Price Barriers in the Dow Jones Industrial Average

Author(s): R. Glen Donaldson and Harold Y. Kim

Reviewed work(s):
Source: The Journal of Financial and Quantitative Analysis, Vol. 28, No. 3 (Sep., 1993), pp.
313-330
Published by: University of Washington School of Business Administration
Stable URL: http://www.jstor.org/stable/2331416 .
Accessed: 18/01/2012 12:33
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
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 [email protected].

University of Washington School of Business Administration is collaborating with JSTOR to digitize, preserve
and extend access to The Journal of Financial and Quantitative Analysis.

http://www.jstor.org

ANALYSIS
OF FINANCIAL
ANDQUANTITATIVE
JOURNAL

Price

Barriers

R. Glen

Donaldson

in the
and

Dow
Harold

Jones

1993
VOL 28, NO 3, SEPTEMBER

Industrial

Average

Y. Kim*

Abstract
This study tests the popular claim that the DJIA's movements around key reference points
affect "investor sentiment" and thus price behavior. It is found that the DJIA's rise and fall
is indeed restrained by "support" and "resistance" levels at multiples of 100 (e.g., 2800,
2900, 3000, etc.) but that, having broken through a 100-level, the DJIA then moves by
more than otherwise warranted. A Monte Carlo study and comparisons with other indices
confirm the significance of these findings. This suggests that some agents may trade on the
basis of the DJIA but does not necessarily suggest that the market is inefficient.

I.

Introduction

The commentaries of financial market analysts frequently imply that certain

values of widely followed asset prices and aggregate market indices hold special
significance for market participants. Perhaps the best example of this can be found
in the unusual amount of attention that the financial press pays to the behavior of
fails
the Dow Jones Industrial Average (DJIA) as it approaches and crosses?or
to cross?multiples
of 100 and especially 1000. The DJIA's reluctance to cross
"the 3000 barrier" during the latter part of 1990 and early 1991, for example,
was seen by some financial commentators as a sign of weakness, with one Wall
Street Journal report claiming that"... the inability ofthe Dow Jones industrials to
push above the psychologically
important 3000 level was a key sign that the stock
market was faltering" (Wall Street Journal, January 2, 1991, p. R3). Conversely,
the DJIA's eventual breaking of the 3000 barrier was seen by some traders as a
sign of strength. The large price increase that occurred as the DJIA rose above
the 3000 mark on June 1, 1991, for example, was explained by one analyst as
follows: "Once we crossed the 3000 mark there were a number of investors who
were finally convinced of the bullish status of this market and decided to jump on
the bandwagon" (Toronto Globe andMail, June 2, 1991, p. B10).
*
Faculty of Commerceand Business Administration,Universityof BritishColumbia,Vancouver,
BC, Canada,V6T 1Z2, and Departmentof Economics, PrincetonUniversity,Princeton,NJ 08544,
respectively. This paper is a substantialrevision of Donaldson (1990). The comments of Giuseppe
Bertola,JohnCampbell,LudgerHentschel,Alan Kraus,EduardoLey, Rick Mishkin,RichardQuandt,
Ken Stewart,Hal Varian,JFQA RefereeWayneJoerding,and seminarparticipantsat the Universityof
BritishColumbia,PrincetonUniversity,the NBER SummerInstitute,andthe 1991 CanadianEconomic
Associationmeetingsarethankfullyacknowledged,as is the financialsupportof PrincetonUniversity's
John M. Olin ResearchProgram.
313

314

Journal of Financial and Quantitative Analysis

Since one would preserve the DJIA's relevant information on relative price
movements with a rescaling of the index, and would lose in the rescaling process
only information regarding the particular numerical values assumed by the index,
it is unclear why some values of the DJIA should be perceived by the market to be
of more importance than other values. The preceding quotations, however, suggest
that some traders nevertheless believe in the special importance of key reference
points in the DJIA. In particular, the preceding assertion that traders may "jump
on the bandwagon" of buying (selling) once the DJIA breaks up (down) through
a "psychologically
important" level suggests that the crossing of one of these
barriers may push the index up (down) more than otherwise
sentiment-induced
warranted. Frequently used phrases such as "support level" and "resistance level"
also imply that, until such time as the psychologically
important barrier is broken,
increases and decreases in the DJIA may be restrained so as to prevent crossing of
the barrier level. A report on this subject in the Wall Street Journal, for example,
states that psychologically
important "... stock index or average levels become
As market barometers approach those levels, stock buyers
sentiment signals.
become less aggressive, fearing a turn in the market, while sellers need less coaxing
to drop their price a notch or two" (Wall Street Journal, April 15, 1991).1
Given the preceding discussion, the goal in this paper is to test the null hy?
pothesis that the behavior ofthe DJIA is not influenced by its proximity to special
levels of the index against the alternative hypothesis that psychologically
impor?
"barriers" that inhibit the rise or
tant levels in the DJIA form sentiment-induced
fall of the index past a barrier support or resistance level but that, having broken
through a barrier, a breakout or bandwagon effect causes the index to move up or
down by more than usual. The results of this study's tests reject the no-barriers
null in favor of the barriers alternative. In particular, it is found that movements
in the DJIA are indeed restrained by barrier support and resistance effects at 100levels in the index (e.g., 2000, 2100, 2200, etc.) but that, having broken through a
100-level, the index rises or falls by more than usual. This anomalous behavior is
not present in the less popular Wilshire Associates 5000 index, which was used as
the main control for the DJIA, nor is it present in any of the computer-generated
indices that were examined. The results are shown to be robust with respect to the
time period investigated and the statistical significance of the results is confirmed
with a Monte Carlo study based on randomized returns and AR(1)-GARCH(1,1)
simulations. The results suggest that some less-than-fully-rational
agents may be
basing their buying/selling decisions on the value ofthe widely reported DJIA, but
do not necessarily suggest that the market is inefficient.

