Copyright © 2020 Springer-Nature.

This version of the article has been accepted for publication, after peer review (when applicable) and is
subject to Springer Nature’s AM terms of use, but is not the Version of Record and does not reflect post-acceptance improvements, or any
corrections. The Version of Record is available online at: https://doi.org/10.1007/s11156-020-00937-2 (see: https://www.springernature.com/gp/
open-research/policies/journal-policies).

Investor Sentiment E¤ects on Share Price

Deviations from their Intrinsic Values Based
on Accounting Fundamentals
Yiannis Karaviasa , Stella Spiliotib and Elias Tzavalisb;
August 24, 2020

Abstract
We investigate the existence of evidence of investor sentiment on share price de-
viations from their intrinsic values across two sentiment regimes of shares market: the
low-to-normal and the excess one. We use the residual income valuation model to cal-
culate the intrinsic values of shares based on accounting fundamentals and we suggest
a panel data threshold model to capture the sentiment regimes of the market, using as
threshold variable alternative investor sentiment indices. The suggested model enables
us, …rst, to endogenously identify from the data the threshold value of a sentiment
index triggering market sentiment regime shifts and, based on it, to examine if the
e¤ects of investor sentiment on share prices across the above two sentiment regimes are
in accordance to the theory. Application of the model to UK data shows that investor
sentiment in‡uences positively share prices in the low-to-normal and negatively in the
excess one. We also show that investor sentiment dominates risk premium e¤ects on
shares characterized by low book-to-market, and dividend- and earnings-to-price ratios.
The above results are consistent with the predictions of the sentiment hypothesis.
JEL classi…cation: G12, G14, G15
Keywords: asset pricing, investor sentiment, risk premium, …rm-speci…c variables,
panel data, cointegration, threshold model.
The authors would like to thank George Constantinides, Theodore Sougiannis, Vas-
silis Sogiakas, and the participants of the 12th CEF -BMRC Conference on Macro and
Financial Economics, 2016, Brunel University, London and the 22nd International Panel
Data Conference, Curtin University, Perth for their helpful comments and suggestions.
Stella Spilioti acknowledges …nancial support from the Research Center of Athens Uni-
versity of Economics and Business under the grant no EP-11268001.
* Corresponding author. a): University of Birmingham, i.karavias@bham.ac.uk. b):
Athens University of Economics and Business, Spilioti@aueb.gr. c): Athens University
of Economics & Business, E.Tzavalis@aueb.gr.

1
1 Introduction

There is a long debate in the …nance literature over whether investor sentiment, such as
excessive optimism, fears and moods, a¤ect signi…cantly share returns and/or prices (see,
Statman (2011), for a survey).1 According to the sentiment hypothesis, deviations of share
prices from their fundamental values should re‡ect behavioral investor attitudes (i.e., psycho-
logical factors) which in‡uence investors’reaction to market and/or company information.
These attitudes tend to lead to overpriced, or underpriced, shares, depending on the senti-
ment level (or regime) of shares market. In particular, the sentiment hypothesis predicts that
abrupt adjustments of share prices to their intrinsic values and bubbles burst should occur in
periods of excess investor sentiment (see, e.g., Thorp (2004), Brown and Cli¤ (2004, 2005)).
On the other hand, the generic e¤ects of investor sentiment on share prices tend to build up
in periods of lower and normal levels of investor sentiment. These predictions are at odd to
those of the risk premium hypothesis (see, e.g., Shefrin (2001, 2015)). Thus, conditioning on
the above di¤erent sentiment regimes of shares market enables us to distinguish the e¤ects
on investor sentiment on share prices from those of the risk premium, which otherwise are
di¢ cult to be distinguished to each other (see, e.g., Fama and French (2004)).
To empirically examine the validity of the above predictions of the sentiment hypothesis,
in this paper we suggest a panel data threshold model of current market share price deviations
from their intrinsic values using as threshold variable alternative choices of investor sentiment
indices, suggested in the literature. These include the consumer con…dence indicator (see
Schmeling (2009)), the economic sentiment index suggested by Baker and Wurgler (2006)
and the weighted average of di¤erent sectors con…dence indices. The periods of the sample in
which abrupt share price adjustments occur can be identi…ed by the excess-sentiment regime
of the model, while those associated with generic sentiment e¤ects by the low-to-normal-
sentiment regime. The suggested model enables us to estimate from the data, at hand,
the threshold value of a sentiment index above (or below) which investor and share market
sentiment is more likely to switch and, based on this estimate, examine if the sentiment e¤ects
1
Proponents of this behavioral approach to asset pricing include De Bondt and Thaler (1985), Bernard
and Thomas (1989), Lee et al (1991), Barberis et al. (1998), Daniel et al (2001), Brooks and Katsaris (2005),
and Li et al (2008), inter alia.

2
are asymmetric across the two sentiment regimes. The fact that the model can estimate
the threshold parameter endogenously from the data based on an investor sentiment index
provides useful information about the level of shares market sentiment that can trigger price
corrections, or bubbles burst. It also distinguishes our model from others in the literature
assigning ad-hoc dummy variables to capture the e¤ects of investor sentiment on share
prices or returns relying on out-of-sample (exogenous) information (see, e.g., Stambaugh et
al (2012), Lutz (2015), Massa and Yadav (2015), and Zhu and Niu (2016)).2 In addition to
this, the panel data nature of the model also allows us to investigate the e¤ects of investor
sentiment on the cross-section distribution of …rm-speci…c characteristics of share prices, like
the size and book-to-market ratio, at each regime of the model. As aptly noted, …rst, by
Baker and Wurgler (2006), shares with small size and low book-to-market ratio should be
a¤ected by investor sentiment e¤ects due to limits to arbitrage. Our model enables us to
investigate if these e¤ects occur in the low-to-normal or excess investor sentiment regimes.
To our knowledge, this is the …rst paper that address the above questions based on current
market share price deviations from their fundamental values. The prices of currently traded
shares constitute present values of future cash ‡ows discounted by expected (required) share
returns embodying ex ante risk premium e¤ects. This enables us to appraise at which extent
the deviations of the market share prices from their intrinsic values are due to investor
sentiment and/or risk premium e¤ects. Another reason of using share prices, instead of
returns often used in the literature (see, e.g., Shefrin (2015), for a survey), is that returns
are measured by their realized values, ex-post, and thus may not e¢ ciently approximate
their expected values embodied in current share prices when quite persistent and/or big in
magnitude future cash ‡ow shocks occur in future periods (see Ogneva (2012)). Such shocks
may obscure behavioral attitudes of investors in market share prices, when using ex-post
realized returns.
We estimate the suggested model based on price deviations of UK shares listed in the
2
For example, optimistic, or pessimistic, periods of share markets often associated with the peak, or
trough, phases of business cycles are captured by dummy variables based on out-of-sample information
(e.g., o¢ cial government announcements about business cycle phases). This approach may be proved quite
restrictive since not all the peaks (or troughs) of business cycles are associated with optimistic (or pessimistic)
periods of share markets. Note that, since the eminent paper of Fama (1981), share market movements are
understood as being pro-cyclical to business cycle phases.

3
UK stock market from their intrinsic values, over the period 1987-2012. To calculate the
intrinsic value of shares, we employ the residual income model (see, e.g., Ohlson (1995)).
This model relies on accounting fundamentals, like the book value and earnings, which are
publicly available by the yearly …nancial statements of companies. These fundamentals are
positively correlated with future cash ‡ows, thus providing important company information
about share prices (see Patatoukas (2014)). The residual income model facilitates calculation
of the intrinsic values of shares, over a …nite horizon, as it is based on the present value of
abnormal earnings. As competition forces abnormal earnings to zero, one needs estimates
of abnormal earnings only for a few years ahead to explain deviations of share prices from
their book values. Another interesting feature of the residual income model is that it can
exploit analyst earnings forecasts to estimate expected future residual income and, hence,
the intrinsic values of shares. However, note that, since these forecasts may be a¤ected
by investor sentiment (see, e.g., Chen (2011)) and thus may subject to serious forecast
errors (see Hribar and Mclnnis (2012)), in our analysis we employ an exponential smoothing
model to calculate expected company earnings. This model relies on past information of the
actual earnings per company share and their corrections from their analyst forecasts. It is a
model of actual earnings of companies which does not su¤er from the problem of optimistic
sentiment on analyst earnings forecasts, while it can exploit useful information from analysts
in forecasting future actual earnings throughout its error correction term.
The estimation of the model leads to a number of interesting conclusions, concerning
both qualitative and quantitative e¤ects of investor sentiment on share prices. First, it
provides strong evidence of asymmetric investor sentiment e¤ects on share prices between
the two sentiment regimes considered by the model. It is shown that the generic e¤ects of
investor sentiment on share prices are positive in the low-to-normal-sentiment regime. In the
excess-sentiment regime, they become negative and much larger in magnitude than the low-
to-normal regime due to the abrupt adjustments of share prices to their fundamental values.
The excess-sentiment regime is characterized by excessive optimism and it includes periods
of bubbles bursts and cumulative share price adjustments to their fundamental values. These
results are in accordance to the predictions of the sentiment hypothesis, mentioned before.
Second, it is shown that the estimate of the threshold parameter value, distinguishing the

4
two sentiment regimes lies in the top (i.e., 20%) percentile of the empirical distribution of
all the alternative sentiment indices used in our empirical analysis. A …nal conclusion that
can be drawn from the estimation of the model concerns the cross-section distribution of
the investor sentiment e¤ects on share prices, with respect to the …rm-speci…c characteristics
mentioned before. This pattern is found to be signi…cant and in accordance to the predictions
of the sentiment hypothesis for both sentiment regimes.
The paper is organized as follows. Section 2 presents the share valuation model employed
to calculate the intrinsic values of shares. Section 3 discuss possible asymmetric e¤ects of
investor sentiment on share prices, in more details, and it introduces the threshold panel
data model. This section also carries out the empirical analysis of the paper and discuss the
estimation results. Section 4 concludes the paper.

2 The share valuation model

2.1 The residual income model

The residual income model (RIM ) (see, e.g., Ohlson (1995), Feltham and Ohlson (1995),
and Penman and Sougiannis (1998)) is based on the dividend discount model (DDM ), given
as
X1 Et (dit+ )
Pit = , for all i; (1)
=1 (1 + rf )
where Et ( ) denotes the conditional on information set It expectations operator, Pit is the
intrinsic (fundamental) value of a company i share, at time t, dit+ denotes the net dividends
of the company, at time t + , and rf is the risk-free discount rate. Note at this point that
expected dividends in formula (1) are discounted by the risk-free rate rf , as initially assumed
by Ohlson (1995). This is done in order to obtain intrinsic values of shares which are net
of risk premium or investor sentiment e¤ects.3 Thus, the deviations of Pit , calculated based
on (1) or Ohlson’s formula (see (3), below), from their actual, market share prices (denoted
as Pit ), given as Pit Pit , can be also used to test for the existence of time-varying risk
premium e¤ects on share prices Pit , apart from investor sentiment. In order to control for
3
Note that another reason for using rf as a discount factor is that it makes the RIM model internally
consistent with the discounted earnings formula Pit = Eit =rf , used in the literature to evaluate the e¤ects
of investor sentiment on share prices (see, e.g., Barberis et al (1998) and Dechow et al (1999)).

