Journal of Applied Economics. Vol XX, No.

1 (May 2017), 49-73

CONVENTIONAL VIEWS AND ASSET PRICES: WHAT TO

EXPECT AFTER TIMES OF EXTREME OPINIONS?

J. Daniel Aromí *
IIEP Universidad de Buenos Aires

Submitted April 2015; accepted March 2016

This study evaluates the performance of stock market indices after times of extreme opinions.
The underlying conjecture is that extreme opinions are associated to overreactions in the
perception of wealth. The analysis covers 34 countries from 1988 through 2013. In a novel
approach, views regarding economic performance are approximated using content in the
global economic press. Consistent with the overreaction conjecture, stock market indices
are shown to under-perform following extreme optimistic views and over-perform after
pessimistic views. A long-short contrarian portfolio earns 11% annually over the next five
years. This persistent and predictable difference in returns cannot be explained by risk
considerations and cannot be replicated using alternative strategies based on past returns or
past economic growth.

JEL classification codes: G12, G17, D84

Key words: asset prices, opinions, expectations, overreaction.

I. Introduction

In November 2009, the weekly magazine The Economist ran a cover in which the
title “Brazil takes off” was accompanied by a statue of Christ the Redeemer ascending
like a rocket from Rio de Janeiro’s Corcovado mountain. This strong sign of optimism
was later reversed in September 2013 when the cover asked “Has Brazil blown it?”

* Daniel Aromí: Universidad de Buenos Aires, Facultad de Ciencias Económicas, Instituto Interdisciplinario de Economia

Política (IIEP-BAIRES), Av. Córdoba 2122 - 2do (C1120 AAQ), Ciudad Autónoma de Buenos Aires, Argentina; email
aromi.daniel@gmail.com. We would like to thank Daniel Heymann for generous and fruitful discussions. Also, we
would to thank Daniel Chodos for providing data and Marcos Dal Bianco, Luciano Del Corro, and participants at the
2014 AAEP meeting and the seminars held at IIEP-Baires, UCEMA and Universidad del Sur for helpful comments.
50 Journal of Applied Economics

together with a picture of a collapsing statue-rocket.1 As eloquently exemplified by

the pair of covers, opinions regarding economic prospects are regularly seen to
enter stages of high hopes or, in other occasions, periods of intense gloom. Determining
the accuracy of these emergent judgments is a matter of interest. This is because
important economic decisions by private and public actors can benefit from a better
understanding of the information conveyed by conventional views.
Conventional views are understood as a set of beliefs that are broadly shared,
are known to be shared and so on. In this work, the focus is placed on broadly shared
views regarding economic conditions. For example, beliefs regarding future economic
prosperity or the evolution of aggregate productivity. Importantly, the public condition
of this set of beliefs implies that they are observable and can be approximated, for
example, analyzing contents in the economic press.
The underlying assumption behind this work is that the evolution of these public
beliefs are strongly linked to the path of asset prices in an economy. More specifically,
one plausible conjecture is that extreme conventional assessments are associated
to excessive responses. Under this conjecture, the occasional emergence of extreme
shared opinions could be linked to mispricing of broad classes of assets and predictable
errors in saving and investment decisions. Despite its relevance, formal empirical
evidence of this conjecture is hindered by lack of sufficiently comprehensive and
precise measures of conventional views.
In this work, this conjecture is empirically evaluated for the case of financial
assets. The performance of stock market indices is evaluated after times of extreme
optimism and extreme pessimism. The study covers 34 countries from 1988 through
2013. One distinctive aspect of this work is the approximation of conventional
views using content published in the international economic press.
Consistent with the postulated conjecture, the results show that optimism is
followed by lower mean returns and pessimism by higher mean returns. This
difference in performance is highly persistent and economically significant. A long-
short contrarian portfolio earns 11% annually over the next five years. Additionally,
it is found that the performance of sentiment based portfolio strategies cannot be
replicated using information on past returns or past economic growth. Finally, the
findings suggest that changes in anticipated risk levels are not a good explanation
of the reported return differentials.
A natural interpretation of these findings is that the occasional emergence of
conventional views regarding economic prospects generates mispricing for a broad

1 See Economist (2009) and Economist (2013).

 Conventional views and asset prices 51

set of assets. Contributions associated to social learning can help rationalize these
events. The well-established literature on social learning has shown that information
can be aggregated quite inefficiently.2 Complementarily, social learning can lead
to more severe inefficiencies if agents follow simple rules that ignore redundancies
in public information3 or if a subset of actors is too influential.4 Additionally, the
existence of return predictability based on publicly available information suggests
the presence of individuals that process information in an incomplete and correlated
manner.5
The recurrent, persistent and predictable overreactions documented in this study
have important implications that go beyond financial markets. Given the key function
of asset prices in the aggregation of information and the coordination of actions,
the results are relevant for the understanding of macroeconomic dynamics and
public policy that aims for stability.
One key assumption of the current study is the idea that conventional views
regarding economic prospects can be approximated processing information in the
economic press. This idea is supported by the dual role of the media. The economic
press publishes information that reflects and, at the same time, shapes public opinion.
It is worth noting that press content is selected based on journalists’ or editors’ beliefs
regarding dominant public opinions. In this context, confirming public opinions might
be a profitable marketing strategy.6 Additionally, it is reasonable to assume that an
important fraction of the content is forward looking. For example, the inspection of
some distant episode might be linked to its relevance given current opinions.
The reported findings show that empirical studies that focus on subjective states
can advance the understanding of economic dynamics. This is a natural outcome
once it is recognized that economic processes emerge as the result of the co-evolution
of structural and subjective elements. According to this perspective, subjective
elements cannot be inferred through logical deductions and there is value in finding
estimations of these elements that can allow for new insights.
This study is related to a fertile body of theoretical and empirical contributions
that calls attention to the role of expectations, learning and coordination in financial