II.

Data

As the most widely reported of all stock market indices, the DJIA is the
obvious candidate for testing the presence of psychological
pricing barriers. As
the primary benchmark against which to compare the DJIA's behavior, the Wilshire
Associates 5000 (WA) index is also examined. The WA is a more obscure index
than the DJIA and therefore a less likely candidate for barriers. The WA was
1The authorsthankHal Varianfor
bringingthis quotationto theirattention.

Donaldson

and Kim

315

selected over other possible alternatives (e.g., Standard & Poor's 500, Russell
2000, Value Line, etc.) because only the WA has values of comparable magnitude
to the DJIA and has therefore cycled through enough hundred levels to justify a
direct comparison using the frequency distribution methodology described below.
Both the DJIA and WA data are taken from the Dow Jones Tradeline Database. The
limited availability of the WA determines the sample period: October 14, 1974,
To test robustness, various
to May 18, 1990 (the day the data were collected).
subsamples were also studied and DJIA data from as far back as January 1, 1952,
were examined.2
The DJIA begins with a price of 673.5 on October 14, 1974, and ends with
a price of 2819.9 on the last day of the sample. The unconditional mean and
variance of the daily return (log of price relatives) are 0.036 percent and 0.012
percent, respectively. The WA begins at 3581.0 and rises to 5394.0 on August 10,
1981. The index was then rescaled (by Wilshire Associates) to begin at 1399.0 on
August 11, 1981, and end with a value of 3383.7. The WA price series was not
adjusted to remove this discontinuity since the tests are based on the realized price
level of the index. (As shown by the Monte Carlo study below, the differences
found between the WA's behavior and the behavior of the DJIA are not a result of
the break in the WA series.) Excluding the discontinuity day, the unconditional
mean and variance of the WA's daily return are 0.033 percent and 0.007 percent,
respectively. For the one test below that uses returns in addition to the price level,
the WA's daily return for August 11, 1981, was set equal to the unconditional mean
return over the rest of the sample.

III.

Testing

the

Barriers

Hypothesis

In this section, the barriers hypothesis is formalized and three statistical tests
for the presence of pricing barriers in stock indices are presented. Evidence of
barriers at 100-levels in the indices is specifically looked for?e.g.,
2800, 2900,
these are the price levels to which special significance is most
3000, etc.?since
frequently attached by financial commentators. (Tests for a separate barriers effect
at 1000-levels may also be of interest. However, the limited price range of the DJIA
precludes use of the frequency distribution tests employed in this paper.) Results
produced suggest that there may indeed be barriers at 100-levels in the widely
followed DJIA but not in the relatively obscure WA. The strength and robustness
of the results are confirmed in Section IV.
A.

Implications

of Barriers

in the DJIA

Before barriers in the data can be tested for, the hypotheses to be tested
must first be formalized. Unfortunately, there is no abstract economic theory that
says how stock price indices should behave in the presence of sentiment-induced
barriers at "psychologically
important" reference points. There is, however, a wellestablished theoretical literature on the closely related problem of price behavior
in the presence of exogenously enforced barriers to price movements in other types
2The authorsthankEduardoLey for providingthe data for 1952-1974, which he collected from
Pierce(1986).

316

Journal of Financial and Quantitative Analysis

of assets, such as central bank-enforced target zones for currency exchange rates.
Along with the implications of the preceding quotations from financial market
analysts, results from this closely related theoretical literature help in formalizing
the barriers hypothesis for stock indices.
Central banks frequently announce "target zones" or trading ranges for the
foreign exchange value of their currency. The central banks state that they will
enforce these zones by intervening in the market so as to keep the price of their
currency from exceeding the upper barrier of the zone or from falling through
the lower barrier of the zone. If the banks really did enforce these zones, then
the price barriers would be "imperforate" since they would confine the asset price
forever. More often than not, however, the exchange rate is eventually allowed to
break through the barriers. When this happens and the barriers are perforated, a
"realignment" typically occurs during which the asset price moves an unusually
large amount and then a new set of barriers is established that contains the new
price level. An interesting problem for international economists has been the
modeling of exchange rate behavior as the currency's price fluctuates within, and
occasionally breaks out of, its barriers. Even though these exchange rate barriers
exist because of outside intervention, while what barriers (if any) that exist in the
DJIA are presumably the product of investors' sentiment-induced
trading behavior,
results from this literature are nevertheless interesting for this research because they
help in understanding the potential dynamic behavior of the DJIA as it fluctuates
within and breaks out of its supposed barriers.3
Krugman's (1987), (1991) initial investigation of exchange rate barriers, and
subsequent work by Froot and Obstfeld (1989) and Flood and Garber (1991),
assume that the barriers are imperforate and thus that the asset price can never
take on values outside of the target zone. Given the additional assumption that
there is no anticipated excess profit available from holding the asset as its price
fluctuates around various points within its imperforate barriers, Krugman proves
that the amount of time the price spends close to the barriers must be greater than
the amount of time it spends away from the edges of the barrier zone. Thus, the
ergodic distribution of price realizations within an imperforate barrier zone should
be U-shaped, with more frequent price realizations close to the edges of the zone
and less frequent realizations in the center of the zone. In other words, Krugman
proves that an asset price confined by imperforate barriers should, in general, hover
just above its lower barrier (i.e., support level) or just below its upper barrier (i.e.,
resistance level). In terms of barriers in stock indices, this result implies that if the
DJIA were truly constrained to never pass through a 100-level barrier (e.g., if the
index were constrained to always assume values between 00 and 99), then DJIA
3Since the barrier(or target) zone model is both well known and somewhat involved, the full
formalmodel is not replacedhere. The basic natureofthe model, however,is easily summarized.The
model is based on the assumptionthat there is an asset whose price is influencedby two factors: a
fundamental(usuallya Brownianmotion with drift)and a termthatcapturesthe market'sexpectations
aboutfutureprice changes in the presenceof barriers.To derivethe equilibriumbehaviorof the price,
one begins by using Ito's Lemmato solve for the expected rate of change in the price as a functionof
the fundamental.The resultis then plugged back into the pricingequationto producethe second order
differentialequationthat governs the behaviorof the asset's price as it fluctuateswithin, and breaks
out of, the barrierzone. The solution to this equation, and thus price behavior,depends on several
integrationconstants,the values of which dependon whetherthe barrierzone thatconstrainsthe asset's
pricemovementsis perforateor imperforate.Foran excellent review,see BertolaandCaballero(1992).

