Harris Raviv 1993

The Society for Financial Studies
Differences of Opinion Make a Horse Race

Author(s): Milton Harris and Artur Raviv
Source: The Review of Financial Studies, Vol. 6, No. 3 (1993), pp. 473-506
Published by: Oxford University Press. Sponsor: The Society for Financial Studies.
Stable URL: http://www.jstor.org/stable/2961976
Accessed: 04-08-2016 10:49 UTC
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms
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].
Oxford University Press, The Society for Financial Studies are collaborating with JSTOR to
digitize, preserve and extend access to The Review of Financial Studies
This content downloaded from 141.218.1.105 on Thu, 04 Aug 2016 10:49:06 UTC
All use subject to http://about.jstor.org/terms
Differences of Opinion Make
a Horse Race
Milton Harris
University of Chicago
Artur Raviv
Northwestern University
A model of trading in speculative markets is devel-

oped based on differences of opinion among trad-
ers. Our purpose is to explain some of the empir-
ical regularities that have been documented
concerning the relationship between volume and
price and the time-series properties of price and
volume. We assume that traders share common
prior beliefs and receive common information but
differ in the way in which they interpret this infor-
mation. Some results are that absolute price
changes and volume arepositively correlated, con-
secutive price changes exhibit negative serial cor-
relation, and volume is positively autocorrelated.
We develop a model of trading in speculative markets

based on announcements of public information. Our
purpose is to explain some of the empirical regular-
"It were not best that we should all think alike; it is difference of opinion
that makes horse races" (Mark Twain in Pudd'nhead Wilson, 1894). Milton
Harris is the Chicago Board of Trade Professor of Finance and Business
Economics at the University of Chicago. Artur Raviv is the Alan E. Peterson
Professor of Finance, Northwestern University. The authors gratefully
acknowledge the financial support of the Bradley Foundation. Professor
Harris wishes to acknowledge partial financial support from Dimensional
Fund Advisors. The authors also thank Yakov Amihud, Stephen Blough, Peter
DeMarzo, Michael Fishman, Larry Harris, Ronen Israel, Charles Kahn, Robert
Korajczyk, Josef Lakonishok, Mark Mitchell, Katherine Schipper, Bill Schwert,
Chester Spatt (the executive editor), Lars Stole; and Paul Pleiderer, the
referee, for valuable comments, as well as seminar participants at the Federal
Reserve Board, The Economics of Financial Markets Conference at Johns
Hopkins University (April 1991), American Finance Association Meetings
(January 1992), Michigan, Berkeley, Stanford, MIT, Dartmouth, USC, and
Rochester. Address correspondence to Milton Harris, Graduate School of
Business, University of Chicago, 1101 East 58th Street, Chicago IL 60637.
The Review of Financial Studies 1993 Volume 6, number 3, pp. 473-506

? 1993 The Review of Financial Studies 0893-9454/93/$1.50
The Review of Financial Studies / v 6 n 3 1993
ities that have been documented concerning the relationship between

volume and price and the time-series properties of these two vari-
ables. We view this goal as an important step in understanding the
operation of financial markets.
It is apparent that financial markets respond to public announce-
ments of relevant information. This fact is amply documented by
countless event studies that measure the price reaction to announce-
ments of earnings, dividends, new issues, macroeconomic data, and
so on. Trading volume has been used less frequently to measure the
market response to announcements; however, several papers have
shown that trading volume responds to earnings announcements.
Beaver (1968, p. 91), for example, shows that volume in the week of
an earnings announcement is, on average, about 34 percent higher
than average volume over the corresponding nonreport period.1
Explanations for trading volume include tax-driven trading, liquid-
ity trading, portfolio rebalancing, and speculation. We focus on spec-
ulative trading as being the major factor accounting for surges in
market activity following public announcements. Speculative trading
stems, presumably, from disagreements among traders over the rela-
tionship between the announcement and the ultimate performance
of the assets in question. Such disagreements can arise either because
speculators have different private information or because they simply
interpret commonly known data differently.
Since the private information approach has been explored more
extensively in the literature (but without accounting for many of the
stylized facts, see below), we adopt the second approach. Thus, we
assume that traders receive common information but differ in the way
in which they interpret this information, and each trader believes
absolutely in the validity of his or her interpretation. We refer to this
as the assumption that traders have "differences of opinion." It seems
to us that people often share common information yet disagree as to
the meaning of this information, not only in the evaluation of risky
assets but also in the evaluation of political candidates, economic
policies, and the outcomes of horse races. One example is the variety
of opinions among financial analysts and macroeconomists regarding
future movements of interest rates, exchange rates, gross national
product, and stock prices despite the fact that all these analysts have
access to the same economic data. Another example is the fact that
weather forecasts often differ even though all forecasters have access
to the same National Weather Service data.
We assume that traders start with common prior beliefs about the
returns to a particular asset. As information about the asset becomes
1 See also Bamber (1987) and Ziebart (1990).
474
Differences of Opinion
available, each trader updates his beliefs about returns using his own
model (likelihood function) of the relationship between the news
and the asset's returns. We assume that there are two types of risk-
neutral, speculative traders. The two types agree on whether a given
piece of information is favorable or unfavorable, but they disagree
on the extent to which the information is important. Speculators in
the "responsive" group increase (decrease) their probability of high
returns more upon receipt of favorable (unfavorable) information
than those in the "unresponsive" group. Therefore, when the cumu-
lative impact of the past information is favorable, the responsive spec-
ulators value the asset more highly and will own all of it. On the
other hand, when the cumulative impact of the past information is
unfavorable, the unresponsive speculators value the asset more highly
and will own all of it. Thus, trading occurs when, and only when,
cumulative information switches from favorable to unfavorable, or
vice versa.
Our main results are as follows:
* Absolute price changes and volume are positively correlated.

* Absolute changes in the mean forecast of the final payoff and
volume are positively related.
* If speculators overestimate (underestimate) the true quality of
the signal, then consecutive price changes exhibit negative (positive)
serial correlation.
* Volume is positively autocorrelated.
* Volume is larger than usual on average at the opening of the
market after any period in which it is closed (e.g., overnight, week-
ends, and holidays).
The literature on asset markets has, to a great extent, focused on

explaining asset prices with only peripheral attention to trading vol-
ume. Two branches of the literature concern us: rational expectations
asset pricing models and models based on differences of opinion.2
Rational expectations models generate disagreement through private
information. These models generally involve trading among privately
informed traders, uninformed traders, and liquidity (or noise) traders
[see, e.g., Grossman and Stiglitz (1976, 1980), Hellwig (1980), Dia-
mond and Verrecchia (1981), Pfleiderer (1984), Kyle (1985), Admati
and Pfleiderer (1988), Grundy and McNichols (1989), Foster and
Viswanathan (1990, 1993a), Holthausen and Verrecchia (1990), Kim
and Verrecchia (1991a, 1991b), Blume et al. (1991), Wang (1992),
2 Another line of research focuses on volume caused by tax-driven trading strategies. This research
shows, for example, that volume will be higher after a decline in the price of the asset. See, for
example, Lakonishok and Smidt (1986).
475
and Shalen (1993)].3 There are two major differences between the
approach in this literature and our approach. First, in rational expec-
tations models, trade is not generated by public information signals.4
Second, in rational expectations models, disagreement is the result
of private information. Consequently, speculators try to infer the infor-
mation of others from their behavior and from market prices; that is,
agents are influenced by the beliefs of others.5 As a result, the ability
of traders to learn from prices and the behavior of other traders must
be obscured by noise.6 Our differences-of-opinion approach allows
us to ignore learning from market prices and to dispense with noise
traders.
Volume results derived from rational expectations models are dif-
ficult to compare with those from our model, because volume in these
models includes portfolio rebalancing, liquidity, and speculative
trades. Our results concern only speculative trades. Moreover, much
of the volume of trade in rational expectations models is driven by
exogenous shocks, for example, to liquidity traders' demands or to
endowments, or by exogenous assumptions about the volume of trade
amongliquidity traders. Nevertheless, some comparisons can be made.
Similar to our result, Pfleiderer (1984) shows that successive price
changes are negatively correlated, although he attributes this result
in his model to the assumed nature of liquidity trading (see p. 17).
Pfleiderer (1984), Blume et al. (1991), and Wang (1992) show a
positive contemporaneous covariance between volume and absolute
price changes. Shalen (1993) shows that dispersion of beliefs con-
tributes to this covariance. In Pfleiderer's model, however, this result
is due entirely to nonspeculative trading, because the correlation
between speculative trading volume and absolute price changes is
zero (see p. 18). Foster and Viswanathan (1993a) and Wang (1992)
also show that trading volume has positive serial correlation. Admati
3 Diamond and Verrecchia (1981) do not include liquidity traders. Instead, they postulate random
shocks to the endowments of privately informed traders. In this model, volume consists of portfolio
rebalancing and speculative trades. Also, Blume et al. (1991) do not include liquidity traders. In
their model, traders have private information about the risky asset and about the precision of this
information.
Kim and Verrecchia (1991a, 1991b) and Grundy and McNichols (1989) are exceptions. In these
models trade is generated by public information, because traders disagree on its interpretation due
to prior private information. Kim and Verrecchia (1991a, 1991b) show that trading volume is
proportional both to absolute price change and the extent to which the precision of prior, private
information differs across traders. Grundy and McNichols (1989) obtain an expression relating the
volume of trade to the strength of the correlation between the errors in the public and private
signals and the precisions of the signals.
