Estimating Returns

for Asset Allocation

Northfield Asset Allocation Seminar

September 2004
 Expected Returns
u Optimum portfolio weights are very sensitive
to return expectations.
Small changes in expected returns large
weight changes.
u Historic returns are often used, but they are
not usually good predictors of future returns.
Domestic Equity returns are upwardly biased by an
expansion of P/E ratios.
Bond returns are upwardly biased by decreases in interest
rates.
Active returns are upwardly biased because of survivorship
bias, particularly with hedge funds.
 Literature: Historic and Expected Returns

u Bradford Cornell, The Equity Risk Premium, 1999

u William Bernstein, The Birth of Plenty, 2004
u Jeremy Siegel, Stocks for the Long Run
u Dimson, Marsh, & Staunton, Triumph of the Optimists: 101
Years of Global Investment Returns
u Arnott and Bernstein, What Risk Premium is Normal,
Financial Analyst Journal, Mar/Apr 1992
u Dempsey, The Nature of Market Growth, Risk and Return
(DOGMA), Financial Analyst Journal, May/Jun 2002
u Ibbotson & Chen, The Supply of Stock Market Returns,
http://icf.som.yale.edu/pdf/Supply(v5).pdf
u Harvard Business School Case Study: The Harvard
Management Company (2001)
Inflation Adjusted U.S. Stock Market Returns
Jeremy Siegel: Stocks for the Long Haul
Time Period
Time Dividend Compound Total Standard
Period Yield Deviation
Yearly Return Annual Returns

