International Review of Economics and Finance

12 (2003) 385 – 397

Periodically collapsing bubbles in the US stock market?

Martin T. Bohl*
Department of Economics, European University Viadrina Frankfurt (Oder), Grosse Scharrnstrasse 59,
15230 Frankfurt (Oder), Germany

Received 27 November 2001; received in revised form 25 February 2002; accepted 2 April 2002

Abstract

The existence of periodically collapsing bubbles in stock markets, applying the Enders –Siklos
momentum threshold autoregressive (MTAR) model, is empirically investigated in this paper. Using
this nonlinear time series technique, we are now able to analyse bubble-driven run-ups in stock prices
followed by a crash in a cointegration framework with asymmetric adjustment. Therefore, applying
this technique makes possible a deeper insight into the behavior of stock prices than was previously
possible using conventional cointegration tests. Although the results from the subsample 1871 – 1995
cannot be interpreted in favor of the existence of periodically collapsing bubbles in the US stock
market, the findings from the 1871– 2001 sample period indicate their presence.
D 2002 Elsevier Science Inc. All rights reserved.

JEL classification: G12; E44; C32

Keywords: Periodically collapsing bubbles; Asymmetric adjustment; Threshold cointegration

1. Introduction

Since the 1980s, there has been continuous research interest into the phenomenon of
speculative bubbles in share prices, motivated primarily by actual developments in stock
markets and the increasing importance of stock markets for investors. The analyses of
speculative bubbles are intimately related to time series analysis so that advances in

* Tel.: +49-335-5534-2984; fax: +49-335-5534-2959.

E-mail address: [email protected] (M.T. Bohl).

1059-0560/02/$ – see front matter D 2002 Elsevier Science Inc. All rights reserved.
doi:10.1016/S1059-0560(02)00128-4
386 M.T. Bohl / International Review of Economics and Finance 12 (2003) 385–397

econometrics enable an investigation of open questions in this field and their use provide the
opportunity to obtain further insights into the characteristics of stock markets. One potential
example is the class of periodically collapsing bubbles (Evans, 1991) and the development of
econometric techniques designed to capture nonlinear adjustment mechanisms in a cointe-
gration framework (Enders & Granger, 1998; Enders & Siklos, 2001).
Studies investigating the consistency of dividend and stock price data with the market
fundamental hypothesis (Blanchard & Watson, 1982; Shiller, 1981; West, 1987) are
confronted with the difficulty that the contribution of hypothetical rational bubbles to stock
prices are not directly distinguishable from the contribution of unobservable market
fundamentals. As an alternative testing strategy, Diba and Grossman (1984, 1988a) proposed
the use of standard unit root and cointegration tests (Bhargava, 1986; Dickey & Fuller, 1981;
Engle & Granger, 1987) for stock prices and observable fundamentals to obtain evidence for
the existence of explosive rational bubbles. This approach relies on the argument that if stock
prices are not more explosive compared to dividends, then rational bubbles do not exist
because they generate an explosive component into stock price time series. The empirical
evidence reported in Diba and Grossman was not conducive to the conclusion that there are
explosive bubbles in US stock prices.1
Evans (1991) argued that the test approaches put forward by Diba and Grossman are
unable to detect an important class of rational bubbles, namely periodically collapsing
bubbles. The application of standard unit root and cointegration techniques leads, with a high
probability, to incorrect conclusions with respect to the presence of bubbles in stock prices.
Evans’ Monte Carlo simulations show that even in the presence of periodically collapsing
bubbles, stock prices are not more explosive than dividends using standard unit root and
cointegration tests (see also Charemza & Deadman, 1995). The explanation relies on the logic
of standard unit root and cointegration tests; they assume a unit root as the null hypothesis
and a linear autoregressive process under the alternative hypothesis. In the case of
periodically collapsing bubbles, the bubble component is a nonlinear process, which falls
outside the alternative hypothesis. Relying on simulated data with periodically collapsing
bubbles present, the findings of standard unit root and cointegration tests reported in Evans’
study incorrectly show the absence of bubbles in the majority of cases.
Basing his central argument solely on Monte Carlo simulations and highlighting the power
properties of standard unit root and cointegration tests, Evans was not able to provide
empirical evidence as to whether periodically collapsing bubbles are actually present in US
stock prices due to lack of techniques suitable to deal with nonlinear processes in a
cointegration framework. Therefore, the presence of this class of rational bubbles in stock
prices remained an open question.
This paper tries to fill this gap using recent advances in the field of time series modelling.
Applying the momentum threshold autoregressive (MTAR) model (Enders & Granger, 1998;

1
As noted many times in the literature, it is virtually impossible to prove the existence of bubbles. This
limitation holds for all indirect time series tests as well as for direct approaches relying on a particular
structural model.
 M.T. Bohl / International Review of Economics and Finance 12 (2003) 385–397 387

