ARTICLE IN PRESS

Journal of Banking & Finance xxx (2005) xxx–xxx

www.elsevier.com/locate/jbf

Valuation ratios and price deviations

from fundamentals
Jerry Coakley a, Ana-Maria Fuertes b,*

a
Department of Accounting, Finance and Management, University of Essex, Colchester CO4 3SQ, UK
b
Faculty of Finance, Cass Business School, 106 Bunhill Row, London EC1Y 8TZ, UK

Received 27 July 2004; accepted 11 August 2005

Abstract

This paper sheds light on US stock price deviations from fundamentals by analyzing the time-series
dynamics of post-1870 S&P valuation ratios. It employs a non-linear, two-regime framework that
allows for different behavior over phases of the stock market cycle. Persistence in the ratios implies
prolonged price deviations from fundamentals stemming from short run continuation fueled by inves-
tor sentiment during bull markets. However, the pull from fundamentals ensures that valuation ratios
and prices move toward their equilibrium levels in bear markets. Impulse response functions highlight
sluggish adjustment and indicate that the effects of positive shocks are more pronounced and long-
lasting in bull markets. The main conclusion is that, while market sentiment plays an important tran-
sitory role, valuation ratios do mean revert and so prices reflect fundamentals in the long run.

2005 Elsevier B.V. All rights reserved.

JEL classification: C40; G12

Keywords: Fundamentals; Behavioral finance; Investor sentiment; Threshold autoregression

1. Introduction

Do stock prices always reflect fundamentals or can they display short-run and, at times,
seemingly persistent deviations from their long-run equilibrium values? This question has
been at the heart of a debate in financial economics ever since ShillerÕs (1981) seminal

*
Corresponding author. Tel.: +44 207 0400186; fax: +44 207 6316416.
E-mail address: a.fuertes@city.ac.uk (A.-M. Fuertes).

0378-4266/$ - see front matter
2005 Elsevier B.V. All rights reserved.
doi:10.1016/j.jbankfin.2005.08.004
 ARTICLE IN PRESS
2 J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx

study. For instance, Summers (1986) suggests that irrational fads in investor sentiment cre-
ate sustained deviations of stock prices from intrinsic valuations and that rational inves-
tors might not be able to arbitrage away the mispricing because of noise trader risk. More
recently, Shiller (2000) argues that the 1990s hike in prices and valuation ratios was fueled
by investorsÕ irrational exuberance. Likewise, Anderson et al. (2003) argue that US stock
prices deviated from their fundamentals in the post-World War II period and suggest a
role for irrationality. Finally, Lee et al. (2002) argue that market sentiment is a priced sys-
tematic risk that is positively correlated with shifts in excess returns. Bullish changes in
sentiment lead to downward revisions in volatility and higher future excess returns and
vice versa for bearish changes.
The issue of whether stock prices reflect fundamentals has been given new urgency by
the sustained 1990s run up in prices. For example, US price–earnings (P/E) and price–div-
idend (P/D) series rose spectacularly during the course of that decade. The S&P (Standard
and PoorÕs) composite P/E ratio hit an all-time peak of 44.2 in December 1999 that was
more than double its long term historical mean level. Such behavior suggests that prices
can become decoupled from fundamentals for protracted periods and interpreting and the-
orizing such extreme movements poses a challenge for financial economics. Several ratio-
nally based explanations for the recent hike in valuation ratios have been adduced. These
include a decline in the equity premium, shifts in corporate payout policies and falls in the
cost of stock market participation and diversification.1 Other explanations are perhaps
more plausible in a post-Enron world and, among them, factors such as noise trading,
market sentiment and limits to arbitrage are prominent.
This paper makes several contributions to the literature. First, taking account of the
stock market cycle enables us to shed light on the deviations of prices from fundamentals.
We use economic theory to motivate asymmetric behavior driven by bull and bear market
phases and test for it empirically using a two-regime model.2 An interesting alternative
approach is to employ models that allow for structural breaks in the ratios or the equity
premium. This assumes mean reversion around a broken trend. For instance, Carlson
et al. (2002) find that valuation ratios mean revert around a broken trend that shifts
upwards in the early 1990s. Likewise, Manzan (2005) assumes a break in the equity pre-
mium during 1950 in analyzing the behavior of price–dividend ratios.
The present paper takes a different tack. Rather than assuming similar dynamics
around a broken trend equilibrium, we conjecture that valuation ratios exhibit distinct
dynamics around constant long-run equilibrium levels depending on the phase of the stock
market cycle. Our empirical results support contrasting bull and bear market behavior and
indicate no structural break. Crucially, when we reestimate our model excluding observa-
tions from the 1990s, our long-run equilibrium valuation ratio estimates remain
unchanged. In this manner our results can help to reconcile the counterintuitive finding
from linear specifications suggesting a lack of mean reversion or permanent stock price
deviations from fundamentals (Campbell and Shiller, 2001).3

1
See Campbell and Shiller (2001) for a critical evaluation of these and other explanations.
2
For other recent non-linear models in finance see Brooks and Garrett (2002), Lekkos and Milas (2004) and
McMillan (2004).
3
This is also why Ackert and Hunter (1999) and Madsen and Milas (2003) employ regime-switching
formulations of the price–dividend relationship based on managerial dividend behavior and levels of inflation,
respectively, although their focus and models differ from ours.
 ARTICLE IN PRESS
J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx 3

The second contribution is that the paper facilitates tests for the presence of non-fun-
damental factors without having to specify an asset pricing model.4 Campbell and Shiller
(2001) observe that prices rather than fundamentals (dividends or earnings) do most of the
adjustment in bringing the ratios back towards their long-run equilibrium levels. Thus,
they argue that valuation ratios can be used to predict stock price changes and conjecture
that the relationship may be non-linear.5 Moreover, since asset returns – logged price
changes – are related to valuation ratios, then the celebrated theoretical models of Barberis
et al. (1998), Daniel et al. (1998) and Hong and Stein (1999) may shed some light on the
behavior of the latter. These authors stress short-run departures from market efficiency in
stock returns stemming from investorsÕ cognitive biases when updating their forecasts as
Bayesian optimizers or to the interaction of noise and fundamental traders.6 We conjec-
ture that valuation ratios also follow the underreaction–overreaction pattern that stock
returns do in these models.
The third contribution is that we obtain some novel results in the context of the valu-
ation ratio literature. Our analysis reveals a significant role for non-fundamentals in
explaining valuation ratio dynamics over the course of the stock market cycle. The ratios
on average display continuation or underreaction in bull markets which points to the
importance of market sentiment and noise trading. In other words, prices deviate from
fundamentals, sometimes for protracted periods as in the 1990s. By contrast, there is sig-
nificant adjustment towards long-run equilibrium in bear markets as fundamentals reas-
sert themselves. In our framework, this implies that the ratios are mean-reverting
overall despite their locally persistent bull market behavior.
Impulse response functions imply that valuation ratios also exhibit the type of under-
reaction–overreaction time profile postulated in behavioral theories of stock returns and
that the effect of innovations decays slowly. Large positive shocks to the valuation ratios
in bull markets have more pronounced and long-lasting effects than similar shocks in bear
markets. This lends support to the view that, during bull market episodes, fundamentals
carry less weight or that deviations are more long lasting. Our key finding is that the
long-run behavior of stock prices is consistent with fundamentals whilst their short-run
evolution reflects unobserved behavioral factors such as market sentiment. This line of
argument is similar to that of Gallagher and Taylor (2001) in their study of risky arbitrage
in the US log dividend–price ratio 1926–1997 and to the one proposed by Manzan (2005)
in his study of the S&P price–dividend ratio using smooth exponential autoregressions.7
Finally our results on the time-series properties of valuation ratios are germane to an
extensive literature that examines stock return forecastability using regressions of stock
returns on lagged valuation ratios as regressors. This is because the null distribution of
typical test statistics such as the t-ratio changes dramatically if the variables are non-sta-
tionary. Standard inference approaches to ascertaining the forecasting power of valuation

