The Predictability of Stock Market Regime: Evidence from the Toronto Stock Exchange

Author(s): Simon van Norden and Huntley Schaller

Source: The Review of Economics and Statistics, Vol. 75, No. 3 (Aug., 1993), pp. 505-510
Published by: The MIT Press
Stable URL: http://www.jstor.org/stable/2109465
Accessed: 27-02-2018 08:46 UTC

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms

The MIT Press is collaborating with JSTOR to digitize, preserve and extend access to The
Review of Economics and Statistics

This content downloaded from 203.189.120.14 on Tue, 27 Feb 2018 08:46:12 UTC
All use subject to http://about.jstor.org/terms
 NOTES 505

THE PREDICTABILITY OF STOCK MARKET REGIME:

EVIDENCE FROM THE TORONTO STOCK EXCHANGE

Simon van Norden and Huntley Schaller*

Abstract-Are stock market crashes and rallies related to the influence of deviations from fundamentals by im-
deviations from the apparent fundamental share price? Using posing various parameter restrictions on the switching-
a switching-regression framework, we test whether apparent
regression model. In addition, we use the coefficient
deviations help to predict the regime from which the next
period's stock market return is drawn and the magnitude of estimates for the switching regression to calculate the
returns in that regime. We find that the probability of a ex ante and ex post probabilities of a market crash.
collapse rises before most actual crashes. Likelihood ratio The ex post probability of a crash is based on the
tests confirm that regime switches are influenced by apparent
current period return; the ex ante probability is based
deviations.
on the apparent deviation from fundamentals in the
previous period. We find that the ex post probability
I. Introduction shows dramatic spikes which correspond to actual
crashes. More surprisingly, we find that the ex ante
The past decade has seen several large fluctuations
probability of a crash typically rises before a crash,
in asset prices that are not easily explained by news
suggesting that deviations from fundamentals have
about market fundamentals. One possible explanation
some predictive ability for stock market regimes.
for these fluctuations is the existence of speculative
As Flood and Hodrick (1986), among others, have
bubbles in asset prices. In a bubble model, a rational
shown, there is generally a way in which tests for
risk-neutral investor will hold an asset that is overval-
bubbles can be reinterpreted in terms of the process
ued relative to its market fundamental so long as its
driving fundamentals.' While we provide a stochastic
expected rate of return equals the rate of return on a
bubble example as a motivation for how stock market
non-bubbly asset. If there is a chance that the bubble
regimes might arise and why both the regime and the
will collapse, then the expected return on a "bubbly"
returns conditional on the regime would be pre-
asset in the states where the bubble survives must be
dictable, we believe that our primary contribution is a
higher to compensate the investor for the possibility of
new type of data description. If the stylized facts that
collapse. Thus the existence of bubbles would not only
emerge in our empirical work are not idiosyncratic to
account for occasional asset price crashes but also
our data, then later studies may wish to offer and test
rapid run-ups in asset prices before a crash.
competing explanations.
The question we address is whether or not stock
market crashes and the booms that precede them are
II. A Stochastic Bubble Interpretation
related to apparent deviations from fundamentals as a
of Regime Switching
model of bubbles would predict. We begin by showing
how a simple bubble model leads to regime-switching
We begin by considering a simple asset-pricing model
behaviour in stock prices. In one regime, an apparent
where risk-neutral investors choose between holding a
deviation from fundamentals grows from one period to
risk-free asset that yields (1 + r) in period t and a
the next; in the second, the apparent deviation shrinks.
risky stock.2 For both assets to be held in equilibrium,
Linearizing this model gives us a system of three equa-
it must be true that
tions (one for returns in each regime and one for the
probability of the bubble collapsing) that we can esti-
Pt = (1 + r) 1 * (Et(Pt+1) + Dt) (1)
mate using standard switching-regression techniques.
We estimate this model on data for the Toronto where Pt and Dt are the stock's price and dividend at
Stock Exchange from 1956 to 1989. We then test for time t and Et denotes the expectation conditional on