Thus, noisy rational expectations models with private signals cannot be interpreted as models with
common signals and differences of opinion, contrary to the claim in Pfleiderer (1984) and Holthau-
sen and Verrecchia (1990).
6 Milgrom and Stokey (1982) show that speculative trades based purely on differences in private
information cannot occur among risk-averse traders in the absence of noise traders. See Varian
(1989) for a short review of this and similar "no-trade" results.
476
and Pfleiderer (1988) and Foster and Viswanathan (1990) show that
volume will tend to be high in some periods and low in adjacent
periods. If we interpret periods in our model as being the same as
in these two models, then this implication is at variance with our
result that volume is positively autocorrelated. On the other hand, if
our periods are shorter or longer than their periods, there is no con-
tradiction. Also, Foster and Viswanathan (1990) show that volume
can be expected to be lower on Mondays, just the opposite of our
result.7
The second branch of the literature in which trading is induced by
differences of opinion is exemplified by the work of Harrison and
Kreps (1978), Varian (1985, 1989), Blough (1988), Hindy (1989), De
Long et al. (1990), and Kandel and Pearson (1992). Harrison and
Kreps (1978) analyze a model very similar to ours to show that prices
can contain a speculative component in addition to reflecting fun-
damental value. Varian (1985, 1989) focuses on differences in prior
beliefs as opposed to differences in models. He shows "the relation-
ship between the equilibrium price and volume of trade and the
equilibriumprobabilitybeliefs about those assets" (1989, p. 5). Blough
(1988) examines the effect on the informativeness of asset prices
when traders have both private information and differences of opin-
ion. The model of Hindy (1989) includes traders who have different
models and who also receive both public and private information.
He shows, using examples, that this model is capable of producing
expected volumes and price changes that are "positively related,
negatively related for all time periods, or have a relation that changes
from positive to negative, or vice versa, over time" (p. 8). De Long
et al. (1990) include two types of traders, one who correctly antici-
pates future price movements and one whose estimate of future prices
is distorted by an additive, random noise term.8 They show that their
model can account for such apparent anomalies as closed-end fund
discounts and premia and the equity premium and excess volatility
puzzles. The approach in Kandel and Pearson (1992) is very similar
to ours, although the specific models are somewhat different. In both
models, traders receive only public information. The major differ-
ences are that traders in Kandel and Pearson (1992) have both dif-
ferent prior beliefs and different likelihood functions and are risk
averse. We consider only different models. Moreover, Kandel and
Pearson analyze only a two-period model and make different distri-
butional assumptions. Their main results are that volume and absolute
7 Their result follows from the assumptions that private information accrues over weekends, but
public information does not, and that liquidity traders can postpone their trades by at most one
day.
8 See also Shleifer and Summers (1990) for a less technical description.
477
The Review of Financial Studies/ v 6 n 3 1993
price changes are positively related and that volume can be positive
even when price does not change. This last result is in contrast to
our model in which volume occurs only when price changes.9
To summarize, the theoretical literature has only recently begun
to produce results concerning volume. Several articles obtain a result
similar to ours on the relation between volume and absolute price
changes. Pfleiderer (1984) also obtains our result that consecutive
price changes are negatively correlated. Admati and Pfleiderer (1988)
and Foster and Viswanathan (1990) show that trading will tend to
bunch. Foster and Viswanathan (1993a) and Wang (1992) show that
volume is positively autocorrelated. Finally, Foster and Viswanathan
(1990) predict lower volume on Mondays. Our results that do not
appear in the previous literature are that absolute changes in the
mean forecast of the final payoff and volume are positively related,
and volume is larger than usual on average at the opening of the
market after any period in which it is closed. Moreover, our model
produces all these results from a single framework.
Empirical results on volume are fairly abundant. To a large extent,
these are consistent with our implications. A number of studies have
documented empirically the positive correlation between volume and
absolute price changes or price variance over time [see, e.g., Karpoff
(1987), Jain and Joh (1988), Schwert (1989), Gallant et al. (1992),
and Lang et al. (1992)]. Patell and Wolfson (1984) find, consistent
with our result and that of Pfleiderer (1984), that successive trans-
action price changes are negatively autocorrelated. Evidence that
volume tends to be higher in the beginning and end of trading days
is provided by Jain and Joh (1988). Amihud and Mendelson (1987,
1991) find evidence of higher volume at the opening of the market.
Foster and Viswanathan (1993b) also observe a U-shaped intradayvol-
ume pattern and find that trading volume is lower on Mondays.10
Lakonishok and Maberly (1990), however, point out that volume for
individuals is larger on Mondays, and volume for institutions is smaller.
Our prediction of higher volume following a period of market closure
is consistent with the evidence of high volume at market opening
and greater individual trading on Mondays (we discuss this point
further later). Also consistent with our implication, Ziebart (1990)
documents a positive relation between volume and the absolute
change in the mean forecast of analysts. Our prediction of positive
autocorrelation in volume is supported by evidence in Harris (1987).
9 An additional recent model involving differences of opinion is presented in Detemple and Murthy
(1992).
0 Both Foster and Viswanathan (1993b) and Brock and Kleidon (1992), using intraday data, find that
volume and trading costs are positively correlated. This is inconsistent with the prediction of Admati
and Pfleiderer (1988).
478
Section 1 presents the formal model in detail. Section 2 derives

and interprets the main results regarding the evolution of trading
volume and prices. Section 3 briefly considers an extension of the
model to the case in which investors are risk averse. Conclusions are
contained in Section 4. All formal proofs are relegated to the Appen-
dix.
1. Model
Here we describe a model of trade based on differences of opinion,

using the approach discussed in the introduction. There are two
groups of risk-neutral speculators who trade, at dates t = 1,2,..., T,
shares of an asset that makes a single random payment of R at date
T. Shares of the risky asset are traded for a riskless asset. For simplicity,
we neglect discounting, so the rate of return on the riskless asset is
zero.
The final payoff, R, can be either high (H) or low (L, with L < H)
with equal probabilities.1" At each date t = 1,..., new public infor-
mation in the form of a signal, s,, is revealed, after which speculators
update their beliefs and may trade at a price pt. Examples of the
signals are earnings announcements, macroeconomic data, or polit-
ical events. The signals are independent and identically distributed
conditional on the true payoff. The true probability density of a signal
s given a final payoff R E {H, L} is denoted by ao5(s).
To clarify the interpretation of the signal and simplify the analysis,
we make an assumption that guarantees that the true posterior at date
t depends on the history of signals only through the cumulative sig-
nal. Formally, we assume
Fkoas fo r s ?-0,
C(s) = Uo2(-S) 0
Hkob-s for s < 0, (1)
where ao and bo are parameters strictly between 0 and 1, and ko is a
constant required to make or (s) and or (s) density functions.12 Using
Bayes' rule and Equation (1), after observing a history of signals st -
(sl,... St), we find the true posterior probability that R = H is
Hro(St) = = O( = [1 + ]- Hr(m), (2)

Ht 2R=H1. Ht=l CrRO. 0 H
Later, we assume that traders realize that the true prior probabilities are equal. This assumption
makes the analysis much more tractable. We conjecture that all our results go through if traders
share the same probabilities even if they are not one-half.
2 Additivity requires that the likelihood functions be symmetric, that is, that the probability of a
particular signal when the final payoff is high is the same as the probability of minus that signal
when the final payoff is low.
479
where m = s1 + . . . + St is the cumulative signal and 00 = b0/a0. In

other words, conditional independence of the signals over time and
our assumption that the likelihood functions are exponential imply
that signals are additive in the sense that a signal of, say, s = 5 is
equivalent to 5 signals of 1 each (or a signal of 8 and a signal of -3).
Since the posterior depends on the signal history only through the
cumulative signal m, we substitute m for st and drop the t subscript
in the posterior. Obviously, the probability that R = L, given m, is
ro (m) = 1 - 1rH(m).
So that we may interpret larger values of the signal as more favorable
information, we assume that the posterior probability of a high out-
come is increasing in m (i.e., 00 < 1). We may interpret 00 as an inverse
measure of the quality of the signal. For example, if 00 = 0, then any
positive signal results in a posterior that assigns probability 1 to R =
H, and any negative signal results in a posterior that assigns proba-
bility 1 to R = L. Conversely, for 0, = 1, the posterior is independent
of the signal.
This completes the description of the true distributions of payoffs
and signals. We now consider the beliefs of speculators regarding
these elements. All speculators know the correct prior that R can be
either H or L with equal probabilities. Differences of opinion are
generated by assuming that speculators have different models for
interpreting the signals. After observing the signal, each speculator
revises his beliefs regarding the final payoff, R, using Bayes' rule and
his own model (likelihood function) of the relation between signals
and the final payoff. We assume that each speculator is absolutely
convinced that his or her model is correct.13 Indeed, each group
believes the other group is basing its decisions on an incorrect model
(i.e., is irrational in this sense).4 All information, including specu-
lators' models, is assumed to be common knowledge, so speculators
make no attempt to infer anything from the behavior of prices or other
speculators.
The speculators' models have the same functional form as the true
likelihood function but with different parameters. Consequently, the
resulting posteriors also exhibit the same functional form as the true
posteriors. In particular, speculators realize that the signals are i.i.d.