4
In this respect, it is similar to the recent study by Jaggia and Thosar (2004) about the medium-term
aftermarket in high-tech US IPOs.
5
See also Campbell and Shiller (1988b), Fama and French (1988) and Pesaran and Timmerman (1995).
6
See Barberis and Thaler (2002) and Hirshleifer (2001) for recent surveys of behavioural finance and see Fama
(1998) for a critique.
7
Our findings also have elements in common with previous analyses based on non-linear adjustment models
such as those of Schaller and van Norden (2002), Psaradakis et al. (2004), Brooks and Katsaris (2005) and
Guidolin and Timmermann (2005).
 ARTICLE IN PRESS
4 J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx

ratios implicitly assume that they are mean reverting. However, evidence of the latter has
proven quite elusive in the empirical literature. In this sense, the present paper provides a
breakthrough in that not only does it document strikingly different behavior in valuation
ratios during bull and bear market phases but it also indicates that the ratios are indeed
mean-reverting or that prices ultimately reflect fundamentals.
The paper is organized as follows. In Section 2 we outline a theoretical framework for
examining the dynamic properties of P/D and P/E ratios and present our non-linear time-
series model. Section 3 discusses the hypotheses. Section 4 analyses the empirical results
and a final section concludes.

2. A non-linear model of valuation ratios

2.1. Theoretical framework

The dynamic dividend discount model of Campbell and Shiller (1988a,b) not only links
prices to fundamentals but also suggests a relationship between valuation ratios and
expected stock returns. One crucial aspect of their model is that, by imposing a transver-
sality condition, non-rational behavior is ruled out. We allow for it on the basis of the con-
jecture that valuation ratios exhibit behavioral properties similar to those of stock returns.
Non-rational behavior is permitted by means of a non-fundamentals or fads term that
captures the impact of market sentiment or the temporary deviation of prices from
fundamentals.
Consider the definition of log or continuously compounded one-period returns
rtþ1 ¼ lnð1 þ Rtþ1 Þ
lnðP tþ1 þ Dtþ1 Þ  ln P t ; ð1Þ
which can be rewritten as
rtþ1 ¼ ptþ1  pt þ ln½1 þ eðd tþ1 ptþ1 Þ ; ð2Þ
where lower case letters denote logged variables. Using a first-order Taylor approximation
of the latter term around d  p, the average dividend–price ratio, we have
rtþ1  k þ qptþ1 þ ð1  qÞd tþ1  pt ; ð3Þ


where k ¼  ln q  ð1  qÞ ln q1  1 and q ¼ 1=½1 þ eðdpÞ .
The estimated weight of pt+1 in (3) for US data is close to 1 and, hence, that of dt+1
almost 0. The accuracy of the approximation depends on the dividend–price variation
not being too large and this seems to work well for monthly US stock returns. Solving
(3) forward and taking expectations it follows that
( )
X1
pt ¼ k=ð1  qÞ þ Et q ½ð1  qÞd tþ1þj  rtþ1þj  þ Et ðqj ptþj Þ;
j
ð4Þ
j¼0

which can be reformulated in terms of the price–dividend ratio as

" #
X
1
pt  d t ¼ k=ð1  qÞ þ Et q ðDd tþ1þj  rtþ1þj Þ þ Et ½qj ðptþj  d tþj Þ
j
ð5Þ
j¼0

and, in either form, the last term drops out under the transversality condition
limj!1Et(qjpt+j) = 0 and limj!1Et[qj(pt+j  dt+j)] = 0. Eq. (4) or (5), which also holds
 ARTICLE IN PRESS
J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx 5

ex post, is called the dynamic Gordon growth model or dividend-ratio model by Campbell
and Shiller (1988a,b).8
Now replace the accounting identity (5) by
" #
X1
j
pt  d t ¼ k=ð1  qÞ þ Et q ðDd tþ1þj  rtþ1þj Þ þ ut ; ð6Þ
j¼0
j
where Et(q pt+j  dt+j) is subsumed in a mispricing term ut. This framework is rather gen-
eral since it lends itself to a number of interpretations. First, assuming that agents are al-
ways fully rational, the mispricing effect ut will be both stationary and insignificant.9 In
this context, the effect of shocks will be instantaneously eroded and so valuation ratios
should show rapid and hence, unambiguous evidence of mean-reversion. Second, most
empirical studies document highly persistent ratios that are observationally equivalent
to unit root processes. This implies that the mispricing effect ut is non-stationary which,
in turn, leads to the counterintuitive interpretation that prices and dividends can arbi-
trarily drift apart.
In between lies a third interpretation that views ut as a fads or non-fundamentals term
with non-linear or regime-sensitive properties as in Lee (1998). Thus, its behavior can vary
over the cycle appearing very persistent in bull markets but rapidly mean reverting
(towards zero) in bear markets.10 This more plausible interpretation permits a role for
investor sentiment in inducing deviations of prices from fundamentals in the short run
or even for sustained periods like the 1990s bull market. The important point, however,
is that the deviations are not permanent over the full cycle as in the unit root interpreta-
tion. Thus valuation ratios, while appearing persistent, do ultimately adjust towards equi-
librium and so fundamentals matter in the long run.

2.2. Two-regime dynamic model

A model of valuation ratios should be able to reconcile both the persistence found in
empirical studies and the mean reversion implied by the dynamic dividend discount model.
Since standard linear time-series models of valuation ratios conflate bull and bear market
phases, the typical inference from them can be regarded as non-informative in two senses.
On one hand, it reflects the ÔaverageÕ dynamics from distinct market phases. On the other,
it leads to the dichotomy that either valuation ratios never mean-revert or that they con-
tinuously adjust towards their long-run equilibrium levels. The non-linear framework out-
lined below is, by contrast, quite flexible. It captures the notion that the short-run
dynamics of both P/D and P/E ratios may occasionally be dominated by mispricing effects
stemming, for instance, from fads or waves of optimism. The intuition is that the ratios
would tend persistently to rise alongside prices during bull markets such as the late
1990s episode. Formally, this is referred to as local non-stationarity. But sooner or later
there is a correction with the onset of bear markets as the influence of fundamentals is

8
See also Campbell et al. (1997, Chapter 7).
9
Under rational expectations ut = 0 as a particular case.
10
As both remaining terms on the right hand side of Eq. (6) are stationary under plausible assumptions, this
implies that valuation ratios are also globally stationary. We assume that returns are stationary overall as
theoretically argued by Campbell et al. (1997).
 ARTICLE IN PRESS
6 J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx

restored and hence, the ratios fall rapidly alongside prices. This correction ensures overall
or global stationarity in valuation ratios.
Let xt denote a valuation ratio and assume that its current change is determined by the
magnitude of the previous periodÕs deviation from equilibrium, the speed of adjustment
and an iid innovation. This can be represented in the augmented Dickey–Fuller (ADF)
regression form as
X
k
Dxt ¼ qðxt1  lÞ þ bj Dxtj þ et ; et  iidð0; r2 Þ; ð7Þ
j¼1

where l is the long-run equilibrium or fundamental attractor, q is the speed of adjustment