5
the risk premium e¤ects on Pit , in our analysis we will rely on Fama-French’s factors model
(see Fama and French (1993, 1996) and the intertemporal asset pricing framework (see, e.g.,
Cochrane (2005)).4
The RIM share valuation formula can be derived from (1) by the so called clean-surplus
equation, given as
Bit = Bit 1 + Eit dit , for all t, (2)

where Bit is the book value of equity of company i, Eit denotes the earnings of the company
from t 1 to t and Eit dit constitutes the added value of the company net of dividends.
Solving out (2) for dit and substituting into (1) yields the following well known valuation
formula of the RIM :
XT Et (Eit+ rf Bit+ 1)
Pit = Bit + + PiT ; for all i, (3)
=1 (1 + rf )

where PiT is the terminal value of share i.5

The share valuation formula given by (3) assumes that the intrinsic value of a share i,
Pit , depends on the current, t-time book value of its company, Bit , which is a measure of
company’s historical capital invested, and the discounted future abnormal earnings (referred
to as residual income) given as Eit+ rf Bit+ 1. The residual income constitutes the
di¤erence between the company’s earnings Eit+ and the opportunity cost of capital. This
income re‡ects the e¤ects of future earnings from wealth creating activities of the company
on the intrinsic value of its share, at time t. There are two main reasons that one may employ
the RIM valuation formula (3), instead of the DDM one (see equations (1)), to calculate
the intrinsic value of a share (see Lee (1999) and Lee et al (1999), or more recently Penman
(2015)). First, formula (3) relies on values of …rm-speci…c variables, like Bit+ which are
publicly available by the …nancial statements and accounting data of the …rms. Second, it
facilitates calculation of share price Pit over a …nite horizon T , based on forecasts of Bit+
4
Both of these approaches provide a more general framework of pricing risk in shares than the CAPM
(see, e.g., Fama and French (1993)).
5
See, e.g., Penman and Sougiannis (1998). The terminal value PiT is given as

1 Et (Eit+T +1 rie Bit+T )

PiT = T
;
(1 + rie ) (rie gi )

where gi is the growth rate of the company’s earnings.

6
and Eit+ . This can be done under di¤erent assumptions about the terminal value of the
share PiT . One such assumption is that PiT is given by the value of a perpetuity of residual
income Eit+ rf Bit+ 1, with growing earnings, or not. Another sensible assumption about
PiT would be that, after horizon T , the company earns income which is equal to its cost of
capital (i.e., Eit+T +1 = rf Bit+T ), which means that PiT = 0.

2.2 Intrinsic values of share prices based on formula (3)

To calculate the intrinsic values Pit , based on formula (3), and their deviations from their
market values Pit Pit , we rely on data from the London Stock Exchange. Our sample
covers the period from year 1987 to 2012, implying T = 26 yearly observations, and it
includes N = 37 companies from the FTSE 100 index, which have been traded continuously
in the UK stock market during the sample period. Thus, the total number of panel data
observations analyzed is N T = 962 observations. The actual market prices Pit are expressed
in nominal values and have annual frequency. In particular, they are obtained 15 days after
the releases dates of the yearly …nancial statements of the companies of our sample in order
to allow for Pit to absorb any news that are incorporated in the …nancial statements and
accounting data of the companies reported. Thus, our study can be also thought of as
an event study of the …nancial statement announcements on share prices, across di¤erent
investor sentiment regimes.
The intrinsic values Pit are based on data for earnings and book values on the dates that
the yearly …nancial statement releases are made by companies.6 More speci…cally, Bit are
calculated based on data of a company’s balance sheets and Eit are obtained from the pro…ts
6
These data are available on annual basis. The analyst earnings forecasts, extracted from DataStream,
are obtained from the Institutional Brokers’Estimate System (IBES). They are based on combined estimates
of the analysts about a company’s earnings per share that concerns the next …scal year. This is done based
on models-projections and research on the future plans of companies, and they are given on a summary
(consensus) level (i.e., taken as the average of detailed (analyst-by-analyst) forecasts, see also Hughes et al
(2008)). In each year and for every company of our sample, the number of analysts of the IBES basis is
su¢ cient to calculate accurate estimates of earnings forecasts. This number varies mostly from 8 to 30, at
an average of 16 forecasts, and it is available upon request.
Descriptive statistics of these forecasts, over the cross-section and time-dimension of our data, are given
in Table 1. These statistics show that the analyst earnings forecasts also include smaller in terms of size
companies. The correlation coe¢ cient of the analyst earnings forecasts with the actual per share earnings
one period ahead is about 68%, at aggregate level, which means that they contain important information
about future company earnings.

7
and loss accounts. Based on the values of Eit , we can calculate expected future abnormal
PT Et (Eit+ rie Bit+ 1 )
earnings (denoted as AE), de…ned as AE = =1 (1+rf )
, where the expected
values of Eit+ are calculated for T = 5 periods (years) ahead and the forecasts of Bit+ are
obtained as Bit+ = Bit+ 1 + Eit+ Dit+ , where Dit+ denotes the dividend of share i in
period t + (see, e.g., Lee et al (1999)).7 As rf , we employ the three month interest rate
in an annual basis, while Dit+ is estimated using the current dividend payout ratio k, as
Dit+ = Eit+ k.
To calculate the expected earnings Et (Eit+ ), we rely on estimates of the following model:

Eit+ = a Eit + (1 a )(Eit Eita ) + it , for = 1; 2; :::; 5 (4)

where Eita are the analyst earnings forecasts and it is the error term. This is a type of
exponential smoothing model which provides forecasts of actual earnings Eit+ , -periods
ahead, by combining information of the current, actual values of Eit and the analyst earning
forecasts Eita , made one-period back. The right hand side (RHS) regressor term Eit Eita of
model (4) can be thought of as an error correction term taking into account errors (revisions)
made in the analyst earnings forecasts one period back about actual earnings Eit . These
errors can be attributed to excess optimism, as noted in the literature (see the introduction).
Evidence that these errors are signi…cant means that the use of earnings forecasts Eita in
calculating abnormal earnings AE will lead to biased estimates of the intrinsic share prices
Pit , meaning that these prices must not be employed to test the investor sentiment hypothesis
predictions based on share deviations Pit Pit , which is the focus of this paper. We have
found that the inclusion of Eit Eita into model (4) considerably improves the performance of
the model to predict future Eit+ , compared to the autoregressive model. As all exponential
smoothing models, (4) constitutes a particularly useful framework to forecast future levels
of Eit+ based on short time-dimension panel data samples, as in our case.
Model (4) was estimated by the least squares dummy variables (LSDV) panel data
method, allowing for …xed e¤ects. Its coe¢ cient estimates and the observed values of Eit
and Eita were used to provide forecasts of Eit+ and calculate AE over the di¤erent forecast
7
In our analysis, we have set PiT = 0, as we have found that the cross-section average of the residual
income Eit+ rf Bit+ 1 after T = 5 periods ahead is not signi…cantly di¤erent than zero, for all t, and
thus its e¤ect on terminal value PiT and current share price Pit can be treated as negligible.

8
horizons = 1; 2; :::; T , as required by formula (3). The estimation results of the model
support its use in forecasting Eit+ and calculating the intrinsic values Pit . The estimates of
the slope coe¢ cient a were found to vary between 0.85 and 0.60 (as increases), implying
a high degree of earnings persistency, for all i, also noted in the literature (see Chen (2011)).
The less than unity estimates of a mean that analyst earnings forecasts Eita constitute bi-
ased upwards estimates of actual earnings Eit . Thus, they should not be used to provide
forecasts of the actual future earnings Eit+ , as argued before. Instead, Eita was found to
contain useful information about Eit+ throughout the error correction term Eit Eita . This
was justi…ed by the estimates of the variance of the error of model (4), for all , which were
found to substantially decrease when Eit Eita is included in the RHS of the model, as an
additional regressor.
To have a pictorial view of how well the valuation formula (3) captures movements in
market share prices, in Figure 1 we graphically present aggregate values of market prices Pit ,
denoted as P_bar, against their intrinsic values, denoted as P _bar; P_bar is calculated
as the cross-section average of market share prices Pit , at any t, while P _bar as the cross-
section average of Pit , at any t, where Pit are calculated based on (3) and the procedure
described above to obtain forecasts of Eit+ . Inspection of Figure 1 indicates that formula
(3) provides intrinsic values of shares which follow very closely the movements of their traded
(market) prices, over medium or long horizons. The graphs of both aggregate series P_bar
and P _bar seem to follow the pattern of a pair of cointegrated time series, driven by a
common stochastic trend. The two series graphically seem to correspond one-to-one to each
other and have a cointegrating vector (1,-1).8 This can explain the stationary pattern of
the deviations between P_bar and P _bar (i.e., P_bar P _bar), which are graphically
presented in Figure 2.
Evidence of cointegration between the above aggregate series of share prices can be obvi-
ously attributed to the fact that share prices Pit and Pit are cointegrated at individual level,
for all i. This result is formally established in the appendix based on the recent developments
of panel data integration-cointegration analysis. In particular, in the appendix we show that
8
Note that evidence that, at aggregate level, share prices and accounting fundamentals are cointegrated
are provided by Curtis (2011), for the US stock market.

9
Pit and Pit are integrated and cointegrated series, for all i, with cointegrating vector (1,-1).
Note that inference procedures based on panel data integration-cointegration analysis are
more powerful than the single time series ones, since they exploit information across both
the time and cross-section dimensions of the data (see, e.g., Karavias and Tzavalis (2016,
2017)).
Establishing that the individual series Pit and Pit are cointegrated, with cointegrating
vector (1,-1), is a crucial step in our analysis. It means that the RIM constitutes a valid long-
run structural relationship to obtain intrinsic (or fundamental) values of shares and the price
deviations Pit Pit are stationary (see, e.g., Cuthbertson and Nitzsche (2004)). Stationarity
of Pit Pit is also claimed by both the sentiment and time-varying risk premium hypotheses.
More speci…cally, the sentiment hypothesis asserts that positive price deviations Pit Pit
should be temporary and converge to zero when share prices adjust to their fundamentals.
On the other hand, the risk premium hypothesis predicts that prices Pit embody mean
reverting risk premium e¤ects.

[INSERT FIGURES 1 AND 2 ]

3 A panel data threshold model of investor sentiment

e¤ects on share prices
The threshold panel data model that we suggest to investigate the e¤ects of the investor
sentiment on share price deviations Pit Pit is given as follows:

X
J
(d ) (d )
X
K
Pit Pit = ci + cj L dumjt L + k xkt
j=1 k=1

!
X
J X
J
(dL ) (d )
+ j zjit + j (dumjt L zjit ) + SEN Tt IfSIt < qg
j=1 j=1
!
X
J X
J
(dL ) (d )
+ j zjit + j (dumjt L zjit ) + SEN Tt IfSIt qg + uit ; (5)
j=1 j=1

for i = 1; 2; :::; N shares and t = 1; 2; ; :::; T time series observations, where If.g is an index
function which takes value 1 (or zero) if the event fSIt qg (or fSIt < qg) occurs, where

10
SIt denotes an investor sentiment index which constitutes the threshold variable of the
model, and q is the threshold parameter above (or below) which regime switching occurs.
In particular, IfSIt < qg corresponds to the low-to-normal-sentiment regime of the share
market, while IfSIt qg to the excess one. The explanatory variables of the model are
de…ned below, where we discuss the e¤ects of them on share prices, in more details, according
to the sentiment hypothesis.9
The panel data nature of model (5) enables us to examine the validity of both the
generic and cross-section predictions of the sentiment hypothesis across the two sentiment
regimes, within the same econometric framework. The generic e¤ects of investor sentiments
on Pit Pit are common across all shares i, at any time t: These can be captured by the
explanatory variable of the model SEN Tt . This variable is common across all i and it is
de…ned as the percentage change of sentiment index SIt so as to be able to capture the generic
e¤ects of positive, or negative, changes of SIt on Pit Pit , over time. More speci…cally, the
inclusion of SEN Tt in the RHS of the model enables us to formally test if the generic e¤ects
of variable SEN Tt on Pit Pit in the low-to-normal-sentiment regime are positive (i.e.,
> 0), as predicted by the sentiment hypothesis. As argued by Brown and Cli¤ (2005), and
Shefrin (2008), in the low-to-normal-sentiment regime optimistic expectations (i.e., positive
sentiments) about share prices are build up by investors. These sentiments drive market
prices of shares up, constantly. On the other hand, the e¤ects of SEN Tt on Pit Pit are
expected to be negative (i.e., < 0) in the excess-sentiment regime due to sudden share
price corrections to their fundamental values, often occurring in this regime. As aptly argued
by Thorp (2004), the excess-sentiment regime, re‡ecting extreme bullish sentiment levels of
investor sentiment, comes after strong and persistent market run-ups when investors are
fully invested in shares. In this regime, a market downturn is expected to happen. This
will signal an abrupt transition to the low-to-normal-sentiment regime. Estimation of the
threshold parameter q from the data, endogenously, throughout model (5) will indicate the
value of SIt above which share price corrections (i.e., market bubbles burst) are more likely
to occur.
9
Note that the model does not allow for regime shifts in all of its slope coe¢ cients. A more general
speci…cation of the model allowing for this can be also considered, but this is an empirical matter.