1 Our test of the bubble model therefore tests a joint hy-

Received for publication June 18, 1991. Revision accepted
for publication March 4, 1992.
* Bank of Canada, Ottawa, and Carleton University, respec-
tively.
Dr. Schaller's research was assisted by the Financial Re-
search Foundation of Canada. The authors would like to
thank seminar participants at Concordia University, Hendrik
Houthakker and two anonymous referees for helpful com-
ments without implicating them in any errors. The views
expressed here are our own; no responsibility for them should
be attributed to the Bank of Canada.
pothesis; for example, see Cochrane (1991).
2 There is an extensive literature noting restrictions on the
admissibility of bubble solutions in this class of models, in-
cluding Diba and Grossman (1987, 1988) and Obstfeld and
Rogoff (1983, 1986). More recent work by Weil (1988) and
Allen and Gorton (1991) has shown that these restrictions are
not robust to minor changes in the model. As our purpose in
this section is to provide intuition for the presence of switch-
ing behaviour and not to contribute to the theoretical bubble
literature, we have chosen to use this simpler model as an
expository device.

Copyright C) 1993

This content downloaded from 203.189.120.14 on Tue, 27 Feb 2018 08:46:12 UTC
All use subject to http://about.jstor.org/terms
506 THE REVIEW OF ECONOMICS AND STATISTICS

information available at time t. One possible solution This ensure

to this equation defines the fundamental price that the bubble is expected to shrink.
Using these two generalizations together with (4)
Pt* - (1 + r) (kE) *E(Dt+k). (2) gives
k=O

(1+ r B-1 - q(b1) ub)P q(b,)

Et(Bt,)= q(bt) q(b,) withI robability (10)

u(bt) Pt 1 - q(bt).

All other prices are said to be "bubbly," with the size
of the bubble defined as
B,-P,P* (3)
Since all asset prices, bubbly or not, must satisfy (1),
this implies that the bubble must satisfy the condition
We can see that when q(bt) q and u(bt) 0, this
reduces to the Blanchard-Watson process.
It is straightforward to derive the expected excess
returns R in each regime, where excess returns are the
rate of return on the bubbly asset less the rate of
return on the riskless asset

Et(Bt+) = Bt * (1 + r). (4)

Et(Rt+,+S) = ( 1 [(1 + r)bt - u(bt)]
Blanchard (1979) and Blanchard and Watson (1982)
consider a particular stochastic solution to (4). They (11)
suppose there are two states of nature, one where the
Et(Rt+,+C) = u(bt) - (1 + r)bt. (12)
bubble survives (state S) and one where it collapses
Noting that conditional expected excess returns are a
(state C). If the probability of being in state S is some
function of bt, we can take first-order Taylor series
constant q and being in state C implies Bt = 0, then
(4) implies approximations of Et(Rt+1IS) and Et(Rt+1IC) with
respect to bt around some arbitrary value b to obtain:
(1 + r)
E1(Bt+11S) = B,. (5)
q E(Rt+1IS) = Yso + yslbt (13)
The intuition here is that (if B1 > 0) agents expect E(Rt+1IC) = Yco + yclbt (14)
capital losses of B, in state C, which must be balanced
by expected capital gains in state S in order to have where
the required rate of return on the bubble.
We now generalize the Blanchard-Watson model in
two ways. First, we wish to allow the probability q of
1- d q(b) [bt [( ) -()]
being in state S to decrease as the relative size of a
bubble (defined as bt - B,/Pt) grows, so that + I(+) dr- db)
q(b) Ldbt
dq(bt) < o. (15)

du(b) ( )(6
Second, we allow the expected value of the bubble Yci - d b +r)(6
conditioned on collapse to be non-zero, thereby allow-
ing for partial collapses. We assume the expected size
Assuming r ? 0, we can then prove tha
of a bubble in state C, which we define as ut * Pt,
depends on the relative size of the bubble in the yCi < 0.3
previous period, so By dropping the expectations operato
tions (13) and (14), we can rewrite them as

E1(Bt+ 1C) = u(bt) * Pt. (7)

Rs t = yso + ys, i bt + ES,t+i (17)
We further assume that u(*) is a continuous and every-
where differentiable function and that Rc t+l= yco + ycl * bt + EC,t+1. (18)

u(0) = 0 (8) 3 The proof for ylY follows directly from (9). For ylY, we
can use this condition and the fact that 1 2 q(b,) 2 0 to show

1 du(bt) 0.
that the second term in the expression is always non-negative.
(6), (8), and (9) then together imply that the first expression is
db (9) also non-negative, so their sum will be non-negative.

