Stocks versus Bonds: Explaining the

Equity Risk Premium

Clifford S. Asness
From the 19th century through the mid-20th century, the dividend yield
(dividends/price) and earnings yield (earnings/price) on stocks generally
exceeded the yield on long-term U.S. government bonds, usually by a
substantial margin. Since the mid-20th century, hoivever, the situation has
radically changed. In addressing this situation, I argue that the difference
between stock yields and bond yields is driven by the long-run difference
in volatility between stocks and bonds. This model fits 1871-1998 data
extremely well. Moreover, it explains the currently low stock market
dividend and earnings yields. Many authors have found that although both
stock yields forecast stock returns, they generally have more forecasting
power for long horizons. I found, using data up to May 1998, that the
portion of dividend and earnings yields explained by the model presented
here has predictive power only over the long term whereas the portion not
explained by the model has power largely over the short term.

- - ^ ^y he dividend yield on the S&P 500 Index The market earnings yield or earnings to price,
f m has long been examined as a measure of E/P (the inverse of the commonly tracked P/E),
y stock market value. For instance, the well- represents how much investors are willing to pay
known Gordon growth model expresses a for a given dollar of earnings. E/P and D/P are
stock price {or a stock market's price) as the dis- linked by the payout ratio, dividends to earnings,
counted value of a perpetually growing dividend which represents how much of current earnings are
stream: being passed directly to shareholders through divi-
P =
D dends. Studies by Sorenson and Amott (1988), Cole,
(1)
R-G' Helwege, and Laster (1996), Lander, Orphanides,
where and Douvogiannis (1997), Campbell and Shiller
P = price (1998), and others, have found that the market E/P
D = dividends in Year 0 has power to forecast the aggregate market return.
R - expected return Under certain assumptions, a bond's yield-to-
G - annual growth rate of dividends in perpe- maturity, y, will equal the nominal holding-period
tuity return on the bond.^ Like the equity yields exam-
Now, solving this equation for the expected return ined here, the inverse of the bond yield can be
on stocks produces thought of as a price paid for the bond's cash flows
(coupon payments and repayment of principal).
/; = - + G. (2) When the yield is low (high), the price paid for the
Thus, if growth is constant, changes in dividends to bond's cash flow is high (low). Bernstein (1997),
price, D/P, are exactly dianges in expected (or Ilmanen (1995), Bogle (1995), and others, have
required) return. Empirically, studies by Fama and shown that bond yield levels (unadjusted or
French (1988, 1989), Campbell and Shiller (1998), adjusted for the level of inflation or short-term inter-
and others, have found that the dividend yield on est rates) have power to predict future bond returns.
the market portfolio of stocks has forecasting power This article examines the relationship be-
for aggregate stock market returns and that this tween stock and bond yields and, by extension, the
power increases as forecasting horizon lengthens. relationship between stock and bond market
returns (the difference between stock and bond
expected returns is commonly called the equity
Clifford S. Asness is president and managing principal risk premium). I hypothesize that the relative
at AQR Capital Management, LLC. yield stocks must provide versus bonds today is

96 ©2000, Association for Investment Managennent and Research

 Stocks versus Bonds

driven by the experience of each generation of the late 1990s are not comparable to those of the
investors with each asset class. past. Although this assertion may have some merit,
The article also addresses the observation of I will argue that it is largely unnecessary to explain
many authors, economists, and market strategists today's low D/P.
that today's dividend and earnings yields on stocks As did dividend yields, the stock niarket's
are, by historical standards, shockingly low. I find earnings yields systematically exceeded bond
they are not. yields early in the sample period, but as Figure 2
Finally, I report the results of decomposing shows, since the late-1960s, earnings yields have
stock yields into a fitted portion (i.e., stock yields been comparable to bond yields and clearly
explained by the model presented here) and a strongly related (as are dividend yields, albeit from
residual portion (i.e., stock yields not explained by a lower level).-^ Table 1 presents monthly correla-
the model). tion coefficients for various periods between the
levels of D/P and Y and E/P and Y. The numbers
Historical Yieids on Stocks and in Table 1 clearly bear out what is seen in Figures 1
Bonds and 2. For the entire period, D/P and Y were neg-
atively correlated because of their reversals; E/P
As far as yields are concerned, 1927-1998 tells a tale was essentially uncorrelated with Y. For the later
of two periods—as Figure 1 clearly shows. Figure 1 period, however, stock and bond yields show the
plots the dividend yield for the S&F 500 and the strong positive relationship many economists and
yield to maturity for a 10-year U.S. T-bond from market strategists have noted.
January 1927 through May 1998.^ Prior to the mid-
1950s, the stock market's yield was consistently Thus, we are left with several puzzles:
above the bond market's yield. Anecdotally, inves- • Why did the stock market strongly outyield
tors of this era believed that stocks should yield bonds for so long only to now consistently
more than bonds because stocks are riskier invest- underyield bonds?
ments. Since 1958, the stock yield has been below • Why did stock and bond yields move relatively
the bond yield, usually substantially below. As of independently, or even perversely, in the over-
the latest data in Figure 1 (May 1998), the stock all 1927-98 period but move strongly together
market yield was at an all-time low of 1.5 percent in the later 40 years of this period?
whereas the bond market yield was at 5.5 percent, • Perhaps most important, why are today's stock
not at all a corresponding low point. This observa- market yields so low and what does that fact
tion has led m?my analysts to assert that the role of mean for the future?
dividends has changed and that dividend yields in The rest of this article tries to answer these questions.