and et is an innovation or unexpected shock.11 The ADF statistic is a (non-standard) t-ra-
tio on the significance of q. If xt is mean-reverting, then the unconditional expectation
E(xt) plays the role of long-run equilibrium and so
x provides a good proxy for the param-
eter l. When q is zero, then there is no mean-reversion or equivalently, xt is unit root
persistent.12
We generalize the linear model (7) to a parsimonious two-regime model that belongs to
the momentum threshold autoregressive (TAR) class. This is a simple reformulation of an
AR model which allows for bull and bear markets. Formally, let xt evolve as a TAR pro-
cess so that it follows one of two possible regimes. The threshold or transition variable that
determines the regime-switching in the TAR framework of Enders and Granger (1998) and
Hansen (1997) is a simple first or long difference qt(d)
Dxtd. In the present context, the
latter can be a rather noisy measure and so instead we characterize bull and bear market
markets through a moving average of past changes
qt ðw; dÞ
w1 Dxt1 þ    þ wd Dxtd ; ð8Þ
0
where w = (w1, . . . , wd) > 0 are weights summing to 1 and d P 1 is a finite integer lag.
Our reformulation of (7) permits xt to adjust toward its long-run equilibrium differently
depending on the direction of the market – either non-decreasing or decreasing – as
follows:
X
k
Dxt ¼ a þ I t qc ðxt1  lÞ þ ð1  I t Þqr ðxt1  lÞ þ bj Dxtj þ et ; ð9Þ
j¼1

where the speed of adjustment is measured by qc in bull markets and qr in bear markets.
The notation c and r is employed to reflect our conjecture that bull and bear market phases
are characterized by continuation and reversal, respectively. A non-zero parameter a im-
plies that, as is typical in processes with asymmetric adjustment, the long-run equilibrium
level l is not necessarily well proxied by the historical mean of xt.
The switching between bull and bear market phases is modeled through the indicator
function

11
A more general model would allow for a time-varying attractor. Such model can be estimated via the Kalman
filter where lt is treated as a latent variable. This approach goes beyond the scope of the present paper.
12
A less restrictive model is one where et is a martingale difference. This permits heteroskedasticity (non-
constant volatility) which does not invalidate the asymptotic distributions obtained by Dickey–Fuller. However,
ARCH effects may result in unit root tests that reject too often, i.e. are oversized, for finite samples (Kim and
Schmidt, 1993).
 ARTICLE IN PRESS
J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx 7

1 if qt ðw; dÞ P 0;
It ¼
0 if qt ðw; dÞ < 0;
where the transition variable qt(w, d), defined as in (8), obeys one of two schemes. One has
exponentially decreasing (ExpD) weights w1 >    > wd so that the more distant past
changes are given less weight – reflecting gradual memory loss – which is consistent with
feedback models (Campbell and Shiller, 1990). The other scheme has exponentially
increasing (ExpI) weights w1 <    < wd and may be viewed as capturing better the persis-
tence in valuation ratios.13 We let the data reveal the most adequate switching scheme.14

3. Hypothesis testing

Most empirical studies of valuation ratios address the issue of persistence versus mean-
reversion by means of standard unit root tests that do not allow for regime-sensitive
behavior. This restrictive framework may bias the inference and hence, the test outcome
should be interpreted in conjunction with evidence on whether the ratios exhibit non-linear
time dependencies.
The merit of the non-linear dynamic model proposed, Eq. (9), is that it provides a par-
simonious framework for testing classical as well as behavioral hypotheses. On one hand,
it permits a test of the classical-theory prediction that stock markets adjust rapidly to news
and shocks. On the other, we can also test behavioral hypotheses that are related to dis-
tinct market phases. In bull markets, trend-chasing investors or noise traders may drive up
valuation ratios irrespective of fundamental news. We can thus test the underreaction
hypothesis in such a phase. In addition, we can test for adjustment or correction towards
long-run equilibrium in bear markets.
Our novel conjectures relating to the time-series behavior of valuation ratios can be
condensed into two overarching predictions. Different aspects of these predictions are
investigated through tests that are described below in terms of their null (H0) and alterna-
tive hypotheses (HA).

Prediction 1. Valuation ratios behave asymmetrically during bull and bear market phases.
It is important to revisit the persistence versus mean-reversion debate in a more general
framework that allows for both symmetric and asymmetric (regime-sensitive) adjustment.
Hence, we propose:
Test 1A. Unit root persistence or mean reversion?

H0: Valuation ratios have a unit root.

H1: Valuation ratios mean revert (non-)linearly.

We test for this by means of an F-statistic (F1A hereafter) for zero-adjustment in both
bull and bear markets under the null using model (9). The test is non-standard in that the
null distribution of F1A is unknown. Finite sample p-values can be calculated by a boot-

P ðj1Þs
13
The ExpI weights
P are wj = e
(j1)s
/jI, j = 1, 2, . . . , d where s = (d  1)1 and jP
I ¼ je is a normalizing
factor such that j wj ¼ 1. The ExpD weights are wj = e1(j1)s/jD with jD ¼ j e1ðj1Þs . These schemes are
successfully employed in Coakley et al. (2001) to capture fast-up/slow-down unemployment phases.
14
Identification and estimation issues are discussed in Appendix A.
 ARTICLE IN PRESS
8 J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx

strap simulation approach (Berkowitz and Kilian, 2000). This type of threshold unit root
test was introduced by Enders and Granger (1998) and we follow their approach but
resample from the residuals rather than randomly drawing innovations from a normal dis-
tribution. We deploy two bootstrap algorithms. One assumes iid(0, r2) errors while the
other controls for heteroskedasticity ðr2t Þ to accommodate the stylized fact that bear mar-
kets are typically more volatile than bull markets.15
A rejection in the above test would suggest mean reversion but the issue remains of
whether the adjustment mechanism is uniform throughout or regime-sensitive instead.
This motivates:
Test 1B. What type of mean-reversion?

H0: Valuation ratios mean-revert symmetrically.

H1: Valuation ratios mean-revert asymmetrically.

Here the alternative hypothesis seeks to reflect the potentially distinct behavior of the
ratios during bull and bear markets. Test 1B is conducted through the non-linear model
(9) by means of an F-statistic (called F1B) for the null restriction that adjustment is iden-
tical in both phases. This is a standard test and so we can rely on the asymptotic chi-
squared (v2) distribution.16 However, we also calculate bootstrap p-values to control for
heteroskedasticity.
A rejection in the above test would naturally prompt the question of which is the precise
type of dynamics in each regime. To address it, we formulate:
Test 1C. How do the ratios behave during bull and bear markets?

H0: Valuation ratios do not mean revert in bull (bear) markets.

H1: Valuation ratios do mean revert in bull (bear) markets.

Classical, market-efficient theories imply that prices rapidly reflect changes to funda-
mentals and accordingly, we should find unambiguous evidence of mean-reversion in val-
uation ratios both in bull and bear market phases (qc < 0, qr < 0). In contrast, behavioral
theories predict mean reversion in bear markets only (qc = 0, qr < 0). The economic intu-
ition is that market sentiment or fads together with limits to arbitrage will induce short-
run continuation in bull markets and so prices can regularly deviate from fundamentals.
In our two-regime framework, Eq. (9), a valuation ratio exhibits underreaction in bull
markets if it randomly drifts upward (a > 0) with no pull from its long-run equilibrium
(qc = 0). This is difficult to reconcile with classical theories.

Prediction 2. Shocks to valuation ratios have long-lasting or seemingly permanent effects.