13One could assume that speculators are unsure of the true model and revise their beliefs about it
as well as about the final payoff. We neglect this added complication on grounds that the true
model is likely to change sufficiently rapidly that such updating is of negligible importance. Thus,
the period length and the number of periods, T, in our model should be interpreted as fairly small.
Note that absolute belief in a model is not irrational in the Bayesian sense, especially since, in our
formulation, there are no possible realizations of the signal that are inconsistent with either group's
model (i.e., occur with zero probability according to the model).
1 Thus, while we maintain the assumption of rational agents, we drop the assumption that rationality
is common knowledge.
480
conditional on the final payoff, but they do not know the true density
functions ao?(s). Instead, group j (j = 1, 2) believes that the condi-
tional density of a signal s given final payoff R is ai(s), where a3i is
the same as ao with parameters kj, a1, and bj replacing ko, ao, and bo.
As before, aj and bj are between 0 and 1, and we assume Oj0 baj/
< 1 forj E {1, 2}. The posterior probabilities, irJ1(m), derived from
the prior and group j's model of the signal-generating process are
given by the same formula as the true posteriors [i.e., Equation (2)],
with j replacing 0.
Obviously, given the foregoing assumptions, if the groups are to
have models resulting in different posterior beliefs, they must have
different values of 0j. As a convention, we assume that 01 < 02. Con-
sequently, given the previous interpretation of 0, group 1 believes
the signal is of higher quality than group 2 and therefore responds
to a given signal history to a greater extent.15 It follows that group 1
is more optimistic when the cumulative signal is positive and is more
pessimistic when m is negative. When m = 0, both groups revert to
their prior beliefs, namely that the payoff is high with probability 2.
Figure 1 depicts rir(m) for j = 1, 2. We refer to group 1 as the
"responsive" group and to group 2 as the "unresponsive" group.
Two important properties of 1r-(m) that follow from (2) are that
xi, is monotone increasing in m and is concave for m > 0 and convex
for m < 0 (see Figure 1). These properties reflect the facts that larger
cumulative signals indicate greater likelihood of high final payoff and
that the posterior is more sensitive to changes in cumulative signal
when beliefs are more diffuse (i.e., when m is near zero).
We turn now to the speculators' preferences and trading. Recall
that all traders are risk neutral. Risk neutrality is appropriate since
we are interested primarily in volume generated by speculation as
opposed to hedging or life-cycle considerations. Thus, in our model,
there is no trading except for speculative purposes. Risk neutrality
also implies, however, that demand functions are infinitely elastic.
Thus, any trader will seek to buy (sell) an infinite number of shares
at any price below (above) his reservation price. Consequently, to
make the equilibrium well defined, we must assume that there is a
fixed number of shares available. This implies that short sales are not
allowed and that, whenever trading occurs, all available shares will
be purchased by the more optimistic group. Furthermore, as long as
one group remains more optimistic, there will be no further trade.
Trade will occur only when the two groups "switch sides." This fact
is a direct consequence of our risk-neutrality assumption. In a model
15 Note that the two groups have access to exactly the same information; that is, neither group has
more precise information than the other. One group simply believes the signal is more informative
than the other group does.
481
0)
0 ~~~~~Posterior
~~~4% ~~~Behief s
.... .Group 2
Posterior
Behef s
Figure ~ ~ ~ ~ ~ ~ . Gru1
Cumulative Signal, m
Figure 1
The probability of a high final outcome is shown as a function of the cumulative signal, m, for hoth
groups of speculators. Group 1 is assumed to be more responsive to the signal than group 2. For
a positive cumulative signal, group 1 values the asset more highly; the reverse is true for m < 0.
with risk-averse speculators, whenever one group's beliefs change

relative to the other's some trading will occur. For tractability, our
model approximates these trades as negligible.16
The price of the risky security changes to reflect the new reservation
prices of the traders each time information appears. The reservation
price of group j in any period is group fs expectation of next period's
equilibrium price.17 The equilibrium price in any period will be
between the two reservation prices for that period, but its precise
value will depend on how the market is organized. For tractability,
we assume that in every period one group (the same one in each
period) has sufficient market power to offer a price on a take-it-or-
leave-it basis. This price will equal the "price-taking" group's res-
ervation price.18 Thus the price-taking group always engages in trades
that they believe have zero net present value.19 It makes no difference
16 See Section 3 for a discussion of the risk-averse case.
7 This can be proved by backward induction. See Lemma 2 in the Appendix.
"I Notice that we are assuming that the market price equals the price taker's reservation price even
in periods in which there are no gains to trade. This is innocuous, since, in such periods, any price
between the two reservation prices results in no trade.
19 If one assumes that the equilibrium price is competively set in each period, then the price in any
period in which there is trade will equal the reservation price of the buyer. This occurs because
the buyer has infinitely elastic demand, but the supply is bounded. Which group is the buyer
482
which group is the designated price taker as long as it is the same

group in each period. Since either group can be the price-taking
group, we use symbols without a j subscript or superscript (e.g.,
7rH(m), 'YH(S), 0) to denote variables pertaining to the price-taking
group.
One interpretation of this assumption is that one group consists
mainly of large traders (e.g., institutional investors), and the other
group consists mainly of small, individual investors. Under this inter-
pretation, market makers give preferential treatment to the large trad-
ers by filling their orders at the reservation price of the small investors.
The price-taking group's reservation price in any period is that
group's current expectation of the final payoff.20 Thus, if the cumu-
lative signal at date t is m, then the reservation price of the price-
taking group and the price of the risky asset at date t is
Pt= p(m) = HIrH(m) + LIrL(m) = (H - L)7rH(m) + L, (3)
where 7rH is given by (2). Note that, since the beliefs depend on the
signal history only through the cumulative signal, so does the price.
Moreover, as is clear from (3), price changes are proportional to
changes in the posterior of the price-taking group.
2. Equilibrium Volume and Price Dynamics
Having described a simple model based on heterogeneous beliefs,

we now characterize the evolution of prices and volume of trade that
results from this model. Recall that the equilibrium price in any
period is the reservation price of the price-taking group of speculators
in that period, even if at that price no trade takes place. Thus, prices
change every period whether or not trading occurs. Our first result
shows that these price changes are larger on average in periods of
positive volume than in periods of zero volume (see the Appendix
for a formal proof).
Theorem 1. Absolute price changes and speculative volume are pos-

itively correlated.21
changes from period to period. Consequently, each group's reservation price in any period will
involve that group's expectation of the other group's reservation price in the next period. Therefore,
the equilibrium price in any period will involve each group's expectation of the other group's
expectation of the first group's expectation, and so forth. This quickly becomes intractable. We are
grateful to Charles Kahn for clarifying our thinking on this point.
20 As mentioned, one's reservation price today is his expectation of tomorrow's equilibrium price.
For the price-taking group, tomorrow's equilibrium price is the price taker's expectation in that
period of the following period's equilibrium price. The result follows by successive iteration forward
to period Tand applying the "law of iterated expectations." See Lemma 2 for a formal proof.
Covariances and expectations in all theorems are calculated with the true distribution of the signal.
These correspond to the estimates of the covariances and expectations that would be obtained
from a time series generated by the model.
483
a,
0.
I - / ~~~~~~S < -M |-M < S < O | < S < M |S > M I

Cumulative Signal, m M 2M
Figure 2
The probability of a high final outcome is shown as a function of the cumulative signal, m. If the
current cumulative signal is M, signals s < -M result in positive volume. Signals s 2 -M result
in zero volume. Three ranges of such signals are shown. For signals between -Mand M, the change
in probability of high outcome is smaller in absolute value than the change for s = -M. For s >
M, the change in probability is smaller in absolute value than that corresponding to -s.
This result is similar to results in Pfleiderer (1984), Blume et al.

(1991), Kim and Verrecchia (1991a, 1991b), Wang (1992), and Kandel
and Pearson (1992). A number of studies have documented empiri-
cally the positive correlation between volume and absolute price
changes or price variance over time [see, e.g., Karpoff (1987), Jain
and Joh (1988), Schwert (1989), and Gallant et al. (1992)]. Jones et
al. (1991) show that this relationship reflects primarily a relation
between volatility of returns and the number of transactions. This last
result is closer to our Theorem 1, since in our model only the number
of transactions is endogenous, and the size of each transaction is
normalized to unity.22
To understand Theorem 1, recall that trading volume is positive
only when the two groups of speculators "switch sides" and that this
occurs only when the cumulative signal changes sign. We argue that
for any given current price or, equivalently, current cumulative signal,
M, the absolute change in 7rH (and therefore in price) is larger when
averaged across signals resulting in volume than when averaged across
22 Some of the empirical work [e.g., Schwert (1989)] establishes the positive covariance between
volume and price or return volatility, as opposed to absolute price change. We can show numerically,
but not analytically, that such a relationship holds for our model.
484
signals not resulting in volume. To see this, we first argue that for
any current cumulative signal and any current signal that results in
no volume there is a corresponding current signal that results in
positive volume and a larger absolute price change. Suppose that M
> 0. For volume to be positive, the signal must be below -M (see
Figure 2). Consider a signal s for which volume is zero (i.e., s >
-M). First suppose that s is between -M and M. From Figure 2, it is
clear that the price change for any such signal is smaller than the
price change for s = -M. Thus, the price change for s E [-M, M] is
smaller than the smallest price change associated with positive vol-
ume. Second, suppose that s > M. The price change associated with
this zero-volume signal is smaller than that associated with the cor-
responding positive-volume signal -s.
