You are on page 1of 7

EDWARD F.

RENSHAW

Selecting
A Defensive
Portfolio

SINCE THE PUBLICATION OF Markowitz's article ance model is used here to obtain a reasonably effi-
and monograph on "Portfoho Selection" [4, 5], there cient portfolio to combine with a safe asset yielding
have been a number of efforts to simplify the pro- four percent in the period 1947 to 1965. The result-
cedures for estimating an efficient portfolio [6, 8, 9] ing portfolio was then compared with some simu-
and several studies which indicate that there may lated portfolios which were randomly derived on
be some degree of predictive content to portfolios the basis of a model that was suggested by Osborne
that are obtained from historical data. Papers by [2:113]. Our main conclusions were:
Smith [10, 11], Cohen and Pogue [1] have used » that distributions of stock returns which appear to
variations of the Sharpe-Markowitz programs to ob- conform to Osbome's random walk model II can hide a
tain portfolios which have performed so well in sub- significant amount of internal structure
sequent periods as to be possibly inconsistent with • that this structure may be of practical importance in
the random walk hypothesis. helping to identify portfolios that will be more defensive
In a more recent paper Treynor and others [13] in a bear market than a representative price index.
have shown that the risk parameters for stocks in One of the more interesting findings to emerge
one mutual fund have exhibited remarkable stabil- from tliis study was the discovery tliat it is not easy
ity. They were able to use historical parameters to to identify defensive securities using standard statis-
explain about half of the month-to-month variation tical techniques. The ten stocks in the DJIA with
in the price of this fund, relative to a market aver- the least variable returns in the 1947 to 1965 period
age, over a two-year period. The implication is that actually performed less well than the industrial
if we can predict the direction of the market, it average during the bear markets of 1966 and the
should be possible to select a portfolio that will out- first six months of 1969. Those stocks in the DJIA
perform a representative market average. witli the lowest correlation with Standard and
Further evidence in support of parameter stability Poor's industrial index performed even less well, on
is provided by a recent simulation of the Dow Jones the average, than the stocks with the lowest stan-
Industrial Average [7]. A simplified portfolio bal- dard deviations. The ten stocks with the highest

WINTER / 1970 / VOL. XIII / NO. 2 19


TABLE I.—A COMPARISON OF SECURITY PROPORTIONS AND OIHER SELECTED PARAMETERS USED IN PORTFOLIO BALANCE
MODELS. T H E DOW JONES INDUSTRIAL AVERAGE 1947-1965

HealJzerJ Correlation The Least Squares


Standard Security Proportions for Portfolios
Security Average Ke- With S & P's Market Intercept
turn Index or Constant Term 1 3 .5

Eastman Kodak 1,207 ,203 ,442 1,118 ,231 .176 ,182 .128 ,089
General Foods 1.171 ,193 ,397 1,095 .173 .134 .140 .180 .249
General Electric 1.180 ,172 ,697 1,061 .126 ,086 ,054 ,095
United Aircraft 1.249 ,300 ,376 1,137 .123 ,147 ,246 .080 ,020
Procter & Gamble 1,158 ,196 ,456 1,069 .092 ,083 ,074 .201 ,197
Texaco 1,204 ,192 ,793 1.053 .079 ,069 ,033 ,144 ,172
Westinghouse Electric 1,152 ,'239 ,334 1.073 .068 ,072 ,084
Standard Oil/California 1,174 ,180 ,686' 1.052 .057 ,0«3 .029 ,088
Chrysler 1,203 ,387 .275 1,098 .041 ,074 .146 ,024
Aroerican Can 1,105 ,165 .368 1,045 .017 .031 .012 .030
Sears Roebuck 1,190 ,214 .717 1,037 ,034 .023
Standard Oil/N.J. 1,185 ,241 .640 1.032 ,022 .034
Swift 1,092 ,162 .856 1,035 ,008
AT&T 1,097 ,140 .473 1,031 ,000* .096 .028
American Tobacco 1,111 ,205 .411 1,028 .110