Donaldson

and Kim

317

prices in the single digits and nineties would be observed more often than prices in
the middle range between the barriers. Over this study's sample period, however,
the DJIA has assumed values from below 600 to over 2900 and has therefore
obviously passed through, or perforated, several 100-levels. Thus, the theoretical
results for imperforate barrier zones should probably not apply to the behavior of
the DJIA around 100-levels.
More recently, Bertola and Caballero (1992) and Pesenti (1990) consider the
converse case of a perforate barrier zone in which a support level or resistance level
can be crossed, though perhaps with some difficulty. Once a barrier is breached in
this environment, a "breakout" or realignment occurs during which the price ofthe
asset moves an unusually large amount and then a new barrier zone is established
that contains the new price level.
This case appears closer to that of pricing
barriers in stock indices. Bertola-Caballero
prove that, given the sudden change
in prices associated with the barrier-crossing breakout effect (in their case, an
exchange rate realignment), the absence of unexploited excess profit opportunities
implies that the amount of time the price spends close to the barriers must be less
than the amount of time it spends away from the edges of the zone. Thus, the
ergodic distribution of price realizations within a perforate barrier zone should
be hump-shaped, with less frequent price realizations close to the edges of the
zone and more frequent realizations in the center of the zone. Empirically, this
hump-shaped distribution is frequently observed in exchange rate data. In terms
of barriers in the DJIA, this result implies that, if there are in fact perforate barriers
at 100-levels in the DJIA, then the DJIA should exhibit prices near these 100-level
barriers less frequently than it exhibits prices around other levels in the index and,
indeed, this is precisely what the DJIA data reveal below.
The intuition behind this result can be seen in the following example. Suppose
that the DJIA is currently at 2980 and that fundamentals will call for a 19-point
increase with probability 0.5 and a 21-point increase with probability 0.5. Further,
suppose that if the 21 fundamental is observed then, as suggested by the preceding
quotations from market analysts, the market's excitement at crossing the 3000
barrier will produce a "breakout effect" during which excited investors "jump on
the bandwagon" and produce a 30-point increase in the index.4 Given the breakout
effect, for there to be no expected excess profit from holding the asset at the price
2980 (i.e., for the expected increase to still be 20 points) it must be true that, if the
19-point fundamental is observed and the barrier is not broken, then (ignoring the
small rounding error) the market will produce only a 10-point rise in the index.
This restraining ofthe fundamental increase might be called the "resistance effect,"
as explained by one ofthe preceding quotations in which it was stated: "How can
a 'resistance level' exist? Because traders believe it is there_Stock
index or
levels
become
As
sentiment
market
barometers
average
signals.
approach those
stock
become
a
in
less
turn
the
levels,
market, while
buyers
aggressive, fearing
4Recallthequotationfromthe introduction,for example,in whichthe 27-pointincreasethatbrought
the DJIA over the 3000 mark on June 1, 1991, was explained by one analyst as follows: "Once we
crossed the 3000 mark there were a numberof investors who were finally convinced of the bullish
statusof this marketand decided to jump on the bandwagon".{TorontoGlobe and Mail, June2, 1991,
p. B10.)

318

Journal of Financial and Quantitative Analysis

sellers need less coaxing to drop their price a notch or two" (Wall Street Journal,
April 15, 1991).
This type of behavior, the rationality of which is discussed below, has two
First, note that the final price in the preceding example
important implications.
with barriers is either 3020 or 2990, instead of the no-barriers fundamental prices
3001 or 2999. Thus, as suggested by the preceding quotations and the formal
an index with perforate barriers at 100theoretical results of Bertola-Caballero,
levels will exhibit prices less frequently around 100-levels and more frequently
away from 100-levels than would an index without pricing barriers at 100-levels.
Second, since a "breakout effect" occurs when a barrier is crossed, pushing the
price to 3020 instead of 3001 in the preceding example, the percentage rate of
the DJIA as it rises up through (fails down through) a
return?on
change?i.e.,
100-level barrier should be greater than (less than) the return conditional on the
DJIA rising up through (falling down through) other points in the index. These
two implications of barriers in stock indices are tested formally below.
B.