To summarize, we have shown that for any current value of m > 0
and any signal corresponding to zero volume, there is a signal that
induces positive volume and results in a larger absolute price change.
A similar argument holds for m < 0. The result does not quite follow
from this argument, however, for two reasons. First, the distribution
of signals may be such that, for zero-volume signals, large values of
the signal are much more likely than small values, whereas the oppo-
site is true for positive-volume signals. In fact, there is sufficient
symmetry in the distributions assumed here to prevent this. Second,
the intuition given above is for positive correlation given any current
price or cumulative signal. The formal proof shows that the result
can be extended to the unconditional covariance.23
In addition to the positive correlation between volume and absolute
price changes, a positive correlation has also been observed between
volume and absolute changes in mean earnings forecasts. More spe-
cifically, Ziebart (1990) documents a positive correlation between
volume and the absolute percentage change, as a result of earnings
announcements during the year, in the mean of analysts' earnings
forecasts for the coming year. We view these annual forecasts as pre-
dictions of (more or less) long-run performance of the firm. In terms
of our model, earnings announcements correspond to our signal,
st, and analysts' forecasts correspond to the speculators' forecasts of
final returns. A speculator's forecast is measured by his conditional
expectation of final payoff, R. From equation (3), group j's forecast
at date t of the final payoff given cumulative signal m is -x1(m) (H -
L) + L. The mean forecast at date t is therefore
,u(m) (0.5)[irl(m) + ir(m)](H - L) + L.
The absolute change in mean forecast from tto t + 1 given cumulative
23 We are grateful to the referee for suggesting a method of proving this extension.
485
signal m and current signal s is
A\t(m, s) = I,.t(m + s) - ufm)
H- L
= 7rH m + s) XH( m)
+ 7H2 (mr + s) - 72 (m) (.
From the equation, it is clear that revisions in forecasts are related to

changes in the speculator's beliefs about the probability of the high
payoff, -rx. Since price changes are also related to changes in specula-
tor's beliefs in our model, it is not surprising that the price change
result can be extended to forecasts. In particular, we show the fol-
lowing result.
Theorem 2. Absolute changes in the mean forecast of thefinalpayoff

and speculative volume are positively correlated.
Note that no causality is implied by this relationship between vol-

ume and forecasts. Both variables are driven by a third, exogenous
factor, namely the signal.
Our next result shows that our model can be consistent with the
well-documented empirical observation that consecutive transaction
price changes exhibit negative serial correlation [see Patell and Wolf-
son (1984) and the references cited there]. There are two senses in
which consecutive price changes are negatively correlated in our
model: transaction-to-transaction and period-to-period. First, the price
changes accompanying consecutive transactions are of opposite sign.
Second, if the price-taking group of speculators overestimates the
quality of the signal, then the true expected price change in the
current period is negative (positive) whenever the price change in
the previous period was positive (negative). Of course, if the price-
taking group underestimates the quality of the signal, the conclusion
is reversed. The result is stated formally in
Theorem 3. (a) The price changes accompanying consecutive trans-

actions are of opposite sign. (b) Suppose that the price-taking group
of speculators overestimates the quality of the signal (i. e., 0 < O). If
the previous price change is positive (negative), then the price will
decline (rise) nextperiod, on average. On the other hand, if 6 > 00,
positive (negative) price changes will be followed on average by price
increases (decreases).
The intuition for these results is quite simple. Part (a) follows from
the fact that transactions occur only when the cumulative signal crosses
486
zero. Thus, in consecutive transactions, m crosses zero in the opposite

direction, resulting in price changes of the opposite sign. For part
(b), if the previous price change is positive (i.e., favorable infor-
mation was revealed last period), then it is more likely that the cumu-
lative signal is positive. If the price-taking group overestimates the
signal quality, then, when the cumulative signal is positive, this group's
view of the future signals is rosier than reality. That is, the true infor-
mation content of the cumulative signal is less favorable than they
think. Consequently, this group's expectation of next period's price
is higher than its actual average value. Since the price this period is
given by the price-taking group's current expectation of next period's
price, the current price is higher than the true expected value of next
period's price. Thus, the price will decline next period, on average.
The entire argument is reversed when the previous price change is
negative or when the price-taking group underestimates the quality
of the signal.
Having discussed price dynamics, we now turn to results concern-
ing the evolution of volume. Our first such result is that volume is
positively autocorrelated; that is, high-volume periods tend to be
followed by high-volume periods, and similarly for low-volume peri-
ods. Recall that positive volume occurs in our model only when the
two groups of speculators switch sides. If such a switch occurred last
period, this makes it more likely that the two groups have similar
opinions this period. This, in turn, makes it more likely that opinions
will again reverse and trade will occur when the current news is
announced. We state the result formally as Theorem 4.
Theorem 4. Volume is positively autocorrelated.
This result was also obtained by Foster and Viswanathan (1993a)

and Wang (1992) and is consistent with the evidence presented in
Harris (1987). He estimates a median correlation coefficient of 0.586
between the number of transactions in a day and the number in the
previous day. He also finds an autocorrelation in dailyvolume of 0.347.
Our second result concerning trading volume is, roughly speaking,
that if more information is released at a given time volume tends to
be larger. We model an increase in information as the appearance of
two signals in a given period instead of just one and show that volume
is larger on average after such an occurrence than after the release
of a single signal. When two signals are released, the range or variance
of possible news is increased. This makes it more likely that specu-
lators will switch sides following such an event. Formally, we state
this as Theorem 5.
487
Theorem 5. Average volume when two signals are released between

feasible trading dates is larger than when only one signal is released.
Theorem 5 has several empirical applications. For example, the

result implies that volume should be larger than usual on average at
the opening of the market after any period in which it is closed,
provided that information is released at the same rate over the closure
period as over other periods. Jain and Joh (1988) and Foster and
Viswanathan (1993b) show, consistent with Theorem 5, that volume
is larger on average in the first hour the market is open than in other
hours. Amihud and Mendelson (1987) find that volume on the open-
ing transactions of the DowJones stocks constitutes, on average, 5.6
percent of the total daily volume in these stocks and is, on average,
8.4 times greater than the average volume on the closing transactions.
For the Tokyo Stock Exchange, Amihud and Mendelson (1991) find
that opening volume at the morning and afternoon sessions consti-
tutes on average 11.6 percent and 9.2 percent, respectively, of the
total daily trading volume. There is also some evidence that trading
volume among individuals is indeed larger on Mondays, although
volume among institutions is smaller [see Lakonishok and Maberly
(1990)]. Osborne (1962) speculates that the difference in trading
behavior occurs because institutions must wait until Monday to make
trading decisions, whereas individuals make these decisions over the
weekend. Another application of Theorem 5 involves the behavior of
volume around simultaneous announcements by several firms in an
industry. In particular, if two firms have correlated returns, then an
announcement of earnings, or other data, by either has information
about the other. Consequently, the volume of trade in the stock of a
given firm should be higher on days in which at least one other similar
firm (e.g., in the same industry) makes an announcement than on
days when only the given firm does.
3. Risk-Averse Investors
One problem with the model analyzed in the previous two sections
is that, due to risk neutrality of investors, short-sale restrictions are
required and trading is always all-or-nothing. In this section, we exam-
ine a model in which investors are risk averse and short sales are
freely allowed. Our purpose is to see to what extent our previous
results are robust to this modification.
Introducing risk aversion adds considerable complexity to the prob-
lem. To restore a measure of tractability, we make three simplifying
assumptions. First, we assume that investors have identical exponen-
tial (constant absolute risk aversion) utility functions. Second, we
488
assume that investors are myopic in the sense that, in each period,
they chose portfolios as if there will be no further trading. Obviously,
this is a very strong assumption and is counter to the analysis of the
previous sections. These two assumptions allow us to derive demand
functions that are simple, independent of the investor's wealth at
each date, and depend on time only through the cumulative signal.
Third, we assume the asset is in zero net supply; that is, we consider
only prices and volume of pure bets.
With the above assumptions, an investor in group j with current
wealth W and cumulative signal m, who takes the current price of
the risky asset, p, as given, chooses quantities q of the risky asset and
b of the riskless asset to solve the following problem:
max Ei[- err(Rq+b) I m]

q, b
subject to pq + b= W
In this problem, E is the expectation using group j's model, and r

is the common coefficient of absolute risk aversion of all investors.
The resulting demand for the risky asset by group j is then
Qj(P) 1 [nlr-'ii(m)H-pl
= r(H-L) rir( m) pH- L
Equating total demand with zero net supply of the risky asset results
in an equilibrium price given by (3), except that the price-taking
group's 0 in (2) is replaced by 0= (0102) 1/2 Thus, group l's equilibrium
holding of the risky asset in any period in which the cumulative signal
is m is
Q*(m) = pm/[2r(H -L)],

where p = ln(02/01). The volume of trade at date t is simply the
absolute change in group l's holding of the risky asset between tand
t + 1. The previous equation implies that this volume is proportional
to the absolute value of the signal, st, at date t.
Using the foregoing analysis, it is possible to show that most of our
earlier results for the risk-neutral case extend to this case as well.