Enders & Siklos, 2001) for stock prices and dividends makes possible the empirical
investigation of the existence of periodically collapsing bubbles in stock prices in a
cointegration framework.2 By taking into account asymmetries in departures from the
long-run equilibrium relationship, the MTAR technique is designed to empirically capture
the characteristics of periodically collapsing bubbles. We apply this approach to annual and
monthly US time series for the period from 1871 to 2001 as well as for the 1871–1995
subsample on real stock prices and dividends. Furthermore, we provide Monte Carlo
simulation findings as to whether the critical test for symmetry in the Enders–Siklos
approach is sufficiently powerful to detect asymmetry when using the data generating
process outlined by Evans.
The paper proceeds as follows. Section 2 describes the theory behind periodically
collapsing bubbles as outlined in Evans (1991). Section 3 presents the MTAR technique to
capture the behavior of this class of rational bubbles in stock prices together with the findings
of the Monte Carlo study. Section 4 provides the empirical evidence for the US stock market,
and Section 5 concludes.

2. Present value model and Evans’ periodically collapsing bubbles

The Diba–Grossman (1984, 1988a) approach relies on a standard present value model of
stock prices, which can be written as:

1
Pt ¼ Et ðPtþ1 þ Dtþ1 Þ; 0 < ð1 þ rÞ1 < 1: ð1Þ
ð1 þ rÞ

The real stock price Pt is related to next period’s expected stock price and the expected real
before-tax dividend payments Et( Pt + 1 + Dt + 1) according to the constant discount factor
(1 + r)  1. Et denotes the conditional expectations operator. If the transversality condition
holds, then the stock price is equal to the fundamental value Pt = Ft. In turn, the market
fundamentals component of the stock price is equal to the present value of expected real
dividends, discounted by the real interest rate r:
X
1
1
Ft ¼ Et Dtþj : ð2Þ
j¼1 ð1 þ rÞj

In case the transversality condition fails to hold, then the stock price can deviate from the
fundamental value according to the general solution to Eq. (1):
Pt ¼ Ft þ Bt ; ð3Þ
2
Four alternative methodologies were proposed in the literature to overcome the Evans critique: Scacciavillani
(1994) suggested a test based on fractional differencing and Hall, Psaradakis, and Sola (1999) used a Markov-
switching ADF test. In van Norden (1996), a switching regression similar, but not equivalent, to the Markov-
switching ADF approach is applied, and Taylor and Peel (1998) introduced a cointegrating regression RALS DF test.
388 M.T. Bohl / International Review of Economics and Finance 12 (2003) 385–397

where the bubble term Bt satisfies (Eq. (4)):

1
Bt ¼ Et Btþ1 : ð4Þ
ð1 þ rÞ

Following Campbell and Shiller (1987), in the absence of bubbles, Eqs. (2) and (3) imply:

X
1
Pt  r1 Dt ¼ ð1 þ rÞr1 ð1 þ rÞj Et DDtþj : ð5Þ
j¼1

If stock prices and real dividends follow integrated processes of order one, Pt,Dt~I(1), and no
bubbles are present, Pt = Ft, then Pt and Dt are theoretically cointegrated with the
cointegrating parameter r  1. If a bubble is present, the right-hand side of Eq. (5) must be
augmented by the nonstationary process Bt so that Pt and Dt cannot be cointegrated. Hence,
the application of unit root tests on Pt, Dt, DPt, and DDt as well as a test for cointegration
between Pt and Dt seem to allow an investigation of the existence of rational bubbles. The
nonrejection of the null hypothesis of a unit root in Pt and Dt, the rejection of the unit root
hypothesis for DPt and DDt, and the rejection of the null hypothesis of no cointegration
between Pt and Dt are interpretable as evidence against the existence of explosive rational
bubbles (Diba & Grossman, 1984, 1988a).
The main argument put forward by Evans (1991) was that the above procedure incorrectly
leads to the conclusion that speculative rational bubbles do not exist, even when periodically
collapsing bubbles are present. Relying on the work of Blanchard (1979), Blanchard and
Watson (1982), and Diba and Grossman (1988b), Evans’ periodically collapsing bubbles can
be formalized as follows:

Btþ1 ¼ ð1 þ rÞBt vtþ1 ; if Bt  a ð6aÞ

  
ð1 þ rÞ d
Btþ1 ¼ dþ qtþ1 Bt  vtþ1 ; if Bt > a: ð6bÞ
p ð1 þ rÞ

In Eqs. (6a) and (6b), d and a are positive parameters with 0 < d < (1 + r)a and p denotes a
probability. vt + 1 is an i.i.d. positive random variable with Etvt + 1 = 1 and qt + 1 denotes an i.i.d.
Bernoulli process, which takes the value 1 with probability p and 0 with probability 1  p.
Setting p and qt + 1 equal to 1 demonstrates that Eq. (6a) is a special case of Eq. (6b).
The Evans model is a generalization of Blanchard’s (1979) bubble model, where the size of
collapses as well as their probability depends on the size of the bubble. Furthermore, Evans’
periodically collapsing bubbles are always positive and burst after reaching high levels and
the model incorporates partial, rather than total, collapses. Because the multiplicative random
variable vt + 1 is strictly positive, the bubble will never change sign, remains positive, and will
never completely vanish. Before the bubble size reaches the level a (see Eq. (6a)), the
 M.T. Bohl / International Review of Economics and Finance 12 (2003) 385–397 389

probability of collapse is 0 and the bubble grows at the mean rate (1 + r). A high value for the
a parameter results in bubbles with a long initial period of relatively steady slow growth.
Once the bubble size reaches the level a (see Eq. (6b)), the bubble grows at the higher rate
(1 + r)p  1, as long as the eruption continues, because qt + 1 is 1 with the probability p.
Accordingly, the bubble may collapse with a per period probability of 1  p, up to the
positive mean value of d, and the process begins again. Thus, p is the probability of the
continuation of the bubble per period (Evans, 1991).
The discussion of Eqs. (6a) and (6b) has shown that the Evans bubble model satisfies two
theoretically well-established properties of stochastic bubbles (Diba & Grossman, 1988b).
First, there can never be a negative stock price bubble, because a negative bubble would
imply a negative expected stock price, which is not economically possible. Second, Evans’
periodically collapsing bubbles cannot completely burst, because following a complete
collapse bubbles cannot emerge again. Furthermore, the parameter values for d, a, and p
determine the scale of the bubble, the average length of time before a collapse, and the
frequency with which bubbles erupt.

3. Periodically collapsing bubbles, MTAR modelling, and Monte Carlo findings

MTAR models proposed by Enders and Siklos (2001) are designed to empirically capture
the characteristics of periodically collapsing bubbles described above in a cointegration
framework. If periodically collapsing bubbles exist in stock prices, the estimated residuals ût
from the cointegrating regression
Pt ¼ bˆ 0 þ bˆ 1 Dt þ uˆ t ð7Þ
should reflect sequences of increases in stock prices followed by a sudden drop. This
behavior can be captured relying on the regression
X
l
Duˆ t ¼ It r1 uˆ t1 þ ð1  It Þr2 uˆ t1 þ gi Duˆ ti þ et ð8Þ
i¼1
using the indicator variable
8
< 1; if Dût1  t
It ¼ ð9Þ
:
0; if Dût1 < t

with t as the value of the threshold. The MTAR model sets up the null hypothesis of no
cointegration H0: r1 = 0, H0: r2 = 0, and H0: r1 = r2 = 0, where Enders and Siklos (Tables 1 and
2) provide critical values for the corresponding t and F statistics.
If the null hypothesis of no cointegration is rejected, the null hypothesis of symmetric
adjustment, H0: r1 = r2, can be tested using the usual F-statistic. The nonrejection of the null
hypothesis H0: r1 = r2 can be considered as evidence of a cointegrating relationship between Pt
and Dt with linear and symmetric adjustment. Obviously, the Engle–Granger (1987) test is a
390 M.T. Bohl / International Review of Economics and Finance 12 (2003) 385–397

special case of the MTAR model. Enders and Siklos show that for a plausible range of
adjustment parameters, the power of the MTAR test can be many times that of the Engle–
Granger test when there are asymmetric departures from equilibrium.3
The MTAR technique is designed to detect empirically periodically collapsing bubbles.
The theoretical potential for positive, but not negative, bubbles, and the characteristic of stock
price increases relative to dividends before a crash, suggests an asymmetry in the devel-
opment of the residual of the cointegrating regression (Eq. (7)). Periodically collapsing
bubbles are captured via changes in ût  1 above the threshold followed by a sharp drop to the
threshold. In contrast, the path of changes in ût  1 below the threshold does not show bubble
eruptions followed by a collapse.
For example, consider a threshold of t = 0 in Eq. (9). A value of Dût>0 is indicative of a
rise in stock prices relative to dividends followed by a crash, whereas a comparable behavior
of decreases in stock prices relative to dividends Dût < 0 followed by a sharp increase back to
the equilibrium position is not expected. The result is an asymmetry in deviations from the
equilibrium indicating the existence of periodically collapsing bubbles. Accordingly, if the
estimated coefficient r̂ 1 is statistically significant and negative and larger in absolute terms
relative to the parameter r̂ 2, the null hypothesis of symmetric adjustment, H0: r1 = r2, is
rejected. The rejection of this null hypothesis is evidence in favor of the existence of
periodically collapsing bubbles in stock prices.
While the null hypotheses of the conventional Engle–Granger test and the MTAR model
are identical, the alternative hypotheses of both differ in case of a rejection of the null
hypothesis H0: r1 = r2. The feature of testing the null hypothesis of no cointegration against
the alternative of cointegration with MTAR adjustment allows an empirical investigation of
periodically collapsing bubbles. Although this approach constitutes only an indirect test of the
presence of periodically collapsing bubbles, it overcomes the drawback of standard unit root
and cointegration tests as outlined in Evans (1991).
Is the F-test for the null hypothesis of symmetry, H0: r1 = r2, sufficiently powerful to detect
asymmetry in case the data generating process is given by the Evans bubble model?
Answering this question is important because the main contribution of this paper lies in
the investigation of the null hypothesis of symmetry and not in the rejection of the null
hypothesis of no cointegration. Hence, relying on Eqs. (6a) and (6b), we follow fairly closely
Evans’ (1991) Monte Carlo experiment by using the parameter values r = 0.05, a = 1, d = 0.5,
initial Bt = d, and T = 100. We apply 5000 replications and the regressions are implemented
with no deterministic terms.4 The percentage of correct rejections of the null hypotheses