15
The residuals suggest a GARCH(1, 1) structure. The null estimates of model (9) augmented with a
GARCH(1, 1) equation are used, together with random draws from the standardized residuals, to generate 1000
bootstrap samples fxt : t ¼ 1; . . . ; T g. This approach has been shown to yield correctly sized tests (Coakley and
Fuertes, in press).
16
The asymptotic theory for continuous TARs establishes that, under the assumptions ofpstationarity
ffiffiffiffi for xt and
iid innovations, conditional OLS yields estimators ð^a; l ^ which are consistent at rate T and asymptotically
^; bÞ
^; q
normal (Chan and Tsay, 1998). Further details on TAR estimation can be found in Appendix A.
 ARTICLE IN PRESS
J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx 9

In contrast with the efficient markets tenet, behavioral theories predict partial adjust-
ment to news. The latter typically generates momentum which is followed by correction
resulting in the distinctive underreaction–overreaction time profile in stock returns.
Extending this to valuation ratios, we formulate:
Test 2A. Is there underreaction–overreaction to news in valuation ratios?

H0: Valuation ratios respond fully and immediately to shocks.

H1: Valuation ratios follow an underreaction–overreaction time profile.

To our knowledge, this question has not been addressed in the context of valuation
ratios.
It is also interesting to confront the classical hypothesis of rapid adjustment to news
with a behavioral hypothesis in which limits to arbitrage or market sentiment serve to
delay adjustment in the wake of shocks. This leads to:
Test 2B. Is the adjustment to news rapid or sluggish?

H0: The valuation ratios adjust rapidly following innovations.

H1: The valuation ratios adjust slowly following innovations.

Finally, we seek to compare the effects of unexpected shocks to valuation ratios during
bull and bear markets. For this purpose, we formulate:
Test 2C. How pronounced is the impact of shocks during bull and bear markets?

H0: Large positive shocks to valuation ratios have similar effects in both bull and bear
markets.
H1: Large positive shocks to valuation ratios have more marked effects during bull
markets.

The above null hypothesis reflects the classical tenet of linearity. By contrast, the alter-
native reflects the behavioral notion of asymmetric dynamics stemming from the inexora-
ble succession of bull and bear markets. Given that market sentiment fuels momentum
during bull markets, it seems natural to predict that positive shocks lead to more pro-
nounced effects during such phases.

4. Empirical results

4.1. Data and preliminary analysis

The empirical analysis is based on monthly data for the S&P Composite stock price
index, dividends and earnings. We follow the literature and use smoothed earnings to
proxy the long-run path of earnings. In what follows P/D denotes the log of the price–div-
idend ratio which spans the period 1871:01–2001:09 giving T = 1569 observations. The P/
E series 1881:01–2002:1 (T = 1453) is defined similarly using a 10-year moving average of
earnings.17

17
We use the stock market data compiled by Robert Shiller available at http://www.econ.yale.edu/
~shiller.
 ARTICLE IN PRESS
10 J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx

Table 1
Summary statistics for S&P composite index valuation ratios
Series Mean Median St. dev Min. Max. SK KU JB test
P/D 3.127 3.093 0.367 1.978 4.503 0.953 5.164 543.4 [0.000]
P/E 2.688 2.728 0.388 1.565 3.789 0.087 3.438 13.429 [0.000]
SK and KU are the skewness and kurtosis coefficients, respectively. JB is the Jarque–Bera test statistic for the
normality null with asymptotic p-values in brackets. P/D is the logged price–dividend ratio from 1871:01 to
2001:09. P/E is the logged price-earnings ratio from 1881:01 to 2002:01.

Table 1 reports descriptive statistics for the two series and Fig. 1 plots them. The P/D
series, while slightly less volatile than the P/E series, shows some skewness and excess kur-
tosis. The average P/D and P/E (before taking logs) is 21.6 and 15.84, respectively. Fig. 1
shows that both series exhibit a clear business cycle or bull–bear market pattern. One
prominent instance is the sharp rise in the 1990s relative to the historical mean and the
subsequent correction.
Table 2 reports the ADF test results for the full sample and the pre-/post-WWII subs-
amples. The insignificant test statistics indicate that valuation ratios do not mean revert
and so the unit root null cannot be rejected in line with extant evidence.18 However, such
a conclusion is difficult to reconcile with the fact that the ratios cross their sample averages
a large number of times during the period under study – albeit at intervals which vary from
a few months to several years – as illustrated in Fig. 1. Lack of mean-reversion would sug-
gest also that prices and dividends (or earnings) randomly drift apart in the long run or,
equivalently, that they do not cointegrate. Campbell and Shiller (2001) attribute such a
counterintuitive finding to the poor properties (low power) of the ADF and related linear
tests in the context of slowly mean-reverting processes. The interesting question then is
why adjustment is so slow.
If valuation ratios behave differently during bull and bear markets, then the unit root
outcome from standard linear tests may be an artefact of ÔaveragingÕ the speed of reversion
estimates from these two phases. Hence, it seems sensible to question the adequacy of the
linear framework. For this purpose, we compute the BDS statistic of Brock et al. (1996)
and the non-parametric triples statistic of Randles et al. (1980). For large samples, both
statistics are asymptotically standard normal under the linearity null. The BDS statistic,
based on the correlation integral, is designed to reveal hidden patterns that should not
occur in linearly dependent data. The triples test examines skewness in all possible triples
of the first-differenced data, fðDxt1 ; Dxt2 ; Dxt3 Þg, to investigate the conjecture that there is
growth asymmetry or steepness in the underlying distribution of xt.19
Table 3 sets out the results for the above time-series linearity tests. The BDS statistic
clearly suggests non-linear dependence in both valuation ratios. Furthermore, the triples
statistic supports the conjecture of steepness in the ratios and its negative sign is indicative
of slow-up, fast-down dynamics. The latter may reflect the fact that valuation ratios rise

18
See also the insignificant ADF test results in Balke and Wohar (2001) and Lamont (1998) inter alia.
19
For the BDS statistic, we employ the usual specifications of embedding dimension, m 2 {2, 3, 4, 5}, and
distance, e/r 2 {0.5, 1.0, 1.5}, where r is the standard deviation of the AR-filtered data (for details, see Campbell
et al., 1997, Chapter 12). A bootstrap version of the triples statistic to correct for AR(1) dependence in Dxt is
suggested by Verbrugge (1997). In the present case, the bootstrap correction made no qualitative difference to the
results.
 ARTICLE IN PRESS
J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx 11

Panel A. S&P Composite P/D ratio 1871:01 – 2001:09

4.35

3.85

3.35

2.85

2.35

1.85
1871.01
1875.06
1879.11
1884.04
1888.09
1893.02
1897.07
1901.12
1906.05
1910.10
1915.03
1919.08
1924.01
1928.06
1932.11
1937.04
1941.09
1946.02
1950.07
1954.12
1959.05
1963.10
1968.03
1972.08
1977.01
1981.06
1985.11
1990.04
1994.09
1999.02
Panel B. S&P Composite P/E ratio 1881:01-2002:01
4

3.5

2.5

1.5
1881.01
1885.01
1889.01
1893.01
1897.01
1901.01
1905.01
1909.01
1913.01
1917.01
1921.01
1925.01
1929.01
1933.01
1937.01
1941.01
1945.01
1949.01
1953.01
1957.01
1961.01
1965.01
1969.01
1973.01
1977.01
1981.01
1985.01
1989.01
1993.01
1997.01
2001.01

Fig. 1. Valuation ratios and historical means. The figure shows the evolution of the logged monthly price–
dividend (P/D) and price–earnings (P/E) valuation ratios.

gradually but inexorably during bull markets whereas they decline sharply during bear
markets. This vindicates the use of a non-linear TAR framework in analyzing the behavior
of valuation ratios.