Although we were not able to show analytically that volume and
absolute price changes are positively correlated (Theorem 1), we did
verify this result numerically. Given the positive covariance of abso-
lute price changes and volume, our result that absolute changes in
the mean of the forecasts and volume are positively correlated (The-
orem 2) goes through. Since volume is positive after every signal in
the risk-averse model, there is no counterpart to Theorem 3(a).24
24 In the risk-averse model, trade occurs even when the relative optimism of the two groups of traders
does not reverse; that is, trade will occur even as the optimistic group becomes more optimistic.
Thus, there is no simple relation between the sequence of transactions and price changes.
489
Theorem 3(b) on the serial correlation of price changes, however,

goes through as before, with 0 replaced by 6. Expected volume in the
risk-averse model is independent of the cumulative signal and hence
does not exhibit serial correlation as in the risk-neutral case (Theorem
4). Finally, it can be shown that volume following two signals is larger
on average than the volume after only one signal (Theorem 5). Thus,
all but one of our previous results extends to the simplified risk-
aversion model.
4. Conclusions
We provide a model of speculative trading volume and price dynam-

ics. Trading is generated by differences of opinion among traders
regarding the value of the asset being traded. These differences of
opinion result from different interpretations of public information
announcements. Although traders are rational in our model, some
view others as being irrational. Given this lack of the "common knowl-
edge of rationality," all behavior in the model is maximizing. This
article attempts to show that the approach taken here can help to
explain the observed behavior of speculative markets.
We show that absolute price changes and volume are positively
correlated, absolute changes in the mean forecast of the final payoff
and volume are positively related, consecutive price changes exhibit
negative serial correlation, volume is positively autocorrelated, and
volume is larger than usual on average at the opening of the market
after any period in which it is closed. Some of these results explain
stylized facts documented in the empirical literature; others remain
to be tested. Although the results were proved using specific func-
tional forms, we believe them to be robust. Our implications follow
from two basic properties of the model. First, the posterior beliefs of
a speculator are most sensitive to the next signal when his or her
current beliefs are most diffuse. Second, speculative trading is most
likelywhen the beliefs of the traders are most diffuse. Neither of these
properties is unique to the particular functional forms chosen. More-
over, the first feature is a general characteristic of Bayesian updating.
The second should be true of any model of speculative trading based
on differences of opinion.
An alternative approach to modeling differences of opinion is to
assume that all speculators share the same likelihood function but
have different prior beliefs about the final payoff, R. This approach,
however, will not generate sequential trading if priors are only trivially
different in the following sense. Suppose that traders agree on the
set of possible "states of nature" and their probabilities but disagree
about the implication of a given state of nature for the final payoff.
490
For example, both groups may agree that future interest rates can be
either high or low, and they agree on the probabilities but disagree
about the impact of high (or low) interest rates on the final payoff.
Formally, suppose that all speculators agree on the prior distribution
g(co) of the state of nature c and on the conditional distribution of
the signal given the state (the likelihood function) but disagree about
the mapping R = fj(w) relating the state to the final payoff. Thus,
speculators disagree about the prior distribution of the final payoff,
but j's prior beliefs are of the form gj(R) = g(f-l (R)) for all j. In
this case speculators will not trade except at the last date. To see this,
consider the last date. The traders will have different reservation
prices based on different fj functions. These reservation prices will
result in an equilibrium price that is a function only of the past signals.
At the second-to-last date, all speculators agree about the distribution
of the next signal since they agree on the state probabilities and the
likelihood function. Therefore speculators agree about the distribu-
tion of the next price since it depends only on current information
and the distribution of the next signal. Since the only reason for
trading at the second-to-last date is to speculate on the equilibrium
price at the last date, speculators' agreement about the distribution
of the price at the last date implies that they have no reason to trade
at the seocnd-to-last date. This argument can be extended to any
previous date by backward induction. Thus, to obtain sequential trad-
ing over time, one must assume that speculators disagree either about
likelihoods or about the fundamental states of nature.
Appendix
We start by stating several important formulas that are used later. First,
it follows immediately from the definition of 1rj(m) that
7H( m) = ,ajrbm and -(m)=-x(-m) (Al)

am+ lb1,7Haj +bj
Second, the true density of the current signal, given cumulative signal
m, is
r(s I m) = o%(s)THr(m) + o2L(s)ir(m) = O( + s) (A2)
It follows that
Tr(s -m) = r(-s I m),

r(s m m) + T(-s I m) is independent of m,
r(-s l m) > r(-s l m') V m' > m > O, s > O, (A3)

r(s l m) > r(-s l m) Vm>O, s>O.
491
Third, the cumulative distribution function (cdf) corresponding to

r is
[ko[ a a L(m) + O H (m) for s < o,

T(s m) (A4)
{1-ko [aH + ( ] for s > O.
In this expression, a =-ln(ao), fi =-ln(bo), and
1= Cs (as + bs) ds=- + -1 (A5)

ko ? a d
Fourth, let v(m, s) be volume at t + 1 if mt = m and st = s 1 if m

and m + s are of opposite sign, and 0 otherwise. Then the probability
of positive volume at t + 1 (which equals the expected volume at t
+ 1) given cumulative signal mt = m at t is
v)(m) = Pr(vt+1 > 0 mt = m) = EO(vt+1 Mt= m)
(m){ for mr'O (A6)
Here, EO denotes the expectation with respect to the true distribution

(of s given m in this case).
We are now ready to prove several preliminary results. Let Xt(m)
be the unconditional density of the cumulative signal at date t. We
next show that Xt is symmetric with respect to m = 0.
Lemma 1. Xt (m) = Xt (- m) for all m and t.
Proof. The proof is by induction on t. For t = 1, Xt(m) = ((m I 0) =

r(-im I 0) since 7r? (0) = r2O(0) = 2. Now suppose Xt(m) = Xt(-m)
for all t < n. Then
Xn(m) = I (m n- = X)Xn-l(X) dx
_00
fr(m -xI x)Xn-,(x) dx

_00
= f r(m + x -x)Xn_1(-x) dx
_00
= | rf(-m -x x)Xnl(x) dx
400
492
= J'Xn(-m I m.-, = x)Xn-(x) dx = Xn(-m) (A7)

Here we have used the induction hypothesis and the fact that
r(s I -m) = r(-s I m). Q.E.D.
Lemma 2. Assume group 1 is the price-taking group. Then an equi-

librium price sequence is
P,( Mt) = P' (mt) X for t = 1, .. . T, (A8)
where p'(mt) = Hir-(mt) + L-r{ (mt). Also, for every t, m, andj, the
reservation price for the risky asset of group j at date t after cumu-
lative signal m is given by r/(m) = Ej[pt+1(m + s) I m ], where E'
denotes the expectation with respect to group j's posterior beliefs.
Thus, r' (m) = p' (m) and r2 (m) = E2[p'l(m + s) I m ]. Obviously,
if group 2 is the price-taking group, the lemma goes through with all
group indices reversed.
Proof. The proof uses a backward induction argument. Consider date

Tand suppose the cumulative signal at date Tis MT-. Since this is the
last period, group j values the risky asset at p'(mT). Consequently, if
p2(mT) > pl(mT), group 2 will offer to buy as much as group 1 wants
to sell at price pl (mT). If p2 (mT) < pl (mT), group 2 will offer to sell
as much as it has at price pl (mT). Thus, the equilibrium price at date
Tis pl(mT).
Now consider date T - 1. If group 1 purchases d units of the risky
asset at price p, its expected utility from this purchase is
d[E1(pl(mT-1 + ) I mT-1) - p] = d[p(mT-1) -P]

since Ej[irj(m + s) I m] = ir'(m). Thus, at T- 1, group l's reservation
price is pl (mT-l) . Similarly, group 2 investors know that they will be
able to buy or sell the risky asset at date Tat price pl(mT-l + s), so
their reservation price at date T - 1 is E2(pl(mTT1 + s) I mTl).
Consequently, by the same argument as before, the equilibrium price
at date T- 1 ispI(mT-l)
Obviously, the argument can be repeated for t = T -2...
Q.E.D.
Hereafter, we denote the price-taking group's irj, Oj, and p1 by r,

0, and p, respectively. Let
Alr(m, s) = I IH(m + s) - rH(m) J.

Then
IAPtl = Ip(mt+?) - p(mt) I = (H - L)Air(mt, st). (A9)
493
It is easy to check that
Air(m, s) < A7r(m, -s) V m > 0, s> 0, (A1O)

Air(m, s) is monotone decreasing in m V m > O, s> O.
Now, let
qi(m) = EO[AZr(m, s) mt = m, vt+1 = i], for i = 0,1,
q(m) = EO[zr(m, s) mt= m]

Note that qi is independent of t.
Lemma 3. Suppose O0 > 0. For m > O (m < 0), E0fpt+ I mt = m] <

(>)pt [i.e., if the currentprice exceeds (is exceeded by) the uncon-
ditional average price, p(O)], then next period's price will be below
(above) this period's on average.
Proof. This will be shown only for m > 0. The proof for m < 0 is
symmetric. Since
EO[pt+l- pt I mt = m] = (H - L)EO[7rH(m + s) - 7rH(m) I m],

it suffices to show that EO[7i-H(m + s) I m] < lrH(m) for all m > 0.