2 See for example Banerjee (1992) and Bikhchandani et al. (1992).

3 Eyster and Rabin (2010, 2014).
4 Golub and Jackson (2010).
5 Models of cognitive limits or simple responses can be found, for example, in Mullainathan (2002), Hong and Stein
(1999) and Brock and Hommes (1998). Conceptual analyses in this line can also be found in Kindleberger and Aliber
(2011) and Shiller (2005).
6 See for example Mullainathan and Shleifer (2005) and Gentzkow and Shapiro (2010).
52 Journal of Applied Economics

and macroeconomic dynamics. Using theoretical models, learning dynamics about

structural parameters have been explored as a source of aggregate fluctuations.7
Additional sources of non-fundamental volatility are suggested by models with
strategic complementarities and multiple equilibria. In these analyses, under-
determination of equilibrium can be interpreted as an indication that, inevitably,
beliefs are a distinct element in the determination of economic outcomes.8
Complementarily, the literature on herd behavior has shown that information can
be aggregated inefficiently and, as a result, aggregate beliefs and behavior can be
quite idiosyncratic.9 In financial markets, models of limits to arbitrage have shown
that market participants might be unable or unwilling to implement trading strategies
that transmit information to asset prices.10 Importantly, further sources of non-
fundamental volatility emerge once cognitive limits and simple mental models are
allowed for.11
In a related contribution, Dumas et al. (2011) show that some well-known
anomalies observed in international equity markets can be explained by a model in
which agents perceive differences in the precision of information regarding local
and foreign markets.
This article is also connected to the well-known empirical literature that shows
evidence consistent with inefficiencies, excess volatility and overreaction in stock
markets. Most of these analyses focus on firm level returns and use historic information
on asset prices, financial statements and market activity.12 More closely related to
the current study, some contributions focus on the dynamic relationship of aggregate
proxies for investor sentiment and stock returns.13 In these contributions, the proxies
are based on consumer confidence surveys and financial market outcomes. In a
related analysis of international equity markets, Hwang (2011) finds that sentiment
toward countries, as inferred from public opinion surveys, is associated to demand
for securities and distortions in asset prices.

7 For example, see learning models in Sargent (1993), Heymann and Sanguinetti (1998) and Milani (2007).
8 See Diamond (1982), Cooper and John (1988) and Obstfeld (1996).
9 For early contributions see Banerjee (1992) and Bikhchandani et al. (1992). Angeletos et al. (2010) provide a related

model applied to macroeconomic fluctuations.

10 See De Long et al. (1990) and Shleifer and Vishny (1997).
11 See for example models of categorical thinking (Mullainathan 2002), trend chasing (Hong and Stein 1999) and
predictor selection dynamics (Brock and Hommes 1998).
12 See, for example, Shiller (1981), Bondt and Thaler (1985), Lakonishok et alt. (1994), Jegadeesh and Titman (2001),

and Baytas and Cakici (1999). For a comprehensive evaluation of these anomalies see Asness et al. (2013).
13 Baker and Wurgler (2007), Jansen and Nahuis (2003) and Schmeling (2009).
 Conventional views and asset prices 53

There is a growing set of contributions that use information in the press to

describe dynamics in macroeconomic and financial settings. Information in the
press has been used to describe predictive content related to the economic cycle
(Baker et al. 2012, Aromí 2014), to exchange rate volatility (Krol 2014) and to
describe dynamics of consumer confidence (Doms and Morin 2004). Anticipation
of daily stock market returns has been shown by Tetlock (2007) and Garcia (2013)
for the US and Aromí (2013) for the case of Argentina. As in the case of the current
study, but at higher frequencies, the evidence found in those articles is consistent
with overreactions in stock prices.
The rest of the paper is organized as follows. Section II describes the data used
in the study and the way in which media information is summarized. Section III
presents the evidence from portfolio strategies. Investment strategies based on
alternative sources of information are described in section IV. Section V presents
evidence using monthly returns data. Section VI concludes.

II. Data and opinion metrics

The analysis uses two categories of yearly frequency information: financial assets
returns and media content. The first set of data is given by the returns of stock market
indices expressed in dollars. The main source for this data is the World Bank.14 The
data covers 34 countries over 26 years (1988-2013). For the early part of the sample
(1988-1995), for some countries, this data was not available from this source. As a
result, supplementary data was obtained from a private data vendor15 and, in few
cases, from the relevant stock exchange. The sample covers countries that belong to
different regions and display heterogeneous levels of economic development.
Given the value of the stock market index of country i at the end of year t (SMIit),
the annual return in year t for country i is given by the difference of the logs of the
index for years t and t − 1: rit = log(SMI it ) − log(SMI it −1 ). Table 1 shows descriptive
statistics for the return in the sampled countries.
The second type of data is an indicator of conventional views based on content
published in the economic press. More specifically, this indicator was built based
in information published in The Wall Street Journal (1984-2013) and The Economist
(1992-2013). Together with The Financial Times, these publications are among the
three main business publications in the English language.

14 http://data.worldbank.org/indicator/CM.MKT.INDX.ZG.
15 http://www.tradingeconomics.com/
Table 1. Descriptive statistics
54

Country Sentiment Index Annual Return

Mean St. dev. Min. Max. Mean St. dev. Min. Max.
Argentina 0.067 0.008 0.047 0.082 0.076 0.541 -0.821 1.617
Austria 0.052 0.007 0.039 0.067 0.027 0.325 -1.050 0.476
Belgium 0.059 0.006 0.048 0.071 0.046 0.316 -1.079 0.495
Brazil 0.062 0.005 0.051 0.072 0.091 0.628 -1.309 1.356
Chile 0.061 0.009 0.049 0.085 0.117 0.304 -0.528 0.668
Colombia 0.083 0.011 0.065 0.114 0.085 0.380 -0.635 0.765
Czech Republic 0.053 0.010 0.038 0.085 0.046 0.291 -0.616 0.565
Denmark 0.050 0.008 0.032 0.064 0.090 0.257 -0.713 0.445
Finland 0.050 0.007 0.034 0.060 0.075 0.391 -0.844 0.871
Greece 0.068 0.010 0.051 0.089 0.060 0.499 -1.079 1.092
Hungary 0.060 0.011 0.042 0.086 0.077 0.382 -0.994 0.693
India 0.069 0.008 0.053 0.088 0.077 0.372 -1.022 0.663
Indonesia 0.065 0.013 0.046 0.086 0.065 0.571 -1.347 1.261
Ireland 0.061 0.008 0.046 0.084 0.060 0.339 -1.204 0.438
Journal of Applied Economics