This content downloaded from 203.189.120.14 on Tue, 27 Feb 2018 08:46:12 UTC
All use subject to http://about.jstor.org/terms
 NOTES 507

To complete the switching-regression model, we need a
functional form for q(b,) that satisfies (6). We use the
the probability of survival is 'I(yqo) = 97.6%. Thus,
probit form
Pr(regime S) = (F(yqO + Yqj Ib,I) (19)
where F is the standard normal cumulative distribu-
tion function and (6) implies Yqi < 0.
The three equations (17), (18), and (19) form a
standard switching-regression model of the type de-
scribed by Goldfeld and Quandt (1976) and Hartley
(1978), and can be estimated by maximum likelihood
methods. Note that this estimation technique does not
of 4.9% per month (45.4% per annum). On average,
the estimates from our switching regression repro-
duce the stylized fact that returns are characterized by
large but infrequent crashes.
The ex ante probability of being in regime i at time t
is defined as the probability conditioning on bt_1,
which is the size of the apparent deviation in the
previous period. This probability is given by the for-
mula (1(i) (yqo + yql Ibt-11)) Pi' where 1(i) is 1
or -1, depending on the regime i. The ex post proba-
bility in regime i at time t also conditions on the
realized excess return R, and is given by

Rt i vi - V Yi, bt 1 vi
01, ~~~~~~ ~~= p x {or}
pA o. + Rt -yio -(yi, bt-1)) . _ 1 + (1 - p

require assumptions about which regime generated
which observation. Instead, it considers the probability
that either regime may have generated a given observa-
tion and gives an optimal classification of observations
into the underlying regimes, as we will see below.
where 4 is the standard normal probability density
function (pdf) and j # i.
Figure 1 plots the ex ante and ex post probabilities
of a collapse. The ex post probability of collapse is
marked by distinct spikes which generally seem to
coincide with actual market crashes. For example, we

III. Estimation of Switching Models see spikes coinciding with crashes in 1957, 1962, 1969,
1973, 1974, 1979, 1981, 1982 and 1987. What may be
The basic data used below to test for bubbles are more interesting is the behaviour of the ex ante proba-
monthly prices and dividends for the Toronto Stock bility of collapse. Periods when the ex ante probability
Exchange Composite (300, weighted) Index, not sea- of collapse is highest end with a spike in the ex post
sonally adjusted, from Bank of Canada Review. They probability, as can be seen in 1972, 1974-1975, 1982
cover the period from January 1956 to November 1989 and 1987. However, the converse is not always true. In
and are month-end figures. Excess returns on stocks some cases, such as 1969 or 1979-80, there is a spike in
are formed by subtracting the interest rate on 90-day the ex post probability of collapse without a corre-
prime corporate paper. Further details may be found sponding rise in the ex ante probability. Of course,
in van Norden and Schaller (1991). 1979-80 was a period when news arrived about oil
In a Lucas (1978) asset-pricing model where divi- shocks, so it would not be surprising if this was a
dends follow a random walk with drift, the fundamen- period in which news about fundamentals was respon-
tal price can be expressed as P,* = pDt, where p sible
is a for dramatic changes in stock prices.
function of tastes and the parameters of the stochastic With few exceptions, the ex ante probability of col-
process for dividends. Since P, - pD, is the deviation lapse in any given month is relatively small (less than
from fundamentals, we regress P, on D, (taking logs 20%).
toThis tends to obscure the fact that large appar-
reduce heteroscedasticity) and use the residuals as our ent deviations from fundamentals are associated with
measure of bt. large cumulative probabilities of collapse. As figure 2
Table I presents the estimated parameters of (17), shows, the cumulative ex ante probability of collapse
(18) and (19).4 Since bt is equal to zero on average
rose by
above 90% before the 1962, 1969, 1973, 1979 and
construction, Yso and yco give us the average be- 1987 collapses.5 Again there are several spikes in the
haviour of stock returns in each regime. We see that,
conditional on the bubble surviving, we expect an ex-
5In figure 2, we calculate the cumulative probability of
cess return of 0.55% per month (6.4% per annum).
collapse as
Conditional on its collapse, however, we expect losses
1- HI Pk
k = or