Price Level Tests

Let Pt be the value of the stock price index on date t and Mt be the last two
: 100]). For
digits in the integer portion of the price (i.e., Mt = modulo[integer(Pf)
=
=
in
the
index (e.g.,
at
100-levels
Barriers
if
then
34.
1234.56,
Mt
Pt
example,
=
The
atM
00.
a
barrier
thus
become
2000,2100,2200,
etc.)
preceding discussion
of the formal theory and the discussion of practitioner quotes suggest that, if there
are barriers at M = 00 in the DJIA, then the DJIA should close on prices with
M-values in the neighborhood around M = 00 less frequently than it closes on
prices with M-values in the middle range. Of course, in the absence of barriers
when no level of the index is more important than any other, one would expect no
strong pattern in the distribution of closing prices across M-values (evidence in
support of this claim is provided by a Monte Carlo study presented below).
The series of vertical lines in Figure 1 plot the frequency with which the DJIA
closes with its price in each of the 100 M-cells. To provide the eye with a better
indication of the pattern that emerges, this figure also plots as the series of stars
In
joined by the dark line, the five-cell centered moving average of frequencies.
similar fashion, Figure 2 plots the individual frequencies for each of the 100 Mcells and the five-cell centered moving average for the WA index. As expected, the
WA distribution exhibits no strong pattern. However, consistent with the barriers
hypothesis, the DJIA appears to close, on average, less frequently around M = 00
and more frequently away from 00.5
5Another
possible way to observe the barriereffect is to calculate the M-value transitionmatrix
from day t ? 1 to day t; i.e., the probabilityof the occurrenceof Mt given Mt-\. The breakout
effect should produce lower frequencies for M-cells just above 00 for a rising index and just below
00 for a falling index. This matrix was computed and some direct evidence of the breakouteffect
was found. Given Mt-\ = 98-99, for example, the DJIA closes with Mt = 00-01 or 02-03 with
only 5-percent probability,but closes with Mt - 04-05 with 11-percentfrequency. This result is
consistent with the breakoutstory: having broken from 98-99 up throughthe 00 barrier,the index
shoots through the lower M-values to close in M-cells 04-05. There is also some evidence of a
similareffect for downwardmoves. For example, prob(M?= 98-99 \Mt-i - 02-03) = 4%, whereas
prob(Mf= 96-97 |Mf_i = 02-03) = 10%. Although these results generally supportthe barriersstory,
the transitionmatrixis omittedfrom this paperbecause small sample problemspreventedthe drawing

Donaldson

and Kim

319

FIGURE 1
Frequency of DJIA Prices
%

Frequency

0.8

0.6

0.4
0

10

20

M-Value

30
(last

40

60

50

two

Actual Frequency

digits

70
in the

80

90

price)

SCell Moving Average

To test the apparent systematic deviation from uniformity in the DJIA dis?
tribution, f(M) is defined to be the frequency with which an index closes with its
last two digits in cell M, minus 1 percent. A first price level test then involves
for each of the 100 M-cells on a constant and a dummy variable
regressing/(M)
that isolates groups of cells in the neighborhood around M = 00. The regression
is
/(Af)

OL+ pDij + UM\

M = 00,01,...,

99,

of statisticallymeaningfulconclusions. Note howeverthat,since the unconditionalfrequencyis simply

the sum down each column?or acrosseach row?of the conditionaltransitionmatrix,Figure 1, which
plots the unconditionalM-value frequency,completely summarizesthe informationcontained in the
DJIA'sconditional transitionmatrix. This informationsuggests what is observed in Figure 1: less
density aroundM = 00 in the DJIA.

320

Journal of Financial and Quantitative Analysis

FIGURE2
Frequency of WA Prices
Frequency

0.4
0

10

20

M-Value

30
(last

40
two

60

50
digits

70
in the

80

90

price)

5Cell Moving Average

Actual Frequency

where Dy is a dummy variable that isolates cells in the range from / to j, and Um
is a random error. The results of these regressions appear in Table 1, where the
dummies are
?>98-02 = 1 if M > 98

or

M < 02, = 0 otherwise;

D95_04 = l ifM>95

or

M < 04, = 0 otherwise;

D90-09 = 1 if M > 90

or

M < 09, = 0 otherwise.

The preceding discussion suggests that under the no-barriers null hypothesis (3
should be zero, while under the barriers alternative hypothesis (3 should be negative
(i.e., less density around the barrier).
The negative coefficients on the DJIA dummy variables in Table 1 reject the
no-barriers null hypothesis and confirm this study's earlier casual observation that
the DJIA closes less frequently on index values whose last two digits are in the

Donaldson

and Kim

321

TABLE1
Price Level Tests for Density around 00
M=00,...,99
f(M) = a + 0D,j+UM\

f(M): the frequency with which the index closes with its last two digits in cell M, minus 1
percent.
Djj. a 1/0 dummy variable that isolates cells in the range from /to j.
All entries x100 (standard deviations in parentheses).
Test Hq (no barriers): f3 = 0 vs H^ (barriers are present): f3 < 0.
Note the significant negative /3s for DJIA but not for WA (for example, D98_02 reveals that
DJIA closes 0.2353% - 0.0118% = 0 2235% less than expected around the 00-level).
Also note that, as hypothesized, the DJIA barrier effect weakens further away from 00.
neighborhood around the 00. The coefficient on D98_o2> f?r example, implies
that the DJIA on average closes 0.2353% - 0.0118% = 0.2235% less frequently
than expected in each of the five cells around M = 00. This result is confirmed
in the data directly with the finding that the percentage deviations from 1 percent
in the five M-cells around 00 are: 98 = -0.31,
00 = -0.29,
01 =
99 = -0.24,
Standard f-statistics for rejection of the no-barrier null for
?0.26, 02 = ?0.01.
the DJIA index range from greater than 99 percent for D98-02 to over 95 percent
for ?>95_o4 and D90-09. As implied by the barriers hypothesis, rejection of the
no-barriers null becomes weaker as one examines wider ranges around the barrier.
Just as important as the DJIA's rejection of the no-barriers null is the WA's failure
These results
to reject; all f-statistics from the WA regressions are insignificant.
therefore confirm Figure l's and 2's casual observations of a possible barrier at
M = 00 in the widely-followed
DJIA index, but no barrier in the less popular WA
index.
A second price level test examines not just the tails of the frequency distribu?
tion immediately around the barriers, but the entire shape of the distribution. Here,
not only does one need to specify the null hypothesis that the distribution should be
uniform in the absence of barriers, but the alternative that the distribution should
have some particular shape if barriers are present. The work of Bertola and Caballero (1992) and others suggests that a hump-shape is an appropriate alternative.
One can examine this possibility by running the regression,
f(M)

a + (3M + 8M2 + UM',

M = 00,01,...