3
In addition to the MTAR model, Enders and Siklos investigate the properties of threshold autoregressive
(TAR) adjustment, which relies in Eq. (9) on the level of the estimated residuals ût  1 instead of changes in ût  1.
The TAR test has, however, less power than the conventional Engle – Granger test. Furthermore, the TAR
adjustment model is not able to capture the asymmetrically sharp adjustments toward the long-run equilibrium
typical for periodically collapsing bubbles.
4
Because we have little a priori knowledge about the true value of the threshold t, Chan’s (1993) method is
used to consistently estimate this parameter. This involves sorting the estimated residuals in ascending order,
excluding 15% of the largest and smallest values, and selecting from the remaining 70% the threshold parameter,
which yields the lowest residual sum of squares (Enders & Siklos, 2001).
 M.T. Bohl / International Review of Economics and Finance 12 (2003) 385–397 391

Table 1
Monte Carlo simulation results using the MTAR technique
Significance Null Percentage of correct rejections of the null hypothesis for different probabilities (p)
level hypothesis 0.990 0.950 0.850 0.750 0.500 0.250
10% r1 = r2 = 0 0.993 0.993 0.993 0.993 0.994 0.995
r1 = r2 0.643 0.640 0.633 0.622 0.578 0.491
5% r1 = r2 = 0 0.984 0.984 0.985 0.986 0.989 0.992
r1 = r2 0.583 0.579 0.571 0.562 0.522 0.445
1% r1 = r2 = 0 0.947 0.946 0.946 0.947 0.962 0.977
r1 = r2 0.474 0.471 0.465 0.451 0.422 0.369
Each entry is the percentage of instances in which the null hypothesis was correctly rejected. For details on the
Monte Carlo simulation, see Section 3.

H0: r1 = r2 = 0 and H0: r1 = r2 are reported for the 10%, 5%, and 1% significance levels as
well as for different probabilities p [0.99, 0.25].
The results of the Monte Carlo study can be found in Table 1. The overwhelming finding is
that both tests are quite powerful, irrespective of the value of the probability and the
significance level. The null hypothesis of no cointegration, H0: r1 = r2 = 0, is correctly
rejected in nearly all cases. The percentages of rejections increase slightly with decreasing
probabilities p. More importantly, the null hypothesis of symmetry, H0: r1 = r2, is correctly
rejected at the 10%, 5%, and 1% significance levels in about 60%, 54%, and 44% of trials,
respectively. Hence, the F-test for the symmetry hypothesis is sufficiently powerful to detect
asymmetry when the data generating process proposed by Evans is used.5

4. Empirical results

Unit root tests and cointegration approaches are implemented for annual and monthly US
data for stock prices and dividends. The Standard and Poor’s stock price index and the
corresponding dividend time series are in real terms. The time series cover the period 1871–
1999 for the annual data and 1871:1–2001:3 for the monthly data. In addition to these sample
periods, the subsample 1871–1995 is also examined to take into account the extraordinary
behavior of US stock prices since the middle of the 1990s. The time series are taken from
Shiller’s Web site hhttp://aida.econ.yale.edu/~shilleri and a description of the time series can
be found in Shiller (1989, 2000). The empirical results are reported in Tables 2 and 3.
First, following the Diba–Grossman methodology, we investigate the stochastic properties
of real stock prices and real dividends individually implementing the test approaches

5
In addition to the results reported in Table 1, we have implemented Monte Carlo studies for different
parameter values of r, a, d, and T. Moreover, additional simulations are run incorporating an intercept in the
regressions. The findings (not shown) do not change our main conclusion concerning the power properties of
both tests.
392 M.T. Bohl / International Review of Economics and Finance 12 (2003) 385–397