Figure 1. S&P 500 Dividend Yieid and T-Bond Yieid to Maturity, January 1927-
May 1998
Yield (%)
18

16

14 S&? 500 D/P

12

10

1h
10-YearT-Bond
0
27 34 41 48 62 76 83 90 98

March/April 2000 97
 Financial Analysts Journal

Figure 2. S&P 500 Earnings Yield and T-Bond Yield to Maturity, January 1927-
May 1998
Yield (%)
18

27 34 41 48 55 62 69 76 83 90 98

Table 1. Monthly Correlation Coefficients, Various Periods

Correlation of Correlation of
Period D/PandY E/P and Y
Full (January 1927-May 1998) -0.28 +0.08
Early (January 1927-December 1959) -0.23 -0.49
Late (January 1960-May 1998) +0.71 +0.69

Model for Stock Market Yields may be deemed accurate but still pose a dilemma
Researchers have'shown a strong link between for fans of efficient markets.
aggregate dividend and earnings yields and Consider a simple model in which the
expected stock market returns, especially for long required long-term returns on aggregate stocks
horizons. When stock inarket yields are high (low), and bonds vary through time. Expected stock
expected future stock returns are high (low). This returns, £(Stocks), are assumed to be proportional
predictability has two possible explanations that to dividend yields, whereas expected bond
are at least partly consistent with efficient markets returns, E(Bonds), are assumed to move one-for-
(there are many inefficient-market explanations). one with current bond yields; that is,
One, investors' taste for risk varies. When investors £(Stocks}, = a + MD/P,) + Z 0)
are relatively less risk averse, they demand less in
the way of an expected return premium to bear £(Bonds), = (4)
stock market risk. Fama and French (1988, 1989), (where a is the intercept, h is the slope, D/P, is
among others, explored this hypothesis. Two, the dividend yield at time t, and e is an error term). The
perceived level of risk can change even if investors'
hypothesis is that b is positive, so expected stock
taste for risk is constant.
returns vary positively with current stock dividend
I explore the hypothesis that the perceived yields, arid that the e terms are identically and
level of risk can change (although the two hypoth- independently distributed error terms represent-
eses are not mutually exclusive). Note that investor ing the portion of expected returns not captured by
perception of long-term risk need not be accurate the model."*
for this hypothesis to be true. If investor perception
of risk is accurate, then the evidence presented here Now, I assume that expected stock and bond
may be consistent with an efficient market. If inves- returns are linked through the long-run stock and
tor perception of risk is inaccurate but explains the bond volatility experienced by investors. So,
pricing of stocks versus bonds, then the hypothesis £(Stocks),-£{Bonds), - c + da (Stocks), +ea(Bonds),. (5)

98 ©2000, Association for Investment Management and Research

 stocks versus Bonds

The hypothesis is that d is positive whereas e is that the results that follow are robust to alternative
negative. That is, I assume that the expected (or specifications of long-term volatility (i.e., from 10-
required) retum differential between stocks and year to 30-year trailing volatility) and still showed
bonds is a positive linear function of a weighted up significantly when windows as short as 5 years
difference of their volatilities.^ Although Equations were used.
3, 4, and 5 do not represent a formal asset-pricing The regressions in this section are simple linear
model, they do capture the spirit of allowing regressions that do not account for some significant
expected retums to vary through time as a function econometric problems; for example, the following
of volatility. Moreover, they yield empirically test- regressions have highly autocorrelated independent
able implications.*' variables, dependent variables, and residuals. But
Rearranging these equations (and aggregating the goal of these regressions is to initially establish
coefficients) produces the following model: the existence of an economically significant relation-
ship. Because statistical inference is problematic, I
D/P ^ Yo + Yi'^ + Y2O(Stocks) + 73a(Bonds) + ED/RJ. (6)
do not focus on (but do report) the f-statistics. The
Now, the hypothesis is that Ji is positive, Y2 is focus is on the econonric significance of the esti-
positive, and 73 is negative. This model, and the mated coefficients and R^ figures. (Subsequent sec-
precisely corresponding model for E/P, is tested in tions explore the issue of statistical significance £ind
the following section.''Other authors (e.g., Merton report robustness checks.)
1980; French, Schwert, and Stambaugh 1987) have Because I required 20 years to estimate volatility
tested the link between expected stock returns and and the monthly data began in 1926, I estimated
volatility by examining the relationship between Equation 6 by using monthly data from January 1946
realized stock retums and ex ante measures of vol- through May 1998. Before examining this equation
atility.^ However, as these authors noted, realized in full, I first examine the regression of D/P on bond
stock returns are a noisy proxy for expected stock yields only and D/P on the rolling volatility of stock
retums. I believe that linking Equatior^s 3, 4, and 5 and bond markets only for the 1946-98 period (the
and focusing on the long term will reveal a clearer first data points are dividend and bond yields in
relationship between stock n:\arket volatility and January 1946 and stock and bond volatility esti-
expected stock market retums as represented by mated from January 1926 through December 1945;
stock market yield (D/P or E/P).^ the f-statistics are in parentheses under the equa-
tions. The results are as follows;
Preliminary Evidence D/P = 4.10% - 0.03r (7)
To investigate Equation 6,1 defined a generation as (40.72) (-2.26)
20 years and used a sim^ple rolling 20-year annual- (with an adjusted R^ of 0.7 percent) and
ized monthly retum volatility for o(Stocks) and
D/P = 2.02% + 0.14a(Stocks) - 0.07a(Bonds) (8)
o(Bonds).^'^ The underlying argument is that each (11.87) (18.96) (-5.24)
generation's perception of the relative risk of stocks
and bonds is shaped by the volatility it has experi- (with an adjusted R^ of 43.0 percent).^^
enced. For instance, Campbell and Shiller (1998) Equation 7 shows that D/P and Y have a mildly
mentioned (but did not necessarily advocate) the negative relationship for 1946-1998, similar to what
argument that Baby Boomers are more risk tolerant I found for the entire 1926-98 period (Table 1).
"perhaps because they do not remember the extreme Equation 8 shows that a significant amount of the
economic conditions of the 1930s." Another example variance of D/P (note the adjusted R^) is explained
is Glassman and Hassett (1999), who argued in Dow by stock and bond volatility, with D/P rising with
36,000 that remembrances of the Great Depression stock market volatility and falling with bond mar-
have led investors to require too high an equity risk ket volatility. This relationship is economically sig-
premium. nificant. An increase in stock market volatility from
15 percent to 20 percent, all else being equal, raises
A 20-year period captures the long-terni gen-
the required dividend yield on stocks by 70 basis
erational phenonrienon that I hypothesized.^^ The
points (bps). Now, note the estimate for Equation 6:
hypothesis is inherently behavioral because it
states that the long-term, slowly changing relation- D/P = 0.007o + 0,35y +0.23 o(Stocks) - 0.31a(Bonds) (9)
ship between stock and bond yields is driven by the (-0.05) (28,77) (39.51) (-25.69)
long-term volatility of stocks and bonds experi- (with an adjusted R^ of 75.4 percent).
enced by the bulk of current investors. Although I This result supports the hypothesis. The divi-
believe a 20-year period is intuitively reasonable, dend yield is mildly negatively related to the bond
given the hypothesis, I am encouraged by the fact yield when measured alone (Equation 7), but this