,99.

322

Journal of Financial and Quantitative Analysis

Under the null ofno barriers 8 should be zero, while under the barriers alternative
8 will be negative. The results of such a procedure are reported in Table 2. Again,
it appears that the DJIA possesses a barrier while the WA does not.6
TABLE2
f(M)
Index

Price Level Tests for Hump-Shape

= a + PM+6M+ UM; M=00,...,99

R2

DJIA

0.0070
-0.00008
9.80
-0.0931
(0.0023)
(0.00002)
(0.0498)
WA
-0 0058
0.00005
1.00
0.1189
(0.0034)
(0.00004)
(0.0737)
f(M): the frequency with which the index closes with its last two digits in cell M, minus 1
percent.
All entries x100 (standard deviations in parentheses).
Test H0 (no barriers): 6 = 0 vs. H<\(barriers are present): 6 < 0.
Note the significant negative 6 for DJIA but not for WA.

C.

Conditional

Returns

Tests

As a third test for barriers, each index's behavior is studied as it progresses through
various M-cells from one closing price to the next. To conduct this test, one must
the index "return" for date t. Second, the
first calculate Rt = Ln(Pt) ? Ln(Pt-\):
value of Rt is assigned to each of the M-cells implicitly passed by the index on
date t. Thus, if Pt-i = 1492 and Pt = 1497, then the return Rt = 0.0033 would be
assigned to cells M = 93, 94, 95, 96, and 97 since these are the cells through which
the index passes as the price rises from 1492 to 1497. This procedure is repeated for
99)
every day in the sample. Finally, for each of the 100 M-cells (M = 00,01,...,
the average (i.e., mean value) of all the returns that were assigned to that cell is
calculated. This average is defined as RM: the average daily return conditional on
99). The behavior of RM across
having passed through cell M (M = 00,01,...,
the various M-cells forms the basis for the conditional returns tests.
Appealing to the barriers story, it is seen that the existence of a barrier at 100levels in an index implies a negative correlation between RM and M. This occurs
for four reasons. First, if a "breakout effect" really does occur once the DJIA
crosses a barrier, then the buying pressure associated with traders "jumping on the
bandwagon" of optimism once the DJIA rises up through a 00-level will push the
index well past the 00-level. This will result in less frequent closings of the DJIA
just above the 00-level, as observed in the previous section, and will also result
in larger-than-normal positive returns as the index rises up through low M-values
6If one had a theory aboutwhere the peak of the barriersdistributionshould occur, one could also
test the overidentifyingrestrictionthat ?/3/28 = X, where X is the peak value. Since there is no
trulyformaltheorythat says how investorswho believe in psychological price barriersshould behave,
however, it is difficult to state with certaintywhere the peak should occur, other than to say that it
should be away from M = 00 (the authorsthankthe referee for bringingthis point to their attention).
Also note thatfor a single peak to exist it must be truethatthereare no minorbarriersat intermediate
levels, such as 50s (no significantevidence of such a minorbarrierin the DJIA is found).

Donaldson

and Kim

323

such as 01,02,03,
etc. Second, if falling down through a barrier is eonsidered bad
news by the market, then the extra selling pressure associated with a downward
breakout would push the DJIA down by more than otherwise warranted once a 00
barrier is crossed. This will result in less frequent closings ofthe DJIA just below
the 00-level, as observed in the previous section, and will also result in larger-thannormal negative returns as the index falls down through high M- values such as 99,
98, 97, etc. Third if, as suggested above, the barrier restrains movements past a
00 resistance level, then movements up toward high M-values would be restrained
to be smaller than normal. Similarly, declines toward low M-values would be
restrained to be smaller than normal by the barrier support level. Thus, with a
barrier at M = 00, low M-cells will be filled with larger than normal increases and
smaller than normal decreases, while high M-cells will be filled with smaller than
normal increases and larger than normal decreases. The existence of a barrier at
100-levels (i.e., M = 00) in the index thus implies a negative correlation between
Rm and M. Of course, in the absence of barriers, one would expect no significant
correlation between RM and M (evidence in support of this claim is provided by a
Monte Carlo study presented below).
Table 3 presents the results from running the regression,
Rk

a + (3M + UM;

M = 00,01,...,99,

where Um is a random error. From the preceding paragraph, it is seen that under the
no-barriers null hypothesis (3 should be zero, while under the barriers alternative
[3 should be negative. For the DJIA, the /-statistic of -10.805 on (3 strongly rejects
the no-barriers null in favor of the barriers alternative, while for the WA, the null
cannot be rejected. These results are interpreted as support for the existence of
barriers at 100-levels in the popularly tracked DJIA, but not in the more obscure
WA index.
TABLE3
RM

Conditional Returns Tests

= a + /3M+UM;
M = 00,.

Index

,99
R*

DJIA

0.1899
-0.0010
(0.0055)
(0 0001)
WA
0.0654
0.0002
(0.0029)
(0.0001)
RM: the average daily return conditional on having passed through cell M(M= 00,
All entries x100 (standard deviations in parentheses).
Test H0 (no barriers). (3 = 0 vs. H-j (barriers are present): j3 < 0.
Note the significant negative (3 for DJIA but not for WA.

IV.

Evidence

from

Additional

Tests

and

53.90
7.21
,99).