where r is the last period in which PTX < .75 and P,' and P,A
4Model diagnostics are presented and discussed in van are the ex post and ex ante probabilities of survival in peri-
Norden and Schaller (1991). od t.

This content downloaded from 203.189.120.14 on Tue, 27 Feb 2018 08:46:12 UTC
All use subject to http://about.jstor.org/terms
508 THE REVIEW OF ECONOMICS AND STATISTICS

FIGURE 1-.PROBABILITIES OF BUBBLE COLLAPSE

------ Ex Ante
- ExPost

1956 1958 1960 1962 1964 1966 1968 1970 1972 1974 1976 1978 1980 1982 1

FIGURE 2.-CUMULATIVE Ex ANTE PROBABILITIES OF COLLAPSE

Critical Value = .75

1.0

0.8

0.6

0.4

0.2

1956 1958 1960 1962 1964 1966 1968 1970 1972 1974 1976 1978 1980 1982 1984 1986 1988

ex post probability of collapse clustered around the a collapse in the three-month period ending in October
1979-80 oil shock which were not preceded by high 1987 was 67%.
cumulative probabilities of collapse.
To focus on an episode that may be of particular IV. Tests of the Switching Model
interest, we evaluate expected returns on the eve of the
October 1987 crash when bt = 0.437. For this value In ofthis section, we test how well the stochastic bub-
the apparent deviation from fundamentals, our esti- ble model corresponds to the data by imposing param-
mates predict a monthly excess return of -12% condi- eter restrictions on the switching regression that mimic
tional on a collapse6 and an ex ante probability of a a variety of stylized facts about market returns. For
collapse of 29%. The cumulative ex ante probability of example, by setting all the bubble coefficients equal to
zero and imposing equal constants in the two regimes,
6This compares to the actual monthly excess return weof can reproduce the stylized fact of high and low
volatility regimes described in Schwert (1989). For-
-23% for the period from the end of September to the end
of October 1987. mally, this would be the special case of the general

This content downloaded from 203.189.120.14 on Tue, 27 Feb 2018 08:46:12 UTC
All use subject to http://about.jstor.org/terms
 NOTES 509

TABLE I.-ESTIMATES OF THE GENERAL SWITCHING MODEL

AND NESTED NULL HYPOTHESES

General
Switching Volatility Normal Mean
Model Regimes Mixtures Reversion

Yso 0.0055 0.0041 0.0063 0.0043

(0.0024) (0.0021) (0.0025) (0.0021)
Ysi -0.0122 -0.0168
(0.0142) (0.0118)
Yco -0.0492 -0.0189
(0.0352) (0.0149)
Yci -0.1749
(0.0923)
YqO 1.9704 1.1144 1.0412 1.0298
(0.4500) (0.3683) (0.3288) (0.3802)
Yql -3.2317
(1.4547)
O'S 0.0381 0.0366 0.0360 0.0360
(0.0119) (0.0027) (0.0026) (0.0028)
'rc 0.0811 0.0832 0.0775 0.0795
(0.0266) (0.0149) (0.0115) (0.0136)

Log-
Likelihood 700.10 692.90 694.48 693.78
Likelihood Ratio
Test Against General
Switching Model 14.40 11.24 12.64
Distribution under Null: X2(4) X2(3) X2(3)
p-Value 0.0061 0.0115 0.0055

Note: Figures in parentheses are standard errors.