11
 The group of explanatory variables zjit ( j = 1; 2; :::; J), included in the RHS of model
(5), de…ned over the cross-section and time-series observations of the data, consists of …rm-
speci…c variables which include the following: the size (capitalization) of a …rm (denoted
as SIZEit ), the book-to-market value (denoted as (B=M )it ), the earnings-to-price ratio
(denoted as (E=P )it ) and the dividend-to-price ratio (denoted as (D=P )it ). These variables
can control for the e¤ects of the Fama-French (F F -model) risk premium factors on share
price deviations Pit Pit (see Fama and French (1993, 1996), and Brennan and Xia (2001)).
According to the F F -model, the risk premium e¤ects of SIZEit on Pit (and, hence, Pit Pit )
should be positive (which means negative for expected returns), while those of (B=M )it
should be negative.10 The risk premium e¤ects of …rm-speci…c variables (E=P )it and (D=P )it
on Pit Pit are expected to be similar to those of (B=M )it . They will tend to reduce market
prices of shares in order to compensate investors for risk premium e¤ects due to possible
earnings or dividends loses. According to the risk premium hypothesis, the risk premium
e¤ects on Pit , or Pit Pit , should be independent of the sentiment regime of the market.
This means that the estimates of slope coe¢ cients j and j should not di¤er across the two
sentiment regimes of the model. This hypothesis can be formally tested based on model (5).
Apart from the F F -model risk premium factors, in the RHS of model (5) we also include
a group of macroeconomic variables (denoted as xkt (k = 1; 2; ::; K)) to control for time-
varying risk premium e¤ects on Pit Pit due to changes in the business cycle conditions of
the economy (see, e.g., Cochrane (2005)). The variables capturing these e¤ects constitute
common factors, for all shares i. They often include the GDP growth rate (denoted as
GROW T Ht ), the di¤erence between the yields of the long and short term bonds (denoted
as T ERMt ), the in‡ation rate (denoted as IN Ft ), the …rst-di¤erence in the short-term risk-
free rate used as a discount factor in formula (3) (denoted as DFt ), the percentage change
in real exchange rate (denoted as EXCHt ) and the aggregate stock market return (denoted
as M ARKETt ). See the studies of Chen (1991), and Ferson and Korajczyk (1995), or more
recently in Flannery and Protopadakis (2002).11
10
The positive sign e¤ect of the size risk premium factor on share prices implies that small cap shares
generate greater returns than the large cap ones (see also Banz (1981)). This overperformance of the small
cap shares is attributed to an additional risk factor, according to the F F -model.
11
Note that the variable M ARKETt can be also considered as a F F -model risk premium factor. However,
since it captures aggregate movements of share prices, we include it in the group of macroeconomic variable.

12
 (d ) (d )
Finally, the group of truncated variables dumjt L zjit , for all j, where dumjt L denotes a
qualitative dummy which equals unity if zjit takes a value in its lowest distribution percentile
dL (e.g., dL =10%), for all j and t, and zero otherwise, re‡ects the sentiment e¤ects of the
group of the …rm-speci…c variables zjit on share price deviations Pit Pit .12 These truncated
variables can capture small size, no earnings and non-dividend paying, as well as extreme
growth e¤ects on Pit , or Pit Pit , which are often attributed to investor sentiment (see, e.g.,
Wurgler and Zhuravskaya (2002), Abreu and Brunnermeier (2003), and Baker and Wurgler
(2006), inter alia). As argued in these studies, shares with the above features are very costly
to arbitrage, and thus their market prices do not discount investment risk premium e¤ects
in equilibrium. This does not happen with shares characterized by substantial earnings, low
growth, large size and dividend paying, which are often traded in the market and their prices
(d )
are also adjusted for risk premium e¤ects. Thus, the sign of the e¤ects of dumjt L zjit on
Pit Pit should be opposite to those of …rm-speci…c variables zjit , de…ned over the whole
support interval (or in their upper 90th percentile) of their empirical distribution, which
capture the F F -model risk premium e¤ects.
The statistical signi…cance of the truncated variables and their interactions depends
partly on the availability of enough observations in the …rst decile of the sample, that will
minimize standard errors. Given that the panel data model (5) is homogeneous in time, it
is not necessary to have many observations in the lower decile at each year, rather its the
total number of observations in the lower decile over all years that matters. In our data,
we have 96 observations in the lower deciles for each of (E=P )it ; (B=P )it and (D=P )it and
93 observations in the lower decile of SIZEit ; therefore, we have su¢ cient data to identify
cross-sectional e¤ects. Firms can move in and out of the decile every year, as it happens with
the periodic (monthly in Fama and French 1993) portfolio rebalancing in the portfolio-sorts
approach.

12
Note that the choice of the decile that we can truncate the empirical distributions of random variables
zjit in order to isolate possible investor sentiment e¤ects on share prices is an empirical matter. In our
empirical analysis, we have found that our results remain robust to a choice of dL = 5%.

13
3.1 Estimation results
3.1.1 Macroeconomic and investor sentiment data

To estimate model (5), we employ the UK data set of shares prices used in the previous
section to calculate the intrinsic values of shares Pit . The macroeconomic variables employed
in the estimation of the model are de…ned and measured as follows: The growth rate variable
(GROW T Ht ) is based on the UK GDP at constant prices. In‡ation rate (IN Ft ) is based on
the UK consumer price index. T ERMt is the di¤erence between the yield of the 10-year bond
and the three-month interest rate. DFt is the …rst di¤erence of the three-month interest rate.
EXCHt is the percentage change of the real e¤ective exchange rate and the aggregate UK
share market return (M ARKETt ) is based on the F T SE 100 price index. The de…nitions
of the …rm-speci…c variables employed in the estimation of the model are given before (see,
Sections 2 and 3). Regarding the variable measuring SIZEt , this is de…ned as the market
share price Pit times the number of shares in circulation (see Fama and French (1993)). Note
that all macroeconomic variables and the sentiment variable SEN Tt are measured before
the release dates of the yearly …nancial statements of the companies used of our sample.
To see if our estimation results are sensitive to di¤erent measures of investor sentiment,
we have estimated model (5) based on three di¤erent sentiment indices SIt and, hence, mea-
sures of SEN Tt . The …rst is based on the sentiment index SIt provided by the DataStream
for the UK stock market. This index is a weighted average (W A) of the following …ve in-
dividual con…dence indicators: an industrial, service, …nancial service, consumer con…dence,
retail trade con…dence and construction con…dence, and is denoted as W ACI. Compared
to consumer con…dence indicator (CCI), often used in empirical studies to re‡ect sentiment
e¤ects (see, e.g., Schmeling (2009)), W ACI gives a more objective and representative mea-
sure of investor sentiment conditions held in the economy and stock market, at any point of
time t, as it is the average of di¤erent con…dence indicators. We will henceforth denote this
measure of variable SEN Tt , as SEN Tt (W ACI).
The second measure of SEN Tt is based on CCI, de…ned before, and is denoted as
SEN Tt (CCI). Finally, following recent literature (see Baker and Wurgler (2006)), the third
measure of SEN Tt is based on an investor sentiment index constructed as a principal com-
ponent factor (P CF ) of the following economic variables: the market share turnover, the

14
numbers of IPOs within each month, the consumer con…dence, the closed-end fund discount,
the put-call trading volume ratio, the put-call interest ratio and market volatility index.
The de…nition of these variables are given in the Appendix. The sentiment variable based
on P CF is denoted as SEN Tt (P CF ). Note that, due to the non availability of data for
some of the above economic series, the values of variable SEN Tt (P CI) are available only
from year 2000 to 2012.
A graphical presentation of the above three sentiment indices is given in Figure 3. Inspec-
tion of this …gure indicates that, for most periods of our sample, all the above three indices
move very closely to each other, especially W ACI and P CF . The correlation coe¢ cients
among them are found to be very high; they vary between 0.70 and 0.80. Another interesting
feature that can be noted by the inspection of Figure 3 is that the three sentiment indices
seem to capture the well known turning points of the UK share market sentiment (i.e., stock
market peaks and troughs) associated with bubbles burst, or speculative episodes, like those
in years 1987, 1997, 2001-2002 and 2007-2008.

[INSERT FIGURE 3]

Before turning into the discussion of the estimation results of threshold model (5), next
we present some descriptive statistics of all variables of the model, including the dependent
variable Pit Pit . These are reported in Tables 1A,1B and 1C and can reveal whether basic
features of our data are consistent with the predictions of the sentiment hypothesis, discussed
before. In particular, Table 1A presents key descriptive statistics, like the mean, standard
deviation, and max and min values, while Table 1B presents the correlation coe¢ cients
among all the variables. Finally, Table 1C presents correlation coe¢ cients of the …rm-speci…c
variables (E=P )it , (B=M )it , (D=P )it and SIZEit with Pit Pit , variables SEN T (W ACI)t
and SEN T (CCI)t , and indices W ACI and CCI; for which there is data availability for the
whole sample period. These coe¢ cients concern the bottom 10th and upper 90th percentiles
of the empirical distributions of the …rm-speci…c variables zjit . The results for the 90th
percentile may better distinguish the e¤ects of risk premium on Pit Pit , compared to
investor sentiment which dominate shares with very small, or zero, values of zjit .

