Time Diversification and Estimation Risk

Bjorn Hansson and Mattias Persson

The recommendation that investors with long investment horizons tilt their
portfolios toward stocks is cominonplace. We used a nonparametric
bootstrap approach to investigate whether in a mean-variance-efficient
portfolio, the lueightsfor U.S. stocks and U.S. T-bills vary in a systematic
manner with investment horizon. This approach allowed us to analyze the
impact of estimation risk on the optimal weights of stocks and fixed-income
securities. The results show that an investor can gain from time
diversification: The weights for stocks in an efficient portfolio were
significantly larger for long investment horizons than a one-year horizon.

ractitioners in the United States commonly
recommend that an investor with a long
investment horizon—someone saving for
retirement, for example—tilt her or his port-
folio toward stocks and away from fixed-income
secui-ities. This recommendation is an important
example of putting into practice the concept of time
diversification, which generally implies a system-
atic relationship between the portfolio weights for a
particular asset class and investment horizon.
To say that investors with a long horizon
should invest more heavily in stocks than investors
with a short horizon is almost conventional wis-
dom, because the relative riskiness of stocks com-
pared with bonds is supposed to diminish with the
length of the horizon {Lee 1990; Siegel 1994; Thorley
1995). Some well-known economists have attacked
this recommendation^among others, Samuelson
(1994), Bodie, Merton, and Samuelson (1992), and
Jagannathan and Kocherlakota (1996). Their asser-
tions are based on the following theoretical result:
If the return series follow a random walk and the
investors have constant relative risk aversion, then
time diversification cannot exist. As pointed out by
Van Eaton and Conover (1998), however, the two
camps are not analyzing the same type of invest-
ment problem. The attackers are looking at a
dynamic investment problem in which the given
horizon is chopped up into several periods of the
same length, whereas the adherents of time diversi-
fication are analyzing the effects of increasing the
horizon for a buy-and-hold strategy. In such a case,
the optimal weights may vary with the horizon even
Bjorn Hansson is professor of economics and Mattias
Persson is a Ph.D. candidate in economics at Lund
University, Szueden.
if markets behave like a random walk and investors
have constant relative risk aversion (Van Eaton and
Conover; Jagannathan and Kocherlakota).
Our purpose in the study reported here was to
analyze whether mean-variance-efficient portfolio
weights for stocks and T-bills vary significantly
with the investment horizon for a buy-and-hold
strategy. The data in this exercise were real return
data for a well-diversified U.S. stock portfolio and
the short-term, nominally risk-free rate from 1900
to 1997. The analysis presupposed that investors
form optimal investment strategies based only on
historical estimates of the following parameters or
inputs to the optimization problem: means, vari-
ances, and covariances. The model is uncondi-
tional, in the sense that the agents do not explicitly
try to model any possible time-series relationships
among the assets. We implicitly accounted for any
possible time dependencies in the observed return-
generating processes, however, by resampling a
great number of return series from the original data
through a computer-intensive method called
"bootstrapping with moving blocks." This strategy
differs trom that used by Jagannathan and Kocher-
lakota and several others, who constructed statisti-
cal models of asset returns.
The bootstrap approach allowed us to study an
important (but often forgotten) issue, namely, esti-
mation risk, in which the true parameters of the
return distributions are unknown.^ In a mean-
variance context, the inputs to the mean-variance
model are sample estimates only, not the true
parameters. The problem of estimation risk is a very
important empirical problem; Jobson and Korkie
(1981) showed that the biases in mean-variance opti-
mization can be large. In addition, the composition
of the sample efficient portfolios is extremely sensi-
tive to changes in the estimates of the asset means