March/April 2000 99
Financial Analysts Journal

negative relationship is a highly misleading indica- ship between stock and bond volatility over short
tor of how stock and bond yields covary. When I periods. In other words, a missing-variable prob-
adjusted for different levels of volatility, I found lem is not much of a problem if the missing variable
stock and bond yields to be strongly positively was not changing greatly during the period being
related. My interpretation of this regression is that examined (such as in 1960-1998). The problem is
stock and bond market yields are strongly posi- potentially destructive, however, if the missing
tively related and the difference between stock and variable varied significantly during the period
bond yields is a direct positive function of the (such as in 1927-1998).
weighted difference between stock and bond vola- Figure 3 presents the actual market D/P and
tility. Intuitively, the more volatile stocks have been the in-sample D/P fitted from the regression in
versus bonds, the higher the yield premium (or Equation 9. Figure 4 presents the residual from this
smaller a yield deficit) stocks must offer. In any regression (actual D/P minus fitted D/P). For
case, when volatility is held constant, stock yields today's reader, perhaps the most interesting part of
do rise and fall with bond yields. Figures 3 and 4 is the latest results. The actual D/P
at the end of May 1998 (the last data point) is 1.5
Again, these results are economically signifi- percent, a historic low. The forecasted D/P is also
cant. For example, a 100 bp rise in bond yields at a historic low, however—2.1 percent—which is
translates to a 35 bp rise in the required stock a forecasting error of only 60 bps.
market dividend yield, whereas a rise in stock mar-
Simply examining the D/P series leads to a
ket volatility from 15 percent to 20 percent leads to
belief that recent D/Ps are shockingly low. These
a rise of 115 bps in the required stock market divi-
regressions suggest a different interpretation; Given
dend yield.
the recent low bond yields and a low realized differ-
The fact that stock and bond yields are univari- ential in volatility between stocks and bonds, I
ately unrelated (or even negatively related) over would forecast an all-time historically low D/P for
long periods (Table 1) is a result of changes in stocks as of May 1998. The fact that the model does
relative stock and bond volatility that obscure the not forecast the actual low in dividend yield is not
strong positive relationship between stock and statistically anomalous (May's forecast error is
bond yields. The reason stock and bond yields are about 1 standard deviation below zero) and may be
univariately positively related over shorter periods a result of the stories other authors have cited to
(e.g., 1960-1998) is because of the stable relation- explain today's low D/P (e.g., stock buy-backs

Figure 3. Actual S&P 500 Dividend Yield and In-Sample Dividend Yield,
January 1946-May 1998
Yield (%)

46 51 56 61 66 71 76 96 98

Note: In-sample D/P fitted from the regression in Equation 9.

100 ©2000, Association for Investment Management and Research

 Stocks versus Bonds

Figure 4. Regression Residual: Actual D/P minus Fitted D/P, January 1946-
May 1998