15
 The above tables provide a number of results which are worthy of comments. First,
the average value of Pit Pit is positive which means that the market prices Pit tend to
be above their intrinsic values Pit , over the whole sample, which is in accordance to the
sentiment hypothesis. However, the substantial volatility of Pit Pit , and its negative large
minimum values, observed during our sample, mean that there is also high probability that
the sign of Pit Pit can become negative, re‡ecting downward adjustments in market prices
Pit due to risk premium e¤ects. Second, the sign of the average values of SEN Tt (CCI), and
SEN Tt (W ACI) is negative. This can be attributed to the dominance of substantial negative
investor sentiment adjustments occurred during the sample, associated with the …nancial
crises and bubbles burst mentioned before. Third, the degree of correlation of Pit Pit with
the explanatory variables of the model is low, which is consistent with evidence supporting
very small predictability of share returns and/or price changes (see, e.g., Avramova and
Chordiab (2006)). Note that this is also true for the three di¤erent measures of SEN Tt
considered. Although the sign of the correlation between SEN Tt and Pit Pit is positive,
which is in accordance with the sentiment hypothesis, its magnitude is quite low.
The explanatory variables which are found to be at most correlated with Pit Pit are the
…rm-speci…c variables (B=M )it , (D=P )it ; and SIZEit : The sign of the correlation coe¢ cients
of (B=M )it and (D=P )it with Pit Pit is negative, while the correlation of SIZEit with
Pit Pit is positive, as predicted by the F F -model of risk premium. This is true for all
sentiment indices considered and it also holds for the 90th percentile results of the above
…rm-speci…c variables (see Table 1C). However, these results do not hold for the bottom 10th
percentile of the distributions of the above variables. As can be seen from Table 1C, the sign
of the correlation coe¢ cients of (B=M )it and (D=P )it with Pit Pit becomes positive in that
percentile, for both sentiment indices CCI and W ACI considered. The same happens with
(E=P )it , while the correlation between SIZEit and Pit Pit tends to zero. These results are
consistent with the predictions of the investor sentiment hypothesis, stating that shares with
low values of (E=P )it , (B=M )it and (D=P )it are positively a¤ected by investor sentiment
e¤ects.13
13
Note that the much closer to zero value of the correlation coe¢ cient of SIZEit in its bottom 10th
percentile with Pit Pit , compared to its upper 90th percentile, may be attributed to fact that the sentiment
e¤ects captured by this …rm-speci…c variable may not so strong compared to the risk premium ones also

16
 Finally, the results of Table 1C indicate that the three alternative measures of investor
sentiment employed in our analysis are correlated with the macroeconomic variables (espe-
cially GROW T Ht , M ARKETt , EXCHt and IN F Lt ), but not at a level which can cause
multicollinearity problems in the estimation of the model. As we have also found in our
empirical analysis, orthogonalizing SEN Tt to macroeconomic variables does not change our
results.

[INSERT TABLES 1A, 1B and 1C]

3.1.2 Estimates

In this section, we present estimates of threshold model (5). We compare the estimation
results of the model to those of the following linear model:

X
J
(dL ) (d )
X
K X
J X
J
(dL ) (d )
Pit Pit = ci + cj dumjt L + k xkt + j zjit + j (dumjt L zjit )+ SEN Tt +eit ,
j=1 k=1 j=1 j=1
(6)
ignoring threshold e¤ects. The estimates of the linear model (6) are presented in Table 2,
while those of the threshold model (5) in Tables 3A-3B; Table 3A presents the estimates of
the threshold model with regime shifts only in the slope coe¢ cient of variable SEN Tt , while
Table 3B with regime shifts in the slope coe¢ cients of both SEN Tt and the …rm-speci…c
variables of the model. For all the above tables, in columns (I), (II) and (III) we respectively
report estimates of the models based on the three alternative measures of variable SEN Tt
considered (i.e., SEN Tt (W ACI), SEN Tt (CCI) and SEN Tt (P CF )).14
The comparison of the estimation results of models (6) and (5) can show if ignoring
threshold-type asymmetries in the investor sentiment e¤ects on share price deviations Pit Pit
undermines these e¤ects. On the other hand, the comparison of the estimation results of
threshold model (5) with regime shifts in the …rm-speci…c variables to those without (see
captured by this variable.
14
Note that, in Table 3B, we do not present estimates of the threshold model (5) based on the measure
of investor sentiment given the principal component factor (P CF ), due to the no availability of data of this
factor for the whole sample. This does not leave enough sample information, over the time dimension of the
data, to identify the slope coe¢ cients of …rm speci…c variables zjit and dumjt zjit , under the two sentiment
regimes.

17
Tables 3B and 3A, respectively) can shed light on the following questions: First, if the
version of model which also considers regime shifts in the slope coe¢ cients of the …rm-speci…c
variables constitutes the best speci…cation of the data and, second, if the investor sentiment
e¤ects on Pit Pit with respect to the cross-section distribution of the …rm-speci…c features of
shares predicted by the sentiment hypothesis characterizes only the low-to-normal-sentiment
regime, as it is argued before.
To estimate the linear model (6), we rely on the least squares dummy variables (LSDV)
estimation method. This method can control for individual e¤ects of the cross-section units
of our panel data set on Pit Pit , by employing appropriately speci…ed dummy variables in
the intercept term of the model, for all i. The threshold model (5) is also estimated based
on the LSDV method. This method is appropriately modi…ed to estimate the value of the
threshold parameter q endogenously from the data based on a search procedure (see Hansen
(1999)). In particular, the optimum estimate of q is taken as that corresponding to the
minimum value of the minimum residual sum of squares of the model, over di¤erent values
of q, taken from the empirical distribution of threshold variable SIt searched for a threshold,
after trimming out to top and bottom 10% percentile values of this distribution. As shown
by Hansen (1999), this searching procedure provides super consistent estimates of q. Given
these estimates, we can carry out inference on the slope coe¢ cients of model (5) based on
the standard asymptotic theory.

[INSERT TABLES 2, 3A AND 3B]

A number of interesting conclusions can be drawn from the results of Tables 2 and 3A-
3B. First, they clearly indicate that the panel data threshold model (5) captures important
regime-shift type of asymmetries both in the sign and magnitude of the generic e¤ects of
(d )
sentiment variable SEN Tt and the truncated …rm-speci…c variables dumjt L zjit on price
deviations Pit Pit . The signi…cance of these asymmetries can be formally justi…ed by the
LR test statistic reported in Tables 3A and 3B, which tests the null hypothesis H0 : j = j;
(dL ) (dL ) (dL ) (dL )
j = j and = , for all j, against its alternative Ha : j 6= j or j 6= j

or 6= . The values of this test statistic, reported in the tables, clearly reject the above

18
null hypothesis (implying no investor sentiment regime shifts), at the 5% level, against its
alternative hypothesis supporting model (5), with regime shifts. Note that, for variable
SEN Tt , the regime-shift asymmetries are more profound for the full speci…cation of the
model, allowing also for shifts in the e¤ects of …rm-speci…c variables on Pit Pit (see the
di¤erences in the estimates of and between Tables 3A and 3B). This speci…cation is
found to better …t into the data, in terms of the variance of the error term of the model uit .
Second, the coe¢ cients estimates of the model reported in Tables 3A-3B are consistent
with the predictions of the sentiment hypothesis, across the two regimes. They show that
the e¤ects of SEN Tt on Pit Pit are signi…cant, at the 5% level, and are positive in the
low-to-normal-sentiment regime, where optimistic investor sentiments are developed, and
negative in the excess. In the excess sentiment regime, a negative change in SEN Tt means
that will reduce the value of Pit Pit re‡ecting corrections of Pit towards their fundamental
values. The di¤erences in the magnitude of the above sentiment e¤ects between the two
regimes are substantial. For instance, based on the slope coe¢ cient estimates of variable
SEN Tt for W ACI, reported in Table 3B, we can see that a positive change of SEN Tt by
1% leads to an increase of Pit Pit by £ 0.02 in the low-to-normal-sentiment regime, while,
in the excess, implies a decrease by £ 0.12, which is much bigger than £ 0.02. These results
hold for the other two sentiment indices considered. The much bigger in magnitude e¤ects
of SEN Tt on Pit Pit in the excess-sentiment regime can be attributed to the substantial
cumulative corrections of share prices to their fundamental values, occurring within this
regime. The comparison of the above results to those of Table 2 indicates that ignoring the
existence of the above sentiment regime shifts undermine, critically, the e¤ects of SEN Tt
on Pit Pit . It leads to estimates of the e¤ects of SEN Tt on Pit Pit which are quite small
in magnitude and, more importantly, it misses the negative e¤ects of SEN Tt on Pit Pit in
the excess-sentiment regime.
Apart from the generic e¤ects of SEN Tt , evidence supports the in‡uence of share prices
Pit by investor sentiment can be also obtained by the estimates of the slope coe¢ cients of
(d )
the truncated …rm-speci…c variables dumjt L zjit , for dL = 0:10, on Pit Pit . The results of
Table 3A clearly indicate that, for W ACI and CCI, the estimates of the above coe¢ cients
are signi…cant for most of the …rm-speci…c variables considered, at the 5% level, while their

19
signs are positive for (E=P )it , (B=M )it , and (D=P )it , and negative for SIZEit . These results
are consistent with those of Baker and Wugler (2006) and in accordance to the sentiment
(d )
hypothesis.15 Also note that the signs of the slope coe¢ cients of dumjt L zjit are consistent
(dL ) (dL )
across the two sentiment regimes of the model (see estimates of j and j , reported
in Table 3B), as one would expect. The positive sign of the slope coe¢ cients of (E=P )it ,
(B=M )it , and (D=P )it in the excess-sentiment regime means that, the lower the value of
these variables, the stronger the e¤ects of share prices corrections will be. Regarding the
(dL ) (dL )
signi…cance of j and j , across the two regimes, the results of the table indicate that
(dL )
these coe¢ cients (namely j ) are signi…cant in the low-to-normal sentiment regime, at
the 5% or 10% levels, for all the …rm-speci…c variables with the exception of SIZEit . In
(dL )
the excess sentiment regime, j is signi…cant for (D=P )it for W ACI, and (E=P )it and
(B=M )it for CCI.
Third, the estimates of the threshold parameter q, reported in Tables 3A-3B, indicate
that the long-term corrections of share prices to their intrinsic values, occurred in the excess-
sentiment regime, are more likely to happen for values of the empirical distribution of the
sentiment indices in their top percentile (e.g., 15% and 20%, for W ACI and P CF ). As
can be pictorially con…rmed by referring to Figure 3, which presents the three sentiment
indices considered in our analysis against the sample estimates of their threshold parameter
r reported in Table 3A, the corrections of share prices occurring in the excess-sentiment
regime are associated with the end of periods of excessive optimism of the UK share market.
In particular, they cover periods of bubble bursts and/or …nancial crises of this market of
years 1987-1988, 2000, 2004 and 2007-2008, as predicted by the sentiment hypothesis. These
results are consistent with those of Anderson and Brooks (2014), measuring the e¤ects of
growing (or collapsing) bubbles on share returns at individual level.
Fourth, regarding the risk premium e¤ects captured by model (5), our results clearly
support the view that these e¤ects also in‡uence share prices Pit . More speci…cally, the
results of Tables 3A and 3B indicate that the estimates of the slope coe¢ cients of the …rm-
speci…c variables zjit , i.e., j and j, capturing the F F -model risk premium factors, have
15
Note that, for P CF , they have the correct sign and are signifcant, at 5% level, only for (B=M )it . This
result may be attributed to the lack of data for P CF , over the whole sample.

20
the correct sign predicted by the risk premium hypothesis, and they are signi…cant for most
of these variables, at the 5% level. This is a more clear cut result for the case that investor
sentiment is measured by W ACI and CCI. Comparing the results across the two sentiment
regimes (see Table 3B) shows that the slope coe¢ cients of zjit tend to be signi…cant mainly in
the low-to-normal sentiment regime. In the excess one, they are not so important, with the
exception of (B=M )it . This result can be attributed to the cumulative share price corrections
occurring in the excess sentiment regime. These may dominate the risk premium e¤ects on
share prices. For both regimes, the opposite sign of the estimates of j and j to those of
(dL ) (dL )
slope coe¢ cients j and j , capturing investor sentiment e¤ects, means that investor
sentiment tend to reverse the relationship between risk premium and share prices.
Finally, regarding the risk premium e¤ects captured by the group of macroeconomic
variables, the results of the table imply that these e¤ects are also priced in the share market.
The estimates of their slope coe¢ cients k indicate that, with the exception of in‡ation rate,
the remaining variables of the above group have signi…cant explanatory power on Pit Pit .
The signs of these estimates correspond to those reported in the literature (see, e.g., Flannery
and Protopadakis (2002)). The negative sign e¤ects of variables GROW T Ht , T ERMt and
DFt on Pit Pit can be interpreted as capturing state (or cyclical) movements of risk premium
on share prices. On the other hand, the positive in sign e¤ects of EXCHt on Pit Pit can be
attributed to the fact that an increase in e¤ective real exchange rate means an improvement
of the international competitiveness of the domestic economy which will reduce the currency
risk embodied in share prices, for foreign investors holding domestic (UK) shares. Note
that the above results hold independently of the presence, or not, of threshold e¤ects in
the model, adding support to the view that risk premium e¤ects are una¤ected by shifts in
investor sentiment.