switching model where Yso = Yco = Yo, Ysi = Yci =
Yql = 0, SO
Rt+ 1 = YO + Et+ 1 (21)
where
Et+ 1 -N(O, os) with prob q
Et+ '-N(O,oac) withprobl-q.
Coefficient estimates for the volatility regimes case
and a likelihood ratio test of its restrictions are pre-
sented in column two of table 1. Since we reject the
restrictions, we conclude that the regimes differ in
more than just their variances.7 This implies either that
the information contained in the measure of deviations
from fundamentals helps to determine which regime
prevails or that the regimes have different expected
values, or both.
The second possibility we consider is that returns are
well characterized by a mixture of normal distributions
(as suggested by Akgiray and Booth (1987)), but are
unrelated to deviations of stock prices from fundamen-
ity by allowing the variances of returns to be drawn
7 Rejection of this null also implies a rejection of the single
from high and low volatility distributions. Coefficient
regime null since the latter is just the special case where
o-s = o-C. However, we cannot test directly whether we have
one regime or two since the parameters of our alternative
hypothesis are not identified under the null of only one
regime. See Lee and Chesher (1984) for details.
tals, which can be expressed as
R t+ N(yeSOo-S) with prob q
Rt + 1 N(yco, ac) with prob 1-q (23)
for some constant q. This is the special case of the
general model where ysi = ycl = yql = 0. As shown
in column three of table 1, we also reject these restric-
tions, which implies that apparent deviations from fun-
damentals help to determine expected returns in each
regime and/or which regime prevails.
A third possibility is that returns are predictable, but
mean returns do not differ across regimes. For exam-
ple, if we set yso = yco = Yo' Ysi = Ycc = yl, and
yql = 0, we obtain
Rt+1 =yo0 + y, * bt + Et+, (24)
where
Et+i N(O, os) with prob q 25
Et+i '"N(O, ac) with prob 1 - q. (2)
This corresponds to the regression test for mean rever-
sion in stock prices in Cutler, Poterba and Summers
(1991), except that we allow more flexibility for volatil-
estimates for this mean reversion null and a test of its
restrictions are presented in column four of table 1.
Rejection of this null implies that there is more in the
data than simple mean reversion. In particular, yql is