-3

replacing dividends). But these stories might not be whole period, they were close to uncorrelated).
at all necessary. For example, the story of stock buy- When changing stock- and bond-market volatility
backs replacing dividends has been around since at is accounted for in Equation 10, however, the
least the late 1980s (Bagwell and Shoven 1989), yet strong positive relationship between E/P and V is
the average in-sample forecasting error of my model extended to the full period.
for D/P for 1990-1998 is only -9 bps. Apparently,
nothing more than Equation 9 is needed to explain Critique and Further Evidence
recent low dividend yields. The regression results presented in the previous
Running a similar regression for E/P, I obtained section fit intuition and the hypothesis as formal-
the following resuft: ized in Equation 6, but they are certainly open to
E/P= -1.39%+0.96y + 0.49a(Stocks)-0.76o(Boiid.s) (10) criticism. They are in-sample regression results and
(-3,70) (27.33) (29.58) (-21.56) are thus particularly open to charges of data min-
(with an adjusted R^ of 64,8 percent). The model ing. They are level-on-level regressions, which ren-
explains about as much of the variance for earnings ders the ^statistics invalid and makes the high R^
yield as dividend yield. As of the end of May 1998, figures potentially spurious.'^ Worse, they are
the E/P for the S&P 500 was 3.6 percent, corre- level-on-level regressions that use 20-year rolling
sponding to a P/E of 27.8. The forecasted E/P from data and a highly autocorrelated dependent vari-
the Equation 10 regression is 3.4 percent, or a fore- able.'* Because the inference is suspect, stock and
casted P/E ratio of 29.1. Unlike the case for D/P, I bond volatility may have followed a pattern that
am not (even to a small degree) failing to explain explained a secular-level change in dividend and
the recent high P/Es on stocks; rather, one would earnings yield merely by chance.
have to explain the opposite, because according to To examine this possibility. Figure 5 shows the
the model, the May 1998 P/E of 27.8 is slightly lozver rolling 20-year volatilities of the stock and bond
than it should be. markets used in the preceding regressions and the
Again, these results are economically signifi- ratio of stock to bond volatility. Aside from the very
cant: The required earnings yield was moving vir- early and very late years of the period, the ratio of
tually one-for-one with 10-year T-bond yields and rolling 20-year stock volatility to bond volatility
increasing 245 bps for each 5 percent rise in stock was dropping nearly monotonically from 1946
market volatility (all else being equal). Examining through mid-1998. Thus, a hypothesis that fits the
Figure 2 and Table 1 shows that E/P and Y were regression results and Figure 5 is that stock yields
strongly positively correlated only for the later and bond yields are positively related but, exoge-
period of the sample (in the earlier period, they nous to this relationship, the level of stock yields
were actually negatively correlated, and for the has been declining over time.

March/April 2000 101

 Financial Anahfsts Journal

Figure 5. Rolling 20-Year Volatilities of Stock and Bond Markets and Ratio of
Stock to Bond Volatility, January 1946~May 1998
Volatility (%) Ratio
9

''"~~\A RaHo (right axis) -

30

25 • \

20
Stock VoUtility (left axis)
15

- 3

V —.^ - 2
5
--^^'^^ — - _ .

"~ Bond Volatility (left axis)

n 1 1 1

56 61 66 71 76 81 91 96 98

The issue is one of causality. Was the drop in (with an adjusted R^ of 55.1 percent).
the level of stock yields versus that of bond yields The time-trend variables capture much of the
occurring because of changes in their relative expe- effect being studied. That is, the relationship
rienced volatilities (as I hypothesize), or were other between D/P and Y goes from weakly negative
factors causing this drop through time and thus (Equation 7) to strongly positive in the presence of
producing spurious regression results? A 50-year the trend variable—meaning that the expected dif-
regression that L^es 20-year rolling data makes ference between stock and bond yields was declin-
answering this question difficult. So, the next sub- ing through time and, after accounting for tliis trend,
sections attempt to explore this critique. stock and bond yields were positively related. The
volatility-based regression, however, is clearly the
Performance of the Model versus a Time strongest: The adjusted R^ is higher, and the coeffi-
Trend. If the drop in stock yields versus bond cient on bond yields is larger and more significant.
yields is coincidentally, not causally, related to vol- Next, the loglinear time trend is added to
atility, then a time trend might do as well as vola- Equation 9 to see how the volatility variables fare
tility in the regression tests. For ease of comparison, in head-on competition:
recall the results for D/P regressed on bond yields D/P = -10.00% + 0.35y + 0.28a(Stocks) (9a)
and stock and bond volatility; Equation 9 was (-3.98) (28.50) (19.63)
D/P = 0.00% + 0.35y + 0.23a(Stocks)- 0.31o(Bonds), -0.46a(Bonds) + 0.02(Loglinear trend)
(-^).O5) (28.77) (39.51) (-25.69) (-11.87) (3.99)
and the adjusted R^ was 75.4 percent. The next (with an adjusted R~ of 76.0 percent).
equations report similar regressions in which, Clearly, the volatility variables drive out the
instead of stock and bond volatility, either a linear time trend (analogous results held for the linear
time trend) to the point at which the trend's coeffi-
or loglinear time trend was used:
cieiit is slightly positive (the wrong sign). Although
D/P = 6.18% + 0.25(Y) -O.OO(Uneartrend) (11) the nearly monotonic fall in bond versus stock vol-
(51.44) (14.69) (-21.97) atility makes it hard to distinguish between causal-
ity and coincidence for the 1946-98 period, the
(with an adjusted R^ of 43.8 percent) and superiority of the volatility-based model over a
D/P = 27.97% + 0.33(y) - 0.04(Loglinear trend) (12) time trend gives comfort. Analogous results favor-
(32.32) (19.88) (-27.61) ing the volatility model were found for E/P.