• The security proportion was positive but less than one tenth of one percent.

average rates of return in the nineteen-year period


studied did outperform the DJIA in the first half of
Presumed-Defensive Comparisons
1969 but did not in 1966. Table I contains security proportions and other
The defensive quality of the Sharpe-Markowitz selected parameters obtained for stocks in the Dow
portfolio in botli of these bear markets would appear Jones Industrial Average during the nineteen-year
to depend, therefore, on a judicious combination of period from 1947-196.5. The first portfolio was ob-
these three parameters rather than a dominant em- tained by using the Shai-pe-Markowdtz security se-
phasis on any one of the more basic elements which lection program. It would have been the best port-
help to determine an efficient portfolio. folio to have combined with a safe asset yielding
None of the basic parameters was particularly four percent during the original fitting period. (To-
useful as an individual indicator in helping to iden- bin has shown that there may be only one point on
tify defensive securities, but it was discovered that the Sharpe-Markowitz investment frontier that can
the constant terms, or market intercepts, obtained be considered efficient if risk assets can be com-
when the returns for individual securities are re- bined with money or a safe asset yielding a fixed
gressed on the percentage returns for a representa- return per period [12].)
tive market average—are quite useful in helping to Average rates of price appreciation for this port-
predict those stocks which are likely to appear in an folio from 1966 through 1969 are presented in Table
efficient portfolio. The endeavor in this paper is to II and can be compared to the average rate of price
sho^v why this is so and to indicate ways in which appreciation for the DJIA as a whole. The Sharpe-
the market intercept can be redefined so as to ob- Markowitz model (portfolio 1) had an average re-
tain an even better definition of defensive securities. turn which was ten percentage points better than
Our most important conclusion is that it may be the DJIA in 1966 and seven percentage points bet-
possible to simplify the estimation procedures to the ter in thefirsthalf of 1969.
point where most analysts unfamiliar with mathe- The model used to obtain the second portfolio in
matical programming can obtain a reasonably effi- Table I is described at the end of this paper. It is
cient defensive portfolio by simply inspecting the based on a fairly simple algebraic formula that does
data. a remarkably good job of simulating those securities

