You are on page 1of 6

The Stochastic Dependence of Security Price Changes and Transaction Volumes in a Model

with Temporally Dependent Price Changes


Author(s): Thomas W. Epps
Source: Journal of the American Statistical Association, Vol. 71, No. 356 (Dec., 1976), pp.
830-834
Published by: Taylor & Francis, Ltd. on behalf of the American Statistical Association
Stable URL: http://www.jstor.org/stable/2286846
Accessed: 15-10-2016 04:43 UTC
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted
digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about
JSTOR, please contact support@jstor.org.

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

Taylor & Francis, Ltd., American Statistical Association are collaborating with JSTOR to digitize,
preserve and extend access to Journal of the American Statistical Association

This content downloaded from 130.194.20.173 on Sat, 15 Oct 2016 04:43:45 UTC
All use subject to http://about.jstor.org/terms

The Stochastic Dependence of Security Price


Changes and Transaction Volumes in a Model
with Temporally Dependent Price Changes
THOMAS W. EPPS*

A model of security price behavior is presented, which allows for the


observed temporal dependence in price changes from one transaction
to the next and for the stochastic dependence between price changes

and the numbers of shares traded. Tests of the model are conducted

not overseen by specialists charged with moderating


fluctuations in price. Again we find strong evidence for
the mixture model, both for bonds and for stocks, but we

with data for corporate bonds and stocks. The results reveal some

do detect some interesting differences in the two price

interesting differences in the two price processes.

processes.

1. INTRODUCTION
In an earlier paper [4] a model of securities markets

2. MODIFICATION AND TESTING OF THE MIXTURE


MODEL

was introduced which implies that the change in the

To estimate the mixing parameter, -y, in (1.1) by

logarithm of price from one transaction to the next

maximum-likelihood methods,2 one must in practice as-

follows a mixture of probability distributions, with trans-

action volume as the mixing variable. Specifically, under

sume that the Et are independently and identically distributed. An important difficulty with (1.1) is that it

a set of assumptions about the behavior of individuals

implies independence of the yt (conditional on volume)

and the nature of markets,' it was shown that the change

if the Et are independent. As noted in [4], this appears


in the logarithm of price, yt, between the tth and (t - 1)th
inconsistent with the findings of Niederhoffer and Osborne
transactions can be expressed as
[11] that price changes on individual stock transactions
yt

VtEt

,,

are strongly dependent. (Brada et al. [2] and Granger

(1.1)

and Morgenstern [7] have contributed additional evi-

where vt is the number of shares traded in the tth trans-

dence of this phenomenon.) Runs tests and autocorrela-

action, y is a positive constant, and Et is a random disturbance with zero mean and constant variance. As long

tion functions for the transactions in a sampling of the


bonds considered in the present study also reveal a strong

as y is different from zero, the variance of yt (conditional


temporal dependence among the yt.
on v) is a function of volume, which acts as a mixing
An alternative model, which preserves the essential
variable. Tests of the model with individual transactions
features of (1.1) and which lends itself to empirical testdata for twenty common stocks showed -y to be signifiing would suppose that the yt result from an autoregrescantly greater than zero; and an examination of the

residuals, it = ytvt-7 showed that their empirical distributions typically possess thinner tails than the normal.

In this paper, a modified nmodel is presented which ac-

sive, moving-average (ARMA) process, 4 (B)yt = e (B)at,

where b (B) = 1 - 0B- _ . . .- 4pBP, 0 (B) = 1

- 01B- 02B2 ...- OqBq, and Bkyt stands for Yt-k,


k = 1, 2, ..., p. In this case, the at play the role of the

counts for the dependence in the yt, which has been ob- Yt in (1.1), having variance which depends on vt, i.e.,

served by several researchers. The model is tested with

data for corporate bonds as well as with the data for the
20 common stocks. Fundamental differences in the price
processes for bonds and stocks are suspected because of

bonds' limited lifetimes and because trading in bonds is


* Thomas W. tpps is Assistant Professor, Department of Economics, University
of Virginia, Charlottesville, VA 22901. Criticisms of an earlier version of this paper
by an associate editor and two referees are acknowledged with appreciation.
1 Major assumptions are (1) that individuals behave as if they attempt to maximize utility functions which depend on the expected values and variances of their
portfolios' end-of-period values (EPV); (2) that a certain measure of risk aversion
is constant; (3) that markets are perfect; (4) that transactions in any security occur

between two equal groups of traders, whose compositions, in general, change from
one transaction to the next as their members place different interpretations on the

new information which generates the transactions; and (5) that the change which

at = vt7Et . (2.1)

If -y is small-as later results indicate-then at is very

nearly white noise, and Box-Jenkins [1] model-identification techniques applied to the Yt can be used to make

tentative identification of an ARMA model wvhich accounts

satisfactorily for the observed dependence in the yt.