South Korea 0.062 0.005 0.054 0.073 0.077 0.490 -1.171 0.793
Malaysia 0.058 0.011 0.038 0.077 0.060 0.374 -1.309 0.621
Mexico 0.064 0.008 0.053 0.081 0.149 0.363 -0.598 0.775
New Zealand 0.051 0.005 0.044 0.061 0.023 0.279 -0.734 0.470
Norway 0.055 0.007 0.044 0.074 0.056 0.343 -1.079 0.647
Pakistan 0.088 0.008 0.070 0.105 0.071 0.425 -0.968 0.751
Peru 0.072 0.012 0.047 0.102 0.201 0.423 -0.528 1.078
Philippines 0.070 0.008 0.055 0.087 0.060 0.450 -0.968 0.859
Table 1. (continued) Descriptive statistics
Country Sentiment Index Annual Return
Mean St. dev. Min. Max. Mean St. dev. Min. Max.
Poland 0.063 0.008 0.053 0.090 0.109 0.583 -0.868 2.202
Portugal 0.059 0.011 0.043 0.078 0.025 0.278 -0.755 0.399
Russia 0.071 0.006 0.058 0.084 0.130 0.765 -1.833 1.345
Singapore 0.052 0.008 0.040 0.068 0.061 0.324 -0.755 0.571
South Africa 0.071 0.010 0.056 0.090 0.075 0.264 -0.545 0.445
Spain 0.058 0.008 0.045 0.077 0.065 0.246 -0.562 0.438
Sweden 0.053 0.006 0.043 0.069 0.083 0.312 -0.755 0.536
Taiwan 0.058 0.004 0.048 0.066 -0.012 0.352 -0.821 0.610
Thailand 0.068 0.011 0.047 0.097 0.044 0.527 -1.561 0.904
Turkey 0.073 0.007 0.060 0.086 0.055 0.618 -0.968 1.267
Venezuela 0.068 0.012 0.041 0.095 -0.014 0.500 -1.079 0.779
Vietnam 0.083 0.008 0.061 0.099 -0.105 0.503 -1.139 0.385
Average 0.063 0.008 0.048 0.082 0.065 0.412 -0.948 0.802
Note: The Sentiment Index was constructed using content from The Wall Street Journal (1984-2013) and The Economist (1992-2013). Annual Return is the dollar return of each country stock market
index and was provided by the World Bank (1988-2013) and, in few instances, by tradingeconomics.com.
Conventional views and asset prices
55
56 Journal of Applied Economics

As discussed in the introduction, there are reasons to believe that conventional

views can be approximated through content published in the media. On the other
hand, summarizing media content to generate sentiment indices is a challenging
task. Below a description of the path followed in this work is provided.
Opinions transmitted in the press regarding different countries are summarized
computing the frequency of words with negative content in a relevant subset of
text. This is a simple approach that has proven successful in other contexts.16
The first step in the construction of the indices involves identifying a list of
keywords associated to each country: name of country, capital city17 and demonym.
Next, for each year in the sample period, the set of articles in which at least one of
these keywords is present is identified. For each of these articles, the portions of
text that are sufficiently close to a keyword associated to any country are selected.
More specifically, the selection corresponds to words that are up to 10 words before
or 10 words after one of the keywords associated to any country.18 The strings of
text associated to country c and year t are merged forming a selection of text Kct.
This concludes the text extraction stage.
An indicator of sentiment for the relevant country for each year is generated
computing the frequencies of words with negative content. Following the seminal
contribution by Tetlock (2007), the list of negative words is built using the negative
valence category from the Harvard IV dictionary. The dictionary was procured
from General Inquirer, a website that provides tools for content analysis of textual
data.19
It must be noted that there are other lists of negative words that can be used in
this analysis. In a relevant antecedent, Loughran and McDonald (2011) develop a
list of negative words for financial contexts. Their analysis shows that informational
gains can be attained using a more precise list of words. On the other hand, lists of
words that are generated after the date of the sample run the risk of incorporating
forward looking bias. As a result, despite the potential gains that could result from

16 In line with the findings reported in Tetlock (2007), indices constructed using words with positive content lack
information on future stock returns. This can be explained by asymmetries in the information content of positive and
negative words as found in the natural language processing literature (Garcia et al. 2012).
17 In the case of Brazil, Spain, India, Philippines and South Africa big cities that can be unambiguously linked to the
country are also included (e.g., Cape Town for South Africa).
18 Following usual practice in content analyses, stop words (common words with no relevant content) were eliminated

before neighboring text was extracted. Additionally, it is worth noting that variations in which 5 or 50 neighboring
words were selected lead to very similar results.
19 http://www.wjh.harvard.edu/~inquirer/homecat.htm.
 Conventional views and asset prices 57

considering alternative list of words, in the main part of the analysis we select the
most cautious path and use the list originally employed in Tetlock (2007). Alternative
lists are also evaluated as part of the sensitivity analysis.
The original list of negative words from General Inquirer includes 2291 words.
In order to improve the precision of the indices, this original list was expanded to
include plural noun forms, different verb tenses and adverbs. This procedure resulted
in a list of 5364 words.
Let Tct be the number of words in Kct, the selected text corresponding to year t
and country c and let Nct be the number of times a negative word is detected in Kct.
Then, the corresponding sentiment index is given by sct = Nct / Tct. A higher number
is associated to more pessimistic views while a lower number is associated to more
optimistic assessments. Thus, the indicator could be labeled as a negative sentiment
index. Throughout the document, at the risk of some confusion, the shorter expression
sentiment index will be used. The construction of this indicator involves the selection
of text comprising approximately 23 million words out of which more than 1.5
million correspond to words classified as negative words.
Table 1 shows important cross-country differences in the level of the sentiment
index. The average value of the sentiment index ranges from 0.05 in the case of
Denmark to 0.088 for the case of Pakistan. The standard deviation ranges from
0.004 in the case of Taiwan to 0.013 in the case of Indonesia. Looking more
meticulously, it is found that the minimum value for the index is 0.032 and corresponds
to Denmark in 1989. On the other hand, the maximum value is 0.114 and corresponds
to Colombia during 2003, a year in which the country experienced weak economic
performance and high levels of violence. These observations constitute a first
indication suggesting that the index is able to capture important economic
developments in the selected countries.
As a first step in the analysis, it could be instructive to observe the evolution of
the indices during the sample period. Figure 1.A presents the average value of
sentiment index for the 34 countries from 1984 through 2013. The path of the index
suggests that there is a close association with the evolution of the global economy.
For example, by a small margin, the minimum value is observed in 2007, on the
final year of a period known as the great moderation. On the other hand, the maximum
value is observed in 2010, in the aftermath of the global crisis originated in the U.S.
and during a critical phase of the European crisis. In addition, a peak observed in
2001-2003 can be linked to the burst of the dot-com bubble, crises in emergent
markets and the conflicts after 9/11. A period of low values is observed in the mid-
nineties, a period of average economic bonanza.
58 Journal of Applied Economics

Figure 1. Sentiment index

A. Average for 34 countries

B. Greece

Note: Shaded regions in panel A correspond to years with world GDP growth below 2%.