September/October 2000 55
Financial Analysts journal
(Best and Grauer 1991). Furthermore, Broadie (1993)
indicated that estimates of efficient portfolios are
optimistically biased predictors of actual efficient
portfolios (i.e., they overestimate expected returns
and underestimate the variance of reaims).^ We
focus on the asset weights rather than the means and
the variances of the efficient portfolios.
We not only analyzed the existence of time
diversification, but we also tested whether time
diversification is significant in a statistical sense
(i.e., whether significant statistical differences exist
between the optimal weights for different invest-
ment horizons).^ We were able to do so because we
generated empirical distributions of the optimal
weights by using a nonparametric bootstrap. There-
fore, we were not constrained by the fact that our
data contained few nonoverlapping investment
horizons of, for example, 10 years. Instead, via
resampling, we could construct the empirical dis-
tributions for the optimal weights and thus use the
sample information to the utmost.
Data
Our monthly data for U.S. stocks and T-bills from
1900 to 1997 were in the form of two total-return
indexes: The stock data were a total-return index
based on the S&P 500 Index and the bill index pro-
vided the returns from holding 90-day U.S. T-bills.
From these two indexes, we constructed series of
continuous returns. We made the assumpHon that
investors are interested only in real consumption
possibilities and deflated the return series by the
seasonally adjusted U.S. Consumer Price Index cal-
culated by the U.S. Bureau of Labor Statistics. Thus,
the optimization problem involves a trade-off
between the expected value of real returns and the
variance of real returns.
Method
In this section, we describe the optimization method
used to find the asset weights of the efficient portfo-
lios and how we applied the bootstrap approach to
construct empirical distributions for asset weights.
Efficient Portfolios. Investors were assumed
to use the mean-variance criterion when forming
their optimal portfolios, and the investment hori-
zons we considered were 1 year, 5 years, and 10
years.^ The investors look for the portfolio weights
X that minimize the following trade-off between
variance and expected return for a particular
investment horizon:^
©2000, Association for Investment Management and Research
56
Min x'y X -lixx
subject to I'x = 1
x>0,
where
X = the weight vector
y - the variance-covariance matrix
^i = the vector of expected returns for the
investment horizon under consider-
ation
X = "the price of risk" or the marginal rate of
substitution of expected return for vari-
ance
I = a vector of ones
A nonnegativity restriction is imposed on the
weights because all asset positions are assumed to
be positive for long investment horizons. An inves-
tor with low risk aversion will choose a portfolio
with a high lambda, whereas an extremely risk-
averse investor will choose a portfolio with the
smallest possible variance, the so-called "minimum-
variance portfolio," which has a lambda equal to
zero. Lambda in an interval of 0 to 10 was used in
the optimization because this interval captures effi-
cient portfolios from the all-stock portfolio to the
minimum-variance portfolio located at lambda
equal to zero. In our comparison of portfolios from
different investment horizons, we use portfolios
with the same lambda, which implies that the port-
folios will differ with respect to expected return and
variance but will have the same risk price.
Keep in mind that for two assets, it is not pos-
sible to compare weights for the same return target.
The procedure we used gives the same portfolio
weights for all horizons because the resampled
average returns are almost identical for different
horizons.
The Bootstrap Approach. Tiie bootstrap
method, introduced by Efron (1979), is a computer-
intensive method for estimating the distribution of
an estimator or a statistic by resampling the data at
hand. In this study, we used a nonparametric mov-
ing block bootstrap introduced by Carlstein (1986)
aiid Kiinch (1989), in wliich serial dependence, as
weii as cross-sectional correlation, is preserved
within the blocks. We used a nonparametric boot-
strap because a parametric form gives inconsistent
estimates if the structure of the serial correlation is
misspecified or not tractable. The assets were drawn
cross-sectionally, so they belonged to the same time
period in the original series. Tlius, we never con-
structed a five-year relationship between stocks and
bills that did not exist in the original series.