Tests of the modified mixture model were conducted
with individual-transactions data for 42 corporate bonds
2 An alternative estimation procedure using regression was considered in [4] but
rejected because of substantial biases that would exist even in large samples.

a Journal of the American Statistical Association

a given news shock produces in the extent of disagreement (about expected EPV)

December 1976, Volume 71, Number 356

between the two groups of traders tends to increase with the magnitude of the

Applications Section

average change in EPv-the average being taken over all traders in the market.

830

This content downloaded from 130.194.20.173 on Sat, 15 Oct 2016 04:43:45 UTC
All use subject to http://about.jstor.org/terms

Security Price Changes and Transaction Volumes 831


1. Likelihood Estimates of Parameters of (2.1)-(2.2) and SummaryStatistics for Residuals forForty-Two Bonds

Bondg
GAC

ja

(600)

&2X

.137*

105a

.750*

11~~~~~~~~~~~~~~~~~~~~~~~~~~ .

Oa

.242*

Oa

.626*

SRb

12.83*

KC

a97d

11.40*

J (600) 139* 1.849* .239* .552* 13.29* 12.91*


NTC (600) .046 1.441* .535* .750* 8.87* 5.28*

x2e

1.54*

58.3*

1.52*
1.45*

58.8*
69.1*

OT (600) -.005 1.428* .151 .390* 10.39* 6.06* 1.49* 176.1*


RDR (600) .088* 1 343* .417* .663* 7.22* 4.06* 1.36* 90.4*

PN
C

(4497)

(180)

CGI

.026*

.117*

(107)

.662*

.230*

.453*

.179

.568*

5.75

4.66*
3.22

-.871

5.93*

CMI (126) .064 1.531* -.418


CR (145) -.058 4.722* -.451

-.225
-.323

5.38
4.38

(170)

.799*

11.04f

-.840*

CTC

.625*

.722*

.019

.727*

.532*

.876*

1.79*

304.3*-

1.62*

20.02*

3.70

3.42
2.76

5.74

1.70

28.86*

1.37*
1.42*

27.7*
44.0*

3.65

1.74

14.7

ESL (118) .336* 1.903* -.287 -.131 5.03 4.14* 1.03* 84.4*
FAS (196) .120 3.497* -.289 .025 6.36* 3.46 1.78 6.5
G

(339)

.114*

.978*

-.636*

-.546*

6.24

3.94*

1.33*

171.5*

GO (115) .259* .437* -.398* .013 6.34* 6.10* 1.23* 54.2*


KMG (105) .131 1.784* -.860* -.731* 5.40 3.33 1.68 23.3*
LIT (99) -.064 5.780* -.777 -.731 6.39* 5.18* 1.56f
MDS (236) -.070 4.318* -.257 .153 7.74* 5.36* 1.53*
MGM (134) .067 8.811* -.865* -.730* 7.58* 7.75* 1.41*
NTT (187) .081 .537* .428* .698* 7.17* 5.15* 1.35*
RCA (146) .074 1.305* .427* .739* 6.73* 5.11* 1.42*

42.9*
67.7*
23.1*
41.4*
20.4*

RHR (87) .265* 2.423* -.842 -.862* 5.70 5.33* 1.09f 46.8*
SAA (206) -.029 7.928* .108 .276 6.12** 4.17* 1.44* 48.8*
UAL (215) .378* .876* - 277* -.127 7.80* 6.00* 1.35* 82.2*
UMM (172) -.021 1.798* .480* .763* 7.55* 5.19* 1.56* 24.6*
CMB

(86)

.156*

1.678*

-.684*

-.465

5.18

4.24*

1.43f

10.5

CWE (130) .005 1.168* .414* .763* 5.03 3.10 1.69 15.1
DOW (73) .233* .515* .776* .812* 4.52 3.62 1.13f 61.7*
F (171) -.047 1.851* .170 .623* 6.00** 3.40 1.97 8.6
FCR

(104)

GMA

.013

(154)