Additional preliminary insights can be gained considering the case of a single

country. Figure 1.B shows the sentiment index for Greece. Particularly suggestive
is the drop in the index, that is, the increasingly optimistic views observed from
2002 through 2006. In 2006 the index for Greece reaches a historic low of 0.0508.
Four years after, in 2010, the index reaches a historic high of 0.0887. This extreme
pattern is suggestive of overreaction in at least one of these instances. These
 Conventional views and asset prices 59

observations based on anecdotal evidence are informative and suggests the plausibility
of the overreaction conjecture. Formal empirical evaluations of this conjecture are
implemented in the next sections.

III. Performance of contrarian portfolio strategies

In this section, contrarian portfolio strategies are implemented and their performance
is evaluated. Each year, past values of the sentiment index are used to construct a
portfolio of countries associated to past optimistic views and a portfolio of countries
associated to past pessimistic views. Under the overreaction conjecture, it is expected
that the first portfolio will experience inferior returns and the second portfolio will
experience superior returns.
The sorting of countries is implemented using the average value of the sentiment
index computed using four-year moving windows. It is expected that averaging the
value of the index over multiple years will reduce noise in the identification of
extreme views. Additionally, averaging is compatible with the focus of this study
on low frequency, highly persistent dynamics.

A. Baseline estimates

As shown in the previous section, the mean values of the indices associated to
different countries show ample variation. Hence, unless some correction is
implemented, there are some countries that could be systematically associated to
optimistic states or pessimistic states simply due to stable high or low values in the
sentiment index. This suggests that adjusting for differences in mean values is
needed if the intention is to capture changes in shared views about countries’
prospects instead of reflecting countries’ permanent characteristics. The indices are
adjusted using the historical values of the index for the corresponding country.
Additionally, the indices associated to different countries express ample variation
in terms of standard deviation. This could be due to more volatile perspectives or to
differences in noise of the index due to heterogeneous levels of coverage of sampled
countries. Taking into account these differences, in the exercise below the indices are
adjusted using only historical information so any look-ahead bias is avoided.
( )
More specifically, the standardized sentiment index ŝct is given by sˆct = sct − sct / sctv
where s–ct is a weighted average of past values and svct is a measure of past volatility.
t −1 t −1
Adjustment parameters are given by sct = ∑sckω (t − k ) and sctv = ∑ sck − sck ω (t − k ),
k =t0 k =t 0
60 Journal of Applied Economics

where t0 is the first period of the sample and the weighting function ω (t – k) is a
t −1
decreasing function which satisfies ∑ω (t − k ) = 1, where ω (t + 1) = (3/2)ω (t), that
k =t0
is, the weight decreases with distance at a constant 33% rate. The weighting function
captures the idea that more recent observations are more informative while the
expanding window allows for more informed adjustments.20 To secure for informed
adjustment of the index value, the analysis is restricted to observations for which there
exist at least four years of historic sentiment data. Finally, the indicator used to construct
the portfolios for year t is the average value of the standardized index in the four most
t
recent years sct* = ∑ sˆ ck .
k =t − 3
For each year, two portfolios are built identifying the top and bottom deciles
using the adjusted index s*ct. The portfolio associated to optimism, or Portfolio 1,
is composed by the stock market indices of the countries that belong to the bottom
decile. Similarly, the portfolio associated to pessimism, or Portfolio 2, is composed
by the stock market indices of the countries that belong to the top decile. Let rtP1(t–l)
be the average return in year t for the stock indices that belong to Portfolio 1 built
in year t – l. In the same fashion, rtP2(t–l) is the equivalent indicator for Portfolio 2.
The analysis below will focus on the returns of these portfolios for different values
of the lag parameter l. If extreme assessments are associated to excessive reactions,
it is expected that Portfolio 1, the optimistic portfolio, will underperform and Portfolio
2, the pessimistic portfolio, will show superior performance.
In addition to comparing average returns, in this section a formal test for abnormal
returns is implemented. A simple empirical model that includes a market factor is
proposed. The statistical model is given by the following equation:

rt P(t −l ) = α + β rt m + εt ,

where rtP(t–l) is the return of portfolio P (t − l ) ∈{P1(t − l ), P 2 (t − l )} in year t, rtm is

the average return for all sampled countries in year t and εt is an error term. Following
a standard procedure in the asset pricing literature, the estimate for the parameter
α is interpreted as the abnormal return of the relevant portfolio. Additionally, similar
calculations are computed for a long-short portfolio strategy in which Portfolio 1,
the optimistic portfolio, is the short position and Portfolio 2, the pessimistic portfolio,
is the long position. Standard errors and associated t-statistics are corrected for
heterocedasticity.21

20 Unreported exercises show that the results are not sensitive to changes in the weighting function.
21 Package car in platform R was used to estimate robust standard errors.
Table 2. Performance of sentiment based portfolios
A. Return and performance for different lags in portfolio formation

Lags Portfolio 1 (Optimism) Portfolio 2 (Pesssism) Portfolio 2 - Portfolio 1

Mean St. Dev. Alpha t-stat. Mean St. Dev. Alpha t-stat. Mean St. Dev. Alpha t-stat.
0 0.087 0.302 0.038 0.9 0.121 0.435 0.049 0.9 0.034 0.337 0.011 0.1
1 0.045 0.367 -0.035 -1.3 0.210 0.372 0.134 2.7 0.165 0.245 0.169 2.7
2 0.011 0.340 -0.055 -2.2 0.058 0.387 -0.011 -0.2 0.047 0.249 0.044 0.7
3 -0.002 0.297 -0.062 -2.2 0.074 0.401 -0.001 -0.0 0.077 0.298 0.061 0.8
4 -0.016 0.310 -0.079 -2.5 0.126 0.295 0.068 1.7 0.142 0.167 0.147 2.9
5 0.002 0.348 -0.067 -1.7 0.155 0.336 0.088 2.7 0.153 0.258 0.155 2.5
6 0.007 0.324 -0.040 -1.3 0.072 0.379 0.022 0.4 0.066 0.279 0.022 0.4