Indices

The tests of the preceding section indicate the existence of barriers at 100levels in the widely followed DJIA index. The goal in this section is to examine
the strength of the results.

324
A.

Journal of Financial and Quantitative Analysis

Simulation

Results

The potentially nonstandard nature of the distributions involved in Section

III's tests leads to the conducting of Monte Carlo simulations to determine the
true statistical significance of the DJIA results. Two types of simulations are run:
one based on randomized actual returns (a quasi-bootstrap method) and another
and
based on modeling returns as an AR-GARCH process. The methodologies
below.
are
the calculated p-values
presented
For the return-randomization "bootstrap" simulation, the actual returns of the
DJIA were randomly reordered and then a new sequence of prices was produced
using the true starting value of the DJIA and the subsequent price levels implied
by the reordered returns series. The regressions in Tables 1, 2, and 3 were then
run on the simulated price series to obtain parameter estimates and standard errors
for the key parameters. This procedure was repeated 1000 times to obtain dis?
tributions from which the significance of this paper's results can be determined.
The distributions produced are roughly symmetric but fat-tailed. Although weakened, the coefficients on the DJIA dummy variables D98-02,?)95-04> and D90-09
from Table 1 's regression generally remain significant, with p-values of 0.009,
0.130, and 0.216, respectively. The 8 in Table 2 exhibits a p-value of 0.107. The
DJIA average return regression M-value coefficient from Table 3 remains highly
significant with a p-value of 0.008.
A second Monte Carlo simulation is based on the modeling of the DJIA daily
return as an AR(1)-GARCH(
1,1) process.7 Following Akgiray (1989), the process
is described

as
~

*,|/,_i
=

pt
ht

(f)o+ (f)iRt-i,

a0 + aie2_{
=

et

F(pt,ht),

Rt-

+ (3\ht-u

({)q- (f)\Rt-\,

> 0, and f3\ > 0. Conditional on the information set /,_i,

where a0 > 0,a\
from
distribution F(-, ?) with conditional mean pt and conditional
the
is
drawn
Rt
variance ht.
The parameter vector z = [a0, a\, (3\, <po,<j)\] is estimated assuming normality
function was
of the conditional distribution. Maximization of the log-likelihood
version
2.0.
Standard
in
GAUSS
module
MAXLIK
the
386,
using
accomplished
errors of the point estimates were calculated using the estimated information matrix
at the optimum. Parameter estimates are as follows (standard errors in parentheses),
Rt

0.000415
(0.000142)

+ 0.0678/?, _ 1,
(0.0173)

7See Engle (1982), Bollerslev (1986), and Akgiray (1989) for discussions of autoregressivecon?
ditionally heteroskedastic(ARCH) and GARCH processes. The particularspecificationused in this
paperfollows work by Akgiray (1989).

Donaldson

ht

0.00000329
(0.00000073)

+ 0.0796<_
(0.00835)

and Kim

325

l + 0.892/z,_ {.
(0.0133)

All parameter estimates are statistically significant. The unconditional variances

of e and/?, o2 = a0/(l ? a\ ? (3\) and o\ = o2/(l ? 4>2X),
respectively, are 0.0114
R accords well with the
of
and
0.0115
the
unconditional
variance
percent;
percent
value estimated from the raw data of 0.012 percent. The near-integrated nature of
the estimated GARCH process is in line with other empirical studies, implying the
presence of a high degree of volatility persistence (see Lumsdaine (1990)); this
paper's sample parameter estimates indicate an a\ + (3\ = 0.971.
As with the randomized-returns "bootstrap" simulation, 1000 simulated price
series were constructed based on the initial DJIA price and subsequent prices
implied by the returns processes generated using the estimated AR-GARCH pa?
rameters. Again, the Monte Carlo distributions are symmetric but possess vary?
ing degrees of excess kurtosis. The p-values for Table 1 's DJIA dummy coef?
ficients implied by the AR-GARCH simulation are 0.008, 0.090, and 0.164 for
^98-02^95-04.
andZ)9o-o9, respectively, which are slightly more significant than
the corresponding p-values from the randomized-returns
simulation. The p-value
for 8 in Table 2 is 0.069. For Table 3's average return regression, the M-value
coefficient is again highly significant with a p-value of 0.002. The results of this
section thus appear to confirm the statistical significance of Section III's results.
B.

Simulated

Price Series

To ensure that Section III's results are not simply a product ofthe general shape
traced by the DJIA's price sequence over the time period studied, the regression
tests are performed on several computer-generated
series that resemble the DJIA in
The
that
the same starting and ending
first
series
is
an
series
has
shape.
exponential
values and mean return as the DJIA. A second series was produced by multiplying
the DJIA by the arbitrarily chosen constant 1.40. As stated in the introduction,
such a rescaling of the price series retains the DJIA's relevant returns information
(and shape), but loses information on the particular values of the DJIA's digits.
A third price series was created by starting on the true value of the DJIA and
using subsequent price levels implied by a returns series constructed as in the
AR-GARCH simulation described above. In order to be sure of reproducing the
general shape of the DJIA, the first AR-GARCH series the computer produced
that possessed the following characteristics was chosen: i) an ending value within
a 100-point band around the DJIA's ending value, ii) a minimum value no less
than that of the DJIA, iii) a global maximum value no greater than 50 points above
the DJIA's maximum, and iv) a maximum over the first 2000 observations no
greater than the DJIA's local maximum over this interval. Finally, a fourth series
was produced in the same manner as above except the bootstrap simulated returns
were used instead of the AR-GARCH returns.
Tables 4, 5, and 6 contain the results from regression tests similar to those
conducted in Tables 1, 2, and 3 above. The absence of significantly negative
dummy coefficients in Table 4, significantly negative 8s in Table 5, and significantly

326

Journal of Financial and Quantitative Analysis

f(M): the frequency with which the index closes with its last two digits in cell M.
Djj. a 1/0 dummy variable that isolates cells in the range from /'to j.
All entries x100 (standard deviations in parentheses).
Note that, unlike the DJIA in Table 1, the (3s in this table are not significantly negative.
negative M-value coefficients in Table 6 reveals that there is no evidence of DJIAstyle barriers in the generated indices. It is therefore concluded that neither the
general shape of the DJIA nor the statistical properties of the DJIA's returns process
can account for the DJIA's apparently anomalous behavior around 100-levels.
C.