4.2. Prediction 1: Asymmetries in bull and bear markets

The non-linear TAR model (9) that allows for distinct adjustment in bull and bear mar-
ket phases is fitted to the valuation ratios (see Appendix A for details on the estimation).
Table 4 summarizes the results.
 ARTICLE IN PRESS
12 J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx

Table 2
ADF test results
Series Sample period T k q ADF test LB(12)
P/D 1871:01–2001:09 1569 15 0.0062 2.047 [0.267] 0.476
1871:01–1944:12 888 15 0.0290 3.931 [0.002] 0.298
1945:01–2001:09 681 11 0.0042 1.197 [0.677] 2.729
P/E 1881:01–2002:01 1453 15 0.0068 2.344 [0.158] 0.216
1881:01–1944:12 768 13 0.0131 2.714 [0.072] 0.303
1945:01–2002:01 685 11 0.0047 1.407 [0.579] 1.877
Results based on model (7) estimated by OLS. The augmentation lag k is selected by testing down from
kmax = 18. The null hypothesis in the ADF test is unit root persistence (q = 0) against mean reversion (q < 0);
one-sided finite sample p-values in brackets. LB(12) is the Ljung–Box statistic for autocorrelation in the residuals.

Table 3
Tests for linear dependence in valuation ratios
Series BDS test Triples test
e/r m
2 3 4 5
P/D 0.50 4.747 5.734 6.860 7.608 4.338
[2.1 · 106] [9.8 · 109] [6.9 · 1012] [2.8 · 1014] [1.4 · 105]
1.00 5.571 6.472 7.457 8.000
[2.5 · 108] [9.6 · 1011] [8.8 · 1014] [1.2 · 1015]
1.50 6.720 7.824 8.640 9.054
[1.8 · 1011] [5.1 · 1015] [5.6 · 1018] [1.4 · 1019]
P/E 0.50 3.301 4.561 5.789 7.347 5.660
[9.6 · 104] [5.1 · 106] [7.1 · 109] [2.0 · 1013] [1.5 · 108]
1.00 3.781 4.934 6.088 6.929
[1.6 · 104] [8.0 · 107] [1.1 · 109] [4.2 · 1012]
1.50 4.482 5.900 6.941 7.657
[7.4 · 106] [4.2 · 109] [3.9 · 1012] [1.9 · 1014]
The BDS test is applied to the residuals of an AR(p) model with p = 16. The triples test is applied to the ratios in
first-differences (Dxt). Asymptotic p-values in brackets. The null hypothesis is that the time-series dependence or
dynamics in valuation ratios is linear.

The best-fit transition function, Eq. (8), is based on ExpI weighting with d^ ¼ 13 and 3
lags for the P/D and P/E series, respectively. The BDS test applied to the TAR residuals
yields some borderline rejections in the P/D case but the order of magnitude of the test
statistics is, nevertheless, notably smaller than in the case of the linear AR residuals (cf.
Table 3). For the P/E series there is no evidence of neglected non-linearities. Overall this
evidence suggests that the TAR model captures quite well the non-linear dependence in the
valuation ratios.
Using the TAR estimates, Fig. 2 classifies the P/D and P/E data into regimes based on
whether the threshold variable is non-decreasing as in bull markets or decreasing as in bear
markets. Interestingly, the proportion of observations identified as pertaining to bull
(54%) and bear (46%) phases is similar for both the P/D and P/E ratios. The slightly higher
proportion in bull markets is consistent with the secular upward trend in stock prices.
Although there are alternative definitions and measures of bull/bear market periods,
Table 4
Non-linear model estimates and inference results

J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx
Series k l w d a Continuation Reversal F2A F2B r2 ARCH(1) BDS0.5
c r
q %T q %T ARCH(3) BDS1.0
Panel A: full sample
P/D 15 2.69 ExpI 13 0.00338 0.0024 54 0.0142 46 5.485 6.766 1.57 · 103 0.243 1.864
(1.65 · 103) (3.34 · 103) (4.32 · 103) (0.048) (0.009) (0.622) (0.062)
[1.93 · 103] [3.08 · 103] [6.49 · 103] [0.052] [0.038] 13.246 1.961

ARTICLE IN PRESS
(0.004) (0.050)
P/E 15 2.19 ExpI 3 0.00362 0.0021 54 0.0126 46 6.260 7.003 1.59 · 103 0.143 0.719
(1.78 · 103) (3.39 · 103) (3.64 · 103) (0.025) (0.008) (0.705) (0.472)
[1.84 · 103] [3.43 · 103] [4.41 · 103] [0.028] [0.044] 3.981 0.483
(0.264) (0.629)

Panel B: pre-1993 sample

P/D 15 2.68 ExpI 3 0.00588 0.0102 53 0.0200 47 8.100 3.566 1.60 · 103 0.104 2.290
(1.94 · 103) (4.89 · 103) (5.01 · 103) (0.001) (0.059) (0.747) (0.022)
[2.41 · 103] [5.25 · 103] [6.87 · 103] 0.195 1.736
(0.900) (0.083)
P/E 15 2.18 ExpI 3 0.00471 0.0050 52 0.0166 48 7.326 6.058 1.64 · 103 0.056 1.630
(1.94 · 103) (4.08 · 103) (4.35 · 103) (0.009) (0.014) (0.813) (0.103)
[2.03 · 103] [4.35 · 103] [5.33 · 103] 3.710 0.193
(0.295) (0.847)
The table shows the conditional OLS estimates and diagnostics of TAR model (9). White heteroskedasticity-robust (brackets) and OLS (parentheses) standard errors
for qc and qr. %T is the percentage of observations in each regime. The F2A and F2B statistics are designed to test for the unit root (qc = qr = 0) and linearity (qc = qr)
hypotheses, respectively. In parentheses iid bootstrap p-values and in brackets ARCH bootstrap p-values for F2A. Asymptotic (parentheses) and ARCH bootstrap p-
values (brackets) for F2B. r2 is the modelÕs squared standard error. ARCH(r) is EngleÕs LM test for no heteroskedasticity up to order r. BDS statistics for m = 2,
r ¼ f0:5; 1:0g with asymptotic p-values in parentheses.
e=^

13
 ARTICLE IN PRESS
14 J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx

Fig. 2. Bull and bear markets for valuation ratios. The figure shows the sample P/D and P/E ratios classified into
bull and bear markets according to a non-linear TAR model.

Fig. 2 shows that the present model can capture these phases reasonably well.20 For
instance, both the 1929 Wall Street (bear) market crash and the sustained bull market
of the 1990s are appropriately identified in both the P/D and P/E histories.
The TAR model facilitates a test for the hypothesis of unit root persistence, qc = qr = 0,
against a flexible alternative that permits linear/non-linear mean reversion (Test 1A). The
test statistics reported in Table 4 are significant at the 5% level on the basis of either the iid
bootstrap or ARCH bootstrap p-values from 1000 replications. This suggests that shocks