102 ©2000, Association for Investment Management and Research

 Stocks versus Bonds

Rolling Regression Forecasts. I formed roll- The volatility-based Model 3 was nearly unbi-
ing out-of-sample forecasts of D/P starting with ased over the 1966-98 period, had the lowest abso-
January 1966. (I began in 1966 because I needed the lute bias of any of the five models, and had the
20 years from 1926 to 1946 to estimate volatility and lowest standard deviation of forecast error. The out-
the 20 years from 1946 to 1966 to formulate the first of-sample rolling regressions thus support the
superiority of the volatility model, although again,
predictive regression.) The regressions used an
the time-trend models are somewhat effective when
"expanding window" that always started in January compared with the more naive Models 1 and 2.
1946 and went up to the month before the forecast.
For comparison purposes, I formed these fore- Earliet' Data. Thebest response to many statis-
casts based on five models. Model 1 attempted to tical problems is extensive out-of-sample testing—
forecast D/P by using only the average D/P (so the that is, tests with data for a previously unexamined
forecast of D/P on January 1966 was the average period. All of the tests so far used monthly data for
D/P from January 1946 through December 1965). the commonly studied period commencing in 1926.
Model 2 attempted to forecast D/P by using a For the tests reported in this section, I used earlier
rolling regression on bond yields only. Model 3 data. Although perhaps not as reliable as the mod-
ern data, annual data on the aggregate stock and
used a rolling version of the complete model from
bond markets are available for as early as 1871.'^
Equation 9 (a regression on bond yields, stock vol-
In addition to simply using new data points,
atility, and bond volatility). Model 4 and Model 5 examining the older information provides an
corresponded to rolling versions of, respectively, advantage that is specific to this study. In Figure 6,
the linear trend model in Equation 11 and the log- the new data are used to plot the ratio of rolling 20-
iinear trend model in Equation 12. Table 2 presents year stock market volatility to rolling 20-year bond
the results of these out-of-sample forecasts. market volatility over the entire 1891-1998 period.^"^

Tabie 2. Out-of-SampIe Forecasts, January 1966-May 1998

Mode! Average Forecasting Error a(Forecasting error)
1. Using averaj^t? D/P 0.9770
2. Using regression on Y 0.29 1.38
3. Usii*g the full niodt'l 0,14
4. Using linear tinu- trend 0,71
5. Using loglinear time trend 0,54 0.62

Figure 6. Ratio of Roiling 20-Year Stock Market Voiatiiity to Bond Market

Voiatiiity, January 1891-May 1998

1891 00 1998

March/April 2000 103

Financial Analysts Journal

Recall that one problem with testing the E/P = 2.907o + 1.68y + 0.25o(Stocks) (20)
hypothesis for 1946-1998 was that the volatility (1.05) (3.13) (3.15)
ratio declined nearly monotonically. Figure 6 shows -2.23a(Bonds)
that the new data preserve this property for this (-4.50)
same time period but that the 1891-1945 period
(with an adjusted R^ of 31.5 percent).
reflects no monotonic trend. Thus, if the model
These regressions provide bad news and good
works for 1891-1945, or 1891-1998, a spurious time
news. The bad news is that some of the regression
trend is not driving the results. I found that divi-
coefficients are very different for the 1891-1945
dend yields also trended down strongly over the period from what they were for the 1946-98 period.
1946-98 period but appear much more stationary Apparently, the (admittedly simple) model is not
when viewed over the entire 1891-1998 period (this completely stable over time. Given changes in the
figure is available upon request). world economy from 1871 to 1998, to think that the
As a data check, before examining the pre-1946 coefficients would be completely stable is perhaps
data, I reexamined the 1946-98 period with the new wildly optimistic.^•'The good news is that, although
annual data set. The following are annual regres- over the 1891-1945 period the stock market's D / P
sions for the already-studied 1946-98 period: and E / P were univariately weakly positively
related to Y (see Equations 17 and 19), this relation-
D/P = 4.12%-0.04y , (13)
ship became much more strongly positive when I
(10.78) (-0.65)
allowed for changing relative stock and bond mar-
(with an adjusted R'- of-1.1 percent); ket volatilities (as in the completely separate 1946-
98 period). This relationship was, as my hypothesis
D / P = -1.15% + 0.29y + 0.24o(Stocks) (14)
forecasted, a strong positive function of the previ-
(-1.64) (6.07) (8.03)
ous 20 years' relative stock versus bond volatility.
- 0.16a(Bonds)
Finally, I present the regressions for D / P for
\ --- -, the full 1891-1998 period. For comparison, I also
(with an adjusted R^ of 66.0 percent; present full-period tests of the time-trend variables
E / P = 6.98% + 0.13y (15) (the E / P results were highly analogous for all
(7.57) 1(0.95) regressions):^^

(with an adjusted R^ of -1.8 percent); D/P = 5.20% - O.uy (21)

(17,79) (-2,53)
E/P = -3.12% + 0.857 + 0.46a(Stocks) (16)
(-1.64) " (6.07) (8.03) (with an adjusted R^ of 4.8 percent);
- 0.40O (Bonds) D/P = 5.90% + 0.03y - O.OOLinear trend (22)
(17.32) (0.42) (-3.54)
(with an adjusted R^ of 48.9 percent).
(with an adjusted R^ of 14.1 percent);
Although not precisely the same as the monthly
regressions presented earlier, the annual regres- D/P = 7.75% - 0.06y - 0.07LogIinear trend (23)
sions on the new data set are similar enough to be (6.09) (-0.91) (-2.06)
encouraging. (with an adjusted R^ of 7.6 percent);
Now, consider the results for these same
D/P = 1,98% + 0.26y + 0.14o(Stocks) (24)
regressions for the earlier 1891-1945 data:
(2.96) (3.52) (4.95)
D/P = 2.60%+0.77r (17) - 0.29o(Bonds)
(2.70) (2.72) (-5.65)
(with an adjusted R^ of 10.6 percent); (with an adjusted R^ of 35.5 percent).
D/P = -1.65% + 1.36y +0.19a(Stocks) (18) The earlier data and the full-period data
(-].18) (5.00) (4.75) strongly support the central tenet of the hypothesis:
Without adjusting for volatility and with or with-
- 0.53o(Bonds) out a time trerid (Equations 21-23), either a nega-
(-2.10) tive or flat relationship appears between D / P and
(with an adjusted R^ of 35.7 percent); bond yields over the entire period. After adjust-
ment for relative stock and bond volatility, this
E/P = 4.20%+ 1.06y (19) relationship is strongly positive (Equation 24).
(2.20) (1.90) Unlike the 1946-98 results, these results are clearly
(with an adjusted R^ of 4.6 percent); present in the absence of a significant trend in the