3.1.3 Robustness of our results

In this section we carry out two exercises to check the robustness of our estimation results.
First, to see if they characterize our data, descriptively, in Table 4 we present values of the
correlation coe¢ cients of price deviations Pit Pit with explanatory variable SEN Tt , and
the …rm-speci…c variables (E=P )it , (B=M )it , (D=P )it and SIZEit in their bottom 10th and

21
upper 90th percentiles of their distributions. This is done across the two investor sentiment
regimes of model (5), identi…ed by the data, based on sentiment indices W ACI and CCI,
which cover the whole period of the sample. The results of this table are consistent with
those obtained by the estimates of the model (see Table 3B). They indicate that there exist
a clear cut change in the sign of the correlation coe¢ cients of SEN Tt with Pit Pit across
the two regimes of the model, while those of the …rm speci…c variables (E=P )it , (B=M )it and
(D=P )it in their bottom 10th percentile have the sign predicted by the sentiment hypothesis.

[INSERT TABLE 4]

The second robustness exercise concerns with the frequency of the data. To see if our
results remain robust to a higher frequency of the data, we have re-estimated model (5)
based on quarterly data which increase the time-dimension of the panel data to T = 104.
Since the …nancial statement announcements are made only annually, we have maintained
their year values as quarterly observations after each year’s announcement, until the next
year’s one (see also Baker and Wurgler (2006)). Based on these observations we have cal-
culated intrinsic prices Pit . The actual prices Pit are selected after the …nancial statement
announcements, as is done for the year frequency of the data, thus preserving the event
study nature of our analysis conditional on the di¤erent regimes of the share market. Se-
lecting actual prices at quarters before the announcements will produce data which lack the
announcement information, or may otherwise include information on events irrelevant to the
…nancial statement. Balance sheet items have also been interpolated in order to get quar-
terly observations, as in Baker and Wurgler (2006). Thus, we have used quarterly values for:
SENT(CCI), SENT(WACI), GDP, in‡ation, exchange rate, risk-free rate and market return.

[INSERT TABLES 5A AND 5B]

The results of the above exercise are reported in Tables 5A-5B; Table 5A presents results
for the case that no regime switching is allowed for …rm-speci…c variables, while 5B for the
full speci…cation of the model. That is, the results of Tables 5A and 5B correspond to those

22
of Tables 3A and 3B, respectively. Note that the tables present results for the sentiment
indices CCI and W ACI, since the data for P CF are not available for the whole sample
period. The comparison of the estimation results of Tables 5A and 5B to those of 3A and
3B, respectively, indicate that there are no important di¤erences in the estimates of the slope
coe¢ cients and the threshold parameter of the model between these two sets of results. This
is true for both the sign and magnitude of the coe¢ cient estimates. An interesting result of
(d )
Table 5B is that the slope coe¢ cient of the truncated …rm-speci…c variable dumjt L zjit , for
zjit = SIZEit , now clearly becomes signi…cant, at the 5% level, for both sentiment indices
(d )
considered. This variable and dumjt L zjit , for zjit =(B=M )it , seem to capture the investor
sentiment regimes, in the excess sentiment regime.

4 Conclusions
Investor sentiment o¤ers an alternative explanation of the deviation of share prices and/or
expected returns from their fundamental values. In this paper, we examine if the e¤ects
of investor sentiment on share price deviations from their intrinsic values are asymmetric
and dependent on the level of shares market sentiment, as is predicted by the sentiment
hypothesis. To address these questions, we suggest a panel data threshold model of the
above share price deviations using as threshold variable an investor sentiment index. The
suggested model enables us to formally test both the cross-section and generic predictions
of the sentiment hypothesis on share prices within the same framework. It also allows us
to endogenously estimate from the data the threshold value of publicly available sentiment
indices above (or below) which market sentiment regime shifts are more likely to occur. To
calculate the intrinsic values of share prices, we rely on the residual income share valuation
model, which relies on publicly available values of accounting fundamentals. The validity of
this model as a structural relationship is tested based on panel data cointegration analysis.
The paper applies the model to a panel data set of shares listed in the UK stock market
over the period 1987-2012 and provides a number of results which demonstrate signi…cant
investor sentiment e¤ects on share prices. First, we show that ignoring the existence of
di¤erent market sentiment regimes in the UK share market undermines the e¤ects of investor

23
sentiment on share prices. Second, we provide clear-cut evidence that generic e¤ects of
investor sentiment on share prices tend to occur in the low-to-normal-sentiment regime,
where investor and market sentiments are build up and tend to be optimistic. In periods
of excess optimism, we …nd that the e¤ects of investor sentiment on share prices become
negative, since abrupt corrections of share prices to their fundamental values tend to occur.
Third, we show that bubble bursts and/or …nancial crises, like those in years 1987-88, 2000,
2004 and 2007-2008, tend to occur in the excess sentiment regime. Finally, another interesting
conclusion that can be drawn from the results of the paper is that shares with small values of
…rm-speci…c characteristics, like the book-to-market value, and the dividend- and earnings-
to-price ratios are strongly a¤ected by investor sentiment. This is true for both sentiment
regimes considered by our model.

References
[1] Abreu, D., and Brunnermeier, M.K., (2003), "Bubbles and crashes", Econometrica, 71,
173-204.

[2] Anderson, K. and Brooks, C. (2014), "Speculative bubbles and the cross-sectional vari-
ation in stock returns", International Review of Financial Analysis, 35, 20-31.

[3] Avramova, D., and Chordiab, T., (2006), "Predicting stock returns", Journal of Finan-
cial Economics, 82, 387–415.

[4] Goizueta Business School, Emory University, Atlanta, GA 30322, USA

[5] Barberis, N., Shleifer, A., and Vishny, R., (1998), “A model of investor sentiment”,
Journal of Financial Economics, 49, 307-343.

[6] Baker, M., and Stein, J.,C., (2004), “Market liquidity as a sentiment indicator”, Journal
of Financial Markets, 7, 271-299.

[7] Baker, M., and Wurgler J., (2006), "Investor sentiment and the cross- section of stock
returns", Journal of Finance, 61, 1645-1680.

24
 [8] Baker, M., and Wurgler J., (2007), “Investor sentiment in the stock market”, Journal
of Economic Perspectives, 21, 29-152.

[9] Banz, R., W., (1981), "The relationship between return and market value of common
stocks", Journal of Financial Economics, 9, 1981.

[10] Bernard, V., and Thomas, J., (1989), “Post-earnings announcement drift: delayed price
response or risk premium?”Journal of Accounting Research, (Suppl.), 27, 1-36.

[11] Breitung, J., and Das, S., (2005), " Panel unit root tests under cross-sectional depen-
dence", Statistica Neerlandica 59, 414–433.

[12] Brennan, M., and Xia, Y., (2001), "Assessing Asset Pricing Anomalies," Review of
Financial Studies, Society for Financial Studies, 14, 905-942.

[13] Brooks, C. and Katsaris A. (2005), "Trading Rules from Forecasting the Collapse of
Speculative Bubbles for the S&P500 Composite Index", Journal of Business, 78, 2003-
2036.

[14] Brown, G., and Cli¤, M., (2004), “Investor Sentiment and the Near-Term Stock Mar-
ket”, Journal of Empirical Finance, 11, 1-27.

[15] Brown, G., and Cli¤, M., (2005), “Investor Sentiment and Asset Valuation”, Journal of
Business, 78, 405-440.

[16] Chen, N. F., (1991), "Financial investment opportunities and the macroeconomy", The
Journal of Finance, 46, 529-554.

[17] Chen, K., (2011), "Investor Sentiment and the Valuation relevance of Accounting In-
formation", WP, Singapore Management University.

[18] Claus, J. and Thomas, J. (2001), "Equity Premia as Low as Three Percent? Evidence
from Analysts’Earning Forecasts for Domestic and International Stock Markets", Jour-
nal of Finance, 56, 1629-1666.

[19] Cochrane, J. H., (2005), Asset Pricing, University of Chicago.

25
[20] Cuthbertson, K. and Nitzsche D (2004), Quantitative Financial Economics: Stocks,
Bonds and Foreign Exchange, 2nd Edition, Wiley

[21] Curtis, A., (2012), "A fundamental-Analysis-Based Test for Speculative Prices", The
Accounting Review, 87, 121-148.

[22] Daniel K., Hirshleifer, D., and Subrahmanyam A., (2001), “Overcon…dence, Arbitrage,
and Equilibrium Asset Pricing”, Journal of Finance, 3, 921-965.

[23] De Bondt, W., and Thaler, R., (1985), “Further Evidence on Investor Overreaction and
Stock Market Seasonality”, Journal of Finance, 42, 557-581.

[24] Dechow, P., Hutton A., and Sloan R.G. (1999), "An Empirical Assessment of the Resid-
ual Income Valuation Model", Journal of Accounting and Economics, 26, 289-324.

[25] Easton, P.D., (1985), "Accounting earnings and security valuation: empirical evidence
of the fundamental links", Journal of Accounting Research, 23, 54-77.

[26] Fama, E., (1981), "Stock returns, real activity, in‡ation, and money", American Eco-
nomic Review, 71, 545-565.

[27] Fama, E., and French, K., (1993), "Common risk factors in the returns on stocks and
bonds", Journal of Financial Economics, 25, 23-49.

[28] Fama, E., and French, K., (1996), “Multifactor Explanations of Asset Pricing Anom-
alies”Journal of Finance, 51, 55-84.

[29] Fama, E., and French, K. (2004), “The Capital Asset Pricing Model: Theory and
Evidence”, Journal of Economic Perspectives, 18., 25-46.

[30] Feltham, G., and Ohlson. J., (1995), “Valuation and Clean Surplus Accounting for
Operating and Financial Activities”, Contemporary Accounting Research, 1, 689-731.

[31] Ferson W. E., and Korajczyk, R. A., (1995), “Do Arbitrage Pricing Models Explain the
Predictability of Stock Returns”, Journal of Business, 68, 309-349.

26
[32] Flannery, M. J., and Protopopadakis, A, (2002), "Macroeconomic factors do in‡uence
aggregate stock returns", Review of Financial Studies, 15, 751-782.

[33] Forbes, W., (2009), Behavioral Finance, Wiley & Sons.

[34] Glushkov, D., (2006), “Sentiment Beta”, Working Paper, University of Pennsylvania,
The Wharton School.

[35] Hansen B., (1999), "Threshold e¤ects in non dynamic panels: estimation, testing and
inference", Journal of Econometrics, 93, 345-368.

[36] Harris, R.D.F., and Tzavalis, E., (1999), "Inference for unit roots in dynamic panels
where the time dimension is …xed", Journal of Econometrics, 91, 201-226.

[37] Hribar, P. and Mclnnis, J., (2012), "Investor Sentiment and Analysts Earnings Forecast
Errors", Management Science, 58, 296-307.

[38] Hughes, J., Liu, J and Su, W., (2008), "On the relation between predictable market
returns and predictable analyst forecast errors", Review of Accounting Studies, 266-291.

[39] Im, K.S., Pesaran, H.M., Shin, Y., (2003), "Testing for unit roots in heterogeneous
panels", Journal of Econometrics, 115, 53 - 74.