1.167*

.159*

.636*

.504*

.887*

.490*

5.58**

.870*

4.59

3.82

1.61f

7.7

2.95

1.63*

8.8

JCP (60) .459* .223 .215 .339* 5.16* 4.53 1.00f 55.9*
NCR (256) .099* 1.067* .265* .538* 7.75* 6.57* 1.49* 51.6*
NJT (152) -.077 4.088* .024 .521* 11.30* 25.00* 1.41* 73.3*
s (93) .131 1.220* .334* .658* 5.69* 4.56* 1.29f 20.2*

SCB (129) .014 .747* .313 .542* 6.39* 4.71* 1.45* 21.1*

SFF (107) .170* .706* -.227 .090 5.18 3.61 1.55f 7.48
SGN (255) .131* .743* .283* .641* 5.48 3.32 1.69* 18.5**
SN (72) .068 3.012* .044 .541* 4.74 2.83 2.00 12.4
SWT (170) .021 1.096* *575* .839* 7.13* 5.82* 1.28* 51.7*
T

(1001)

UCL

Average
a

.314*

(96)

.109

Entries

.009*

-.096

.076*

1.358*

.370*

.253

12.48*

.538

12.09*

5.63**

1.00*

4.35*

1366.*

1.14f

32.3*

1.886

with

asterisks

are

more

than

two

standard

errors

fromn

zero.

bStudentized range. One and two asterisks imply rejection of normality hypothesis at .05 and .10 levels. (David, et al. [3]).

c Sample kurtosis. Entries with asterisks are more than two standard errors above 3.0, where var(k) = 24/n. (Kendall and Stuart, [9, p. 243 J).
d Characteristic exponent, computed as in Fama and Roll [5, 6]. Entries with asterisks are significantly less than 2.0 at .05 level, based on sampling distributions in [6].
Chi-square test for normality, based on 12 classes (11 df). One and two asterisks imply rejection of Ho at .05 and .10 levels.
Significance cannot be determined from available tables.
x Sample sizes in parentheses.

identified in the first column of Table 1 by the ticker

ARMA model. Autocorrelation and partial autocorrelation

symbols of the stocks of the issuing companies.3 Data for

functions computed from the yt series for the five bonds

the first five bonds in Table 1 (for which large samples

showed that they could be well explained by either

had been obtained) were used to identify an appropriate

ARMA (1, 1) or ARMA (0, 2) models.4 Comparison of the


residuals from fitting the general ARMA (1, 2) model and
on
a subease, the ARMA (1, 1) model, showed that the latter

I A detailed listing of the actual bonds comprising the sample will be furnished
request. The first 25 bonds were drawn at random from a population of 147 New
York Stock Exchange (NYSE) bonds in which at least five transactions occurred on

each of two randomly selected trading days in January 1971. The remainder were

drawn from a population of 78 NYSE bonds which had asked prices equal to at least
90 percent of par and of which at least 100,000 were outstanding as of the first trad-

lag day in 1971. For the first five bonds, the first 600 transactions in 1971 were
recorded. For the sixth bond, PN, all transactions in the first 10 months of 1971 were
usd; and for the remaining bonds all transactions occurring in January 1971 were
recorded. The average number of years until maturity for the 42 bonds was slightly

performed at least as well for all five bonds.

The parameters y, a2, 4, and 0 of the model given by


(2.1) and
yt - oYt-i + at - at- (2.2)
were estimated simultaneously by finding the maximum

more than 20.0. Eight bonds had remaining lifetimes of 3 to 6 years; 25, of 15 to 25
years; and 9, of 25 to 39 years.

4ARMA (p, q) models contain p autoregressive and q moving-average parametes.

This content downloaded from 130.194.20.173 on Sat, 15 Oct 2016 04:43:45 UTC
All use subject to http://about.jstor.org/terms

832 Journal of the American Statistical Association, December 1976


2. Characteristics of Empirical Distributions of Sums of Maximum Likelihood Residuals for Bond PNa
SRb

Number of Sample

elements size Natural Random Natural Random Natural Random Natural Random
in sum order order order order order order order order
1 4474 11 .05f 11.05a 4.66* 4.66* 1.79* 1.79* 304.3* 304.3*

2 2237 8.96f 8.70a 4.23* 3.92* 1.68* 1 .9f 172.2* 142.6*


5 894 7.41* 6.78 3.88* 3.24 1.68* 1.80* 64.9* 63.0*

8 559 8.05* 8.05* 4.80* 3.61 * 1.62* 2.00 51.5* 33.7*


10 447 7.56* 5.91 4.70* 3.00 1.66* 1.88 41.9* 37.4*
16 279 6.90* 5.81 4.10* 3.06 1.63* 2.00 26.2* 22.9*