Table 2
Unit root tests
Time series Sample ADF l KPSS
Annual data
Pt 1871 – 1995 0.04 0 1.80***
1871 – 1999 4.63 1 1.61***
Dt 1871 – 1995  0.09 0 2.22***
1871 – 1999 0.82 0 2.28***
DPt 1871 – 1995  10.91*** 0 0.19
1871 – 1999 0.69 2 0.56**
DDt 1871 – 1995  9.87*** 0 0.12
1871 – 1999  9.82*** 0 0.24

Monthly data
Pt 1871 – 1995 0.07 5 12.11***
1871 – 2001  0.39 8 9.54***
Dt 1871 – 1995  1.23 5 14.94***
1871 – 2001  1.39 5 15.97***
DPt 1871 – 1995  14.95*** 4 0.23
1871 – 2001  9.33*** 7 0.51**
DDt 1871 – 1995  10.47*** 4 0.06
1871 – 2001  10.55*** 4 0.05
Pt denotes the level of real stock prices, Dt the level of real dividends, D the first difference operator, ADF the
augmented Dickey and Fuller (1981) statistic, and KPSS the statistic suggested by Kwiatkowski et al. (1992). The
lag lengths of the ADF test l are selected according to Hall (1994). For the KPSS test, the lag length is l = 4 for
annual time series and l = 7 for monthly time series (Schwert, 1989). The time series are taken from Shiller’s Web
site hhttp://aida.econ.yale.edu/~shilleri critical values are from MacKinnon (1991) and Sephton (1995).
* Denotes significant statistics at the 10% level.
** Denotes significant statistics at the 5% level.
*** Denotes significant statistics at the 1% level.

proposed by Dickey and Fuller (1981) as well as Kwiatkowski, Phillips, Schmidt, and Shin
(1992), hereafter ADF and KPSS. We use MacKinnon’s (1991) and Sephton’s (1995)
response surface estimates to calculate critical values. Results for the undifferenced time
series of real stock prices Pt and dividends Dt as well as for the time series in first differences
DPt and DDt are presented for both sample periods. The lag lengths l for the ADF tests are
determined by the Hall (1994) procedure, while Schwert’s (1989) formula l = int(4(T/100)1/4)
is used for the KPSS tests. The ADF tests are performed with a constant term and the KPSS
tests investigate the null hypothesis of level stationarity. Hereafter, for all test statistics, we
use the conventional 10%, 5%, and 1% significance levels.
The ADF and KPSS statistics in Table 2 can be interpreted in favor of the existence of one
unit root in the levels of both time series. The ADF tests cannot reject the null hypothesis of a
unit root in the undifferenced real stock price and dividend time series. With only one
exception, the ADF tests reject the null hypothesis of a unit root in both differenced time
series. These findings are generally supported by the KPSS tests, which reject the null
hypothesis of level stationarity for the time series in levels in all cases. Furthermore, with two
 M.T. Bohl / International Review of Economics and Finance 12 (2003) 385–397 393

exceptions, the KPSS tests cannot reject this null hypothesis for the time series in the first
difference.6 The exceptions refer to the first difference of the real stock price time series in the
samples 1871–1999 and 1871–2001. In case of the annual time series, the findings are
interpretable in favor of the existence of a unit root in DPt, while the ADF and KPSS statistics
for the monthly time series provide an inconclusive result.
Our empirical results are broadly in accordance with the findings in Campbell and Shiller
(1987), Diba and Grossman (1988a), Saltoglu (1998), and Timmermann (1995) for sample
periods starting in the 1870s and ending in the 1980s. Following the methodology put
forward by Diba and Grossman, our findings can be interpreted by and large in favor of the
absence of speculative bubbles in the US stock market. Nevertheless, when looking at the
stock price time series for the sample including the extraordinary behavior of US share prices
since the middle of the 1990s, the empirical results are not so clear cut.
Next, we test for cointegration between real stock prices and dividends using the
conventional Engle–Granger
P (1987) approach based on Eq. (7) and the auxiliary regression
Dût = rût  1 + il = 1giDût  i + et. Lag lengths l are selected according to statistically sig-
nificant coefficients on the lagged values Dût  i. Panel A in Table 3 reports the findings of the
Cointegrating Regression Durbin–Watson (CRDW) tests and the Cointegrating Regression
ADF (CRADF) statistics. While for the 1871–1995 subsample there is evidence in favor of a
cointegrating relationship, the test statistics for the sample including the period of rapid share
price increases cannot reject the null hypothesis of no cointegration. This holds for the annual
and the monthly time series.
Furthermore, Johansen’s (1988, 1991) maximum likelihood procedure is performed and
the lag lengths are selected according to the criteria of serially uncorrelated residuals using
LM-type tests for first-order and fourth-order autocorrelations LM1 and LM4. Relying on
trace test statistics, the evidence for the 1871–1995 subsample is again in favor of a
cointegrating relationship between real stock prices and real dividends. As well, the evidence
of the annual data for the period from 1871 to 1999 supports the existence of a cointegrating
relationship. However, when looking at the results for the monthly time series for the 1871–
2001 period, the trace statistic cannot reject the null hypothesis of no cointegration and the
estimated cointegrating coefficient b̂ 1 shows a dramatic increase. Moreover, according to the
LM4 test, the residuals contain fourth-order autocorrelation.7
In sum, the conventional Engle–Granger and Johansen cointegration tests indicate that real
stock prices and dividend time series are cointegrated in the subsample 1871– 1995.
Furthermore, for this subsample, the estimated long-run cointegrating coefficients are stable
across the implemented cointegration techniques and comparable to the estimated values in
Campbell and Shiller (1987), Diba and Grossman (1988a), Saltoglu (1998), and Taylor and
Peel (1998) on US data. Therefore, the findings of the conventional cointegration analysis are