[40] Jorgensen, B., Lee, Y., and Yoo, Y., (2011), "The Valuation Accuracy of Equity value
Estimates inferred from Conventional Empirical Implications of the Abnormal Earnings
Growth Model: US Evidence", Journal of Business Finance and Accounting, 28, 446-
471.

[41] Karavias, Y., Spilioti S. and Tzavalis E. (2016), “A comparison of Investors’Sentiments

and Risk Premium E¤ects on Valuing Shares”, Finance Research Letters, 17, 2016, 1-6.

[42] Karavias Y., and Tzavalis E., 2014. Testing for unit roots in short panels allowing for
structural breaks. Computational Statistics and Data Analysis, 76, 391-407.

[43] Karavias Y., and Tzavalis E., 2016. Local Power of Fixed-T Panel Unit Root Tests With
Serially Correlated Errors and Incidental Trends. Journal of Time Series Analysis, Wiley
Blackwell, 37, 222-239, 03.

27
[44] Karavias Y., and Tzavalis E., 2017. Local power of panel unit root tests allowing for
structural breaks. Econometric Reviews, 36, 1123-1156.

[45] Karavias, Y, Tzavalis, E., 2019, Generalized …xed-T panel unit root tests. Scand J
Statist; 46: 1227–1251.

[46] Lamont, O.A., and Thaler, R., (2003),"Can the market add and subtract? Mispricing
and tech stock carve-outs", Journal of Political Economy, 111, 27-268.

[47] Lee, C., Shleifer, A., and Thaler, R. (1991), “Investor Sentiment and the Closed-End
Fund Puzzle”, Journal of Finance, 46, 75-109.

[48] Lee, C., Myers, J. and Swaminathan, B., (1999), “What is the fundamental Value of
the Dow?”, Journal of Finance, 693-741.

[49] Lee, C., (1999), “Accounting-Based Valuation: Impact on Business Practices and Re-
search”, Accounting Horizons, 13, 413-425.

[50] Lemon M. and Portniaguina E., (2006), “Consumer Con…dence and Asset Prices: Some
Empirical Evidence”, Review of Financial Studies, 19, 1499-1529.

[51] Levin, A., Lin, C.F., and Chu C-S. J., (2002), "Unit root tests in panel data: asymptotic
and …nite-sample properties". Journal of Econometrics, 108, 1-24.

[52] Lutz, C., (2015), "The asymmetric e¤ects of investor sentiments", Macroeconomic Dy-
namics, 1-27.

[53] Massa, M., and Yadav, V., (2015), "Investor Sentiment and Mutual Fund Strategies",
Journal of Financial and Quantitative Analysis, 4, 699-727.

[54] Li, X., Mi¤re, J., Brooks, C. and O’Sullivan, N. (2008), "Momentum pro…ts and time-
varying unsystematic risk", Journal of Banking & Finance, 32, 541-558.

[55] Neal R., and Wheatley, S. M., (1998), “Do Measures of Investor Sentiment Predict
Returns?”Journal of Financial and Quantitative Analysis, 33, 523- 547.

28
[56] Ogneva, M. (2012), "Accrual Quality, Realized Returns and Expected Returns: The
importance of Controlling for Cash Flow Shocks", Accounting Review, 87, 1415-1444.

[57] Ohlson, J., (1995), “Earnings Book Values and Dividends in Security Valuation”, Con-
temporary Accounting Research, 11, 661-687.

[58] Patatoukas, P., (2014), "Detecting News in Aggregate Accounting Earnings: Implica-
tions for Stock Market Valuation", Review of Accounting Studies, 19, 134-160.

[59] Pedroni, P., (1999), “Critical Values For Cointegration Tests in Heterogeneous Panels
With Multiple Regressors”, Oxford Bulletin of Economics and Statistics, 61, 653-670.

[60] Pedroni, P., (2001), " Purchasing power parity tests in cointegrated panels" Review of
Economics and Statistics, 83, 727-731.

[61] Penman, S., (2005), " Discussion on ’On Accounting-Based valuation Formulae’ and
’Expected EPS and EPS Growth as Determinants of Value’, Review of Accounting
Studies, 10, 367-378.

[62] Penman, S., (2015), "Valuation Models: An Issue of Accounting Theory", Ch 11 in

Routledge Companion to Financial Accounting Theory (edited by S. Jones), Routledge.

[63] Penman, S., and Sougiannis, T., (1998), “A comparison of dividend, cash ‡ow, and
earnings approaches to equity valuation”, Contemporary Accounting Research, 15, 343-
383.

[64] Scheinkman, J., A., and Xiong, W., (2003), “Overcon…dence and Speculative Bubbles”,
Journal of Political Economy 111, 1183-1220.

[65] Shefrin, H., (2001), Behavioral Corporate Finance, Journal of Applied Corporate Fi-
nance, 14, 113-126.

[66] Shefrin, H., (2008), Risk and Return in Behavioral SDF-based Asset Pricing Models,
Journal of Investment Management, 6, 4-22.

[67] Shefrin, H., (2015), “Investors’Judgments, Asset Pricing Factors and Sentiment”, Eu-
ropean Financial Management, 21, 205-227.

29
[68] Shefrin, H., and Statman M. (2003), “Style of Investment Expectation”, The Handbook
of Equity and Style Management, Goggin and Fabozzi eds. New York: Wiley.

[69] Stambaugh, R., Yu, J., and Yuan, Y. (2012), "The Short of it: Investor Sentiment and
Anomalies, Journal of Financial Economics, 104, 288-302.

[70] Statman, M., (2011), What Investors Really Want, McGraw-Hill.

[71] Schmeling, M., (2009), Investor Sentiment and Stock Returns: Some International Ev-
idence”, Journal of Empirical Finance, 16, 394-408.

[72] Zhu, B. and Niu F., (2016), "Investor Sentiment, Accounting Information and Stock
Price: Evidence from China", Paci…c-Basin Finance Journal, 38, 125-134.

[73] Wurgler, J., and Zhuravskaya E. (2002), “Does Arbitrage Flatten Demand Curves for
Stocks”, Journal of Business, 75, 4, 583-608.

5 Appendix
In this appendix, we …rstly present the results of our panel data integration-cointegration
analysis for share prices Pit and their intrinsic values Pit . Secondly, we give the de…nitions
of the economic variables used to construct the economic sentiment index (SIt ) based on
principal component analysis.

5.1 Panel data integration-cointegration analysis

5.1.1 Unit root tests

Before testing if Pit and Pit are cointegrated series, we …rst need to establish that both of
these series are integrated series of order one, denoted as I(1). To this end, we conduct a
number of alternative panel data unit root testing procedures, suggested in the literature.
These include the following test statistics: Im’s et al (2003) (denoted as IP S), Levin’s et
al (2002) (denoted as LLC), Breitung’s and Das (2005) (denoted as B-t) and Harris’s and

30
Tzavalis (1999) (denoted as HT ).16 The last test has higher power than the three tests
for panels whose time dimension is short (…nite). The test of Breitung and Das (2005) is
appropriate for large T panels and it is robust to cross-sectional correlation of the panel
autoregression model error terms. All the above test statistics allow for linear trends in the
auxiliary regression testing for unit roots, so as to treat symmetrically deterministic trends
in series Pit and Pit under the null and alternative hypotheses.
The results of our panel data unit root analysis are reported in Table A1. The table
presents values of the above statistics. Their p-values (error type I of rejecting the null
hypothesis of a unit root) are reported in brackets. The results of the table clearly indicate
that Pit and Pit constitute I(1) series, for all i. The values of all the test statistics reported
in the table can not reject the null hypothesis of a unit root at 5%, or at lower levels. The
p-values rejecting the above null hypothesis are all equal to one for all the test statistics,
which mean that the probability of rejecting the null hypothesis of unit root falsely equals
one.

Table A1: Panel unit root tests

HT LLC IP S B-t
Pit 2.40 [1.00] 1.28 [1.00] 0.99 [1.00] 4.37 [1.00]
Pit 4.10 [1.00] 1.44 [1.00] 2.76 [1.00] 9.71 [1.00]
Eit =rf 7.21 [1.00] 42.45 [1.00] 12.99 [1.00] 20.63 [1.00]

Notes: The table presents panel unit root tests for share prices Pit and Pit . HT stands for
Harris and Tzavalis’ (1999) panel unit root test, IP S for Im’s, Pesaran and Shin (2003), LLC
for Levin’s, Lin and Chu (2002) and, …nally, B -t stands for Breitung’s and Das (2005). All the
above test statistics allow for deterministic trends in their auxiliary unit root regression. P-values
of rejecting the null hypothesis of unit root are in brackets.

5.1.2 Cointegration analysis

Our cointegration analysis relies on recent developments of panel data econometrics. Com-
pared to single time series cointegration analysis, panel data cointegration methods can
provide more robust and powerful inference about cointegration between series Pit and Pit ,
for all i, since they are based on disaggregated data. Using panel data sets, we can avoid
16
Recent, more generalized versions of the HT tests allowing for serial correlation (see Tzavalis and
Karavias 2014, 2019)) were also applied, but the results are not presented as the conclusions remain the
same.

31
smoothing out possible di¤erences in the stochastic trends of individual series Pit , or Pit ,
which may be proved very important for erroneously accepting the null hypothesis of coin-
tegration between these series. The results of this analysis are reported in Table A2. In
particular, the table presents values of cointegration test statistics and estimates of the slope
coe¢ cient of the following panel data cointegrating regression:

Pit = ai + bPit + it ; for i = 1; 2; 3; :::; N and t = 1; 2; :::; T , (7)

where it denotes the error term. The slope coe¢ cient estimates reported in the table are
based on the fully modi…ed least squares (FMLS) and dynamic least squares (DLS) panel
data cointegration methods suggested by Pedroni (2001), allowing for heterogenous error
terms it . The cointegration test statistics reported in the table are those of Pedroni (1999).
They are de…ned as panel- and panel-t, and they are based on the residuals of regression
(7).

Table A2: Cointegration results

Pit = ai + bPit + it ; i = 1; 2; :::; N and t = 1; 2; :::; T
Estimates FMLS DLS
b 1:07 (0:12) 1:09 (0:04)
Wald(1) 0:56 [0:57] 1:51 [0:11]
panel-t 6:92 [0:00]
panel- 4:88 [0:00]

Notes: The table presents the results of our cointegration analysis. Pit are calculated based
on Ohlson’s formula (3) and analyst earnings forecasts. Standard errors are in parentheses and
p-values in the brackets. Wald(1) is the Wald test statistic of null hypothesis b=1. This statistic is
distributed as a chi-squared random variable with one degree of freedom. FMLS stands for the fully
modi…ed least squares (FMLS) estimator, while DLS for the dynamic least squares (DLS) estimator
(see Perroni (2001)). The order of the dynamic terms assumed in these estimators is set to one.
P anel- and panel-t are Pedroni’s (1999) cointegration test statistics based on the residuals of the
cointegrating regression. These statistics can test the null hypothesis of no cointegration against
its alternative of cointegration. N A means not applicable.

The results of the table clearly indicate that share prices Pit and their intrinsic values
Pit obtained through formula (3) constitute a pair of cointegrated series with a long-run
coe¢ cient b = 1, implying a cointegrating vector between Pit and Pit given by (1,-1). None of
the cointegration tests reported in the table can accept the null hypothesis of no cointegration
between Pit and Pit , while the null hypothesis b = 1 can not be rejected by the Wald test

32
statistic reported in the table, denoted as Wald(1). These results mean that price deviations
Pit Pit constitute stationary series, for all i. We have found that the success of the RIM
to provide share prices Pit which closely follow the long-run movements of their market
counterparts can be attributed to the fact that this model relies on book values Bit . In
particular, we have found that movements in Bit determine to a large extent those in market
prices Pit and their intrinsic values Pit . Share prices Pit are found to be also cointegrated
with their book values Bit , for all i. These results are not reported for reasons of space.