104 ©2000, Association for Investment Management and Research

 Stocks versus Bonds

ratio of stock to bond market volatility and despite earlier regression tests, which were admittedly an
any changes in the world economy from 1871 to econometric nightmare, nonoverlapping observa-
1998. In fact, unlike the volatility-based model, the tions were used for Figure 7, and the autocorrela-
time trends utterly fail to resurrect the positive tion of both the dependent and independent series
relationship between stock and bond yields over was close to zero. '"^ Thus, any need for econometric
the full period. When I used the data for 1946-1998, corrections (e.g., first differencing) was avoided.
I introduced the issue of distinguishing whether The problem now is that I have only six obser-
the volatility-based model was spuriously sup- vations, so the tests might lack power, but this is
ported because the changes in relative volatility
approximated a time trend. The earlier and full- not the case. The f-statistic of the regression line is
period evidence powerfully indicates that it is the +7.64, and the adjusted R~ is 92.0 percent. With six
time trend whose efficacy is spurious for 1946- observations, a f-statistic must exceed +2.45 to be
1998, not the volatility-based model. significant at a p value of 2.5 percent in a one-tailed
test. Clearly, the /-statistic for this test is well past
Full-Period Scatter Plots. As a final and per- this level of significaiice.
haps most compelling test, I examined nonoverlap- As a robustness check, I recreated Figure 7 but
ping 20-year periods from 1878 until 1998.1 report starting 10 years later (resulting in only five obser-
the results for the resulting six observations in vations over this period). The results are in Figure
Figure 7. Figure 7 plots the ratio of annualized 8. This figure is even more striking than Figure 7
monthly stock market volatility over correspond- (the f-statistic in Figure 8 is +12.46, and the adjusted
ing monthly bond volatility for the 20 years ending R^ is 97.5 percent). Note from Figure 6 (the graph
before the labeled year against the excess of stock
market earnings yields over bond yields for the of the rolling volatility ratio) that two peaks are
year in question. I chose earnings yields for this visible in the ratio of stock to bond volatility. These
investigation because the evidence is that they are peaks roughly correspond to the right side of,
directly close to being comparable to bond market respectively. Figures 7 and 8. In both cases, the
yields whereas dividend yields move as a damp- model fits these extreme observations exception-
ened function of bond yields (that is, the coefficient ally well (that is, the largest volatility ratio corre-
on y in Equation 10 is nearly 1.0, which makes the sponded to the largest end-of-period gap of stock
simple difference relevant to examine). earnings yield over bond yield). Also note that
Figure 7 clearly supports the model: The these two periods (the 20 years ending in 1918 and
greater stock volatility is versus bond volatility, the the 20 years ending in 1948) share no overlapping
higher E/P musi be versus Y. In contrast to the observations, yet the model fits both perfectly.

Figure 7. Ratio of Annualized Monthly Stock Market Volatility to Correspond-

ing Monthly Bond Voiatiiity versus Excess of Stock Market Earnings
Yield over Bond Yield, 1871-May 1998
E/p-r(%)
16
1918
"14 h-

12

10 -
1938 ^

6 -
1978
4
• 1958
2 1898
0

-2 - •
-A 1998
5 10 15

Ratio of 20-Year Stock to Bond VolatiJity

March/April 2000 105

 Financial Analysts Journal

Figure 8. Ratio of Annualized Monthiy Stock Market Volatility to Correspond-

ing Monthly Bond Volatility versus Excess of Stock Market Earnings
Yield over Bond Yield, 1881-May 1998
E/p-y(%)
10

8 -
1908 ^ x *. • " ^ 1948

4 - 1928 ^

0
^ ^ ^ 1968

-2
1988
1 J 1

4 6 8 Ifl
Ratio of 20-Year St(x:k to Bond Volatility

Finally, for completeness, I present in Table 3 Market Predictability

the adjusted R^ and f-statistics for each of eight Researchers have found that variables D/P and E/P
possible regressions on nonoverlapping periods have power to forecast aggregate stcx:k market
for which I have six 20-year data points (each row returns. Moreover, this power appears to increase as
in Table 3 presents the results of a regression that time horizon lengthens (e.g., Fama and French 1988,
differs by one year in its starting and ending point 1989). I tested this finding for 1946-1998 using pre-
from the prior/next row). Only one of these eight dictive regressions of excess monthly and annual-
regressions produced results well below tradi- ized 5- and 10-year compound S&P 500 returns on
tional levels of significance, and even in this case, aggregate D/P (f-statistics on all multiperiod regres-
the sign is correct.^*^ sions were adjusted for overlapping observations
We believe these nonoverlapping tests are com- and heteroscedasticity). Here are the findings:
pelling evidence, irrespective of the econometric S&P monthly return = -0.56% + 0.32D/P (25)
problems with our earlier tests, that following long- (-1.03) (2.38)
term periods of high (low) stock market volatility
relative to bond market volatility, the required yield (with an adjusted ?? of 0.7 percent);
on stocks is relatively high (low) versus bonds. S&P 5-year return = ^.13% + 4.09D/P (26)
(-0.88) (4.77)
(with an adjusted R^ of 56.1 percent);
Table 3. Statistics for Eight Regressions
S&P 10-year return = -1.443% + 3.22D/P (27)
Period Adjusted K^ /-Statistic
(-0.38) (4.34)
1891-1991 88.5% +6.28
1892-1992 73.7 +3.87 (with an adjusted K^ of 58.7percent).
1893-1993 81.0 +4.72 Equations 25-27 verify the findings of other
1894-1994 45.6 +2.28 authors that D/P has weak, but statistically signif-
1895-1995 99 +1.25 icant, power for forecasting monthly returns and
1896-1996 91.5 +7.42
strong statistically significant power for forecasting
1897-1997 +4,36
longer-horizon returns.
1898-1998 92.0 +7.64
Mean mi +4.73 Now, a new predictive variable, D/P(Error), is
Median +4.54 introduced. It is the in-sample residual term from
Note: Each row presents the results of a regression that differs by the regression of D/P on Y, a(Stocks), and a(Bonds)
one year in its starting and ending point from the prior/ next row. for the 1946-98 period (Equation 9). It represents the