6
In addition to the test results contained in Table 2, ADF tests with a constant term and a linear time trend in
the alternative hypothesis and KPSS tests analysing the null hypothesis of trend stationarity were performed.
Generally, the findings of these tests confirm the results presented.
7
In addition to different cointegration procedures, alternative lag lengths selection criteria were applied. The
findings are qualitatively identical to the results presented in Table 3.
394 M.T. Bohl / International Review of Economics and Finance 12 (2003) 385–397

Table 3
Cointegration tests
Panel A: Engle – Granger results
2
Frequency Sample b̂1 CRDW CRADF R l
Annual 1871 – 1995 32.86 0.43 *  3.60* * 0.89 0
1871 – 1999 43.62 0.23 0.76 0.77 1, 3
Monthly 1871 – 1995 35.53 0.03  4.25* * * 0.83 1, 5
1871 – 2001 49.77 0.01  2.23 0.65 1, 4, 5

Panel B: Johansen procedure (trace test)

Frequency Sample b̂1 r=0 r1 LM1 LM4 l
Annual 1871 – 1995 35.16 13.61 * 0.06 2.15 3.38 1
1871 – 1999 24.93 70.73* * * 0.55 1.97 3.66 1
Monthly 1871 – 1995 38.22 31.67* * * 0.15 2.46 5.15 1, 2, 3
1871 – 2001 118.50 6.76 1.00 2.24 15.26* * * 1, 2, 3

Panel C: MTAR technique

Frequency Sample t̂ r̂1 r̂ 2 F̂C F̂A l
Annual 1871 – 1995 10.74  0.30 (3.14)* * *  0.15 (2.18) * * 7.30* * 1.56 0
1871 – 1999 19.06 0.60 (8.29)* * *  0.20 (3.47) * * * 45.89* * * 88.42* * * 1
Monthly 1871 – 1995 0.52  0.02 (4.13)* * *  0.01 (1.96) * * 10.27* * * 2.41 1, 5
1871 – 2001 2.69  0.01 (4.18)* * * 0.002 (0.69) 8.97* * * 12.95* * * 1, 4, 5
b̂1 represents the estimated cointegrating parameter, CRDW the Cointegrating Regression Durbin – Watson
2
statistic, CRADF the Cointegrating Regression ADF statistic, R the coefficient of determination, l the lag length,
r the number of cointegrating vectors, LM1 and LM4 the LM-type tests of first-order and fourth-order
autocorrelated residuals, t̂ the estimated threshold (Chan, 1993), and r̂1 and r̂2 the estimated parameters of the
MTAR model. t statistics are in parentheses. F̂C and F̂A denote the F-statistics for the null hypothesis of no
cointegration and symmetry critical values are from Enders & Siklos (2001), Engle & Yoo (1987), MacKinnon
(1991) and Osterwald-Lenum (1992).
* Denotes significant statistics at the 10% level.
** Denotes significant statistics at the 5% level.
*** Denotes significant statistics at the 1% level.

interpretable as evidence against the existence of speculative bubbles in the US stock market
over the subsample 1871–1995. This evidence is insensitive to alternative model specifica-
tions and test approaches and is in accordance with the previous literature. However, for the
sample 1871–2001, the empirical evidence is against the existence of a cointegrating
relationship between US stock prices and dividends. According to the Diba–Grossman
methodology, the findings of the Engle–Granger and the Johansen approach support the
existence of explosive rational bubbles in US stock prices.
Relying on the conventional test approaches, we do not have evidence to the existence of
periodically collapsing bubbles in stock prices. The univariate MTAR model, put forward by
Enders and Granger (1998), is applied to the times series DPt and DDt individually to provide
a first piece of evidence on linear or asymmetric adjustment mechanisms. While the findings
(not shown) for the annual time series do not reveal asymmetries interpretable in favor of
 M.T. Bohl / International Review of Economics and Finance 12 (2003) 385–397 395