B. Return and performance of mean sentiment based-portfolios (1 through 5 lags)

Portfolio 1 Portfolio 2 Portfolio 2 - Portfolio1
Mean 0.010 0.121 0.111
St. dev. 0.324 0.337 0.176
Alpha -0.059 0.051 0.110
Conventional views and asset prices

t-stat. -2.7 1.8 2.5

Notes: Panel A: Portfolio 1 has equal sized long positions in the stock indices of the countries of the most optimistic decile. Portfolio 2 has equal sized long positions in the stock indices of the countries
in the most pessimistic decile. For the model: rt P(t −l ) = α + β rt m + εt for l = 0,1,…,6, alpha is the estimated constant, rt P(t −l ) is the return of the corresponding sentiment-based portfolio (Portfolio 1,
Portfolio 2 or Portfolio 1 – Portfolio 2), rtm is the market return and εt is an error term. Panel B: mean sentiment-based portfolio corresponds to the average of the portfolios with 1 through 5 lags in
5
portfolio formation year. The associated return is: r P(t ) = ∑rt P(t −l ) / 5 . The t-statistics are corrected for heterocedasticity.
l =1
61
62 Journal of Applied Economics

Table 2.A shows the results for multiple values of the lag parameter l. One
notable feature is the poor performance of Portfolio 1. From two years and up to
six years after portfolio formation, the mean return of this portfolio is below 1.1%.
More dramatically, for the case of three and four year lags, the mean return is
negative. Similar conclusions emerge from observing the estimation of the market
factor model. For example, four years after portfolio formation, the estimated
abnormal return is -7.9%.
On the other hand, the performance of Portfolio 2, the portfolio associated to
pessimism, takes the opposite direction. For example, one year after portfolio
formation, the mean return is 21%. High values are also observed for the mean
returns four and five years after portfolio formation. The estimations of factor models
provide similar results. The abnormal returns one, four and five years after portfolio
formation are estimated to be above 6.8% and statistically different from zero. In
contrast, two and three year lags result in estimated abnormal returns that are slightly
negative but statistically null.
The last panel in Table 2.A describes the return associated to the long-short
portfolio strategy. Lags of one, four and five years result in positive abnormal returns
of at least 15%. These estimations are suggestive of overreactions that are gradually
corrected years after the extreme opinions are identified. Lags of two, three and six
years show positive but statistically insignificant abnormal returns. At this stage, it
is not clear whether the differences in performance for different number of lags are
simply noise or reflect a stable property of the return reversal process. Another
interesting observation is that the results suggest a weak contemporaneous relationship
between conventional views and stock index returns.
In terms of summarizing the results of this exercise, it is convenient to provide a
description of the performance of the portfolio strategies when the returns associated
to different values of the lag parameter l are combined. With this objective, the average
5
return r P(t ) = ∑rt P(t −l ) / 5 is computed for portfolios P (t − l ) ∈{P1(t − l ), P 2 (t − l )} . In
l =1
other words, the new portfolios are built combining, with equal weights, the portfolios
that exploit information with one through five lags. It is expected that this portfolio
will allow for a more precise assessment of the association between lagged sentiment
and return differentials.
As shown in Table 2.B, the mean return of Portfolio 1 is 1% while the mean
return of Portfolio 2 is 12%. Suggesting that risk considerations are unlikely to
provide a satisfactory explanation, the estimated standard deviations for each portfolio
are very similar. The estimated abnormal returns are -5.9% for the case of Portfolio
 Conventional views and asset prices 63

1 and 5.1% in the case of Portfolio 2. The associated long-short portfolio strategy
results in a highly statistically significant abnormal return equal to 11%.
Figure 2 shows the returns of this long-short portfolio strategy. The figure shows
a difference in performance that seems to be quite stable throughout the sample
period. As can be observed, there are only four years in which the return is negative
and in none of these years the return is below -20%. In contrast, the annual return
is above 20% in six occasions.
The results presented above are consistent with the proposed conjecture of
excessive responses associated to extreme opinions. Optimism is followed by
underperformance and pessimism is followed by over-performance. This difference
is highly persistent and stable over the sampled period. Risk considerations are
unlikely to provide a satisfactory explanation.

B. Sensitivity analysis

In this subsection we provide information on several exercises in which aspects of

the original exercise are modified. It is understood that this alternative exercises
might shed light on the robustness of the previous results and the directions that

Figure 2. Returns of the long-short portfolio (based on four years of publications’ data)

Notes: The bar corresponding to year t represents the average return on year t of the portfolios built using the sentiment indices
of years t-5 through t-1.
64 Journal of Applied Economics

can allow for informational gains. In general, the exercises show that the results
are quite robust. Also, there seems to be little room for informational gains.
Table 3 provides information on the returns of the average contrarian portfolio
under four types of modifications of the original exercise. The first analysis evaluates
alternative sizes for the moving windows over which average sentiment is computed.
As panel A shows, little changes are observed when the original size of 4 years is
modified in the direction of shorter or longer periods.
The second examination deals with the selection of words. In the original
exercises, words that are at a distance equal or lower than 10 from relevant keywords

Table 3. Characterization of returns in sensitivity analysis exercises

A. Size of moving window (years)
3 4 5
Mean 0.110 0.111 0.119
St. dev. 0.177 0.176 0.172
Alpha 0.108 0.110 0.113
t-stat. 2.4 2.5 2.8
B. Maximum distance of selected words
5 10 50
Mean 0.122 0.111 0.107
St. dev. 0.140 0.176 0.161
Alpha 0.122 0.110 0.112
t-stat. 3.0 2.5 2.6
C. Sentiment dictionary
Harvard IV Dictionary Loughran & McDonald (2011) Mohammad & Turney (2013)
Mean 0.111 0.065 0.110
St. dev. 0.176 0.187 0.154
Alpha 0.110 0.063 0.112
t-stat. 2.5 1.3 2.5
D. Weight adjustment parameter
1 1.5 2
Mean 0.110 0.111 0.092
St. dev. 0.175 0.176 0.165
Alpha 0.105 0.110 0.092
t-stat. 2.5 2.5 2.1
Notes: Return 5of the mean portfolios computed averaging the returns of the portfolios associated to 1 through 5 lags in formation
year: r P(t ) = ∑rt P(t −l ) . Alpha is the estimated constant for the model: r P(t ) = α + β rt m + εt for l = 0,1,…,6. Where rtm is the
l =1
market return and εt is an error term. t-statistics are corrected for heterocedasticity.
 Conventional views and asset prices 65