20
See Pagan and Sossounov (2003) for a recent discussion of the issues in identifying bull and bear markets.
 ARTICLE IN PRESS
J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx 15

to the valuation ratios are ultimately transitory and so the unit null hypothesis is rejected.
This finding seems more plausible than the lack of mean reversion suggested by the earlier
ADF tests. Relatedly, a simple Monte Carlo simulation suggests that TAR-based unit root
tests have better power properties than ADF tests to detect overall mean reversion in
regime-sensitive processes.21
The rejection of the unit root null prompts the question of the type of mean reversion in
valuation ratios. Accordingly, we test the null hypothesis of linear adjustment against the
alternative of asymmetric or regime-sensitive adjustment (Test 1B). The test statistics in
Table 4 reject linear mean-reversion at better than the 5% level. Hence, there is evidence
to suggest that the behavior of valuation ratios cannot be adequately analyzed using linear
models as is typical in the literature.
The TAR estimation results reported in Table 4 can shed light on the nature of the
adjustment dynamics in bull and bear market phases. The speed of adjustment coefficients
suggest that the pull from the attractor in the bull market phase, qc, is much weaker than
that during the bear market phase, qr. Indeed, the t-statistics indicate that the former is
insignificant in bull markets, or that qc = 0, but that the drift term a is significantly posi-
tive. Thus the null hypothesis (zero-adjustment) in Test 1C cannot be rejected and there is
evidence of continuation in bull markets. This outcome is difficult to reconcile with risk-
based theories. However, it is consistent with behavioral theories that recognize the influ-
ence of investor sentiment or the interaction between noise traders and rational investors
in generating momentum in stock returns following unexpected good news. It is also in
line with behavioral notions such as investorsÕ trend chasing (Hong and Stein, 1999),
biased self-attribution (Daniel et al., 1998) or conservatism (Barberis et al., 1998).
By contrast, the t-statistic for the adjustment coefficient in bear markets (qr) is signifi-
cantly negative for both the P/D and P/E ratios. Therefore the no-adjustment null in Test
1C is rejected in favour of reversal or correction during bear markets. This provides the
intuition behind the overall stationarity of the ratios despite the presence of continuation
in bull markets, namely, sooner or later the ratios switch to a bear market phase as rational
investors come to dominate the market. Thus prices reflect fundamentals in the long run.
However, it may be argued that the above inferences are biased due to the protracted
bull market episode that characterized the ratios from the mid-1990s. As a robustness
check, the observations from 1993:01 onwards are excluded and the analysis is repeated.22
The threshold effect (qr  qc) remains statistically significant and the percentage of obser-
vations classified as bull and bear market phases remain unchanged. An interesting result
is that the estimate for the attractor l is essentially identical to that obtained over the
whole sample, indirectly implying no structural break in the long-run equilibrium.23 This
vindicates our approach of interpreting valuation ratio dynamics around a constant long-
run equilibrium as price deviations from fundamentals.

21
Pseudo-data for the valuation ratios is constructed from the TAR estimates in Table 4 and iidN ð0; r ^2 Þ
innovations where r ^2 is the residual variance. We generate 1000 time series of length T = 1500 + T0 and discard
the first T0 = 500 observations to minimize the (zero) initialization effects. The 5% nominal significance level is
adopted. The rejection frequency or empirical power thus obtained is 89% (F1A) and 67% (ADF) for the P/D
model and 90% (F1A) and 57% (ADF) for the P/E model.
22
We also discarded the whole post-1990 sample and the results were qualitatively similar.
23
This issue is formally corroborated by means of the Zivot and Andrews (1992) breakpoint test. The null is a
unit root and the alternative is mean reversion around a broken trend at an endogenous date. The test statistics at
3.94 (P/D) and 3.37 (P/E) do not support the hypothesis of a shift in fundamentals during the sample period.
 ARTICLE IN PRESS
16 J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx

4.3. Prediction 2: Shocks have long-lasting effects

Having documented that valuation ratios are regime-sensitive but overall mean revert-
ing or equivalently, that stock prices are driven by fundamentals in the long run, we now
investigate the effects of a one-off shock on the future evolution of the ratios. Generalized
impulse response (GIR) measures are obtained for this purpose. The GIR represents the
average time profile or response of a variable to random shocks. A Monte Carlo simula-
tion approach is used to compute two GIRs for each valuation ratio, one associated with
shocks that arrive during bull markets and another one for bear markets (for details, see
Appendix A). The maximum horizon is set to 150 months and all GIRs are normalized so
that the initial effect of the shock is unity.
The GIR estimates in Fig. 3 reveal several interesting patterns which are difficult to
rationalize on the basis of risk-based, classical theories alone. First and contra the null
hypothesis in Test 2A, valuation ratios clearly do not respond fully and immediately to
shocks. Instead the GIRs replicate the typical underreaction–overreaction time profile
posited in behavioral theories of stock returns. It is evident that the ratios display
short-run underreaction or continuation in the wake of the shocks. Indeed, the GIRs reach
a peak a full 12 months after impact. Such a delayed response of prices to fundamentals is
clearly at odds with the immediate one envisaged in classical theories under rational expec-
tations. Second, the hypothesis that the effect of a shock on the valuation ratios decays
rapidly is not supported either (Test 2B). In fact, the GIRs reflect very sluggish adjustment
after the initial 12-month peak taking a further 20–30 months for the effect of a large
shock to revert even to its initial level.
Finally, there is evidence to suggest that shocks arriving during bull markets do not
have the same effect as those impacting during bear markets (Test 2C). The GIRs reveal
that the impact of large positive shocks is asymmetric, having more pronounced and long-
lasting effects if they arrive in a bull market phase.24 For instance, the P/D ratio takes
some 43 months to revert to its initial response level in the wake of a positive shock that
arrives during a bull market. This is a year (38%) longer than in the case of its bear market
counterpart. Moreover, following a bull market shock the P/D ratio peaks at some 1.8
times its initial response. This is some 13% above the corresponding peak following a bear
market shock. The difference between the two responses is in line with behavioral theories.
More specifically, it can be explained by the greater influence of investor sentiment and
noise trading alongside the diminished role of fundamentals during bull markets. This is
ruled out under classical, efficient-market theories.

4.4. Summary and implications

Standard linear tests typically suggest that valuation ratios are unit root processes or
equivalently, that mean-reversion is absent. This implies permanent mispricing in a
dynamic dividend discount model or that prices and fundamentals randomly drift apart
with no tendency to come back in line. Our findings suggest that this counterintuitive evi-

24
It is possible that the large number of common augmentation terms needed in the TAR model to absorb
residual autocorrelation (k = 15) may be obscuring the difference in the GIRs for the two phases. Although we do
not pursue the idea here, we should note that a more complex non-linear model where the full parameter vector
0
(q, b) is regime-sensitive may reveal even a more distinct impact of shocks in bull and bear markets.
 ARTICLE IN PRESS
J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx 17

P/D ratio
2.0
Initial condition
Bull market Bear market
1.6

1.2

0.8

0.4

0.0
20 40 60 80 100 120 140 time

P/E ratio
1.8
Initial condition
1.6 Bull market Bear market

1.4

1.2

1.0

0.8

0.6

0.4

0.2
20 40 60 80 100 120 140 time

Fig. 3. Generalized impulse response of valuation ratios. The figure shows the time profile of the ratios following
a one-off shock of size twice the variance of the TAR model residuals. The solid line plots the valuation ratio
response when the shock occurs (time = 0) during a bull market. The hatched line plots the valuation ratio
response when the shock occurs during a bear market.

dence is no more than an artefact of the poor properties (low power) of these simple tests
when mean-reversion is non-linear or regime-sensitive. The latter is all the more plausible
given the succession of bull and bear market phases over the course of the stock market
cycle.
Our two-regime framework reveals that valuation ratios regularly behave like random
walks with an upward drift or that stock prices become decoupled from fundamentals in
bull markets due to market sentiment. This finding, which dovetails neatly with non-fun-
damentals explanations of the spectacular rise of valuation ratios in periods such as the
1990s, helps to explain the lack of mean-reversion in valuation ratios typically documented
in the literature. Moreover, our results reveal that this persistent behavior is sooner or
later reversed when the adjustment of the ratios toward their equilibrium levels is re-
instated in bear markets. This ensures the overall stationarity of the ratios or more intu-
itively, that prices and fundamentals become realigned one more.
 ARTICLE IN PRESS
18 J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx

To sum up, the evidence suggests that the long-run behavior of valuation ratios is con-
sistent with fundamentals whilst their short-run evolution can be driven by behavioral fac-
tors such as market sentiment or fads. The asymmetric impact of shocks to valuation
ratios in bull and bear markets has potential implications regarding the predictability of
future stock returns on the basis of current and past values of these ratios. In particular,
the performance of existing approaches to forecasting stock returns may improve once
such regime-sensitive properties are taken into account.