periodically collapsing bubbles for both periods, the test results for the monthly stock price
times series DPt show statistically significant adjustment mechanisms, which are in
accordance with the hypothesis of periodically collapsing bubbles. Interestingly, the real
dividend time series DDt displays this kind of asymmetries as well.
We now turn to the findings for the MTAR model, which are reported in Panel C of Table
3. Shown are the estimation results for Eq. (8) including the estimated parameters r̂1 and r̂2
and the associated t statistics for the null hypotheses H0: r1 = 0 and H0: r2 = 0. Table 3 also
contains the F-statistics F̂C and F̂A, which test the null hypothesis of no cointegration, H0:
r1 = r2 = 0, and the symmetry property, H0: r1 = r2. In addition, the consistently estimated
attractor t̂ following Chan’s (1993) method is shown in Table 3.
When looking at the results in Panel C, the estimated parameters for deviations above and
below the threshold are statistically significant negative, at least at the 5% level, and the F̂C
statistics reject the null hypothesis of no cointegration for the 1871–1995 subsample. While
the point estimates for the parameters r̂1 are higher in absolute terms compared to the
estimated r̂2 coefficients, the F̂A statistics, however, cannot reject the null hypothesis of
symmetric adjustment. Against the background of our findings on the univariate MTAR unit
root tests for the monthly time series DPt and DDt, this result can be explained by a
synchronised asymmetric behavior across the two time series.
The evidence from the samples 1871–1999 and 1871–2001 do not correspond to the
findings for the 1871–1995 subsample. Both F̂C tests reject the null hypothesis of no
cointegration, and both F̂A statistics reject the symmetry null hypothesis. For the annual time
series, the r̂1 parameter is statistically significant positive, which cannot be interpreted as a
valid error correction mechanism. However, when looking at the results for the monthly time
series, the r̂1 parameter is statistically significant negative, while the r̂2 parameter is not. This
evidence is favorable for the hypothesis that periodically collapsing bubbles are present in US
share prices over the 1871–2001 sample period.
In sum, the evidence provided by the MTAR cointegration tests can be interpreted in favor
of the absence of periodically collapsing bubbles in the US stock market for the subsample,
excluding the rapid share price increases since 1995. The null hypothesis of no cointegration
is rejected and there is no indication of an asymmetric development of the residuals
introduced by run-ups in stock prices followed by a crash. However, the evidence for
monthly time series for the 1871–2001 sample supports the existence of periodically
collapsing bubbles in US stock prices.

5. Conclusion

This article introduces the MTAR cointegration model proposed by Enders and Siklos
(2001) to empirically investigate periodically collapsing bubbles (Evans, 1991) in annual and
monthly US stock prices. Evans stressed that even in the presence of periodically collapsing
bubbles, stock prices will appear to be integrated and cointegrated with their fundamentals so
that standard integration and cointegration tests are not able to detect this class of rational
bubbles. The Enders–Siklos MTAR model is a generalization of the Engle and Granger
396 M.T. Bohl / International Review of Economics and Finance 12 (2003) 385–397

(1987) two-step procedure allowing a formal test of rational speculative bubbles, which
eventually burst after reaching high levels. Technically, the bubble component can be taken
into account as a nonlinear process in the alternative hypothesis. Our Monte Carlo simulation
findings show that the MTAR approach provides a sufficiently powerful test to detect
periodically collapsing bubble behavior when the actual data generating process is given by
the bubble model put forward by Evans.
Relying on the MTAR technique, the empirical findings contained in this paper are
interpretable in favor of the absence of periodically collapsing bubbles in the US stock market
over the subsample 1871–1995. Deviations from the long-run equilibrium do not seem to
indicate an asymmetric adjustment to the long-run relationship. This finding supports the
results contained in Taylor and Peel (1998) who investigate the bubbles hypothesis for the US
stock market relying on a different testing approach and they were not able to confirm the
existence of periodically collapsing bubbles.
However, the evidence for the sample including the rapid share price increases since the
middle of the 1990s is interpretable in favor of the existence of periodically collapsing
bubbles in US stock prices. According to the results for the 1871–2001 sample, in the short-
run US stock prices exhibit run-ups followed by crashes, while in the long-run US share
prices adhere to fundamentals.

Acknowledgements

Comments on an earlier version of this paper by seminar participants at the Deutsche

Bundesbank are gratefully acknowledged as were the helpful comments by two anonymous
referees, Ralf Ahrens, Dirk Plankensteiner, and Stefan Reitz. Furthermore, the author is
indebted to Oleksandr Maksymenko, Shauna Selvarajah, and Sergiy Yermak for their
excellent research assistance.