The sample of real continuous returns on
stocks and bills, R, was grouped into k overlapping
blocks of 60 months. We chose a block length of 60
months because it is probably long enough to pick
up most forms of possible time dependencies. The
blocks were then resampled with replacement b
times until a series R* with the same length as R was
obtained, which was equivalent to constructing a
realization or trajectory for stocks and bills for each
drawing or R*. From each resampled R*, we calcu-
lated a variance-covariance matrix, V*, and an
expected return vector, \i*, for each investment
horizon from nonoverlapping holding-period
retuims. We then used \* and [i* as inputs to the
mean-variance optimization and obtained the opti-
mal portfolio weights, x*. We repeated this proce-
dure 1,000 times. In the end, we had a set of
bootstrapped observations for each optimal portfo-
lio in the mean-variance optimization and every
investment horizon.
The empirical distribution of the weights, based
on the bootstrap samples, allowed us to draw infer-
ences about the weights. We constructed 90 percent
confidence intervals based on the percentiles of the
distribution of the assets weights. We ordered the
observations in ascending order. That is, the 5 per-
cent percentile, x*^^\ is the 50th ordered value of the
replications and j"f^""' is the 95 percent percentile
and the 950th ordered value of the replications. We
Figure 1. Orlginai Total-Return Series and Trajectories for Stocks and BiMs
(iognormal scale)
1,000
Time
Stocks Low
Stocks High
Stocks
September/October 2000
Time Diversification and Estimation Risk
obtained an indication of the magnitude of the esti-
mation risk from the intervals because the confi-
dence region displayed the degree of uncertainty
associated with the efficient frontier. Our bootstrap
approach to measuring estimation risk is an alterna-
tive to the confidence regions of Jobson (1991) and
the methods proposed in Michaud (1998).
Analysis
Before discussion of time diversification and esti-
mation risk, we will first take a brief look at the
data. Figure 1 depicts our original index series for
stocks and bills and two bootstrap realizations of
total returns for stocks and bills, which imply four
trajectories. The bootstrap realizations illustrate the
trajectories when the weight in stocks was high and
low, respectively: "Stocks High" and "Bills Low"
are from one bootstrap realization; "Stocks Low"
and "Bills High" are from another realization.
These trajectories resemble the original index
series. As Figure 1 shows, a high weight in stocks,
"Stocks High," does not have to be associated with
an extreme positive realization for stocks. In this
case, the high weight in stocks is connected with an
extreme negative realization for bills. The same is
true for the relationship between "Stocks Low" and
"Bills High," but in that case, the risk-free asset had
extremely low volatility.
Bills Low
Bills High
Bills
57
Financial Analysts Journal

Time Diversification. To discuss the exist-
ence of time diversification (i.e., whether any signif-
icant differences would exist between the weights
for short versus long horizons), we start by simply
comparing average weights. Then, we will turn to
proper statistical inference.
Table 1 reports the average weights and the
90 percent confidence intervals for the weight in
stocks for various efficient portfolios and invest-
ment horizons. (The results for bills are not pre-
sented because they are an exact mirror image of
the results for stocks.) Only results for lambda in
the interval of 0 to 1 are reported because for higher
lambda, most of the efficient portfolios consisted
of 100 percent stocks.'' The average weight in
stocks increases with investment horizon, whereas
(and as a result) the average weight in bills falls
with investment horizon. We found the largest
relative differences in average weights for portfo-
lios with low lambdas. Comparison of the average
weights for 1- and 10-year investment horizons
with lambda equal to zero (i.e., the minimum-
variance portfolio) shows that the average weight
increases from 2.9 percent in stocks to 11.3 percent
in stocks with the longer horizon. For these hori-
zons, when lambda equals 1, the relative weight
increase at the long horizon is smaller.
In fact, if time diversification exists, the largest
relative difference in the asset weights will always
be found at low lambdas, at least when only two
assets are being considered. For a lambda larger
than some X**, the efficient portfolio will always
consist of 100 percent stocks. If time diversification
exists, the implication is that the X** at the 10-year
horizon is smaller than the X** at the 1-year invest-
ment horizon. For many portfolios, having the
same relative increase in weight would not even be
possible because the weight has an upper bound of
1 as a result of the prohibition of short sales.
The most direct test of time diversification is to
compare the weights from the same trajectory. To
decrease the effect of estimation errors, we matched
our weights from each bootstrap realization (trajec-
tory) and counted the number of cases in which the
5-year stock weight, .Y5, and 10-year stock weight,
.rlO, was above the corresponding 1-year weight, xl.
This procedure produces low estimation errors
because extreme one-year weights are usually asso-
ciated with extreme weights for the longer horizons.
The result is presented in Figure 2. Note that the
number of cases when, for example, the stock weight
for the five-year horizon is larger than the corre-
sponding one-year weight ranges from 700 to 865—
that is, in more than half of the drawings. The critical
values are 521 and 537 for, respectively, the 95 per-
cent and 99 percent level. Thus, significant differ-
ences appear for all levels of lambda. The greatest
significance is encountered for an absolute risk aver-
sion around 0.5, which is a quite moderate level. The
significance is always higher for the 5-year horizon
because the 10-year horizon has larger estimation
errors as a result of having fewer nonoverlapping
observations. Therefore, we found a systematic rela-
tionship between the investment horizon and the
allocation to stocks in mean-variance-efficient port-
folios: Investors with long investment horizons
should have relatively more stocks than investors
with short horizons.