From (A2),
E0[7rH(m, s) I m]
= ,) iHr(m + s)r(s I m) ds
lrH(mH 0 jj + s)o-% (s)

= irTO(m) ds
k m [ 0 (M + )ds + i ? (m+s) ds
=k0ir% (m) J O ) + ds d(ll

krH [l(m + s) ao +rH(m - s)b1 d (A
i%7Hn)I + I0 7
Suppose that we show
H ( 0[H(m + s)aO +rH(m- s)b 1
< lrH(m) (a + ba ) for all s)> O. (A12)

Then integrating both sides and using (A5) gives the result. There-
fore, it suffices to establish (A12), which can be rewritten as
494
lrH(m + s) (s + m)
7ro,Km + s)
+ 7r m s) -ls)(-s)lr%(m)
+ lrH(m l(m - s) < 7rH(m), (A13)
by dividing by as + bs and using (Al). To show (A13), note that it

is satisfied as an equality at O = 0. Therefore, it suffices to show that
the left-hand side of (A13) is strictly decreasing in 0O for m > 0. But
it is easy to check that z(00) =r%(s)1r (m)/Ir%(m + s) is decreasing
in 00 for m > 0 and s> 0. Also, 7rH(m + s) > 7rH(m - s) for s> 0,
and
7r0(s)7r%(m) + s)r(m . (A14)

ir (m + s) ir%(m - s)
Therefore, denoting the left-hand side of (A13) by Z(0O), we have
(since lrH iS independent of 0O),
Z'(0O) = z'(0O)[rH(m + s) - lrH(m-s)] < 0. Q.E.D.
Lemma 4. Suppose g and h are monotone functions on 9 with either

both increasing or both decreasing and with strict monotonicity on
a set ofpositive Lebesgue measure. LetX be an absolutely continuous
random variable. Then Cov[g(X), h(X)] > 0.
Proof. We can assume, without loss of generality, that Eg(X) = 0. We
also consider only the case in which g and h are increasing, since the
other case is symmetric. Then there exists an xo such that g(xo) = 0.
Now g(x) > 0 implies that x > xo which in turn implies h(x) '
h(xo). Similarly, g(x) < 0 implies that h(x) ? h(xo). Moreover, at
least one of the inequalities on h must be strict over a set of values
of x with positive probability. Therefore, g(x) h(x) > g(x) h(xo) with
strict inequality for a set of positive probability. Consequently,
Cov[g(X), h(X)] = E[g(X)h(X)] > h(xo)Eg(X) = 0. Q.E.D.
Lemma 5. Suppose g and h satisfy the hypotheses of Lemma 4for x

> 0, and g, h, and the density f ofX are symmetric with respect tQx
0. Then Cov[g(X), h(X)] > 0.
Proof. By symmetry, g and h have the same monotonicity for x '
0. Now
E[g(X)h(X)] = F(O)E[g(X)h(X) I X ' 0]

+ [1 - F(0)]E[g(X)h(X) X ' 0]
> F(O)E[g(X) I X ' O]E[h(X) I X ' 0]
495
+ [1 - F(O)]E[g(X) I X ' O]E[h(X) I X ' 0]

by Lemma 4 applied to each conditional expectation. Now, by sym-
metry, it is easy to show that E[g(X) I X ' 0] = E[g(X) I X ' 0] =
E[g(X)]. The result now follows from the preceding inequality.
Q.E.D.
Lemma 6. Suppose g(m, s) and h(m, s) are two functions with the
following properties: g(m) E?[g(m, s) I m] and h(m) = E?[h(m,
s) I m] are symmetric about m = 0, both strictly monotone in the
same direction for m > 0, and Cov[g(m, s), h(m, s) I m] ' Ofor all
m with strict inequality for m =# 0. Then Cov[g(fi, s), h(m, s)] > 0.
Proof. Let g = E?g(m), h = E?hG(m). Then it is easy to check that
Cov[g(m5, s), h(m, s)] = EO[g(mG, s)(h(m, s) -(m))]

+ EO[g(mi, s)(h(m) -h)].
But the first term is the expectation over m- of the covariance con-
ditional on m,. Since the conditional covariance is assumed to be
strictly positive everywhere (except at zero), the first term is positive.
The second term is simply Cov[g(m), h(mz)]. That this is positive
follows from Lemma 5 and the symmetry of Xt (Lemma 1). Q.E.D.
Lemma 7. q1(m) ' qo(m) for all m with strict inequality for m 7# 0.
Proof It is easy to check that q1(0) = qo(0). We provide a detailed

proof of strict inequality only for m < 0. The proof for m > 0 is
symmetric.
First, vt+1 > 0 if and only if st > - mt. Using (A9), it suffices to show
that
r(s I m) ds
1 - T(-m I m)
> m Alffm sr(s I m) ds
> 00 A-,(m, s) T(- m I i) (A15)
It is easy to check that
i|rH(m + s) - rH(m)I < |L7rH(0) -rH(m)I,
for all m < s< -m,
and
IrH(m + S) - lrH(m)I < IH(m -s) - H(m)I,
for all s < m.
496
Therefore,
fm A( s (s I m) ds
_00 T(-m I m)
- rH(m - - r(s I m) ds
s00 T(-m I m)
T(-m I m) - T(m I m)
+l H XH(Om + s) - 7rH(Tm) I m)
J-m -T(-m T(-s I mm)m) d
+ HTH(O) - lrH(m)I T(T(m I m) , (A16)
using a change of variable in the integral. Comparing this to the left-

hand side of (A15), we see that, on each side, we are taking the
expectation of an increasing function of s, namely IlrH(m + s) -
7rH(m) l. The cdf on the left-hand side is simply
1( Im
F (x) = T( 1m- T(-
T(-m II m)
m) for x >- - m. (A17)
The cdf on the right-hand side, G(x), has a mass point of
Go =T(-m I m) - T(m I m) (A18)

T(- mlIm)
atx= -m, and
x(sInm) ds rm r (s I m) ds
G(x)=Go0+ f 4 I = Go +JT(mI)
J^m T(- m I m) Jx T(- m I m)
_T(- m I mn) - T(-x I mn)
T( m)-in T(-x n) m) for x > -im. (A19)
It is easy to check that G is in fact a cdf.

To establish (A15), is suffices to show that F stochastically domi-
nates G in the first-order sense; that is,
F(x) < G(x) for -m x < oo.
Rearranging terms, we see that this is equivalent to
T(-m I m)[l - T(x I m)]> T(-x I m)[l - T(-m I m)].
Substituting from (A4) into the preceding inequality and rearranging

shows that this inequality is equivalent to
ax(am - 1)/a + bo(bm - 1)/f > 0. (A20)

497
This is true, however, since a, and b0 are less than 1 and m < 0.
Q.E.D.
Lemma 8. Cov[v, Airtl mt = m] ' Ofor all m with strict inequality

for m #A 0.
Proof.
Cov[vt, A7rt I Mt= m] = EO[v(m, ?)Ar(m, s) I m] - v(m)q(m)

= T)(m)[q1(m) - q(m)]
' 0 with > for m # 0,
by Lemma 7, since q(m) is a weighted average of q1 (m) and

qo (mi). Q.E.D.
Theorem 1. Cov[v0A, s), A-rx(, s) > 0.
Proof. It easy to check from (A6) that v(m) = v(-im) and that v' is
strictly decreasing in m for m ' 0. Similarly, that q(m) = q(- m)
follows from (A23). To apply Lemma 6, we must show that, for m' >
m > 0, the following expression is positive:
q(m) - q(m') = f [Air(m, s) - Ar(m', s)]r(s I m) ds
+ [Air(mi', s)[r(sI m) - r(sI m')] ds. (A21)
In fact, we can show that each integral in this expression is positive:

Ti
A7r(m', s)[ [r(s I ) - r(s m')] ds
-00
f Ar(m', s)[r(s ) -s m i -(s m')] ds
+ f Air(m', -s)[r(-s m)-( s - in' m')] ds
> J 7('-s)[;r(s I m) - (s I m')] ds
+ f A7r(m', s)[T(-s I i) - r(-s I m')] ds
0,
where the inequality follows from (A3) and (A10), and the final
equality follows from (A3).
498
Now let d(m, s) = rH(m + s) - rH(m). Then
f00[Air(m, s) - Ar(m', s)]r(s m) ds

'00
= f [d(m, s) - d(m', s)]r(s I m) ds
+ f [d(m', -s) - d(m, -s)]r(-s I m) ds
> [d(m, s) - d(m', s) + d(m', -s) - d(m, -s)]r(-sI m) ds,
where the equality follows from a change of variable and the inequal-
ity uses (A3) and (AlO). Therefore, it suffices to show that d(m, s)
- d(m, -s) is decreasing in m for s and m > 0. This is equivalent
to showing that lrH(m + s) - rH(m - s) is decreasing in m for s and
m > 0, or, since
7H(x) = (-ln0)rHr(x)[1 - 7rH(X)],

that
lrH(m + s)[1 - lrH(m + s)] < rH(m - s)[1 - 7rA(m- s)].
This last inequality is easy to check.

This shows that the first integral of (A21) is positive and completes
the proof that q is decreasing in m for m > 0. The theorem now
follows from Lemmas 6 and 7. Q.E.D.
Lemma 9. E 0[Af(m, s) I mt = m, vt+ I > 0] ' E0[Ap(m, s) I Mt = m,

Vt+ =0], for all t and m, with > for m 7# 0.
