to central banks and finance ministries, and of epidemics to health ministries. There is increasing interest in the use of search engine data to provide such 'nowcasts' of social and economic indicators. However, people may search for a phrase because they independently want the information, or they may search simply because many others are searching for it. We consider the effect of the motivation for searching on the accuracy of forecasts made using search engine data of contemporaneous social and economic indicators. We illustrate the implications for forecasting accuracy using four episodes in which Google Flu Trends data gave accurate predictions of actual flu cases, and four in which the search data over-predicted considerably. Using a standard statistical methodology, the Bass diffusion model, we show that the independent search for information motive was much stronger in the cases of accurate prediction than in the inaccurate ones. Social influence, the fact that people may search for a phrase simply because many others are, was much stronger in the inaccurate compared to the accurate cases. Search engine data may therefore be an unreliable predictor of contemporaneous indicators when social influence on the decision to search is strong.
Book
This book was originally published by Macmillan in 1936. It was voted the top Academic Book that Shaped Modern Britain by Academic Book Week (UK) in 2017, and in 2011 was placed on Time Magazine's top 100 non-fiction books written in English since 1923. Reissued with a fresh Introduction by the Nobel-prize winner Paul Krugman and a new Afterword by Keynes’ biographer Robert Skidelsky, this important work is made available to a new generation. The General Theory of Employment, Interest and Money transformed economics and changed the face of modern macroeconomics. Keynes’ argument is based on the idea that the level of employment is not determined by the price of labour, but by the spending of money. It gave way to an entirely new approach where employment, inflation and the market economy are concerned. Highly provocative at its time of publication, this book and Keynes’ theories continue to remain the subject of much support and praise, criticism and debate. Economists at any stage in their career will enjoy revisiting this treatise and observing the relevance of Keynes’ work in today’s contemporary climate.
Book
The global financial crisis has made it painfully clear that powerful psychological forces are imperiling the wealth of nations today. From blind faith in ever-rising housing prices to plummeting confidence in capital markets, "animal spirits" are driving financial events worldwide. In this book, acclaimed economists George Akerlof and Robert Shiller challenge the economic wisdom that got us into this mess, and put forward a bold new vision that will transform economics and restore prosperity. Akerlof and Shiller reassert the necessity of an active government role in economic policymaking by recovering the idea of animal spirits, a term John Maynard Keynes used to describe the gloom and despondence that led to the Great Depression and the changing psychology that accompanied recovery. Like Keynes, Akerlof and Shiller know that managing these animal spirits requires the steady hand of government--simply allowing markets to work won't do it. In rebuilding the case for a more robust, behaviorally informed Keynesianism, they detail the most pervasive effects of animal spirits in contemporary economic life--such as confidence, fear, bad faith, corruption, a concern for fairness, and the stories we tell ourselves about our economic fortunes--and show how Reaganomics, Thatcherism, and the rational expectations revolution failed to account for them.