20 California Management Revkto


which typically appear in the Sharpe-Markowitz TABLE n . — A V E R A G E R A T E S OF PHICE APPRECIATION FOR
model. One reason this portfolio did not do quite so THE Dow J O N E S INDUSTRIAL AVERAGE; T H E SHARPE-
MARKOWITZ M O D E L , AND OTHER PORTFOLIO
well during the bear markets of 1966 and 1969 is
BALANCE MODELS. SELECTED PERIODS,
the extra weight that was given to stocks such as 1966-1969
United Aircraft, V^'estinghouse, and Chrysler—all of
which had comparatively low correlations with S & Portfolios
P's industrial index during the original fitting period TUT A
but subsequently fiuctuated widely in relation to 1 2 3 4 5
the market average. In the last two instances it
1966 .811 .911 .883 .892 .924 .911
would have been quite reasonable to have expected
1967 1.152 1 .150 1 .182 1.219 1 .118 1 .101
a much higher index correlation in subsequent peri- 1968 1.043 1 .000 1 .001 .970 1 .028 1 .049
ods since similar type companies such as Ceneral 1969 .848 .918 .885 .850 .953 .980
Electric and Ceneral Motors had much higher index 1966-69 .964 .995 .988 .983 1 .006 1 .010
correlations during the original fitting period than
Westinghouse and Chrysler. The lower correlations rates than at higher rates.) See the least squares
for these two companies may, in the final analysis, market intercepts which are presented in the fourth
have been largely the result of temporary adversi- column of Table I. Security proportions for portfo-
ties that are infrequent and largely random in char- lio 3 were obtained by simply summing the index
acter. (One way to adjust the correlation coeffi- values for all securities with a positive index of effi-
cients for disturbances that are presumably ran- ciency and computing relative index proportions.
dom might be to sort firms on the basis of different This portfolio did not perform quite as well, on
industries and then regress the observed correlation the average, as the first two portfolios, yet its bear
coefficients on some indicator of intrinsic risk such performance in both instances was closer to the first
as market share or aggregate sales. The calculated two portfolios than to the performance of the DJIA.
correlation coefficients could then be substituted for (See Table II). Was this an accident, or is there a
the actual coefficients to obtain a normalized risk theoretical basis for supposing that a portfolio
parameter that is less influenced by historical acci- based on those stocks with the highest market inter-
dents that are peculiar to individual firms.) cepts should be more efficient than random selec-
The third portfolio in Table I is based on a dis- tion? Consider a linear regression of the form:
covery that the ten stocks which appeared in the Xi = a + bX, (2)
Sharpe-Markowitz portfolio also had the ten high- Let Rj equal the arithmetic average return on the
est constant terms when the returns for individual /•-th security and M equal the arithmetic average
securities were regressed on the percentage returns return for a representative market average. Since a
for S & P's industrial average. Where Ij equals an least squares regression must pass through the mean
index of efficiency, a; equals the market intercept, values for both tlie independent and dependent vari-
and i equals a rate of interest at which the portfolio ables, we can rewrite equation (2) so that the re-
manager can lend without risk, we can use the fol- gression coefficients are a function of expected
]o^ving criterion to rank stocks on the basis of effi- returns:
ciency:
Rj = Qj + hM (3)
Ij = fl, - i (1)
The slope coefficient in a linear regression is equal
When i was set to equal 1.04-tlie same rate of to the correlation coefficient, Cj, times the standard
interest used to obtain portfolios 1 and 2-we were deviation of the dependent variable, Sj, divided by
able to determine uniquely the ten stocks which ap- the standard deviation of the independent variable
peared in the Sharpe-Markowitz portfolio by
merely selecting those stocks with an index of effi-
ciency greater than zero. (This outcome may be 5 = C/ — (4)
partly an accident. Positive values for equation (1)
could be expected to do a better job of simulating Substituting equation (4) into (3) and solving
the Sharpe-Markowitz efficient set of portfolios at for tlie constant term, we learn that a, will he higher
lower and more defensive borrowing or lending —other things equal—the higher the expected re-