1802-
6.4% 7.0% 16.9%
1870
1871-
5.2% 6.6% 16.8%
1925
1926-
4.6% 7.2% 20.4%
1997
 Historic Returns
u When calculating historic returns, even
long histories show returns that are
much higher than what probably can be
expected in the future.
u The US is the surviving economy and
was an emerging market in 1802, the
point at which real per capita GDP
growth began to grow at 2% for what is
now the developed world.
 Capital Asset Pricing Model
u Expected Return is:
Expected Return = Risk Free Rate + * Market Return
u What is the Risk Free Return for a reasonable
investment horizon:
Risk Free Return = Inflation + Embedded
Productivity
Market Return: The equity risk premium
Beta = Systematic risk, proportional to covariance
between asset and market portfolio.
 What is the Market Portfolio
u It should include the entire US stock market. The Russell
3000 or Wilshire 5000 are reasonable proxies. The MSCI
World index does not include small cap stocks.
u It should include corporate bonds, since the market should
be (mostly) independent of the markets capital structure.
Analogous to Modigliani Miller.
u It should include foreign stocks since they are part of a
well diversified portfolio.
u It should not include government or agency bonds, since
they are related to the supply and demand for current vs.
future consumption. See Robert Fuhrman, Northfield
Newport Conference 2004.
 Expected Returns Using CAPM Model
Rf = 4%, Rm = 3.5%
Expected
Asset Class Beta
Return
US Bonds 0.12 4.42
Non US Bonds 0.19 4.67
US High Yield Bonds 0.34 5.19
US Value Stocks 0.99 7.47
Non US Stocks, Developed 1.23 8.31
Emerging Market Stocks 1.26 8.41
US Growth Stocks 1.39 8.87
 Bayes-Stein Adjustment
u The reduction in errors arising from Bayesian
shrinkage estimators is a mathematically provable
result.
u Bayesian shrinkage jointly minimizes the errors in the
return expectations over the portfolio, rather than
trying to minimize the errors in each asset class
return expectation separately.
u The inadmissibility of sample mean as an estimator
for multivariate (portfolio) problems was proven by
Stein in 1955.
 Bayes-Stein Methods
uWe need to distinguish between
using these techniques:
To explicitly compensate for error
maximization in mean-variance
optimization.
For improving the quality of return
forecasts that we make. This technique
provides better numbers for what we
actually believe, but requires providing
more information than just historical data.
Literature: Bayesian Adjustment
u Compensating for Error Maximization
http://www-gsb.uchicago.edu/fac/arnold.zellner/more/CURRENT-
PAPERS/bayshrin.pdf
Jorion, International Portfolio Diversification with Estimation Risk,
Journal of Business, July 1985
Jorion, P., Bayes-Stein Estimation for Portfolio Analysis, Journal
of Financial and Quantitative Analysis, September 1986
Jorion, P., Bayesian and CAPM Estimators of the Means:
Implications for Portfolio Selection, Journal of Banking and
Finance, June 1991
u Requires Additional Data:
Black and Litterman, Global Portfolio Optimization, Financial
Analyst Journal, 1992
He and Litterman, The Intuition Behind Black-Litterman Model
Portfolios, Goldman Sachs
 Bayesian Priors
u There are really two dimensions of priors.
1. Whether the prior is diffuse or centralized.
2. Whether the prior is uninformative or
informative or somewhere in between: semi-
informative.
u Imagine we have some historical data for 3
asset classes: stocks, bonds, cash.
u The sample period return for stocks is -3%,
bonds 4%, cash 9%.
 Diffuse Priors
uA diffuse prior would be one where the prior
could take on a variety of values.
u Imagine we just picked some random number
between -3 and 9. This would be a diffuse
and uninformative prior.
u This is the technique Markowitz and Ussman
used in their recent paper that compared
Bayesian methods and re-sampling.
 Semi Informative Prior
uA slightly informative prior would be to
assume that the prior was random, but
normally distributed between 0 and 9
(implying mean 4.5)
u It is economically irrational for any
investor to invest in a risky asset with
negative expected (excess) returns.
 Centralized Prior
u Lets say the government regulated interest
rates on bonds to always be 6%.
u In such a case, it might be sensible to always
use 6% as a prior. The prior would have a
single value rather than a distribution.
u Grinolds alpha scaling rule of thumb is very
similar to a Bayesian shrinkage where W = IC
and Up = 0 (a single fixed value).
 Time Horizon
u While we normally talk about long term
returns in asset allocation, real world
investors usually have intermediate
horizons over which they view things.
u For example, a 30 year zero coupon
Treasury bond has wildly volatile annual
returns but the return over 30 years is
known with a high degree of certainty if
we hold one bond all the way.
Jorions Bayes-Stein Adjustment
All Bayesian estimates have the form:
Ubs = W (U) + (1-W) Up
where:
Ubs = the Bayes adjusted expectation of return
W = is the weight given to the return based on the
data we observe
U = return based on the data we observe (history)
Up= the return prior
 Stein Shrinkage Estimator
Jorions Technique estimates W:
W = 1 - [(n-2) / (T-n-2)] 1 / [(U-Up)T Q-1(U-Up)]
where:
T = number of time periods of data
n = number of asset return time series
Q = covariance matrix of return time series
 Single Value Prior
u Jorions method that has a single value prior
for all asset classes addresses estimation risk
in optimization.
u Our actual expectations have not changed
based on the Bayesian adjustments
u This is the way our beliefs ought be
presented to an optimizer. Rationale:
When our beliefs about returns are uncertain, the
uncertainly makes the asset classes less distinct
from one another.
Efficient return estimators shrink the return
spreads amongst the asset classes.
Jorions Bayes-Stein Adjustment
u The prior is the historical return on
minimum variance portfolio formed of all
the assets, excluding a risk free asset.
u If you have risky assets, the minimum
return you would expect from any one of
them individually would be the return
that could be achieved by investing in
the portfolio that yields the lowest risk.
 Bayesian Adjustment: Risks
u Once we've adjusted the means away from
the sample means, the standard deviations
will increase slightly
u The adjusted correlations may also differ
slightly from the sample values. In practice,
correlations do not change when using two
significant digits.
u From an optimization standpoint, we are
biasing the covariance matrix to a slightly
more conservative posture.
 Jorions Adjustment 1979-2004
Historic Bayesian
Asset Class Risk
Return Return
Cash 4.94 4.93 0.58
US Bonds 8.62 7.60 4.54
Non US Bonds 10.60 9.04 10.46
US Equity 13.18 10.91 15.83
Non US Equity 10.08 10.08 17.76
 Currency Hedge Non US Bonds
Asset Class Historic Bayesian Risk
Return Return
Cash 4.93 4.93 0.58
US Bonds 8.62 7.61 4.54
Non US Bonds 8.10 7.23 3.66
US Equity 13.18 10.92 15.83
Non US Equity 10.08 10.09 17.76
 Bayesian Adjustment
u Bayesian Adjustment reduces expected
return for bonds by about 1%.
u Bayesian Adjustment reduces expected
return for stocks by about 3%.
u US $ depreciation caused Non US
Bonds.
2% extra return
Increase in volatility 3.6 to 10.5
 Informative Priors
u Black-Litterman uses an informative
prior, where we are bringing a lot of new
information into the problem: in this
case the CAPM.
u We estimate what we think the values of
the returns should be, given each
assets weight in the global market
portfolio
 Black-Litterman Model
u This model is used for active asset allocation
at Goldman-Sachs: as described in He and
Litterman, The Intuition Behind Black-
Litterman Model Portfolios,
u The models key action is the estimation of
the vector of implied returns based on
estimations of the:
Markets risk tolerance
The weights of each asset in the market portfolio
The covariance between the assets in the market
 Black-Litterman Model
u An assets expected returns should vary from
the implied returns depending on the relative
outlook for that asset and the confidence in
the active prediction.
u The implied returns are calculated:
=w/
is the vector of implied returns
w is the portfolio weight vector
is the covariance matrix
is the markets risk tolerance parameter (RAP)
 Black-Litterman Return Estimation
Weight in
Asset Class Implied Return
Market Portfolio
Cash 0 4.56
US Bond 20 5.00
Non US Bond 30 5.11
US Stock 50 8.02
Non US Stock 20 7.89
 What About Actively Managed Funds?

u Use style analysis vs. benchmarks for which

you have good expected return estimates.
u Use the style factor weighted average of the
asset class returns to establish a baseline
expected return.
u Adjust the alpha for selection (you are not
likely to pick something that hasnt been
good) and survivorship bias.
u Add the alpha to the expected returns due to
beta.
 Conclusions
u Historic returns, even over long time histories,
are poor predictors of future (and hence
expected) returns.
u Bayesian adjustment accounts for the
uncertainty of return forecasts. Jorions
Bayes Stein adjustment and Black-Litterman
are both examples of Bayesian adjustment.
u CAPM and Black-Litterman are equilibrium
models to estimate expected returns from
historic covariance data.