20 223 8.50* 5.92 6.16* 3.08 1.50* 2.00 31.9* 16.4


24 186 7.30* 4.74 5.95* 2.53 1.60* 2.00 21.0* 17.7**
30 149 8.11* 5.42 6.25* 3.06 1.67** 1.77 18.2** 12.6

36 124 7.17* 4.97 5.61* 2.65 1.72 2.00 13.0 19.3


40 111 6.14* 5.10 5.75* 2.53 1.55* 2.00 21.5* 11.0
46 97 5.92* 4.52 5.11* 2.44 1.90 2.00 9.0 8.9
48 93 6.56* 5.06 4.63* 3.10 1.65f 1.83 15.8 11.4
50 89 7.42* 5.01 6.50* 2.60 1 .58f 2.00 8.3 5.8
52 86 6.93* 4.56 5.32* 2.52 1.85 2.00 9.6 11.0
54 82 5.57** 5.08 3.74 3.17 1.56f 1.69 16.0 6.1
56 79 5.58** 5.20 4.18* 3.15 1.65f 1.96 10.4 5.8
58 77 6.36 4.27 4.75* 2.28 1.41f 2.00 9.0 9.0

60
62
64
*

74
72
69

7.15*
6.65*
5.97*

Entries

in

4.91
3.91
5.03

"Natural

5.77* 3.25 1.68 1.87 6.4


5.33* 2.00 1.79 2.00 7.9
4.16 2.99 1.77 2.00 14.3
order"

column

are

for

6.7
7.2
6.8

residuals

ordered

random order.

hx,d,e., See footnotes to Table 1.

of the log-likelihood function5

tistics in the table reflect these properties: the Studentized-range test leads to rejection of the normality hy(n -2) n
pothesis in 27 of the 42 cases; virtually all the sample
2 t=3
kurtoses are greater than 3.0; and estimates of characn
t
teristic exponents are significantly less than two in most
- - {V 2EY[, Oti(y1 - , (2.3)
cases. Thus, if forced to choose between the normal dis2 t=3 j=l
tribution and an infinite-variance member of the stable
where n is the number of transactions
thewould
sample.
class of distributions, in
the latter
seem better able
Parameter estimates for the 42 bonds are shown in
to explain the behavior of bonds residuals. There is, of
Table 1, along with estimates of standard errors derivedcourse, no reason besides convenience to confine attention
from the information matrix. Estimates of -y have theto members of the stable class, since the disturbances are
expected sign in 33 cases, of which 19 are more than twonot presumed to be aggregates of other random variates.
standard errors from zero. None of the estimates with
As a test for stability, distributions of nonoverlapping
wrong signs is significant. In addition, 25 estimates of ksums of up to 64 residuals for Bond PN were examined.
and 30 estimates of 0 are more than two standard errorsAs shown in Table 2 by the summary statistics in columns
from zero. Model (2.1)-(2.2) thus seems to be a useful
headed "Natural Order," there is no sign of convergence
way of characterizing the price process for bonds.
to the normal. The sample kurtosis behaves erratically,
Residuals from the model were computed from
and while the d values appear also to be erratic, the
fluctuations are not larger than would be expected due to
sampling error (see [6, Table 2]). Thus, the stablet = Vt 8E (y - 06- 1)
t-20
Paretian law might be a reasonably good model for the
disturbances.
For most of the bonds the empirical distributions
of the
a more plausible explanation for the fact
residuals are characterized by excessive frequency However,
near
that stathe distributions of sums of the residuals do not
the mean and in the tails (beyond ?E 3a). Summary
converge to the normal is that some specification error
inwas
Model
(2.1)- (2.2) has caused them to be heterogeneous
5 To reduce sizeable computational co8ts, the summation in brackets
evaluated from t - 20 to t. If! 0I S 0.8, as it appears to be in each case, the simplification
or
dependent.6
While Box-Pierce statistics computed from
introduces an error in the summation of no more than one percent. Remaining sums
in (2.3) were evaluated from 23 to n, and the factor involving n in the first term was
taken to be (n - 22)/2. For the first five bonds in Table 1, the log-likelihood
6 The funccomplexity of the estimation problem makes purely exploratory changes in
tion was computed over a grid of admissable values of -y, c0, and 0, andthe
was
found
model
an to
infeasible way of deciding whether a specificaticn error has been com-

be quite well-behaved.

mitted. A referee suggested that the formulation of the model in terms of logarithms