Tabie 1 . Average Weights and Confidence Intervais at the 90 Percent Levei forthe Weights in Stocks, S
Investment Horizon Irwestment Horizon Investment Horizon
1 Year 5 Years 10 Years

X •• •<**]
r 1
V. [..'-] •<-) •<•-'*)
s[.v*" "'J
•

1 0.249 0.682 1.000 0.418 0.825 1,000 0,351 0.815 1,000

0,90 0.230 0,634 1.000 0,385 0.787 1,000 0,324 0,781 1.000
0.80 0.212 0.579 1.000 0,350 0,739 1,000 0,302 0,741 1,000
0.75 0.202 0,549 1.000 0.333 0.712 1,000 0,292 0.718 1,000
0.70 0.190 0.517 0.981 0,317 0,682 1,000 0.278 0,693 1.000
0.60 0.170 0.450 0.845 0.288 0.615 1.000 0,238 0.635 1.000
0.50 0.149 0.380 0.699 0.252 0.537 1.000 0.182 0,567 1.000
0.40 0,127 0.309 0.549 0.208 0.450 0.847 0.134 0.486 1.000
0.35 0,117 0.274 0,482 0.186 0,405 0.750 0.103 0.441 0.989
0.30 0.106 0.238 0.413 0.155 0.359 0.656 0.081 0.393 0,874
0.25 0.097 0.202 0,347 0,123 0.312 0.559 0.040 0.344 0.774
0.20 0.086 0.167 0,280 0.082 0.265 0.478 0,000 0.295 0,669
0,15 0.069 0,131 0.214 0,036 0.218 0,415 0,000 0.245 0,577
0.10 0.047 0.096 0.152 0,000 0.172 0,350 0,000 0.196 0,475
0.05 0.017 0.060 0.103 0.000 0.130 0.288 0,000 0.151 0,400
0 0,000 0,029 0.075 0.000 0.095 0.252 0.000 0,113 0,341