Proof. Using (4) and the fact that 7r1 (m + s) - 7r (m) < 0 if and
only if r2 (m + s) - 7r2(m) < 0,
AM(m, s) = H- L [1|1r(m + s) - r1 (m)
+ 1 -K2 (m + s) - X2 (m)|]. (A22)

Therefore, it suffices to show
El[ Ir4,(m + s) - 7rX(m)I Imt= m, vt+1 > 0]

> EO[ I 7r.;(m + s) - rj,(m) IMt= m, Vt+1 = 0],
for j = 1,2. But this is exactly (A15) with r-f instead of 1rH. The proof
of Theorem 1 goes through without modification. Q.E.D.
499
Lemma 10. Cov[Ai t(m, s), Air(m, s) I m, = m] ' Ofor all m with strict
inequality for m # 0.
Proof. Using (A22), the proof of Lemma 8 goes through. Q.E.D.
Theorem 2. Cov[Ap0(t, s), Air(A, s)] > 0.

Proof. The same proof as for Theorem 1 (using Lemma 10 instead of
Lemma 8) goes through if we show that EO[A4u(m, s) I m] is symmetric
with respect to m = 0 and decreasing for m > 0. This follows from
the proof for q in Theorem 1 and (A22). Q.E.D.
Theorem 3. (a) The price changes accompanying consecutive trans-

actions are of opposite sign.
(b) If00> 0, then E [Apt I Apt,> 0] < 0 and E0fApt AptP 1 <
0]> O.IfOo<0, thenE iApt I Apt-1> 0]> OandE 0[AptI APt-1<
0]< 0.
Proof. Part (a) was explained in the text. For part (b), we prove only
the first inequality; the proofs of the others are symmetric. The first
inequality is equivalent to EO[Apt I st-, > 0] < 0. To prove this
inequality, first recall that r(s I -im) = r7-s I m), so
E0[APt I -im]
=(H-L) E [ mH(- + s)-xrH(-im) Imr]
= (H- L) lrjH(-m + s) - 7rH(-m)]T(s m -m) ds
_00
= (H -L) f [lrH(-in + s) -7lrH(-in)Yr(-s I in) ds

J00
= (H - L) [7rH(-m - t) - rH(-m)]r(t m i) dt
J-
=(H-L)f [rL(M + t) -rL(M)]T(t I m) dt

-,0
=-EO[AptI m] (A23)
Now,
EO[Apt I St-l > 0]
EO[Apt I m]Xt(m I st-I > 0) dm
500
= EO[Apt I m]Xt(m I st-1 > 0) dm
+ EO[Apt m -]t(-m I st-1 > 0) dm
= J E0[AptI m][Xt(mI st1>0) -Xt(-mI st_1>O)]dm, (A24)
where Xt(m I st-1) is the conditional density of mt given st-1. By

Lemma 3, EO[Apt I m] < 0 for m > 0, so it suffices to show that
Xt(m I St1 >0) > X(-m I st1 >0) foreverym> Oand t> 0.
Consequently, it also suffices to show that
Xt(m I st-1) > Xt(-m I st-,) for every m > 0, st1 > 0, and t > 0.
But, this inequality is equivalent to
r(s I m)Xt(m) > r(s I m)Xt(-m)

or
r(s l m) >T(-S l m) foreverym> Oands> 0,
using Lemma 1. This inequality follows easily from the facts that
ofH(s) > of- (s) for s > 0 and ri (m) > 2 > ir2 (m) for m > 0. Q.E.D.
Theorem 4. Cov[vt+ , vt] > Ofor all t.
Proof We first show that EO(vt+1 I vt > 0) > EO(vt+l I vt = 0). To

show this, we begin by calculating the expectations conditional on
mt-, = n for any n. First suppose n > 0. Then
EO(vt+l I vt > 0, mt-1 = n)
= EO(v+ I m, < 0, m,_1 = n)
= n,=i,i1=n)r(m - n In) dm
Jw EO(vt+l I mt = m, mt-, = n) Ttn|n
00 ~~~~~~T(- n In)
o r(m - n n) dm
= I ro (m)a-m
0- TT(-n n)
N (n)
T (-n In) J-cxa am (m - n) din
T(-n I n) (a + )
(A25)
(a + O
501
The third equality follows from (A6), the fourth equality from (A2),
the fifth from the definition of o-?, and the last from (Al) and (A4).
A similar calculation shows that this formula is also valid for n < 0.
Since the expression does not depend on n, EO(vt1 I vt > 0) =ko
(a + O).
Similarly, for n > 0,
EO(v,+l I vt 0, mt-I = n)
= EO(vt+1 mt > 0, mt1 = n)
00 r()m(m -n n) dm
= 4K)bg iT(-n n)
l0 (n) (0o
l-T(-n
1 T(-n n Jbmo-%(m
I n) 0 b d -n) dm
IH bbn--dm + bman+n d
_l-T(-
___________F
n I n) [ 1a1 f3J
= ~ ~~~ +,
1 - T(- n I n) a +
ab[n(a + + 1]. (A26)
It can be shown that this expectation is symmetric with respect to n

- 0; that is, E0(vt+1 | vt= 0, mt+I = n) = E0(vt+1 I vt= 0, mt+? I
n).
It suffices to show, therefore, that
r(n) b [n(a + f) + 1] < 1, for every n > 0, (A27)
or
ay-n + b-n-n (a + )> 2. (A28)
At n = 0, this is an equality, and it is easy to check that the left-hand

side increases with n for n ' 0.
Now
Cov[vt+1, vt] = EO[vt+l I vt > O]Pr[vt > 0] -E[vt+,]EO[vt]

> EO[vt+1]Pr[vt > 0] - E[vt+]EO[vt] = 0,
where the inequality follows from the inequality established in the
first part of the proof. Q.E.D.
We next calculate the density and cdf of the sum of two signals.
502
Let s- and - be two draws from the true distribution of s- conditional

on R = R E {H, L}. Denote the true density of s1 + - by w( I R).
First, suppose that R = H. Then, for t > 0,
w(t | H) = oH(t - s)oaH (s) ds
= ko[Lf aO-saO ds + fs bO-taO ds + fs a-sbO-s dsl
? ~~ ?(a +: (A29)
For t < 0,
w(t I H) =* kL b-tb6s ds + bO-taOds
+ f a-b-s dsl
For ta+ '<or?Ond0L

= =(2 ktt + - . (A30)
For R = L, one simply reverses the roles of a0 and b0, since o-?L is o%H
with a0 and bo interchanged. The formulas can be stated compactly as
0k2aII(2+ 'tA for t0 and R-H
w(t | R) = H (A31)
( 2 ort 0 and R = H
It is now easy to check, using integration by parts, that
?(+a+f + t) fort-0
w(tI | H = (A32)
0 : ( a +: ) for t < 0,
Also, W(t I L) is the same as W(t I H) with ao and bo interchanged.

We may now prove our next major result.
Theorem 5. E (v+2 I trade is not allowed at t + 1)> ER (v+1.
Proof: We prove this conditional on m = m for any m. The result

then follows by integrating over m. First suppose m < 0. Then
503
EO(vt,, I mt = m and trade is not allowed at t + 1)

7r0(m) a-mk 2 1 2
=- _~~~+ -m
ae ae a + 0
7+rL(m)bmkm(1 2 m)
- rH(m)a-mko( k +
- lr%(m)a-m(l - mko). (A33)
The first equality uses the fact that vt+2 > 0 if and only if st + st+1 >
-m and (A32) to average 1 - W(-m I R) over R = H,L, using
7rH(m) and 7r2(m). The second equality uses (Al) and (A5). Now
EO(vt+l I mt= m) = 1 - T(-m I m) = ir%(m)(ao)m.
The result for m < 0 follows since 1 - mko > 1. The proof for m >
0 is symmetric. Q.E.D.
References
Admati, A., and P. Pfleiderer, 1988, "A Theory of Intraday Patterns: Volume and Price Variability,"
Review of Financial Studies, 1, 3-40.
Amihud, Y., and H. Mendelson, 1987, "Trading Mechanisms and Stock Retturns: An Empirical
Investigation," Journal of Finance, 42, 533-553.
Amihud, Y., and H. Mendelson, 1991, "Volatility, Efficiency, and Trading: Evidence from the Jap-
anese Stock Market," Journal of Finance, 46, 1765-1789.
Bamber, L. S., 1987, "Unexpected Earnings, Firm Size, and Trading Volume Around Quarterly
Earnings Announcements," Accounting Review, 62, 510-532.
Beaver, W., 1968, "The Information Content of Annual Earnings Announcements," Empirical Research
in Accounting, Journal of Accounting Research (Suppl. to vol. 6), 67-92.
Blough, S., 1988, "Differences of Opinion and the Information Value of Prices," working paper,
The Johns Hopkins University.
Blume, L., D. Easley, and M. O'Hara, 1991, "Market Statistics and Technical Analysis: The Role of
Volume," working paper, Cornell University.
Brock, W. A., and A. W. Kleidon, 1992, "Periodic Market Closure and Trading Volume: A Model of
Intraday Bids and Asks," Journal of Economic Dynamics and Control, 16, 451-489.
De Long,J. B., A. Shleifer, L. H. Summers, and R. J. Waldmann, 1990, "Noise Trader Risk in Financial
Markets," Journal of Political Economy, 98, 703-738.
Detemple, J., and S. Murthy, 1992, "Intertemporal Asset Pricing with Heterogeneous Information
and Beliefs," working paper, Columbia University.