This content downloaded from 130.194.20.173 on Sat, 15 Oct 2016 04:43:45 UTC
All use subject to http://about.jstor.org/terms

Security Price Changes and Transaction Volumes 833


3. Likelihood Estimates of Parameters of (2.1)-(2.2) and Summary
Statistics for Residuals for Twenty Common Stocks

Stockg *a &2 X 105a Oa &a SRb kc &.97d x2e


oi (215) - .009 1.268* .479 .649* 6.24** 4.07* 1.55* 46.1 *
ARC (295) .107* .461 * -.365* -.021 4.56 2.23 2.00 41.7*
J (606) .133* .172* .149* .650* 4.92 2.48 2.00 26.6*
T (1077) .123* .157* .125* .680* 5.89 2.53 2.00 65.1*
MHC (272) .152* .992* .001 .306* 5.52 2.88 1.79 12.03

EMR

SOB

(201)

(202)

.042

-.012

.589*

.642*

.395

.220

.598*

.355

4.57

5.99

2.94

3.81*

1.56*

1.45*

KO (201) .164* .436* -.238 .071 9.04* 7.93* 1.97


WLA (208) -.034 .386* -.284 .058 5.22 2.99 2.00
EK (506) .147* .425* .271* .562* 5.18 2.23 2.00
PIN

(215)

JPM
SOH

.023

(304)
(232)

1.589*

.054
.051

.364

.767*
.604*

.561*

.640*
-.335

4.62

.712*

-.276

2.81

5.26

5.56

1.50*

2.10

3.41

73.6*

155.0*
26.7*
54.4*
49.2*

11

1.5*

2.00

245.3*

1.45*

172.1*

DNB (194) -.038 .918* -.535 -.381 6.42* 4.51 * 1.28* 95.7*
uc

(298)

.042

.868*

.191

.421

4.83

2.60

1.80

61.8*

AA (367) .050 .748* .784* .689* 7.27* 3.20 1.75* 255.9*


ABT (235) .130* .962* -.235 -.001 4.82 2.40 2.00 29.0*
SLB (180) .093 .336* -.723* -.607* 4.58 2.94 1.45* 102.7*
RXM (225) .032 1.270* .869* .825* 6.55* 3.33 1.76** 60.9*
RT (231) .085* .732* .157 .304 5.34 2.51 2.00 82.7*
Average

.067

a*b-c.d,e,g

See

.716
footnotes

to

Table

1.

3. CONCLUSIONS

the PN residuals for 12, 24, and 36 lags do not imply


rejection of the independence hypothesis at even the .10

The mixing model, augmented to account for temporal

level, the runs test-a more appropriate test in view of

dependence among price changes, performs well in tests

the probable nonhomogeneity of the residuals-gives

with bond and stock transactions data, offering further

strong indication of temporal dependence.7 Furthermore,

support of the hypothesis that the variance of the price-

when the residuals are placed in random order and the

change process is random and determined by transaction

sums test repeated, there is clear indication of rapid con-

volume. There are, however, some noteworthy differences

vergence to the normal,8 as shown by summary statistics


in columns labeled "Random Order" in Table 2.9

Table 3 shows the results of applying Model (2.1)- (2.2)

between the price processes in the two markets. Tables 1


and 3 show that the averages of the estimates of -y and a2

are considerably smaller for stocks than for bonds, in-

dicating that the shocks at in (2.1) possess larger condito the stock transactions data studied in [43. Again the
tional variance in the market for bonds. This difference
modified mixing model is well supported by the data:
undoubtedly reflects-in addition to the relative "thin16 of the 20 estimates of -y have the expected sign, and
ness" of the bond market10-the presence of specialists in
eight of these are more than two standard errors from
the market for stocks. Specialists may in effect reduce
zero. As the summary statistics suggest and direct exboth -y and c2 by keeping the bid-asked spread small and
amination of the empirical distributions shows, distribu-

helping to absorb large orders, which otherwise could not

tions of the residuals do not have thicker tails than the

be executed quickly without substantial price fluctuation.

normal, on average.

The specialists' actions may also help to explain the rela-

tively thin tails in the distributions of the et for stocks,


of prices may have been responsible for the nonconvergence of the distributions of
sums. Most studies of the price process have dealt with changes in log price because
of the presumption that traders concern themselves with the proportionate effects,

rather than the absolute effects, of price changes. (For discussion of this point, see
[7, 12, 10].) However, as a test of the suitability of the log formulation in the
present case, the model was respecified in terms of changes in absolute price and
refitted with the data for bond PN. Estimates of -y, ,, and 0 were virtually identical
to those obtained with logs, and the distributions of sums of residuals behaved in
the same manner.