Robustness

The final concern of this study is for the sensitivity of the results with respect
to the sample period and the particular starting and ending values of the DJIA
index. To ensure that the results are not dependent on these factors, the begin?
ning and ending prices from the sample were iteratively truncated and the tests
were rerun. This procedure was performed 30 times to produce starting values
between 608 and 674 and ending values between 2645 and 2831: the results were
not altered appreciably. The sample was also shortened by deleting up to 1000
observations, 100 observations at a time. The smaller sample size weakened the
f-statistics somewhat, but in all cases the dummy regression coefficients from Ta?
ble 1 's regressions remained negative, as did the 8 from the regression in Table

Donaldson

and Kim

327

TABLE5
Price Level Hump Tests on Simulated Data
= a + (3M+SM + UM\ M = 00,..., 99

f(M)

f(M): the frequency with which the index closes with its last two digits in cell M, minus 1
percent.
All entries x100 (standard deviations in parentheses)
Note the lack of significantly negative 6s.
TABLE6
Conditional Returns Tests on Simulated Data
M = 00_99
RM = a + /3M+UM;

99).
RM: the average daily return conditional on having passed through cell M(M = 00,.
All entries x100 (standard deviations in parentheses)
=
Test H0(no barriers): (3 0 vs. /-/-((barriersare present)- (3 < 0.
Note that, unlike the DJIA in Table 2, the ps in this table are not significantly negative
2. The M-value

coefficient in Table 3's returns regression also remained highly

significant throughout.
To examine further the robustness of the results, the DJIA series was extended
back to January 1, 1952, and all tests were rerun. As Tables 7, 8, and 9 reveal,
the 1952-1990
results are very similar to the 1974-1990
results. To check the
significance of these results, the following AR(1)-GARCH(1,1)
specification is
estimated for the DJIA for 1952-1990,
Rt

0.000317
(0.000069)

+ 0.1412/?,_1,
(0.0107)

328

Journal of Financial and Quantitative Analysis

ht

0.00000085
(0.00000013)

+ 0.0697<_
(0.00490)

{ + 0.920/zr_ x.
(0.0059)

Monte Carlo results based on this specification produced p-values of 0.015,0.117,

and 0.344 for Table 7's dummies Dg%_02,?>95-04> and D90-09, respectively, a pvalue of 0.020 for Table 8's 8, and a p-value of 0.004 for Table 9's slope parameter.
A randomized-return simulation produced p-values of 0.018,0.154,
and 0.342 for
Table 7's dummies 1)98-02^95-04^
and D90--09, respectively, a p-value of 0.031
for Table 8's 8, and a p-value of 0.005 for Table 9's slope parameter. It is therefore
concluded that the barriers phenomenon is a robust feature of the DJIA.

f(M): the frequency with which the index closes with its last two digits in cell M.
Df.-:a 1/0 dummy variable that isolates cells in the range from / to /.
All entries x100 (standard deviations in parentheses).
Test H0(no barriers): (3 = 0 vs. H^(barriers are present): (3 < 0.
Note the significant negative /?s for the 1952-1990 DJIA.

f(M): the frequency with which the index closes with its last two digits in cell M, minus 1
percent. All entries x100 (standard deviations in parentheses).
Test /-/0(no barriers): 6 = 0 vs. /-^(barriers are present): 6 < 0.

V.

Conclusions

The possible existence of barriers to price movements in the daily closing

values of the DJIA have been tested for. Consistent with the economic theory of
price behavior in the presence of exogenously imposed barriers to price movements

Donaldson

and Kim

329

TABLE9
Conditional Returns Robustness Test
M = 00,.. ,99
RM = a + (3M+UM;
Index

(3

R2

DJIA 1952-1990

-0.0009
58.5
0.1339
(0.0043)
(0.0001)
Rm the average daily return conditional on having passed through cell M(M = 00,.. , 99).
All entries x100 (standard deviations in parentheses).
Test H0(no barriers): (3 = 0 vs. /-/-((barriersare present): (3 < 0.
Note the significant negative /3, as suggested by the barriers hypothesis and by the author's
earlier DJIA results.
(e.g., Bertola and Caballero (1992)) and consistent with the claims of financial
market commentators, it is found that the DJIA closes fewer times, on average, on
index values in the neighborhood of 100-levels.
It is also found that the DJIA's
conditional returns are negatively correlated with associated M-values. Together,
these results imply that the DJIA's rise and fall is restrained by barrier support and
resistance levels but that, having broken out through a barrier, the DJIA moves up
or down by more than usual. The statistical significance of the results is confirmed
and AR(1)-GARCH(1,1)
Monte Carlo simulations.
through randomized-returns
The anomalies do not appear in the less popularly-tracked WA index, nor do they
series that were studied. The results
appear in the numerous computer-generated
are not dependent on sample size, beginning and ending values, or the general
shape of the index. The DJIA's anomalous behavior is thus interpreted as an
indication that, as suggested by many financial market commentators, there may
well be pricing barriers at 100-levels in the widely followed DJIA index.
This paper closes with a discussion of the implications of price barriers for
market rationality and efficiency.
Standard finance theory suggests that it is the
information on relative price movements (i.e., returns) that should be of interest
The authors know of no theory that states that
to rational market participants.
special levels of some arbitrarily scaled index should be of particular importance
in a totally rational market. The finding of barriers in the DJIA therefore suggests
two possibilities.
The first possibility is that the market is totally rational and that,
contrary to the understanding of standard finance theory, certain values of widely
followed indices are in fact of genuine significance.
(Is it possible, for example,
that rational agents who find continuous portfolio rebalancing costly might use the
numerical value of the widely reported DJIA as an easily obtainable signal that
A second,
they should engage in optimal but infrequent portfolio rebalancing?)
and perhaps more likely, possibility is that some traders are in fact less than fully
rational and may indeed be using the DJIA's highly publicized crossing of, or
failure to cross, 100-level reference points as sentiment signals on which to base
their buying/selling decisions. Even if this second possibility is true, however, the
finding of a barrier at DJIA 100-levels, does not necessarily imply that the entire
market is irrational. Froot et al. (1992) and DeLong et al. (1990), for example,
present models in which rational traders follow fundamentally irrational trading