References

Bhargava, A. (1986). On the theory of testing for unit roots in observed time series. Review of Economic Studies,
53, 369 – 384.
Blanchard, O. J. (1979). Speculative bubbles, crashes and rational expectations. Economics Letters, 3, 387 – 389.
Blanchard, O. J., & Watson, M. (1982). Bubbles, rational expectations, and financial markets. In Wachtel P. (Ed.),
Crises in the economic and the financial structure ( pp. 295 – 315). Lexington: Lexington Books.
Campbell, J. Y., & Shiller, R. J. (1987). Cointegration and tests of present value models. Journal of Political
Economy, 95, 1062 – 1088.
Chan, K. S. (1993). Consistency and limiting distribution of the least squares estimator of a threshold autore-
gressive model. Annals of Statistics, 21, 520 – 533.
Charemza, W. W., & Deadman, D. F. (1995). Speculative bubbles with stochastic explosive roots: the failure of
unit root testing. Journal of Empirical Finance, 2, 1453 – 1463.
Diba, B. T., & Grossman, H. I. (1984). Rational bubbles in the price of gold. NBER Working Paper No. 1300.
Diba, B. T., & Grossman, H. I. (1988a). Explosive rational bubbles in stock prices? American Economic Review,
78, 520 – 530.
 M.T. Bohl / International Review of Economics and Finance 12 (2003) 385–397 397

Diba, B. T., & Grossman, H. I. (1988b). The theory of rational bubbles in stock prices. Economic Journal, 98,
746 – 754.
Dickey, D. A., & Fuller, W. A. (1981). The likelihood ratio statistics for autoregressive time series with a unit root.
Econometrica, 49, 1057 – 1072.
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.
Enders, W., & Siklos, P. L. (2001). Cointegration and threshold adjustment. Journal of Business and Economic
Statistics, 19, 166 – 176.
Engle, R. F., & Granger, C. W. J. (1987). Cointegration and error correction: representation, estimation and testing.
Econometrica, 55, 251 – 276.
Engle, R. F., & Yoo, B. S. (1987). Forecasting and testing in cointegrated systems. Journal of Econometrics, 35,
143 – 159.
Evans, G. W. (1991). Pitfalls in testing for explosive bubbles in asset prices. American Economic Review, 81,
922 – 930.
Hall, A. (1994). Testing for a unit root in time series with pretest data-based model selection. Journal of Business
and Economics Statistics, 12, 461 – 470.
Hall, S. G., Psaradakis, Z., & Sola, M. (1999). Detecting periodically collapsing bubbles: a Markov-switching unit
root test. Journal of Applied Econometrics, 14, 143 – 154.
Johansen, S. (1988). Statistical analysis of cointegration vectors. Journal of Economic Dynamics and Control, 12,
231 – 254.
Johansen, S. (1991). Estimation and hypothesis testing of cointegration vectors in Gaussian vector autoregressive
models. Econometrica, 59, 1551 – 1580.
Kwiatkowski, D., Phillips, P. C. B., Schmidt, P., & Shin, Y. (1992). Testing the null hypothesis of stationarity
against the alternative of a unit root. Journal of Econometrics, 54, 159 – 178.
MacKinnon, J. G. (1991). Critical values for cointegration tests. In R. F. Engle, & C. W. J. Granger (Eds.), Long-
run economic relationships ( pp. 267 – 276). Oxford: Oxford University Press.
Osterwald-Lenum, M. (1992). A note with quantiles of the asymptotic distribution of the maximum likelihood
cointegration rank test statistics. Oxford Bulletin of Economics and Statistics, 54, 461 – 471.
Saltoglu, B. (1998). Speed of adjustment to the long-run equilibrium: an application with US stock price and
dividend data. Applied Financial Economics, 8, 367 – 375.
Scacciavillani, F. (1994). Long memory processes and chronic inflation. IMF Staff Papers, 41, 488 – 501.
Schwert, G. W. (1989). Tests for unit roots: a Monte Carlo investigation. Journal of Business and Economic
Statistics, 7, 147 – 159.
Sephton, P. S. (1995). Response surface estimates of the KPSS stationarity test. Economics Letters, 47, 255 – 261.
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. (1989). Stock market volatility. Cambridge, MA: MIT Press.
Shiller, R. J. (2000). Irrational exuberance. Princeton, NJ: Princeton University Press.
Taylor, M. P., & Peel, D. A. (1998). Periodically collapsing stock price bubbles: a robust tests. Economics Letters,
61, 221 – 228.
Timmermann, A. (1995). Cointegration tests of present value models with a time-varying discount factor. Journal
of Applied Econometrics, 10, 17 – 31.
van Norden, S. (1996). Regime switching as a test for exchange rate bubbles. Journal of Applied Econometrics,
11, 219 – 251.
West, K. D. (1987). A specification test for speculative bubbles. Quarterly Journal of Economics, 102, 553 – 580.