were selected. In the modified implementation presented in panel B of Table 3,

distances of 5 and 50 are considered. It is verified that minor informational gains
result from imposing a smaller distance. But the results are not significantly altered
in any of the two cases.
Following Tetlock (2007), the sentiment indices were built exploiting the negative
category in Harvard IV dictionary. Alternative lists of words could have been
exploited. For example, Loughran and McDonald (2011) developed a list of negative
words for financial contexts, while Mohammad and Turney (2013) constructed a
dictionary of words employing a novel form of online collaboration. Panel C shows
that our original exercise displays results that dominate those associated to the list
of negative terms generated by Loughran and McDonald (2011) and are similar to
the ones associated to the list generated by Mohammad and Turney (2013).
Finally, the standardization of individual country sentiment indices is considered.
As described in the previous section, each country index was standardized using
historic values of the average index and a historic metric of variability. In that
exercise, the weight allocated to past values decreased with distance at a rate of
50% according to the rule: ω (t + 1) = (3/2) ω (t). In other words, the weight adjustment
parameter was set equal to 3/2. According to the results shown in panel D, the results
still hold when the rate at which the weight drops with distance is altered. It is
observed that doubling the rate of increment in weights and keeping the weights
constant result in similar properties for the distribution of returns of the contrarian
sentiment long-short portfolio.

IV. Alternative sources of return prediction: past returns or economic

growth

In this work, a novel source of persistent return predictability is identified. Extreme

negative and positive opinions are found to predict expected returns. Shared views
are computed using information distributed in the economic press. One relevant
question is the extent to which similar return predictability could be achieved using
more standard forms of information such as past returns or recent performance in
terms of economic growth. This is a sensible consideration since periods of intense
optimism are likely to be associated to positive returns and growth accelerations
while periods of intense pessimism are commonly associated to negative returns
and poor economic growth. Additionally, conventional opinions are measured with
noise, hence it is plausible that alternative sources of information can result in more
predictability.
66 Journal of Applied Economics

In this section, two alternative sources of information are considered. A portfolio

strategy that takes into account past economic growth is implemented. This portfolio
strategy can be thought to evaluate the extent to which return predictability is explained
by naïve projection of recent economic growth performance. Also, the case of portfolio
strategies based on cumulative returns in the most recent years is considered. This
strategy can be linked to the well-known literature that evaluates the performance of
strategies that bet on previous losers and against previous winners.22 In each case, the
strategy is implemented following the algorithm described in the previous section.
Economic growth data from the World Bank is used to compute, for each year,
the economic growth in the most recent four years.23 Portfolio 1 is associated to the
stock indices of countries in the decile with the largest economic growth. Portfolio
2 is the portfolio associated to the stock indices of countries in the decile with the
lowest economic growth.
Similarly, in the case of strategies that exploit past returns, Portfolio 1 is associated
to the stock indices that show the largest cumulative returns for the previous four
years. Portfolio 2 is the portfolio associated to the stock indices with the lowest
cumulative returns. The stock indices selected correspond to the top and bottom
deciles respectively.
Table 4 describes the returns associated to the two portfolio strategies described
above. For the strategy based on previous economic growth, the mean return
associated to high growth is below the growth return associated to low economic

Table 4. Return and performance of alternative portfolio strategies

GDP growth sorts Past return sorts

High (H) Low (L) L-H High (H). Low (L) L-H
Mean 0.024 0.079 0.055 0.019 0.074 0.055
St. dev. 0.404 0.385 0.353 0.345 0.408 0.252
Alpha -0.052 0.009 0.061 -0.036 0.015 0.051
t-stat. -1.1 0.2 0.7 -1.6 1.2 0.8
Correl. w/sentiment portfolio 0.53 0.66 0.71 0.75 0.57 0.70
Notes: Return of 5
the mean portfolios computed averaging the returns of the portfolios associated to 1 through 5 lags in formation
year : r P(t ) = ∑rt P(t −l ) . Alpha is the estimated constant for the model: r P(t ) = α + β rt m + εt for l = 0,1,…,6. Where rtm is the
l =1
market return and εt is an error term. t-statistics are corrected for heterocedasticity.

22 See Bondt and Thaler (1985) for the seminal contribution and Asness et al. (2013) for a recent comprehensive
evaluation.
23 The data is available at http://data.worldbank.org/indicator/NY.GDP.MKTP.KD.ZG.
 Conventional views and asset prices 67

growth. The estimated return is 6.1% when the market factor model is estimated.
Nevertheless, the estimated parameter is not significantly different from zero. Despite
the difference in terms of return predictability, it must be noted that there is a strong
association between the return of the long short portfolio strategy based on extreme
opinions and the return of the long short portfolio strategy based on economic
growth. More precisely, the correlation coefficient for the returns of the respective
long-short portfolios is 0.71.
Similar results are observed in the case of portfolio strategies that exploit
information on past returns. The mean return for the portfolio of past winners is
1.9% while the mean return for past losers is 7.4%. The difference has the expected
signed but is significantly smaller than the difference observed in the case of sentiment
portfolios. According to the estimated market factor model the abnormal return is
5.1% but statistically it is not significantly different from zero. Despite the difference
in performance, it is clear that the sentiment based strategy and the strategy based
on past returns are strongly associated. The correlation coefficient for the returns
of the long-short portfolios is 0.70.
In summary, it has been found that strategies based on past returns and economic
growth are strongly linked to portfolio strategies based on sentiment indices. On
the other hand, these alternative portfolio strategies are unable to replicate the
performance of the sentiment based strategies. This suggests that the sentiment
indices are able to reflect information on subjective states that cannot be captured
with similar precision by alternative simpler indicators.