This content downloaded from 203.189.120.14 on Tue, 27 Feb 2018 08:46:12 UTC
All use subject to http://about.jstor.org/terms
510 THE REVIEW OF ECONOMICS AND STATISTICS
significantly different from 0, which implies that a large
deviation from fundamentals makes a collapse more
likely.
We conduct two tests for the robustness of our
results. First, maximum likelihood inferences may be
sensitive to our assumption of an i.i.d. normal distribu-
tion of the E's. We recalculate the significance levels by
randomly shuffling b, (thereby destroying any linkBlanchard,
be-
tween bt and R,+1) and simulating the distribution of
the coefficient estimates and likelihood ratio tests un-
der the null hypothesis that the two series are unre-
lated.8 The resulting p-values are 0.1923, 0.0491 and
0.0137 for ysl, ycl and yql, and 0.0128, 0.0128 and
(Lexington,
0.0157 for the three likelihood ratio tests. Second, we
Campbell, J. Y., and R. J. Shiller, "The Dividend-Price Ratio
use an alternative measure of bt, namely, the differ- and Expectations of Future Dividends and Discount
ence between a VAR projection of the dividend-price Factors," Review of Financial Studies 1 (Fall 1988),
ratio and the actual dividend-price ratio.9 This alter-
native measure yields similar coefficient estimates and
stronger rejections of the three null hypotheses: the
marginal significance levels of the resulting likelihood
ratio tests are 0.0082, 0.0175 and 0.0069, respectively.
V. Conclusion
We have found evidence of regime switches in stock
market returns that are influenced by apparent devia-
tions from fundamentals. Our estimates are consistent
with the stylized fact of large but infrequent crashes.
The size of the apparent deviations from fundamentals
has an important influence on both the probability and
expected magnitude of a collapse. For example, using
the apparent deviation just before the October 1987
crash, our estimates suggest a cumulative probability of
collapse of 67% for the three-month period ending in
October and a one-month loss conditional on collapse
of 12%. High ex ante collapse probabilities tend to be
followed by actual crashes. The exceptions seem to be
associated with the major oil price shock of 1979-1980,
a period when news about fundamentals may have
strongly influenced prices.
We believe there are two important directions for
further research. The first is to determine whether our
stylized fact-that apparent deviations from funda-
mental price influence both the regime that prevails in
the stock market and -the size of expected returns-
generalized to other countries, time periods and asset
markets. The second is to see whether a plausible
model of equilibrium asset pricing could explain our
results in terms of switching fundamentals. Whatever
the ultimate interpretation, our paper highlights an
aspect of the behaviour of stock market returns that
has not previously received attention.
8 Details are available from the authors; for another eco-
nomic application of this type of randomization procedure,
see Kim, Nelson, and Startz (1991).
9 See Campbell and Shiller (1988). In their notation, our
bt = SI - 16
This content downloaded from 203.189.120.14 on Tue, 27 Feb 2018 08:46:12 UTC
All use subject to http://about.jstor.org/terms
REFERENCES
Akgiray, Vedat, and G. Geoffrey Booth, "Compound Distri-
bution Models of Stock Returns: An Empirical Com-
parison," Journal of Financial Research 10 (1987),
269-280.
Allen, Franklin, and Gary Gorton, Rational Finite Bubbles,
National Bureau of Economic Research Working Pa-
per No. 3707, May 1991.
Olivier Jean, "Speculative Bubbles, Crashes and
Rational Expectations," Economics Letters 3 (1979),
387-389.
Blanchard, Olivier, and Mark Watson, "Bubbles, Rational
Expectations and Financial Markets," in P. Wachtel
(ed.), Crises in the Economic and Financial Structure
MA: Lexington Books, 1982).
195-228.
Cochrane, J. H., "Volatility Tests and Efficient Markets,"
Journal of Monetary Economics 27 (1991), 463-485.
Cutler, David M., James M. Poterba, and Lawrence H. Sum-
mers, "Speculative Dynamics," Review of Economic
Studies 58 (1991), 529-546.
Diba, Behzad T., and Herschel I. Grossman, "On the Incep-
tion of Rational Bubbles," Quarterly Journal of Eco-
nomics 102 (Aug. 1987), 697-700.
, "Rational Inflationary Bubbles," Journal of Monetary
Economics 21 (May 1988), 35-46.
Flood, Robert P., and Robert J. Hodrick, "Asset Price
Volatility, Bubbles and Process Switching," Journal of
Finance 41 (1986), 831-842.
Goldfeld, Stephen M., and Richard E. Quandt, Studies in
Nonlinear Estimation (Cambridge, MA: Ballinger,
1976).
Hartley, Michael J., "Comment on 'Estimating Mixtures of
Normal Distributions and Switching Regressions' by
Quandt and Ramsey," Journal of the American Statisti-
cal Association 73 (Dec. 1978), 738-741.
Kim, Myung Jig, Charles R. Nelson and Richard Startz,
"Mean Reversion in Stock Prices? A Reappraisal of
the Empirical Evidence," Review of Economic Studies
58 (1991), 515-528.
Lee, Lung-Fei, and Andrew Chesher, "Specification Testing
when Score Test Statistics Are Identically Zero," Jour-
nal of Econometrics 31 (1984), 121-149.
Lucas, Robert E. Jr., "Asset Prices in an Exchange Economy,"
Econometrica 66 (Nov. 1978), 429-445.
Obstfeld, Maurice, and Kenneth Rogoff, "Speculative Hyper-
inflations in Maximizing Models: Can We Rule Them
Out?" Journal of Political Economy 91 (Aug. 1983),
675-687.
Obstfeld, Maurice, and Kenneth Rogoff, "Ruling Out Diver-
gent Speculative Bubbles," Journal of Monetary Eco-
nomics 17 (1986), 349-362.
Schwert, G. William, "Why Does Stock Market Volatility
Change Over Time?" Journal of Finance 44 (1989),
1115-1153.
van Norden, Simon, and Huntley Schaller "The Predictability
of Stock Market Regime: Evidence from the Toronto
Stock Exchange," Bank of Canada Working Paper No.
91-6, Oct. 1991.
Weil, Philippe, "On the Possibility of Price Decreasing Bub-
bles" Harvard Institute of Economic Research Discus-
sion Paper No. 1399, 1988.