5.2 Economic sentiment variables

The de…nitions of the economic variables used to construct the economic sentiment index,
SIt , based on principal component analysis are as follows:

Market share turnover is expressed in terms of trading volume and trading values. Market
share turnover, or more generally liquidity, can be viewed as an investor sentiment index.
Higher turnover predict lower subsequent returns in both …rm-level and aggregate data (see,
e.g., Baker and Stein (2004), Scheinkman and Xiong (2003)). Market turnover by value is
the total value of trades over the month divided by the total capitalization of the London
Stock Exchange (LSE). Market turnover by volume is the total volume of shares traded on
LSE over the month divided by the number of shares listed on the stock exchange.
Numbers of IPOs within each month: First-day return of IPOs is expressed as the dif-
ference between initial trading price and o¤er price divided by o¤er price of the IPO stock.
The IPO market is often viewed to be sensitive to sentiment. More speci…cally, high …rst
day return on IPOs is considered as a measure of investor enthusiasm while the low return
of IPOs is often interpreted as a symptom of market timing (see Baker and Wurgler (2007)).
Consumer con…dence is a business survey data reported by the European Commission
and the European Financial A¤airs. UK respondents express their economic or …nancial
expectations over the next 12 months in the following areas: the general economic situation,
unemployment rate, personal household …nancial position and personal savings.
Closed-end fund discount is the di¤erence between the net asset value of a fund’s se-
curity holdings and the fund’s market price. Many authors based on the closed-end fund

33
discounts in order to measure individual investor sentiment considering that when the dis-
count increases the retail investors are bearish (see Lee et al (1991), and Neal and Wheatley
(1998)).
Put-call trading volume ratio is a measure of market participants’sentiment derived from
options. It equals to the ratio of trading volume of put options by the trading volume of
call options considering a bearish indicator in the stock market. More speci…cally, when the
trading volume of put options becomes large relative to the trading volume of call options,
the sentiment goes up, and vice versa.
Put-call interest ratio is the open interest of put options divided by the open interest of
call options. This ratio could be considered as a preferred measure of sentiment o¤ering a
better predictive power for volatility in subsequent periods, as it may be argued that the
open interest of options is the …nal picture of sentiment at the end of the day or the week.
The market volatility index measures the implied volatility of options and de…nes the
investor’s certainty or uncertainty regarding the volatility. More speci…cally, the higher the
market volatility index is the greater the fear of investors becomes.

34
Figure 1: Average values of market prices Pit (denoted as Pt - P_bar) against their
counterpart intrinsic values based on the RIM (denoted as Pt - P̄_bar).

35
Figure 2: Deviations between Pt and Pt .

36
Figure 3: Alternative sentiment indices against their threshold parameter values estimated.

37
 Table 1A: Basic descriptive statistics
mean Std. Dev. Min. Max.
Pit P it 1:49 3:34 21:23 23:49
GROW T H 2:46 2:48 5:51 5:56
IN F 3:57 1:97 0:52 9:46
T ERM 0:46 1:60 3:04 3:22
EXCH 0:34 5:59 13:81 13:72
DF 0:41 1:74 3:95 4:21
M ARKET 2:10 6:87 16:32 13:06
E=P 0:13 1:33 2:84 26:60
a
E =P 0:41 0:44 0:38 4:10
B=M 0:56 0:26 0:00 4:25
D=P 0:04 0:03 0:00 0:31
SIZE 9386847:51 17338415:24 62654:25 1:82E + 08
SI(W ACI) 102:02 10:13 79:5 124:6
SI(CCI) 100:05 1:14 98:24 101:96
SI(P CF ) 0 2:07 2:70 3:41
SEN T (W ACI) 0:97 9:89 19:49 24; 43
SEN T (CCI) 0:11 0:84 2:38 1:27
SEN T (P CF ) 1:46 4:76 16:9 2:03

Notes: The table presents the mean, standard deviation (Std. Dev.), and the minimum and
maximum (min and max) values of all the variables employed in the estimation of model (5).
Our data consists of T = 26 time series observations (from year 1987 to 2012) and N = 37
companies whose shares are continuously traded in the UK stock market, i.e., N T = 962 panel
data observations. For the measure of investor sentiment P CF , T = 13 due to unavailability of
data before year 2000. The variables GROW T H; IN F; T ERM; DF and M ARKET are in
percentage terms, or di¤erences of them (see T ERM and DF ).

38
 Table 1B: Correlation coe¢ cients
Pit P it GROW T H IN F T ERM EXCH DF M ARKET
Pit P it 1
GROW T H 0:0574 1
IN F 0:0907 0:5058 1
T ERM 0:0549 0:5283 0:1461 1
EXCH 0:1375 0:5120 0:32 0:010 1
DF 0:0405 0:6817 0:1456 0:7525 0:2073 1
M ARKET 0:1095 0:0376 0:1553 0:3165 0:3051 0:0133 1
E=P 0:1074 0:0578 0:0481 0:1028 0:0445 0:1127 0:0266
B=M 0:5140 0:0951 0:0144 0:0537 0:1116 0:1033 0:1407
D=P 0:2385 0:1393 0:1395 0:0525 0:1510 0:0624 0:0866
SIZE 0:1478 0:1510 0:0241 0:0732 0:3139 0:1117 0:3443
SEN T (W ACI) 0:0930 0:5145 0:3754 0:1601 0:7028 0:0922 0:3337
SEN T (CCI) 0:0719 0:1419 0:3144 0:2592 0::5646 0:2050 0:5908
SEN T (P CF ) 0:0525 0:2351 0:0348 0:2376 0:2526 0:1816 0:0050

39
Table 1B (continued): Correlation coe¢ cients
E=P B=M D=P SIZE SEN T (W ACI) SEN T (CCI) SEN T (P CF )
E=P 1
B=M 0:2421 1
D=P 0:0152 0:5124 1
SIZE 0:0787 0:1886 0:3244 1
SEN T (W ACI) 0:0225 0:1017 0:1291 0:3132 1
SEN T (CCI) 0:0019 0:1093 0:0512 0:3543 0:6627 1
SEN T (P CF ) 0:0359 0:0376 0:0152 0:0542 0:2846 0:1497 1

Notes: The table presents the correlation coe¢ cients among all the variables of model (5), including the dependant one. These are
based on T = 26 time series observations (from year 1987 to 2012) and N = 37 companies whose shares are continuously traded in the
UK stock market, i.e., N T = 962 panel data observations. For the measure of investor sentiment P CF , T = 13 due to unavailability
of data before year 2000
 (d )
Table 1C: Correlation coe¢ cients of the truncated …rm-speci…c variables dumjt L zjit
E=P B=M D=P SIZE
Bottom 10th (dL = 0:10) percentile
SI(W ACI) 0:1062 0:0021 0:0960 0:0738
SEN T (W ACI) 0:0625 0:0136 0:0541 0:0758
SI(CCI) 0:0429 0:0372 0:1050 0:0355
SEN T (CCI) 0:0372 0:0095 0:0558 0:1056
Pit P it 0:0786 0:2937 0:1793 0:0171
Upper 90th percentile
SI(W ACI) 0:0451 0:0579 0:1145 0:1069
SEN T (W ACI) 0:0059 0:0985 0:1058 0:1989
SI(CCI) 0:0802 0:1222 0:1063 0:1323
SEN T (CCI) 0:0119 0:1109 0:0571 0:2468
Pit P it 0:0924 0:4749 0:2550 0:164

(d )
zjit ) in their bottom

40
Notes: The table presents correlation coe¢ cients estimates of the truncated …rm-speci…c variables ( dumjt L
10th and upper 90th percentiles of their distribution with the sentiment indices SI (W ACI) and SI(CCI), and their associated sentiment
variables SEN T (W ACI) and SEN T (CCI), as well as the price deviations Pit Pit . These estimates are based on T = 26 time
series observations (from year 1987 to 2012) and N = 37 companies whose shares are continuously traded in the UK stock market, i.e.,
N T = 962 panel data observations. For the measure of investor sentiment P CF , T = 13 due to unavailability of data before year 2000.
 Table 2: LSDV Estimates of linear model (6)
(I) (II) (III)
GROW T H 0:146 ( 3:19) 0:122 ( 2:93) 0:128 ( 2:77) 0:104 ( 2:44) 0:384 ( 2:85) 0:332 ( 2:42)
IN F 0:068 (0:55) 0:112 1:05 0:059 (0:42) 0:097 (0:82) 0:450 (3:00) 0:426 (2:96)
T ERM 0:353 ( 4:39) 0:252 ( 3:55) 0:326 ( 3:73) 0:215 ( 2:57) 0:378 ( 1:80) 0:397 ( 1:90)
EXCH 0:052 (2:99) 0:041 2:88 0:057 (2:29) 0:048 (2:20) 0:120 (4:07) 0:115 (3:94)
DF 0:140 ( 2:14) 0:104 ( 1:53) 0:145 ( 1:96) 0:105 ( 1:38) 0:141 (0:74) 0:048 (0:25)
rM 0:019 (1:46) 0:027 2:22 0:020 (1:60) 0:028 (2:37) 0:039 (1:73) 0:046 (2:08)
E=P 0:001 ( 3:31) 0:001 ( 3:84) 0:001 ( 3:34) 0:001 ( 3:86) 0:001 ( 2:26) 0:001 ( 2:14)
(dL )
dum (E=P ) 0:012 (2:28) 0:0125 (2:28) 0:006 (1:08)
B=M 0:017 ( 2:77) 0:013 ( 2:17) 0:017 ( 2:76) 0:013 ( 2:15) 0:023 ( 3:32) 0:027 ( 3:02)
(dL )
dum (B=M ) 0:047 (2:08) 0:047 (2:07) 0:042 (3:39)
D=P 0:100 ( 1:52) 0:074 ( 1:09) 0:101 ( 1:56) 0:075 ( 1:12) 0:027 (0:31) 0:157 (1:33)
(dL )
dum (D=P ) 0:341 (0:60) 0:337 (0:60) 0:059 ( 0:09)

41
SIZE 0:803 (3:44) 0:617 (2:96) 0:820 (3:42) 0:642 (3:01) 0:326 (1:09) 0:245 (0:76)
dum(dL ) (SIZE) 0:930 ( 2:59) 0:947 ( 2:64) 0:250 ( 0:41)
SEN T (W ACI) 0:009 (2:18) 0:009 (1:96)
SEN T (CCI) 0:001 ( 0:01) 0:036 ( 0:25)
SEN T (P CF ) 0:083 (4:92) 0:078 (4:52)
R2 0:216 0:3174 0:181 0:3170 0:275 0:377
Observations 962 962 481
Sample 1987 2012 1987 2012 2000 2012