V. Performance of contrarian strategies at higher frequencies

So far the analysis has been carried out using annual returns data. This is consistent
with the focus of this work on associations between data with multiple year lags.
As shown in section III, sentiment indices are found to anticipate returns differentials
up to five years after sentiment levels are measured. In favor of annual return data
analyses, it must be noted that they exclude high frequency noise that might hide
long term associations. In addition, comparable stock market data for long periods
is more easily available at annual frequencies.
On the other hand, one shortcoming of low frequency analyses is given by the
relatively small number of observations. Statistical tests are more reliable under a
larger set of observations. Additionally, a monthly analysis can inform about the
short term risk associated to exploiting this long term patterns in the data. In this
section, the analysis of section III is replicated using monthly returns data.
68 Journal of Applied Economics

As indicated, available data covers a shorter time span and a smaller set of
countries. The sample period goes from January 1999 through July 2014. The
number of sampled countries drops from 34 in the annual analysis to 30 in the
monthly evaluation.24 Dollar returns were computed using the stock index denominated
in local currency and the dollar exchange rate for the last day of the month. The
source for monthly stock market indices is Bloomberg. For most countries, exchange
rate data corresponds to the daily series provided by the Federal Reserve Bank of
St. Louis.25 For eight countries, this data was not available from that source and
was obtained from a private data provider.26
Table 5 shows results for the monthly frequency analysis. The average returns
presented in panel A show that the returns following periods of high sentiment (high
pessimism) are, on average, higher than the returns that follow periods of low sentiment
(low pessimism). As in the annual analysis, the difference is the largest for five year
lags. In this case, monthly average returns differ by 0.95%. This gap is similar to that
observed in the annual returns analysis. The findings are also replicated when differences
in volatility are evaluated. For example, portfolios built using five-year lagged sentiment
metrics show no difference in terms of the standard deviation. In both cases, the
standard deviation of monthly returns is approximately 7.4%. Importantly, despite
the differences in data coverage and high frequency noise, the estimated abnormal
returns and associated t-statistics are also in line with the annual return analyses.
Panel B in Table 5 shows the performance of the average portfolio strategy that
combines portfolios built using one through five-year lagged indices. The monthly
return of the portfolio associated to lagged optimism is 0.15%. In contrast, the
monthly return of the portfolio associated to lagged pessimism is 0.96%. For the
portfolio associated to optimism, the estimated monthly abnormal return is above
0.5% and highly significant in statistical terms. The third column shows that the
average monthly return of a long-short portfolio strategy is 0.8%. The associated
abnormal monthly return is 0.79% and statistically significant. This abnormal return
is approximately 10% in annual terms which is similar to the value observed for a
different sample in the annual frequency analysis of section III.

24 The list of countries 30 covered by the monthly returns dataset is given by: Argentina, Austria, Belgium, Brazil,
Chile, Colombia, Denmark, Finland, Greece, India, Indonesia, Ireland, Malaysia, Mexico, New Zealand, Norway,
Pakistan, Peru, Philippines, Poland, Portugal, Singapore, South Africa, South Korea, Spain, Sweden, Taiwan, Thailand,
Turkey and Vietnam.
25 The series can be found at: https://research.stlouisfed.org/.
26 The data from these countries (Argentina, Chile, Indonesia, Pakistan, Peru, Philippines, Turkey and Vietnam) was
provided by OANDA (http://www.oanda.com/).
Table 5. Performance monthly of sentiment based portfolio strategies
A. Return and performance for different lags in portfolio formation

Lags Portfolio 1 (Optimism) Portfolio 2 (Pesssism) Portfolio 2 - Portfolio 1

Mean St. Dev. Alpha t-stat. Mean St. Dev. Alpha t-stat. Mean St. Dev. Alpha t-stat.

1 0.0015 0.0692 -0.005 -2.4 0.0108 0.0776 0.004 1.2 0.0093 0.0583 0.009 2.1

2 0.0012 0.0703 -0.006 -2.7 0.0079 0.0780 0.000 0.1 0.0067 0.0534 0.006 1.5

3 0.0012 0.0708 -0.006 -2.7 0.0080 0.0746 0.001 0.3 0.0068 0.0513 0.007 1.7

4 0.0005 0.0699 -0.007 -2.8 0.0081 0.0713 0.001 0.4 0.0076 0.0506 0.008 1.9

5 0.0033 0.0740 -0.004 -1.5 0.0128 0.0734 0.006 2.0 0.0095 0.0528 0.010 2.5

6 0.0050 0.0728 -0.002 -0.8 0.0110 0.0728 0.004 1.4 0.0060 0.0514 0.006 1.5

B. Return and performance of mean sentiment based-portfolios (1 through 5 lags)

Portfolio 1 Portfolio 2 Portfolio 2 - Portfolio1

Mean 0.0015 0.0096 0.0080

St. dev. 0.0718 0.0770 0.0612

Alpha -0.0056 0.0024 0.0079

Conventional views and asset prices

t-stat. -3.6 1.0 2.5

Notes: Panel A: Portfolio 1 has equal sized long positions in the stock indices of the countries of the most optimistic decile. Portfolio 2 has equal sized long positions in the stock indices of the countries
in the most pessimistic decile. Alpha is the estimated constant for the model: rt P(t −l ) = α + β rt m + εt for l = 0,1,…,6. Where, rt P(t −l ) is the return of the corresponding sentiment-based portfolio (Portfolio
1, Portfolio 2 or Portfolio 1 – Portfolio 2), rtm is the market return and εt is an error term. Panel B: mean sentiment-based portfolio corresponds to the average of the portfolios with 1 through 5 lags in
5
portfolio formation year. The associated return is: r P(t ) = ∑rt P(t −l ) . t-statistics are corrected for heterocedasticity.
l =1
69
70 Journal of Applied Economics

In summary, the analyses that use monthly return data are consistent with return
predictability. Despite the high frequency noise that characterizes monthly returns
and the reduced time and country sample coverage, the differences in return are
statistically significant. The results serve as a robustness check of the exercises
developed using annual return data.

VI. Conclusions

This study proposes a novel metric of conventional views to evaluate instances of

overreaction in financial and macroeconomic contexts. More specifically, the metric
uses content of the international economic press. The analysis focuses on the relative
performance of stock market indices following periods of extreme opinions. The
results show that strong positive views are associated to subsequent lower returns
and strong negative views are associated to subsequent higher returns. The difference
in returns is highly persistent; for example, returns for portfolios constructed using
five year-old information show differences in means that are both statistically and
economically significant.
The evidence is consistent with the occasional emergence of extreme shared
views that result in mistaken valuations of a broad class of financial assets that are
corrected in subsequent years. According to the results, predictable differences in
risk cannot explain the differences in returns.
This evidence is relevant for the interpretation of dynamics in financial and
macroeconomic contexts. The existence of recurrent, persistent and predictable
overreactions has implications for both private actors’ decision making and the
design of public policies that aim for stability. Additionally, this evidence suggests
that there is value in empirical analyses that exploit estimations of subjective states.
Strategies based on past returns or economic growth records are not able to replicate
the results attained when the sentiment indices are used.