106 ©2000, Association for Investment Management and Research

 Stocks versus Bonds

D / P on the S&P 500 in excess or deficit of what I (with an adjusted R^ of 57.1 percent);
would have predicted had I been using this model S&P 10-year return = -3.00% -H 3,61D/P(Fit) (34)
to forecast D / P (i.e., the unexplained portion). The (-0.76) (4.81)
results of the same regression tests as done for Equa-
tions 25-27 on this new variable are as follows (all -\- 2.29D/P(Error)
results of this section were analogous when tested (2.00)
on E/P): (with an adjusted R^ of 61.1 percent).
S&P monthly return = O.f.77a + 1.75 D/P (Error) (28) Clearly, the power of D/P for predicting short-
(4.29) (6.74) run (monthly) S&P 500 returns is driven by D/
P{Error). As horizon lengthens, D/P(Fit) becomes
(with an adjusted R^ of 6.6 percent); more and more important, and at the 10-year
S&P 5-year return - 12.60% + 4.65 D/P (Error) (29) horizon, D/P(Fit) is considerably more important.
(6.50) (3.00) To examine even longer forecast horizons and
over longer periods, I again used annual data back
(with an adjusted R^ of 21.2 percent);
to 1871 and formed D/P(Fit) and D/P(Error) from
S&P 10-year retum = 12.08% + 2,01D/P(Error) (30) Equatiori 24. Recall that the first 20 years are needed
(5.64) (1,35) to estimate volatility, so the following regressions
(with an adjusted R^ of 7.1 percent). are for 1891-1998 (all returns are annualized com-
Comparing the results for D/P(Error) with pound returns):
D / F shows that D/P(Error) has far more predic- S&P annual retum = 18.1% - 1.46D/P(Fit) (35)
tive power than D / P at short (monthly) horizons (1.89) (-0.71)
but far less power at longer horizons.^^ The power -1- 2.89D/P(Error)
of D/P(Error) to forecast short-horizon returns (1.91)
can be interpreted as picking up time-varying risk
(with an adjusted R^ of 2.0 percent);
aversion or, alternatively, as market mispricing (I
leave this decision to future work). In either case, S&P 5-year return = 5.32% + 0.99D/P(Fit) (36)
(0.59) (0,51)
when D/P(Error) is high, stocks are selling for
lower prices than is usual in the same interest rate -H 2.32D/P(Error)
and volatility environment and those low prices (3,67)
indicate higher short-horizon expected returns (with an adjusted R^ of 12.2 percent);
(and vice versa). S&P lO^year return = -1.78% + 2.43D/P(Fit) (37)
Finally, I fo?med D/P(Fit) as the fitted values (-0.21) (1.43)
from regression Equation 9. D/P(Fit) can be inter-
+ 0.81D/P(Error)
preted as the normal dividend yield as forecasted
by the model considering the level of bond yields (1.89)
and stock and bond market volatility. By construc- (with an adjusted R^ of 12.4 percent);
tion, the following relationship holds: S&P 15-year return =-10.89% + 4.24D/P(Fit) (38)
D/P = D/P(Fit) + D/P(Error). (31) (-3.91) (9,72)
By regressiiig stock returns on both D/P(Fit) + 0.18D/P (Error)
and D/P(Error), I decomposed the forecasting (0.43)
power of D/P into a portion coming from fitted
D/P and a portion coming from residual D/P. The (with an adjusted R^ of 33.7 percent);
following regressions were carried out for 1946- S&P 20-year retum = -8.66% + 3.74D/P(Fit) (39)
98 ^^ (-2.36) (5.58)
S&P monthly return = 1.25% - 0.15D/P(Fit) (32) - 0.29D/P(Error)
(2.07) (-0.99) (-2.08)
-H 1.75D/P(Error) (with an adjusted j<^ of 42.2 percent).
(6.74) The estimated coefficients of D / P (Fit) and
(with an adjusted R^ of 6.6 percent); D/P(Error) for each of the forecast horizons
(regression Equations 35-39) are plotted iii Figure
S&P 5-year return = -2.80% + 3.77D/P(Fit) (33)
9. Although annual predictability (Equation 35) is
(-0.56) (3.93)
weak, the short-term predictability present is
+ 4,96D/P(Error) clearly driven by D/P(Error). The story changes
(4.97) dramatically as horizon increases, until at long