58 ©2000, Association for Investment Management and Research


Figure 2. i\Aatched Weights in Stocks
N u m b e r of Observations
1,000
900 .v5 > A-1
809
790
800 740 764
700 735
713
686
600 xlO > xl
500
400
300
200
100
0
0 0.05 0.10 0.15 0.20
Estimation Risk. Turning to the confidence
intervals in Table 1, note that the width of the inter-
val generally increases with lambda. The cause is
what is called "error maximization" (Michaud
1989), which is the dark side of portfolio optimiza-
tion: In search of assets that increase the return and
decrease the risk of a portfolio, the optimization
procedure favors assets with overestimated
expected returns and underestimated risks. Thus,
our result is in line with the well-known result that
portfolios close to the minimum-variance portfolio
have less estimation error (see, for example. Best
and Grauer, Broadie, and Chopra and Ziemba 1993).
Table 1 shows that the interval width generally
increases with the investment horizon. One reason
is the presence of fewer nonoverlapping observa-
tions at longer investment horizons. Figure 3 and
Figure 4 show the ordered weights for stocks for all
three investment horizons and portfolios with
lambda equal to, respectively, 0.1 and 0.5. Compar-
ing the slopes of the curves in Figure 3 shows that
our estimates of the stock weights for the one-year
horizon are quite stable and that the uncertainty in
our estimates increases with the investment horizon.
The number of extreme weights—that is, the niim-
ber of cases when the weight for stocks is 0 or 1—
also increases with the investment horizon. This
phenomenon makes the confidence intervals wider
and thereby makes our estimates more uncertain.
Figure 4 shows that the variation in weights for X =
0.5 is larger than the variation in weights for the
portfolios at ^ - 0.1. The numbers of extreme weights
September/October 2000
Time Diversification and Estimation Risk
847 855 865 863 848
830 838 815
696 '•
655
0.25 0.30 0.35 0.40 0.50 0.60 0.70 0,80 0.90 1.00
Lambda
are also larger for the portfolios with X ^ 0.5. Thus,
the number of extreme weights increases both with
lambda and with investment horizon.
Results with Different Block Lengths.
Finally, we analyzed whether our results were sen-
sitive to the chosen block length.^ Choosing a block
length involves a trade-off. On the one hand, as
block length decreases, the moving block bootstrap
will destroy the time dependency of the data and
average accuracy will decline. On the other hand,
as block size increases, fewer blocks will result, the
pseudodata will tend to look alike, and again, aver-
age accuracy will decline. Berkowitz and Kilian
(1996) found that the performance of the moving
block bootstrap tends to be fairly stable in the
neighborhood of the optimal block size.
We used block sizes of 84 and 120 months to
check the robustness of our results for a block size
of 60 months but found that the results did not
change the overall picture:^ We still concluded that
investors with long investment horizons should tilt
their portfolio weights toward stocks. The optimal
block length, in the sense of taking care of as much
mean reversion in stock returns as possible, is prob-
ably greater than 60 months but below 120 months.
Tentative Expianation. Our results indicate
that the weights for stocks in efficient portfolios are
significantly greater for longer horizons than for a
one-year horizon (and vice versa for bills). A tenta-
tive explanation is that over certain investment
59
Financial Analysts Journal
Figure 3. Quantiie Functions for the Weights in Stocks at >. = 0.1
Weight in Stocks
1.0
100 200 300 400
Number of Bootstrap Observations
horizons, the return-generating process for stocks is
mean reverting and/or the process for bills is posi-
tively autocorrelated. If so, the annualized risk for
stocks should fall with investment horizon whereas
the risk for bills should increase. Table 2 displays
the average mean returns and average annualized
standard deviations for stocks and bills and the
average correlation between stocks and bills from
the 1,000 bootstrap resamples. The risk for stocks
falls from around 20 percent for the one-year hori-
zon to a bit above 17.5 percent for the five-year
horizon; the risk for bills increases from about 5
percent for the one-year horizon to just above 7.5
percent for the five-year horizon. Note in particuiar
that the mean returns for stocks and bills are the
same for the three investment horizons. Because the
return spread between stocks and bills is constant
over the investment horizons, the change in portfo-
lio weights might stem from the fact that the stan-
dard deviation for stocks falls with longer
investment horizons but increases for bills. Thus,
our results may be at least partly explained by these
characteristics of the return-generating processes.
Keep in mind that our results follow from the use of
moving blocks in the bootstrap procedure, which
takes care of time dependency in the data; the
60 ©2000, Association for Investment Management and Research
500 600 700 800 900 1,000
results are not caused by any particular paramet:ric
assumptions of the return-generating processes.
Conclusions
Our main purpose has been to analyze whether
mean-variance-efficient portfolio weights for stocks
and bills vary significantly with investment horizon
in a buy-and-hold strategy. In this analysis, we kept
the risk price, the slope of the efficient frontier, con-
stant while varying the investment horizon from 1
year to 5 years to 10 years. We used real U.S. return
data from 1900 to 1997 for a well-diversified stock
portfolio and a short-term, nominally risk-free rate.
We used a bootstrap approach to analyze the
variation in portfolio weights for stocks and bills
for the different investment horizons—in particu-
lar, a nonparametric moving block bootstrap with
a block length of 60 months in which serial depen-
dence and cross-sectional correlation were pre-
served within the blocks. With this approach, we
could also study the impact of estimation risk on
the optimal weights of stocks and bills.
As far as the question of estimation error is
concerned, we found that estimation error increased
with price of risk, X, and with investment horizon.
The ftrst effect is a result of error maximization.
 Time Diversification and Estimation Risk