W I N T E R / 1970 / V O L . X I I I / N O . 2 21
turn for the /-th security, the smaller its standard securities and the market average to the mean values
deviation, and the lower its market correlation. that were established for the nineteen-year period.
The resulting values for aj were then used in con-
aj = Rj- CA~ \M (5) junction with equation (1) to obtain portfolio 4.
It seems clear from Table I that an emphasis on
Equation (5) is not exactly the same as other bear market performance can lead to important
models that have been used to obtain reasonably ef- changes in portfolio composition. Three companies
ficient portfolios (see equation (6)), but it does en- —C.E., Westinghouse Electric, and American Can-
compass all of the more important parameters. The given proportions ranging from slightly over one to
direction of impact, moreover, is essentially the more than 12 percent in the first three portfolios,
same. We can be fairly confident, therefore, that are not included in portfolio 4. The weight given to
portfolio 3 is not just an accident and that the crite- Eastman Kodak, United Aircraft, and Chrysler is
rion wliich was proposed in connection with equa- also reduced in a fairly significant manner. Other
tion (1) might be of significant value in helping to stocks which did not figure so prominently in the
obtain reasonably efficient defensive portfolios. first three portfolios—such as Procter & Camble,
The most exciting aspect of the aj-i criterion is Texaco, and Standard Oil of California—were given
the possibility of modifying equation (3) so as to considerably more weight.
obtain an even better definition of defensive securi- Three additional stocks which did enter portfo-
ties than would otherwise be obtained. Not much lios 2 and 4 were excluded from 1 and 3. The most
analysis of the data is required to suggest that some dramatic diff^erence in portfolio composition was for
companies with comparatively low correlations with American Telephone—a near zero proportion in
a representative market average—such as United portfolio 2 and a weight of almost ten percent in
Aircraft, Westinghouse Electric, and Chrysler—have portfolio 4.
achieved this low correlation, not as a result of out- Did these changes make a difference in subse-
standing performance in most bear markets, but as quent performance? The differences in overall per-
a result of having performed rather dismally on oc- formance were fairly small, yet it seems likely that
casion when the market in general was having a most investors would have preferred the results of
good year. portfolio 4 over the results of the first three. None of
If our objective is to obtain a defensive portfolio the first three did as well as in the bear markets of
it is questionable whether a comparatively poor 1966 and 1969 or in the lackluster year of 1968. Nor
performance in a bull market ought to be allowed did they perform as weU for the four periods as a
to influence the selection of stocks for a bear mar- whole.
ket. One way to avoid most of this problem and also The main difference between portfolios 4 and 5 is
simplify tlie procedures for obtaining an efficient that instead of basing our market line on the average
portfolio is to start with equation (3) and then ob- bear market performance of 1957, 1960, and 1962,
tain an identified system by choosing one additional we selected that particular bear market year with a
point, besides the mean values for an individual se- median market slope to represent the most typical
curity and the market average, which best describes experience of a security in a bear market situation.
the sensitivity of a particular security to declines in This amounted to considering both the worst and
the market average. best relative performance to be suspect and possibly
This, in any event, was the principle that was unrepresentative of future behavior. The most dra-
used to derive portfolios (4) and (5). For both we matic difference between portfolio 5 and the other
started out with the mean returns for the individual four is American Tobacco, now American Brands.
stocks in the DJIA and S & P's industrial average This particular security was once considered to be
during the nineteen-year period from 1947 to 1965. highly defensive. It has not been able to recover
To obtain portfolio (4) we then computed the aver- fully from the shock of 1962, however, when it lost
age returns for the individual stocks and S & P's more than 40 percent of its market value, despite
average during the bear market years of 1957, 1960, intensive efforts to diversify and improve its finan-
and 1962. A straight line was used to connect the cial image.
average bear market performance of the individual Portfolio 5 did not perform quite as well as in

22 California Management Review


1966, but it was as defensive as 1 through 3 and The expression, So/{Ro-i), normalizes the index
would have to be considered more eflBcient for the values for different securities in relation to that se-
four periods as a whole. The implication, it seems to curity which would be the best to own if a person
me, is that we might be able to train a security ana- were given the all-or-none choice of investing in only
lyst with only rudimentary skills as a mathematician one risk asset. The assumption is that a security
to estimate a reasonably efficient defensive should not be included in a portfolio with more than
portfolio, almost by inspection. If qualitative rea- tv^^o risk assets unless it can qualify to be included in
sons can be advanced for supposing that extreme a two-stock portfolio with that security which would
observations are likely to be umrepresentative of fu- be the best to own in the absence of diversification.
ture behavior there may be a real advantage to first The s^f (Ro-i) hurdle is not as restrictive as the
plotting the historical data and then fitting a mar- constraints that are built into some of the more
ket line which is largely judgmental. elaborate programs for achieving portfolio balance.
More securities will often qualify for inclusion,
under equation (6), than are obtained using the
A Simplified Portfolio Balance Model Sharpe-Markowitz models. This could be a desirable
The model used here to obtain portfolio 2 was de- feature leading to more efficient realized portfolios if
rived by solving tlie two-asset case and noting that secinrity analysts tend to overestimate their ability
a fairly simple formula will tell whether two securi- to differentiate between high and low performers.
ties should be in the same portfolio. The formula For a stock to be included in an efficient portfo-
can be used to compute an index of efficiency that lio, the index value for equation (6) should be posi-
has been shown to generate reasonably efficient tive. Security proportions are determined by sum-
portfolios of more than two risk assets [6]. The fol- ming the positive index values for included securi-
lowing list of symbols will be helpful in interpreting ties and computing relative index proportions. By
the index. varying the interest rate, i, we can generate a set of
portfolios that are reasonably efficient for different
7y = an index of efficiency which has been shown to do a
good job of simulating those stocks which are in- borrowing or lending rates.
cluded in the Sharpe version of the Markowitz
selection program.
Rj = the expected return on the j-th security.
Comments
Sj = the standard deviation of the returns for the j-th. In Markowitz' original monograph it was sug-
security.
Ro = the expected return on the security with the liighest
gested that the semi-variance might provide a bet-
iRj—i)/sj ratio. ter measure of risk than the standard deviation for
SQ — the standard deviation of the returns for the security all observations. Since most investors are not likely
with the highest {Rj—i)/sj ratio. to be bothered by an exceptionally good perfor-
i = a rate of interest at which the portfolio manager can mance in an above-average year, it is somewhat sur-
borrow or lend without risk.
prising that students of portfolio theory have shovim
C; = the correlation coefficient between the returns on
the j-th security and the returns on a representative so little inclination to investigate alternative mea-
stock price index. sures of risk.
Although the data presented in this paper do not
- i) So
provide a very conclusive test of alternative defini-
(6)
{Ro — i) tions, they are consistent with other studies which
Equation (6) can be interpreted in the following suggest that there may be significant structure to
way. The larger the difference (Rj-i) between the changes in security prices and would seem to indi-
return on the /-th security and the lending or bor- cate that it may be worthwhile to give more atten-
rowing rate, the higher the index value and the more
likely it is that the /-th security will be included in
an efficient portfoho. A higher expected return can Edward F. Remhaw is Professor of Economics
and Finance at the State University of New York
be offset, however, either by a larger standardized at Albany. He has written numerotis articles
variance, Sj, or by a higher correlation, Cy, with on the stock market and other topics in public
other securities. and private finance.