330

Journal of Financial and Quantitative Analysis

rules in order to profit from the existence of noise traders and thereby produce
anomalies in the market as a whole.8
Finally, it is important to note that the finding of barriers at DJIA 100-levels
does not necessarily imply that the market is inefficient. In particular, the existence
market does not necessarily imply returns predictabilof a less-than-fully-rational
set that contains the value of the DJIA.9
information
the
weak-form
ity given
Indeed, as proven by Bertola and Caballero (1992), if there truly are barriers at
100-levels in the DJIA then the hump-shaped price distribution present in Figure
1 is in fact implied by the absence of predictable excess returns. It thus appears
that, even if some subset of the market is less than perfectly rational, the Efficient
Market Hypothesis remains a useful paradigm.
References
Akgiray,V. "ConditionalHeteroscedasticityin Time Series of Stock Returns:EvidenceandForecasts."
Journal of Business, 62 (Jan. 1989), 55-80.
Bertola, G., and R. Caballero. "TargetZones and Realignments."AmericanEconomic Review, 82
(June 1992), 520-536.
Bollerslev,T. "GeneralizedAutoregressiveConditionalHeteroskedasticity."Journalof Econometrics,
31 (April 1986), 307-327.
DeLong, B.; A. Shleifer; L. Summers;and R. Waldmann."PositiveFeedbackInvestmentStrategies
and DestabilizingRationalSpeculation."Journal of Finance, 45 (June 1990), 379-395.
Donaldson, R. G. "PsychologicalBarriersin Asset Prices, Rationalityand the Efficient MarketHy?
pothesis."FinancialResearchCenterMemorandum#114, PrincetonUniv. (1990).
ConditionalHeteroscedasticitywith Estimatesof the Varianceof United
Engle, R. F. "Autoregressive
"
KingdomInflation Econometrica,50 (July 1982) 987-1007.
Flood, R., and P. Garber. "The Linkage between Speculative Attack and TargetZone Models of
ExchangeRates."QuarterlyJournal of Economics, 106 (Nov. 1991), 1367-1372.
Froot, K., and M. Obstfeld. "ExchangeRate Dynamics under Stochastic Regime Shifts: A Unified
Approach."Journal of InternationalFinance, 31 (Nov. 1991), 203-230.
Froot, K.; D. Scharfstein;and J. Stein. "Herdon the Street: InformationalInefficienciesin a Market
with Short-TermSpeculation" Journal of Finance, 47 (Sept. 1992), 1461-1484.
Krugman,P."TriggerStrategiesandPriceDynamicsin EquityandForeignExchangeMarkets."NBER
WorkingPaper2459 (1987).
_.
"TargetZones and Exchange Rate Dynamics." QuarterlyJournal of Eco?
nomics, 106 (April 1991), 669-682.
Ley, E., and H. Varian."A Note on the Dow Jones Digits." Univ. of Michiganmanuscript(1991).
Lumsdaine,R. L. "AsymptoticPropertiesofthe Quasi-MaximumLikelihoodEstimatorin GARCH(1,1)
and IGARCH(1,1)Models " HarvardUniv. manuscript(1990).
Pesenti, P. "Perforateand ImperforateCurrencyBands,"Yale Univ. manuscript(1990).
Pierce, P. S. TheDow Jones Averages1886-1985, Illinois: Dow Jones Irwin (1986).
TorontoGlobe andMail, variousissues.
WallStreetJournal, variousissues.
8In particular,Froot et al. (1992) prove that, as long as a significant fraction of the market is
following a "chartist"rule, it may pay perfectlyrationaltradersto also chart,even thoughthey realize
thatchartingis fundamentallyirrational.DeLong et al. (1990) demonstratea similarresultfor the case
in which noise tradersfollow a positive feedbacktradingrule.
9This claim was checked by examiningthe averagereturnRt conditionalon Mt- \. No significant
relationshipwas found between Rt and Mt-\. In particular,knowing what value the DJIA closed at
yesterday(i.e., Mt-\) does not help predict the returnRt for today. (Note that, since Rjy for day / is
calculatedconditionalon cell M having been passed duringday / (note past tense), while Rt measures
the expectedfuturereturncalculatedconditionalon yesterday'svalue of M, observinga barrier'seffects
ex post in Rjy is perfectly consistent with not observing effects on expectedreturnsRf. In particular,
the absenceof barriereffects in expectedreturnsdoes not negatethis study'spreviousresults,it simply
implies that the marketis efficient.) Ley and Varian(1991) presentadditionalevidence that suggests
thatthis paper'sfindingof barriersin the DJIA does not imply marketinefficiency.