Figure 4. Quantiie Functions for the Weights in Stocks at >. = 0.5

Weight in Stocks
1.0

0 100 200 300 400 500 600 700 800 900 1,000

Number of Bootstrap Observations

which implies that the optimization framework
chooses assets with overestimated expected returns
and imderestimated risks. The second effect is partly
a result of fewer nonoverlapping observations at
longer investment horizons than at shorter horizons.
To analyze whether any significant statistical
differences would show up in the stock weight for
the different investment horizons when lambda
was kept constant, we used a method that is quite
efficient in taking estimation errors into consider-
ation: We matched the weights from our bootstrap
samples and counted the number of cases in which
the longer horizon had a larger weight in stocks
than the one-year horizon. For all lambdas, the
stock weight was significantly larger for the longer-
horizon portfolios. Thus, we conclude that our evi-
dence supports the existence of time diversification:
The weights for stocks in efficient portfolios are
significantly higher for long investment horizons
than for a one-year horizon.
We would like to thank seminar participants at Lund
University, the Finance and Macro workshop at EPRU,
Copenhagen Business School, the annual meetings of the
Southern Finance Association 1998 and the European
Financial Management Association 1999, and the
Research Department at the Federal Reserve Bank of
Atlanta for vaiuabie comments.

Tabie 2. Average Mean Returns, Standard Deviations, and Correlations

Investment Horizon Investment Horizon In\'estinent Horizon
lYear 5 Years 10 Years
Standard Standard Standard
Asset Mean Deviation Correlation Mean Deviation Correlation Mean Deviation Correlation
Stocks 6-09'Ji. 20.09% 0.15 6.09% 17.69% 0.24 6,09% 17,40% 0.23
Bills 0.95 5.09 0.95 7.64 0,95 7,53
Note: Standard deviation is annualized. Correlation is the correlation between stocks and bills.

September/October 2000 61
Financial Analysts Journal

Notes
1. More specifically, estimation risk means that the joint prob-
ability density function of security returns,/{r| 6), is not
completely known. Included are the situation when the func-
tional form of/is unknown and the more common case in
which the functional form of/is presumed to be known but
the values of the parameter vector, 9, are unknown. Bawa,
Brown, and Klein (1979) provided a good overview of this
problem and its implications for optimal portfolio selection.
2. Broadie used a Monte Carlo approach similar to the one
used by Jobson and Korkie but imposed a restriction on
short selling.
3. Liang, Myer, and Webb (1996) used the bootstrap method
in a somewhat similar manner for studying the real estate
weight in a nriixed portfolio.
4. The data came from Taylor (1998).
5. We also have results for a 20-year investment period but
they did not in any way change the overall results.
6. Markowitz (1992) and Sharpe (1991) used this optimization
method in an expected utility framework. Our minimization
problem is exactly the same as the optimization for an
investor with a negative exponential utility function when
the returns are jointly normally distributed. In this case,
lambda is the absolute risk tolerance. But our optimization
method is an excellent approximation even if the investor
has some other utility function and/or the returns are not
jointlynormallydistributed (see Levy and Markowitz 1979),
7. The results for lambdas greater than 1 are available from
the authors.
8. Li and Maddala (1997) and Berkowitz and Kilian (1996)
found that a moving block bootstrap is sensitive to the
selection of biock length.
9. The results for block lengths of 84 and 120 months are
available from the authors.