Notes: This table presents LSDV estimates of the linear model (6). These estimates are based on T = 26 time series observations (from
year 1987 to 2012) and N = 37 companies whose shares are continuously traded in the UK stock market, i.e., N T = 962 panel data
observations. For the measure of investor sentiment P CF , T = 13 due to unavailability of data before year 2000. Columns (I), (II) and
(III) present estimation results for the three alternative measures of the sentiment variable SEN Tt considered, i.e., SEN T (W ACI),
SEN T (CCI) and SEN T (P CF ), respectively. t-ratio statistics are reported in parentheses. R2 is the coe¢ cient of determination.
 Table 3A: Estimates of model (5) without regime shift in …rm-speci…c variables
(I) (II) (III)
GROW T H 0:130 ( 2:68) 0:115 ( 2:58) 0:358 ( 2:49)
IN F 0:097 (1:75) 0:138 (2:15) 0:423 (2:19)
T ERM 0:343 ( 4:24) 0:236 ( 2:67) 0:309 ( 1:87)
EXCH 0:061 (3:56) 0:052 (2:62) 0:093 (3:32)
DF 0:128 ( 3:36) 0:120 ( 2:92) 0:036 ( 0:16)
rM 0:029 (2:40) 0:024 (1:86) 0:051 92:57)
E=P 0:001 ( 2:87) 0:001 ( 2:90) 0:001 ( 1:83)
(dL )
dum (E=P ) 0:011 (1:96) 0:011 (1:98) 0:005 (0:63)
B=M 0:013 ( 4:31) 0:013 ( 4:19) 0:025 ( 4:35)
dum(dL ) (B=M ) 0:047 (3:51) 0:048 (3:55) 0:042 (2:50)
D=P 0:075 ( 1:82) 0:083 ( 2:00) 0:119 (1:44)
(dL )
dum (D=P ) 0:316 (1:25) 0:356 (1:41) 0:031 (0:08)
SIZE 0:621 (2:17) 0:595 (2:07) 0:366 (1:05)
(dL )
dum (SIZE) 0:899 ( 1:74) 0:896 ( 1:74) 0:228 ( 0:22)

Sentiment regime: SEN T (W ACI) SEN T (CCI) SEN T (P CF )

IfSI t < qg 0:021 (1:99) 0:160 (1:47) 0:161 (1:42)
IfSI t qg 0:087 ( 2:82) 0:675 ( 2:59) 0:207 ( 1:98)
Threshold parameter q 107:40 100:67 2:53
95% CI of q [107:07; 110:1] [99:69; 101:08] [1:85; 3:20]
LR test stat. 10:91 8:60 7:986
Observations 962 962 481
Sample 1987 2012 1987 2012 2000 2012

Notes: The table presents estimates of a version of model (5) without allowing for threshold
e¤ects in the …rm-speci…c variables. These are based on T = 26 time series observations (from
year 1987 to 2012) and N = 37 companies whose shares are continuously traded in the UK stock
market, i.e., N T = 962 panel data observations. For the measure of investor sentiment P CF ,
T = 13 due to unavailability of data before year 2000. Column (I) presents results for the case
where the sentiment index SI is based on W ACI , while Columns (II) and (III) for the cases that
it is based on CCI and P CF , respectively. LR is the likelihood ratio test statistic that there is no
threshold e¤ects in slope coe¢ cients and , i.e., H0 : = , against its alternative Ha : 6= .

42
 Table 3B: Estimates of the full speci…cation of model (5)
(I) (II)
GROW T H 0:1479 ( 2:87) 0:0897 [ 1:69]
IN F 0:0931 (1:80) 0:0768 [1:07]
T ERM 0:3354 ( 4:15) 0:2727 [ 2:92]
EXCH 0:0585 (3:38) 0:0704 [3:21]
DF 0:1214 ( 3:13) 0:1338 [ 3:19]
rM 0:0296 (2:38) 0:0257 [1:99]
IfSI t < qg IfSI t > qg IfSI t < qg IfSI t > qg
E=P 0:0017 ( 2:76) 0:012 ( 0:33) 0:0018 [ 2:89] 0:0018 [ 3:61]
dum(dL ) (E=P ) 0:0159 (1:92) 0:0952 (0:95) 0:0102 [1:66] 0:1510 [3:76]
B=M 0:0122 ( 3:84) 0:0196 ( 3:26) 0:0121 [ 3:69] 0:1319 [ 2:99]
(dL )
dum (B=M ) 0:0584 (3:31) 0:0145 (0:48) 0:0495 [2:43] 0:03737 [2:24]
D=P 0:1249 ( 2:76) 0:1027 (1:24) 0:1121 [ 2:44] 0:0207 [ 0:30]
dum(dL ) (D=P ) 0:3022 (1:18) 0:7306 (2:24) 0:3803 [1:47] 0:4048 [1:42]

43
SIZE 0:5887 (1:96) 0:9550 (1:15) 0:3070 [1:02] 1:0288 [2:50]
(dL )
dum (SIZE) 0:8716 ( 1:65) 1:6476 (0:87) 0:5065 [ 0:87] 1:5243 [ 1:42]
SEN T (W ACI) 0:0208 (1:95) 0:1296 ( 2:96)
SEN T (CCI) 0:0206 (1:94) 0:7180 ( 2:63)
Threshold parameter q 107:4 100:85
95% CI of q [106:32; 108:10] [100:52; 101:00]
LR test stat. 16:08 7:49
Observations 962 962
Sample 1987 2012 1987 2012
Notes: The table presents estimates of model (5), allowing also for threshold e¤ects in the …rm-speci…c variables. These are based on
T = 26 time series observations (from year 1987 to 2012) and N = 37 companies whose shares are continuously traded in the UK
stock market, i.e., N T = 962 panel data observations. Column (I) presents results for the case where the sentiment index SI is based
on W ACI , while Columns (II) for the cases that it is based on CCI . LR is the likelihood ratio test statistic of null hypothesis H0 :
(dL ) (d ) (d ) (d )
j = j; j = j L and = , for all j, against its alternative Ha : j 6= j or j L 6= j L or 6= .
 Table 4: Correlation coe¢ cients of the explanatory variables with Pit Pit across the two sentiment regimes
Pit Pit
IfSI(W ACI) < q = 107:40g IfSI(W ACI) > q = 107:40g IfSI(CCI) < q = 100:85g IfSI(CCI) > q = 100:85g
bottom 10% upper 90% bottom 10% upper 90% bottom 10% upper 90% bottom 10% upper 90%
E=P 0:0743 0:0907 0:0203 0:0162 0:0798 0:0974 0:0244 0:0248
B=M 0:2839 0:4211 0:0614 0:1859 0:2173 0:345 0:1777 0:2942
D=P 0:1527 0:2226 0:0843 0:1078 0:1062 0:1683 0:1368 0:1761
SIZE 0:0179 0:1394 0:0004 0:0783 0:0212 0:0925 0:0012 0:1299
SEN T (W ACI) 0:0550 0:0680
SEN T (CCI) 0:0527 0:0645

Notes: The table presents values of the correlation coe¢ cients of the …rm-speci…c variables in their 10th and 90th percentiles of their
distribution and the sentiment variable SEN T with price deviations Pit Pit across the two sentiment regimes of model (5), implied by
sentiment indices W ACI and CCI , respectively. These values are based on T = 26 time series observations (from year 1987 to 2012)
and N = 37 companies whose shares are continuously traded in the UK stock market, i.e., N T = 962 panel data observations.

44
 Table 5A: Estimates of model (5) without regime shift in …rm-speci…c variables, using quarterly data
(I) (II)
GROW T H 0:152 ( 0:89) 0:173 (1:00)
IN F 0:099 (1:53) 0:094 (1:50)
T ERM 0:419 ( 3:67) 0:414 ( 3:43)
EXCH 0:007 (0:44) 0:019 ( 1:07)
DF 0:205 ( 4:08) 0:168 ( 3:44)
rM 0:017 (1:58) 0:002 (0:17)
E=P 0:001 ( 2:75) 0:001 ( 2:80)
(dL )
dum (E=P ) 0:013 (2:07) 0:013 (2:07)
B=M 0:012 ( 3:47) 0:011 ( 3:38)
dum(dL ) (B=M ) 0:053 (3:74) 0:053 (3:76)
D=P 0:059 ( 1:34) 0:065 ( 1:50)
(dL )
dum (D=P ) 0:146 (0:53) 0:175 (0:63)
SIZE 0:620 (1:93) 0:550 (1:70)
(dL )
dum (SIZE) 1:097 ( 1:98) 1:045 ( 1:88)

Sentiment regime: SEN T (W ACI) SEN T (CCI)

IfSI t < qg 0:011 (1:02) 0:320 (1:47)
IfSI t qg 0:071 ( 2:15) 0:207 ( 0:89)
Threshold parameter q 107:3 100:45
95% CI of q [107; 107:5] [100:20; 100:46]
LR test stat. 5:43 3:90
Observations 3; 848 3; 848
Sample 1987Q1 2012Q4 1987 2012

Notes: The table presents estimates of a version of model (5) without allowing for threshold
e¤ects in the …rm-speci…c variables. These are based on quarterly time series observations (from
1987Q1 to 2012Q4) of the measures of sentiments W ACI and CCI , the macroeconomic variables
of the model, and the prices Pit . For the intrinsic values of shares Pit , we have maintained their
year values as quarterly observations after each year’s …nancial statement announcement, until the
next year’s one. Column (I) presents results for the case where the sentiment index SI is based on
W ACI , while Columns (II) and (III) for the cases that it is based on CCI and P CF , respectively.
LR is the likelihood ratio test statistic that there is no threshold e¤ects in slope coe¢ cients and
, i.e., H0 : = , against its alternative Ha : 6= .

45
 Table 5B: Estimates of the full speci…cation of model (5), using quarterly data
(I) (II)
GROW T H 0:166 (0:97) 0:167 (0:92)
IN F 0:101 (1:55) 0:096 (1:49)
T ERM 0:441 ( 3:77) 0:397 ( 3:23)
EXCH 0:000 ( 0:02) 0:004 ( 0:20)
DF 0:219 ( 4:28) 0:169 ( 3:35)
rM 0:014 (1:27) 0:002 (0:21)
IfSI t < qg IfSI t > qg IfSI t < qg IfSI t > qg
E=P 0:0017 ( 2:66) 0:011 ( 0:04) 0:002 ( 2:75) 0:021 ( 0:89)
(dL )
dum (E=P ) 0:012 (1:97) 0:008 (0:22) 0:005 (0:61) 0:046 (1:81)
B=M 0:010 ( 3:02) 0:016 ( 2:95) 0:009 ( 2:45) 0:012 ( 3:05)
(dL )
dum (B=M ) 0:056 (2:92) 0:049 (2:73) 0:109 (3:57) 0:039 (2:56)
D=P 0:094 ( 1:99) 0:034 (0:45) 0:095 ( 1:57) 0:049 ( 0:91)
dum(dL ) (D=P ) 0:020 (0:07) 0:435 (1:43) 0:124 (0:41) 0:189 (0:65)

46
SIZE 0:337 (0:93) 1:829 (2:70) 0:446 (0:97) 0:634 (1:44)
(dL )
dum (SIZE) 0:063 ( 1:07) 3:772 ( 2:62) 0:542 ( 0:70) 1:837 ( 2:14)
SEN T (W ACI) 0:010 (0:86) 0:099 ( 2:61)
SEN T (CCI) 0:337 (1:43) 0:050 ( 0:20)
Threshold parameter q 107 99:98
95% CI of q [103:05; 107:3] [99:32; 100:10]
LR test stat. 18:73 12:66
Observations 3; 848 3; 848
Sample 1987Q1 2012Q4 1987Q1 2012Q4
Notes: The table presents estimates of model (5), allowing also for threshold e¤ects in the …rm-speci…c variables. These are based
on quarterly time series observations (from 1987Q1 to 2012Q4) of the measures of sentiments W ACI and CCI , the macroeconomic
variables of the model, and the prices Pit . For the intrinsic values of shares Pit , we have maintained their year values as quarterly
observations after each year’s …nancial statement announcement, until the next year’s one. Column (I) presents results for the case where
the sentiment index SI is based on W ACI , while Columns (II) for the cases that it is based on CCI . LR is the likelihood ratio test
(d ) (d ) (d ) (d )
statistic of H0 : j = j ; j L = j L and = , for all j, against its alternative Ha : j 6= j or j L 6= j L or 6= .