7 The number of runs in the series of 4,474 residuals was 1,565, yielding a value
of the test statistic (distributed as standard normal under Ho) of -20.1.
8 See Hsu, et al. [8] for discussion of the randomization technique as a test for
stability.

which imply lower (conditional) probabilities of very

large price changes (measured in standard-deviation


units).

Based on the estimates of -y and a2 for bonds, it appears


that the variance of the at process (conditional on vt)
declines as a bond approaches maturity. For the eight
bonds in the sample with remaining lives of six years or

less, the average estimates of -y and a2 were 0.032 and


1.16 X 10- versus corresponding averages of 0.127 and
2.09 X 10- for the remaining bonds in the sample (all

9 As pointed out in [4], predicating the estimation of y on the assumption of


normal disturbances may bins in favor of the normality hypothesis the inferences
drawn about the disturbances from the maximum-likelibood residuals. A referee
has suggested that, for comparison, likelihood estimation of y be undertaken by
assuming that the et follow the Cauchy distribution-a suggestion which I plan to

10 A "thin" market is one in which the number of traders is small. The variance
of f 2) is inversely related to the number of traders, as shown by the definition of

pursue in a subsequent study.

f in [43.

This content downloaded from 130.194.20.173 on Sat, 15 Oct 2016 04:43:45 UTC
All use subject to http://about.jstor.org/terms

834 Journal of the American Statistical Association, December 1976


for the Mixture-of-Distributions Hypothesis," Econometrica,

of which had at least 15 years until maturity)." Such a

tendency for the parameters y and o2 to decrease over


time -(which implies an error in the specification of Model
(2.1)-(2.2)) might account in part for the thick tails of

the distributions of the Et for bonds.


[Received March 1975. Revised May 1976.]

REFERENCES

44 (March 1976), 305-21.

[5] Fama, E.F. and Roll, R., "Some Properties of Symmetric


Stable Distributions," Journal of the American Statistical
Association, 63 (September 1968), 817-36.

[6] , "Parameter Estimates for Symmetric Stable Distributions," Journal of the American Statistical Association, 66 (Jun
1971), 331-8.

[7] Granger, C.W.J. and Morgenstern, O., Predictability of Stock


Market Prices, Lexington, Mass.: D.C. Heath & Company,
1970.

[1] Box, G.E.P. and Jenkins, G.M., Time Series Analysis: Forecasting and Control, San Francisco: Holden-Day, Inc., 1970.

[8] Hsu, D., Miller, R.B., and Wichern, D.W., "On the Stable
Paretian Behavior of Stock-Market Prices," Journal of the
[2] Brada, J., Ernst, H., and Van Tassel, J., "The Distribution of
American Statistical Association, 69 (March 1974), 108-13.
Stock Price Differences: Gaussian After All?" Operations Re[9] Kendall, M.G. and Stuart, A., The Advanced Theory of Statistics,
search, 14 (March-April 1966), 334-40.

[3] David, H.A., Hartley, H.O., and Pearson, E.S., "The Distribution of the Ratio, in, a Single Normal Sample, of Range to
Standard Deviation," Biometrika, 41 (December 1954), 482-93.
[4] Epps, T.W. and Epps, M.L., "The Stochastic Dependence of
Security Price Changes and Transaction Volumes: Implications

Vol. I, 2nd ed., New York: Hafner Publishing Co., 1969.

[10] Moore, A.B., "Some Characteristics of Changes in Common


Stock Prices," in P. Cootner, ed., The Random Character of
Stock- Market Prices, Cambridge, Mass.: M.I.T. Press, 1964.
[11] Niederhoffer, V. and Osborne, M.F.M., "Market Making and
Reversal on the Stock Exchange," Joural of the American
Statistical Assbciation, 61 (December 1966), 897-916.

11 The relationships of y and O2 with maturity (if they do exist) are apparently
nonlinear: correlations between ' and years to maturity (T) and between a2 and T
for all 42 bonds were only 0.095 and 0.042, respectively.

[12] Osborne, M.F.M., "Brownian Motion in the Stock Market,"


Operations Research, 7 (March-April 1959), 145-73.

This content downloaded from 130.194.20.173 on Sat, 15 Oct 2016 04:43:45 UTC
All use subject to http://about.jstor.org/terms