References
Bawa, V.S., S.J. Brown, and R.W. Klein. 1979, Estimation Risk aud Lee, W.Y. 1990. "Diversification and Time, Do Investment
Optimal Portfolio Choice. Amsterdam: North-Holland. Horizons Matter?" lotirnal of Portfolio Management, vol, 16, no. 3
Berkowitz, J., and L. Kilian. 1996. "Recent Developments in (Spring):21-26,
Bootstrapping Time Series." Finance and Economics Discussion Levy, H.,and H,M. Markowitz. 1979, "ApproximatingExpected
Scries, no, 45, Federal Reserve Board. Utility by a Function of Mean and Variance." American Economic
Best, M,, and R. Grauer. 1991, "On the Sensitivity of Mean- Rei'iew. vol. 69, no. 3 (June):308-317.
Variance Efficient Portfolios in Asset Means: Some Analytical
and Computational Results." Revieio of Financial Studies, vol. 4, Li, H., and G.S. Maddaia. 1997. "Bootstrapping Cuintegrating
no, 2 {Su mm er):315-342, Regressions." journal of Econometrics, vol. SO, no. 2
(October ):297-318.
Bodie, Z,, R,C. Merton, and W.F. Samuelson. 1992. "Labor
Supply Flexibility and Portfolio Choice in a Life Cycle Model," Liang, Y., F.C.N. Myer, and ].R. Webb. 1996. "The Bootstrap
journal of Economic Dynamics and Control, vol. 16, no. 3/4 (July/ Efficient Frontier for Mixed-Asset Portfolios." Real Estate
October) :427-449. Economics, vol, 24, no. 2 (Summer):247-256.
Broadie, M. 1993. "Computing Efficient Frontiers Using Markowitz, H,M, 1992. Portfolio Selection—Efficient
Estimated Parameters." Annals of Operations Research: Financial Diversification of Investments. Oxford: Basil Blackwell.
Engineering, vol, 45 (Special lsaue):21-58,
Michaud, R.O, 1989, "The Markowitz Optimization Enigma: Is
Carlstein, H. 1986. "The Use of Sub Series Values for Estimating
theVarianceof a General Statistic from a Stationary Sequence." Optimized Optimal?" Financial Analysts journal, vol. 45, no. 1
Annals of Statistics, vol, 14, no. 3 (July):1171-95. Qanuary/February):31-42.
Chopra, V,K., and W.T. Ziemba, 1993. "The Effects of Error in . 1998. Efficient Assel Management—A Practical Guide to
Means, Variances, and Covariancos on Optimal Portfolio Stock Portfolio Optimization and Assel Allocation. Boston, MA:
Choice." Journal of Portfoiio Management, vol. 19, no. 4 Harvard Business School Press.
(Winter) :6-ll. Samuelson, P,A. 1994, "The Long-Term Case for Equities and
Efron, B. 1979. "Bootstrap Methods: Another Look at the How It Can Be Oversold." journa! of Portfolio Management,
Jackknife," Annals of Statistics, vol. 7, no, 1 (January): 1-26. vol, 21, no. 1 (Fa]l):15-24,
Jagannathan, R,, and N,R. Kocherlakota, 1996. "Why Should Sharpe, W.F. 1991, "Capital Asset Prices with and without
Older People Invest Less in Stocks Than Younger People?"
Negative Holdings." journal of Finance, vol, 46, no. 2 Gune):
Federal Reserve Bank of Minneapolis Quarterly Review, vol. 20, no. 3
(Summer):ll-23, 489-510.
Jobson, J.D, 1991. "Confidence Regions for the Mean-Variance Siegel, J.J. 1994, Stocks for the Long Run. Chicago, IL: Irwin.
Efficient Set: An Alternative Approach to Estimation Risk." Taylor, Bryan. 1998, Guide to the Global Financial Database.
Rei'iewof Quantitative FinanceandAccoiinting,vo\.'[,no.3:235-257. Alhambra,CA: Global Financial Data (www.globalfindata.com),
Jobson, J.D., and B. Korkie. 1981. "Estimation for Markowitz Thortey, S.R. 1995. "The Time-Diversification Controversy."
EificieniPortMioa." journal of the American Statistical Association,
vol. 75, no. 371 (September) :544-554. Financial Analysts journal, vol. 51, no. 3 (May/June):68-76.
KUnch, H.R. 1989. "The Jacknife and the Bootstrap for General Van Eaton, D.R., and J.A. Conover, 1998. "Misconceptions about
Stationary Observations." Annals of Statistics, vol. 17, no. 3 Optimal Equity Allocation and Investment Horizon." Financial
(July):1217-41. Analysts journa!, vol. 54, no. 2 (March/Aprit):52-59,

©2000, Association for Investment Management and Researcli