WINTER / 1970 / VOL. XIII / NO. 2 23


tion to the defensive qualities of a variety of portfo- 6. E. F. Renshaw, "Portfolio Balance Models in Per-
lio balance models. spective: Some Ceneralizations that Can Be Derived
from the Two Asset Case," Journal of Financial and
A not insignificant advantage of the a, criterion
Quantitative Analysis (June 1967), 123-149.
from a statistical point of view is that we obtain an
objective measure of defensiveness for all securities 7. , and V. D. Renshaw, "Simulating the Dow
Jones Industrial Average: A Further Test of tlie Random
regardless of whether they appear in an efficient
Walk Hypothesis," Financial Analysts Journal, forthcom-
portfolio or not. This variable could then be in- ing.
cluded in a multiple regression along with other
8. W. F. Sharpe, "A Simplified Model for Portfolio
technical and financial variables in the hopes of ob- Analysis," Management Science (Januaiy 1963), 277-
taining even better predictions of actual price be- 293.
havior during bear market periods. 9. , "A Linear Programming Algorithm for Mu-
tual Fund Portfolio Selection," Management Science
REFERENCES (March 1967), 499-510.
10. K. V. Smith, "A Transition Model for Poitfolio
1. K. ]. Cohen and J. A. Pogue, "An Empirical Eval- Revision," Journal of Finance (Sept. 1967).
uation of Alternative Portfolio-Selection Models," Jour-
nal of Business (April 1967), 166-193. 11. , "Alternative Procedures for Revising In-
vestment Portfolios," Journal of Financial and Quantita-
2. P. H. Cootner, The Random Character of Stock tive Analysis (Dec. 1968).
Market Prices (Cambridge, iMass.; MJ.T. Press), 1964.
12. J. Tobin, "Liquidity Preference as Behavior To-
3. K. A. Eox, Intermediate Economic Statistics (New wards Risk," Review of Economic Studies (Feb. 1958),
York: Wiley), 1968. 65-86.
4. H. M. Markowitz, "Portfolio Selection," Journal of 13. J. L. Treynor, W. W. Priest, L. Fisher, and C. A.
Finance (March 1952), 77-91. Higgins, "Using Portfolio Composition to Estimate
5. , Portfolio Selection: Cowles Foundation Risk," Financial Analysts Journal (Sept.-Oct. 1968),
Monograph 16 (New York: Wiley), 1959. 93-102.

24 California Management Review

You might also like