References

Angeletos, George, Guido Lorenzoni, and Alessandro Pavan (2010), Beauty contests
and irrational exuberance: A neoclassical approach. Working Paper 15883, NBER.
Aromí, J. Daniel (2013), Media news and the stock market: evidence for Argentina.
IIEP-Baires. mimeo.
Aromí, J. Daniel (2014), Estados de opinión y ciclo económico: un estudio empírico
para Argentina, IIEP-Baires, mimeo.
 Conventional views and asset prices 71

Asness, Clifford, Tobias Moskowitz, and Larry Pedersen (2013). Value and momentum
everywhere. Journal of Finance 68: 929-85.
Baker, Michael, and Jeffrey Wurgler (2007). Investor sentiment in the stock market.
Journal of Economic Perspectives 21: 129-52.
Baker, Scott, Nicholas Bloom and Steven Davis (2012). Measuring economic policy
uncertainty. Research Paper No. 13-02, University of Chicago Booth School of
Business.
Banerjee, Abhijit (1992), A simple model of herd behavior. Quarterly Journal of
Economics 107: 797-817.
Baytas, Ahmet and Nusret Cakici (1999). Do markets overreact: International
evidence. Journal of Banking & Finance 23: 1121-44.
Bikhchandani, Sushil, David Hirshleifer and Ivo Welch (1992). A theory of fads,
fashion, custom, and cultural change as informational cascades. Journal of
political Economy 100: 992-1026.
Brock, William, and Cars Hommes (1998). Heterogeneous beliefs and routes to
chaos in a simple asset pricing model. Journal of Economic Dynamics and
Control 22: 1235-74.
Cooper, Russell, and Andrew John (1988). Coordinating coordination failures in
Keynesian models. Quarterly Journal of Economics 103: 441-63.
De Bondt, Werner and Richard Thaler (1985). Does the stock market overreact?
Journal of Finance 40: 793-805.
De Long, J. Bradford, Andrei Shleifer, Lawrence Summers and Robert Waldmann
(1990). Positive feedback investment strategies and destabilizing rational
speculation. Journal of Finance 98: 379-95.
Diamond, Peter (1982). Aggregate demand management in search equilibrium.
Journal of Political Economy 90: 881-94.
Doms, Mark, and Norman Morin (2004). Consumer sentiment, the economy,
and the news media. Working Paper 2004-09, Federal Reserve Bank of San
Francisco.
Dumas, B., K. Lewis, and E. Osambela (2011). Differences of opinion and international
equity markets. Working Paper No. 16726, NBER.
Economist (2009). Brazil takes off. November 14.
Economist (2013). Has Brazil blown it? September 28.
Eyster, Erik, and Matthew Rabin (2010). Naive herding in rich-information settings.
American Economic Journal: Microeconomics 2: 221-43.
Eyster, Erik, and Matthew Rabin (2014). Extensive imitation is irrational and
harmful. Quarterly Journal of Economics 129: 1861-1898.
72 Journal of Applied Economics

Garcia, Diego (2013). Sentiment during recessions. Journal of Finance. 68: 1267-
300.
Garcia, David, Antonios Garas, and Frank Schweitzer (2012). Positive words carry
less information than negative words. EPJ Data Science 1: 1-16.
Gentzkow, Matthew, and Jesse Shapiro (2010. What drives media slant? Evidence
from US daily newspapers. Econometrica 78: 35-71.
Golub, Benjamin, and Matthew Jackson (2010). Naive learning in social networks and
the wisdom of crowds. American Economic Journal: Microeconomics 2: 112-49.
Heymann, Daniel, and Pablo Sanguinetti (1998). Business cycles from misperceived
trends. Economic Notes 27: 205-32
Hong, Harrison, and Jeremy Stein (1999). A unified theory of underreaction,
momentum trading, and overreaction in asset markets. Journal of Finance 54:
2143-84.
Hwang, Byoung-Hyoun (2011). Country-specific sentiment and security prices.
Journal of Financial Economics 100: 382-401.
Jansen, W. Jos, and Niek Nahuis (2003). The stock market and consumer confidence:
European evidence. Economics Letters 79: 89-98.
Jegadeesh, Narasimhan, and Sheridan Titman (2001). Profitability of momentum
strategies: An evaluation of alternative explanations. Journal of Finance 56:
699-720.
Kindleberger, Charles, and Robert Aliber (2011). Manias, panics and crashes: a
history of financial crises. Palgrave Macmillan.
Krol, Robert (2014). Economic policy uncertainty and exchange rate volatility.
International Finance 17: 241-56.
Lakonishok, Josef, Andrei Shleifer, and Robert Vishny (1994). Contrarian investment,
extrapolation, and risk. Journal of Finance 49: 1541-78.
Loughran, Tim, and Bill McDonald (2011). When is a liability not a liability? Textual
analysis, dictionaries, and 10-Ks. Journal of Finance 66: 35-65.
Milani, Fabio (2007). Expectations, learning and macroeconomic persistence.
Journal of Monetary Economics 54:2065-82.
Mohammad, Saif M., and Peter Turney (2013). Crowdsourcing a word–emotion
association lexicon. Computational Intelligence 29: 436-465.
Mullainathan, Sendhil (2002). Thinking through categories. MIT. mimeo.
Mullainathan, Sendhil, and Andrei Shleifer (2005). The market for news. American
Economic Review 95: 1031-53.
Obstfeld, Maurice (1996). Models of currency crises with self-fulfilling features.
European Economic Review 40: 1037-47.
 Conventional views and asset prices 73

Schmeling, Maik (2009). Investor sentiment and stock returns: Some international
evidence. Journal of Empirical Finance 16: 394-408.
Shiller, Robert (1981). Do stock prices move too much to be justified by subsequent
changes in dividends? American Economic Review 71: 421-36.
Shiller, Robert (2005). Irrational exuberance, 2nd ed. Princeton University Press.
Shleifer, Andrei, and Robert Vishny (1997). The limits of arbitrage. Journal of
Finance. 52: 35-55.
Tetlock, Paul, (2007). Giving content to investor sentiment: The role of media in
the stock market. Journal of Finance 62: 1139-68.