5. Conclusions

This paper documents asymmetries in the time evolution of valuation ratios employing
a non-linear model that allows for distinct bull and bear market phases. Our analysis of a
sample of monthly S&P price–dividend and price–earning ratios post-1870 yields some
interesting findings. First, the ratios exhibit short-run continuation or underreaction pos-
sibly fuelled by irrational exuberance in bull market phases such as the late 1990s episode.
Such dynamic behavior, namely, randomly drifting well above their long-run equilibrium
level, is difficult to reconcile with risk-based, classical theories. However, it is consistent
with recent behavioral contributions in financial economics that underline the role of
investor sentiment and noise trading. Second, non-linear impulse response functions that
simulate the evolution of the ratios following a one-off innovation indicate sluggish adjust-
ment and that the effects of positive shocks are more pronounced in bull market phases
due to the influence of market sentiment. Both of these findings help to explain the coun-
terintuitive lack of mean reversion in the ratios that is typically inferred from standard unit
root tests. Finally and reassuringly, we find evidence of reversal or significant adjustment
toward equilibrium levels in bear markets. Such mean reversion in valuation ratios implies
that stock prices ultimately reflect fundamentals in our framework.
What do we learn from these findings? On the one hand, they vindicate ShillerÕs
(1981) and recent research indicating that prices can exhibit substantial short-run devia-
tions from fundamentals due to the role of market sentiment, noise traders and limits
to arbitrage. Our novel time-series framework reveals that the recognition of asymmetric
dynamics over the cycle (bull and bear markets) is crucial for reconciling such appar-
ently persistent deviations and the overall mean reversion in valuation ratios. Thus, our
results not only underline the importance of noise trading and market sentiment in the
short run but also corroborate that prices reflect fundamentals in the long run. Further
extensions are desirable. It may be important to acknowledge this regime-sensitive behav-
ior in forecasting models that use valuation ratios as predictors of future stock returns.
This is an important issue that we intend to examine in future research. Another develop-
ment of our non-linear framework would be explicitly to model the duration of bull and
bear market phases as stochastic processes in order to forecast turning points in the
market.

Acknowledgments

An earlier version of this paper circulated under the title ‘‘Continuation and reversal in
US valuation ratios’’. It was presented at the EFA 2003 Conference in Glasgow, the SCE
2002 Conference on Computing in Economics and Finance in Aix en Provence, the Royal
 ARTICLE IN PRESS
J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx 19

Statistical Society 2002 Workshop on New Developments in Empirical Finance in London

and at seminars at the Bank of England, Cass Business School, University of Essex, Uni-
versity of Kent and University of Manchester. We thank participants and Ian Garrett, Lu-
cio Sarno and Marno Verbeek for comments. We are also grateful to the Managing Editor
Giorgio Szegö and an anonymous referee for detailed comments that resulted in a substan-
tial improvement in the exposition of the paper.

Appendix A. TAR estimation issues

The Granger and Teräsvirta (1993) specific-to-general strategy for non-linear time-ser-
ies modelling is utilized. Accordingly, the linear AR model nested in (9) for qc = qr = q is
fitted to the ratios and we follow Ng and Perron (1995) in using a 0.10-level testing down
approach to identify k from kmax = 18. This specification is then generalized by allowing
0 0
for two phases through q = (qc, qr) . The TAR parameter vector (l, q, b, r2) where
0
b = (b1, . . . , bk) is estimated as follows.
Assume known parameters (w, d) for the transition variable. Rather than taking the
sample mean,
x, as the attractor or steady-state parameter l, this is estimated by a grid
search. The motivation is that
x is a biased estimator of l for asymmetric processes.
The remaining parameters are estimated sequentially by OLS conditional on each grid
point yielding ^ aðlÞ, q ^
^ðlÞ, bðlÞ and residual variance r^2 ðlÞ for l 2 C where C is a set of
plausible values. Our grid search for l is conducted between the .10(s.10) and .90(s.90)
quantiles of xt with step size kG such that s.90 s
m
.10 P kG where m is a positive integer defin-
ing G P 100 grid points. This approach yields reasonably accurate estimates as shown in
Coakley et al. (2003). The attractor estimator is

l ^2 ðlÞ
^ ¼ arg min r ðA:1Þ
l2C

and the estimates of the remaining parameters are given by q ^ð^ ^ lÞ.
lÞ and bð^
Under Gaussian innovations, this estimation procedure is statistically equivalent to
maximum likelihood. By iterating the above grid search approach, a TAR model is fitted
for different combinations of weighting scheme w = {ExpI, ExpD} and delay parameter
1 < d 6 dmax. The Akaike information criterion is used to select best (w, d). We adopt
dmax = 24 months.
Impulse response functions are estimated to measure the response of the ratios to a one-
off shock vt. The time profile provided by these functions, xt+n, n = 0, 1, . . . , N, is examined
for the presence of behavioral features. We employ the generalized impulse response
(GIR) function introduced by Koop et al. (1996) for the analysis of non-linear time series.
A practical problem in estimating GIRs for TAR processes is that a closed-form solution
does not exist for multiple-step-ahead forecasts. However, by conceptualizing the GIR as a
T
random variable on the probability space of the data fxt gt¼1 , a Monte Carlo simulation
facilitates estimates as follows.
For each combination of history ht1 = {xt1, xt2, . . .} and shock vt that act as initial
conditions, use the TAR parameter estimates and N randomly selected future shocks
N
{et+1, . . . , et+N} to generate K sets of forecasts for the shocked system, fxtþn ðht1 ; vt Þgn¼0 .
N
Generate K sets of benchmark forecasts fxtþn ðht1 Þgn¼0 using a randomly sampled shock,
 ARTICLE IN PRESS
20 J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx

et, and the same initial history and random future shocks as in the earlier forecasts. The
GIR is defined as
GIRðn; ht1 ; vt Þ ¼ E½xtþn jvt ; ht1   E½xtþn jht1 ; n ¼ 0; 1; . . . ; N ðA:2Þ

and thus it is history- and shock-specific. It is approximated by averaging the difference

between the two types of forecasts over the K replications. This is repeated for J different
combinations of initial history and shock, and the GIR is obtained by averaging the J indi-
vidual draws. As n ! 1, the theoretical GIR(n, ht1, vt) approaches zero for mean-revert-
ing processes.
The bull market GIR is obtained by randomly drawing the initial history (with replace-
ment) from the sample such that qt(w, d) P 0 when the shock vt arrives at time t. We pro-
ceed analogously for bear markets using qt(w, d) < 0 as the initial condition. The future
shocks (and time t shock for the benchmark forecasts) are randomly drawn from the
TAR residuals ^et . As time passes, t + 1, . . . , t + N, the process may switch from a bull to
a bear market phase and vice versa according to the dynamic evolution dictated by the
TAR model parameters and future shocks. The initial shock is defined as vt ¼ 2^ r where
^ is the standard error of the TAR model. We draw J = 100 histories for each regime
r
and conduct K = 500 Monte Carlo replications for each history.