Diamond, D. W., and R. E. Verrecchia, 1981, "Information Aggregation in a Noisy Rational Expec-
tations Economy," Journal of Financial Economics, 9, 221-235.
Foster, F. D., and S. Viswanathan, 1990, "A Theory of the Interday Variations in Volume, Variance,
and Trading Costs in Securities Markets," Review of Financial Studies, 3, 593-624.
504
Foster, F. D., and S. Viswanathan, 1993a, "The Effect of Public Information and Competition on
Trading Volume and Price Volatility," Review of Financial Studies, 6, 23-56.
Foster, F. D., and S. Viswanathan, 1993b, "Variations in Trading Volumes, Return Volatility and
Trading Costs: Evidence on Recent Price Formation Models," Journal of Finance, 48, 187-211.
Gallant, A. R., P. E. Rossi, and G. Tauchen, 1992, "Stock Prices and Volume," Review of Financial
Stuidies, 5, 199-242.
Grossman, S. J., and J. E. Stiglitz, 1976, "Information and Competitive Price Systems," American
Economic Review, 66, 246-253.
Grossman, S. J., and J. E. Stiglitz, 1980, "On the Impossibility of Informationally Efficient Markets,"
American Economic Review, 70, 393-408.
Grundy, B. D., and M. McNichols, 1989, "Trade and Revelation of Information through Prices and
Direct Disclosure," Review of Financial Studies, 2, 495-526.
Harris, L., 1987, "Transaction Data Tests of the Mixture of Distributions Hypothesis," Journal of
Financial and Quantitative Analysis, 22, 127-14 1.
Harrison, J. M., and D. M. Kreps, 1978, "Speculative Investor Behavior in a Stock Market with
Heterogeneous Expectations," QuarterlyJournal of Economics, 93, 323-336.
Hellwig, M., 1980, "On the Aggregation of Information in Competitive Markets," Journal of Eco-
nomic Theory, 22, 477-498.
Hindy, A., 1989, "An Equilibrium Model of Futures Markets Dynamics," working paper, MIT.
Holthausen, R. W., and R. E. Verrecchia, 1990, "The Effect of Informedness and Consensus on
Price and Volume Behavior," Accounting Review, 65, 191-208.
Jain, P. C., and G-H. Joh, 1988, "The Dependence between Hourly Prices and Trading Volume,"
Journal of Financial and Quantitative Analysis, 23, 269-283.
Jones, C. M., G. Kaul, and M. L. Lipson, 1991, "Transactions, Volume and Volatility," working paper,
University of Michigan.
Kandel, E., and N. D. Pearson, 1992, "Differential Interpretation of Public Signals and Trade in
Speculative Markets," working paper, University of Rochester.
Karpoff, J. M., 1987, "The Relation between Price Changes and Trading Volume: A Survey," Journal
of Financial and Quantitative Analysis, 22, 109-126.
Kim, O., and R. E. Verrecchia, 1991a, "Trading Volume and Price Reactions to Public Announce-
ments," Journal ofAccounting Research, 29, 302-32 1.
Kim, O., and R. E. Verrecchia, 1991b, "Market Reactions to Anticipated Announcements," Journal
of Financial Economics, 30, 273-309.
Kyle, A., 1985, "Continuous Auctions and Insider Trading," Econometrica, 53, 1315-1335.
Lakonishok, J., and E. Maberly, 1990, "The Weekend Effect: Trading Patterns of Individual and
Institutional Investors," Journal of Finance, 45, 231-243.
Lakonishok, J., and S. Smidt, 1986, "Capital Gains Taxation and Volume of Trading," Journal of
Finance, 41, 951-974.
Lang, L. H. P., R. H. Litzenberger, and V. Madrigal, 1992, "Testing Financial Market Equilibrium
under Asymmetric Information," Journal of Political Economy, 100, 317-348.
Milgrom, P., and N. Stokey, 1982, "Information, Trade and Common Knowledge," Journal of
Economic Theory, 26, 17-27.
Osborne, M. F. M., 1962, "Periodic Structure in the Brownian Motion of the Stock Market," Oper-
ations Research, 10, 345-379.
505
Patell, J. M., and M. A. Wolfson, 1984, "The Intraday Speed of Adjustment of Stock Prices to Earnings
and Dividend Announcements," Journal of Financial Economics, 13, 223-252.
Pfleiderer, P., 1984, "The Volume of Trade and the Variability of Prices: A Framework for Analysis
in Noisy Rational Expectations Equilibria," working paper, Stanford University.
Schwert, G. W., 1989, "Why Does Stock Market Volatility Change over Time?," Journal of Finance,
44, 1115-1153.
Shalen, C. T., 1993, "Volume, Volatility and the Dispersion of Beliefs," Review of Financial Studies,
6, 405-434.
Shleifer, A., and L. H. Summers, 1990, "The Noise TraderApproach to Finance,"Journal of Economic
Perspectives, 4, 19-33.
Varian, H. R., 1985, "Divergence of Opinion in Complete Markets," Journal of Finance, 40, 309-
317.
Varian, H. R., 1989, "Differences of Opinion in Financial Markets," in Courtenay C. Stone (ed.),
Financial Risk: Theory, Evidence and Implications, Proceedings of the EleventhAnnual Economic
Policy Conference of the Federal Reserve Bank of St. Louis, Kluwer, Boston, 3-37.
Wang, J., 1992, "A Model of Competitive Stock Trading Volume," working paper, MIT.
Ziebart, D. A., 1990, "The Association Between Consensus of Beliefs and Trading Activity Sur-
rounding Earnings Announcements," Accounting Review, 65, 477-488.
506

Harris Raviv 1993

Uploaded by

Copyright:

Available Formats

Harris Raviv 1993

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Harris Raviv 1993

Uploaded by

Copyright:

Available Formats

The Society for Financial Studies

Differences of Opinion Make a Horse Race

A model of trading in speculative markets is devel-

We develop a model of trading in speculative markets

The Review of Financial Studies 1993 Volume 6, number 3, pp. 473-506

ities that have been documented concerning the relationship between

1 See also Bamber (1987) and Ziebart (1990).

* Absolute price changes and volume are positively correlated.

The literature on asset markets has, to a great extent, focused on

Section 1 presents the formal model in detail. Section 2 derives

Here we describe a model of trade based on differences of opinion,

Hro(St) = = O( = [1 + ]- Hr(m), (2)

where m = s1 + . . . + St is the cumulative signal and 00 = b0/a0. In

with risk-averse speculators, whenever one group's beliefs change

16 See Section 3 for a discussion of the risk-averse case.

7 This can be proved by backward induction. See Lemma 2 in the Appendix.

which group is the designated price taker as long as it is the same

Pt= p(m) = HIrH(m) + LIrL(m) = (H - L)7rH(m) + L, (3)

2. Equilibrium Volume and Price Dynamics

Having described a simple model based on heterogeneous beliefs,

Theorem 1. Absolute price changes and speculative volume are pos-

I - / ~~~~~~S < -M |-M < S < O | < S < M |S > M I

This result is similar to results in Pfleiderer (1984), Blume et al.

,u(m) (0.5)[irl(m) + ir(m)](H - L) + L.

The absolute change in mean forecast from tto t + 1 given cumulative

signal m and current signal s is

A\t(m, s) = I,.t(m + s) - ufm)

+ 7H2 (mr + s) - 72 (m) (.

From the equation, it is clear that revisions in forecasts are related to

Theorem 2. Absolute changes in the mean forecast of thefinalpayoff

Note that no causality is implied by this relationship between vol-

Theorem 3. (a) The price changes accompanying consecutive trans-

zero. Thus, in consecutive transactions, m crosses zero in the opposite

Theorem 4. Volume is positively autocorrelated.

This result was also obtained by Foster and Viswanathan (1993a)

Theorem 5. Average volume when two signals are released between

Theorem 5 has several empirical applications. For example, the

max Ei[- err(Rq+b) I m]

In this problem, E is the expectation using group j's model, and r

Q*(m) = pm/[2r(H -L)],

Theorem 3(b) on the serial correlation of price changes, however,

We provide a model of speculative trading volume and price dynam-

7H( m) = ,ajrbm and -(m)=-x(-m) (Al)

r(s I m) = o%(s)THr(m) + o2L(s)ir(m) = O( + s) (A2)

Tr(s -m) = r(-s I m),

r(-s l m) > r(-s l m') V m' > m > O, s > O, (A3)

Third, the cumulative distribution function (cdf) corresponding to

[ko[ a a L(m) + O H (m) for s < o,

{1-ko [aH + ( ] for s > O.

In this expression, a =-ln(ao), fi =-ln(bo), and

1= Cs (as + bs) ds=- + -1 (A5)

Fourth, let v(m, s) be volume at t + 1 if mt = m and st = s 1 if m

v)(m) = Pr(vt+1 > 0 mt = m) = EO(vt+1 Mt= m)

(m){ for mr'O (A6)

Here, EO denotes the expectation with respect to the true distribution

Lemma 1. Xt (m) = Xt (- m) for all m and t.

Proof. The proof is by induction on t. For t = 1, Xt(m) = ((m I 0) =

fr(m -xI x)Xn-,(x) dx

= J'Xn(-m I m.-, = x)Xn-(x) dx = Xn(-m) (A7)