References

Ackert, L.F., Hunter, W.C., 1999. Intrinsic bubbles: The case of stock prices. Comment. American Economic
Review 89, 1372–1376.
Anderson, D.C., Darrat, A.F., Zhong, M., 2003. Do US stock prices deviate from their fundamental values?
Some new evidence. Journal of Banking and Finance 27, 673–697.
Balke, N.S., Wohar, M.E., 2001. Why are stock prices so high? Dividend growth or discount factor? Federal
Reserve Bank of Dallas Working Paper 00-01.
Barberis, N., Thaler, R., 2002. A survey of behavioral finance. NBER Working Paper No. 9222.
Barberis, N., Shleifer, A., Vishny, R., 1998. A model of investor sentiment. Journal of Financial Economics 49,
307–343.
Berkowitz, J., Kilian, L., 2000. Recent developments in bootstrapping time series. Econometric Reviews 19, 1–48.
Brock, W.A., Dechert, W., Scheinkman, J., LeBaron, B., 1996. A test for independence based on the correlation
dimension. Econometric Reviews 15, 197–235.
Brooks, C., Garrett, I., 2002. Can we explain the dynamics of the UK FTSE 100 stock and stock index futures
markets? Applied Financial Economics 12, 25–31.
Brooks, C., Katsaris, A., 2005. A three-regime model of speculative behaviour: Modelling the evolution of the
S&P 500 Composite. Economic Journal 115, 767–797.
Campbell, J.Y., Shiller, R.J., 1988a. Stock prices, earnings and expected dividends. Journal of Finance 43, 661–
676.
Campbell, J.Y., Shiller, R.J., 1988b. The dividend–price ratio and expectations of future dividends and discount
factors. Review of Financial Studies 1, 195–228.
Campbell, J.Y., Shiller, R.J., 1990. Market volatility and investor behavior. AEA Papers and Proceedings 80, 58–
62.
Campbell, J.Y., Lo, A.W., MacKinlay, C., 1997. The Econometrics of Financial Markets. Princeton University
Press, Princeton, NJ.
Campbell, J.Y., Shiller, R J., 2001. Valuation ratios and the long-run stock market outlook: An update. Cowles
Foundation Discussion Paper No. 1295, Yale University.
Carlson, J.B., Pelz, E.A., Wohar, M.E., 2002. Will the valuation ratios revert to their historical means? Some
evidence from breakpoint tests. Journal of Portfolio Management (Summer), 23–35.
Chan, K.-S., Tsay, R.S., 1998. Limiting properties of the least squares estimator of a continuous threshold
autoregressive model. Biometrika 85, 413–426.
 ARTICLE IN PRESS
J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx 21

Coakley, J., Fuertes, A.-M., in press. Tests for sign and amplitude asymmetries using threshold autoregressions.
Journal of Economic Dynamics and Control.
Coakley, J., Fuertes, A.-M., Zoega, G., 2001. Evaluating the persistence and structuralist theoris of
unemployment from a nonlinear perspective. Studies in Nonlinear Dynamics and Econometrics 5, 179–202.
Coakley, J., Fuertes, A.-M., Pérez, M.T., 2003. Numerical issues in threshold autoregressive modeling of time
series. Journal of Economic Dynamics and Control 27, 2219–2242.
Daniel, K., Hirshleifer, D., Subrahmanyam, A., 1998. Investor psychology and security market under- and
overreactions. Journal of Finance 53, 1839–1885.
Enders, W., Granger, C.W.J., 1998. Unit-root tests and asymmetric adjustment with an example using the term
structure of interest rates. Journal of Business and Economic Statistics 16, 304–311.
Fama, E.F., 1998. Market efficiency, long-term returns, and behavioral finance. Journal of Financial Economics
49, 283–306.
Fama, E., French, K., 1988. Dividend yields and expected stock returns. Journal of Financial Economics 22, 3–
27.
Gallagher, L.A., Taylor, M.P., 2001. Risky arbitrage, limits of arbitrage, and nonlinear adjustment in the
dividend–price ratio. Economic Inquiry 39, 524–536.
Granger, C.W.J., Teräsvirta, T., 1993. Modelling Nonlinear Economic Relationships. Oxford University Press,
Oxford.
Guidolin, M., Timmermann, A., 2005. Economic implications of bull and bear regimes in UK stock and bond
returns. Economic Journal 115, 111–143.
Hansen, B.E., 1997. Inference in TAR models. Studies in Nonlinear Dynamics and Econometrics 2, 1–14.
Hirshleifer, D., 2001. Investor psychology and asset pricing. Journal of Finance 56, 1533–1598.
Hong, H., Stein, J.C., 1999. A unified theory of underreaction, momentum trading, and overreaction in asset
markets. Journal of Finance 54, 2143–2184.
Jaggia, S., Thosar, S., 2004. The medium-term aftermarket in high-tech IPOs: Patterns and implications. Journal
of Banking and Finance 28, 931–950.
Kim, K., Schmidt, P., 1993. Unit root tests with conditional heteroskedasticity. Journal of Econometrics 59, 287–
300.
Koop, G., Pesaran, M.H., Potter, S.M., 1996. Impulse response analysis in nonlinear multivariate models.
Journal of Econometrics 74, 119–147.
Lamont, O., 1998. Earnings and expected returns. Journal of Finance 53, 1563–1587.
Lee, B.S., 1998. Permanent, temporary, and non-fundamental components of stock prices. Journal of Financial
and Quantitative Analysis 33, 1–32.
Lee, W.Y., Jiang, C.X., Indro, D.C., 2002. Stock market volatility, excess returns, and the role of investor
sentiment. Journal of Banking and Finance 26, 2277–2299.
Lekkos, I., Milas, C., 2004. Time-varying excess returns on UK government bonds: A non-linear approach.
Journal of Banking and Finance 28, 45–62.
Madsen, J.B., Milas, C., 2003. The price–dividend relationship in inflationary and deflationary regimes. Mimeo,
Department of Economics and Finance, Brunel University, London.
Manzan, S., 2005. Nonlinear mean reversion in stock prices. Mimeo, University of Leicester.
McMillan, D.G., 2004. Nonlinear predictablility of short-run deviations in UK stock market returns. Economics
Letters 84, 149–154.
Ng, S., Perron, P., 1995. Unit root test in ARMA models with data dependent methods for the selection of the
truncation lag. Journal of the American Statistical Association 90, 268–281.
Pagan, A.R., Sossounov, K.A., 2003. A simple framework for analysing bull and bear markets. Journal of
Applied Econometrics 18, 23–46.
Pesaran, M.H., Timmerman, A., 1995. Predictability of stock returns: Robustness and economic significance.
Journal of Finance 50, 1201–1228.
Psaradakis, Z., Sola, M., Spagnolo, F., 2004. On Markov error-correction models, with an application to stock
prices and dividends. Journal of Applied Econometrics 19, 69–88.
Randles, R., Flinger, M., Policello, G., Wolfe, D., 1980. An asymptotically distribution-free test for symmetry
versus asymmetry. Journal of the American Statistical Association 75, 168–172.
Schaller, H., van Norden, S., 2002. Fads or bubbles? Empirical Economics 27, 335–362.
Shiller, R.J., 1981. Do stock prices move too much to be justified by subsequent changes in dividends? American
Economic Review 71, 421–436.
Shiller, R.J., 2000. Irrational Exuberance. Princeton University Press, Princeton, NJ.
 ARTICLE IN PRESS
22 J. Coakley, A.-M. Fuertes / Journal of Banking & Finance xxx (2005) xxx–xxx

Summers, L.H., 1986. Does the stock market rationally reflect fundamental values? Journal of Finance 41, 591–
602.
Verbrugge, R., 1997. Investigating cyclical asymmetries. Studies in Nonlinear Dynamics and Econometrics 2, 15–
22.
Zivot, E., Andrews, D.W.K, 1992. Further evidence on the Great Crash, the oil-price shock and the unit-root
hypothesis. Journal of Business and Economic Statistics 10, 251–270.