March/April 2000 107

 Financial Analysts Journal

Figure 9. Estimated Coefficients of D/P(Fit) and D/P(Error) for Each Forecast

Horizon, 1891-1998
Regression Coefficient
5

4 -

3 -

2 -

1 -

0 --

-I -

10 20
Horizon (years)
Note: Ail returns are aniiuaiized compound retiinvs.

horizons (15 years and 20 years), D/P(Fit) is Do Stock Yields Have Farther to
clearly addiiig considerable predictive power
whereas D/P(Error) is adding none. Figure 9 tells Fall?
a clear story that at short horizons, D/P(Error) is Many have wondered lately why the market is
what counts but at long horizons, what counts is currently selling at such a historically low D/P and
D/P(Fit). (Analogous results held for E/P.) E/P (or high P/D and P/E). In particular, in the
To sum up, the forecasting power of D/P can be book Dow 36,000, Classman and Hassett came to an
decomposed into the forecasting power of D/P(Pit) extreme conclusion. They argued that the reason
and D/P(Error). In the model, D/P(Fit) is the nor- stock prices seem so high relative to measures such
mal or expected dividend yield, and D/P(Error) is as dividends and earnings is that the expected (or
interpreted as the D/P in excess (or deficit) of nor- required) return on the stock market is going down
mal. Evidence presented here indicates that D/P as investors realize that the stock market is less risky
itself forecasts stock returns at both long and short in relation to the bond liiarket than previously
horizons but for different reasons. D/P(Eit) fore- thought. Furthermore, they reasoned that this fall
casts long-horizon stock returns but has almost no in expected returns is not over yet and concluded
power for the short term. D/P(Error) forecasts that it will not stop Lmtil stock and bond market
short-horizon stock returns but has little power for expected returns are equal (a point at which, by
the long term. their calculations, the Dow will reach approxi-

108 ©2000, Association for Investment Management and Researcii

 Stocks versus Bonds

mately 36,000). Part of their reasoning sounds much safe stocks really are, is actually proceeding much
like the arguments advanced here. Well, part of it is, as it has throughout financial history. According to
and part of it is not. the model, investors have repriced stocks to reflect
Their first conclusion is 100 percent consistent a lower perception of stock market risk, but any
with this article: the coticlusion that stocks have low farther drop in the required return on stocks (and
yields now because they are perceived to be less concurrent rise in stock prices) must come from a
risky versus bonds than historically normal. In fact, further reduction in actual stock volatility (versus
my central thesis is that the return required by inves- bond volatility) or a reduction in bond yields. If
tors to own stocks versus bonds varies directly with investors have been all along implicitly using the
the perceived relative risk of the two assets (for relationship hypothesized here to price stocks (as
which I used their respective rolling 20-year volatil- the data strongly support they have since at least
ities as proxies). I believe that my model, coupled 1891), then they have acted consistently in recently
with currently low bond market yields and a low raising the price of equities. But we can expect no
perceived risk of stocks versus bonds, entirely more such rises unless either interest rates or real-
explains, within the bounds of statistical error, ized relative volatility change.^** The model dis-
today's low yields on stocks (and, according to the cussed here suggests that unless the inputs to the
model, the low long-term expected returns that model change, any repricing of equities is approx-
come with low yields). Thus, my work strongly imately complete.
supports one aspect of the argument in Doit' 36,000, Finally, if the model is accurate, a belief that a
namely, that stock market expected returns versus near-term windfall profit of about three times your
bonds have come down as investor perceptions of money is currently available in the broad stock
the relative risk of stocks versus bonds have market, a belief held by Glassman and Hassett, is
changed. dangerous. First, investors who believe in the
My conclusions differ, however, from the next windfall possibility may overallocate to stocks.^^
conclusion of Dow 36,000. Glasstnan and Hassett Second, short-term pricing errors induced by
extrapolated the trend in lowered return-premium believers in this argument (or "bubbles") can be
expectations to continue, but my model offers them dangerous to the real economy. Third, and perhaps
no support. The authors of D'ozo 36,000 stated that most worrisome, if the model presented and tested
the fall in stock expected returns is not over yet and in this paper is correct, the belief that stocks stand
will not be complete until the expected return on to receive a one-time enormous windfall profit is
stocks is the same as bonds (presumably not yet the not simply wrong, it is backward. The low stock
case) because the c«jthors believe that stocks are no yields of today are fully explained by the model,
riskier than bonds in the long term. This hypothesis meaning that the forecast of short-term stock
is quite provocative. If stocks are no riskier than returns is about average.•^^ Moreover, if the conclu-
bonds, then stock prices should rise as investors sion here is true that the best forecasting variable
realize stocks are currently priced as if they are for long-term stock returns is the absolute level of
more risky. Now, much debate involves the long- stock yields, then today's low yields (both D/P and
run risk of stocks versus bonds, and to review or E/P) point to a poor forecast for the long-term
settle this matter is not the province of this paper.^^ return on stocks.
However, much of the reason behind the current
prominence of this debate in the first place is how
different today appears from the past (i.e., today's Conclusion
historically high stock prices versus dividends or Each of the puzzles stated at the beginning of this
earnings). My conclusion is that, in fact, tbe struc- article can be resolved by using the model provided
ture of the world really is not much different today; in Equation 6 for the required yield on stocks. Con-
only the inputs to the model have changed. In other sider the first question: Why did the stock market
words, stock yields (and required returns) have strongly outyield bonds for so long only to now
always moved with bond yields, and the relative consistently underyield bonds? The model states
difference between them has always been a func- that (1) the higher bond yields are and (2) the higher
tion of their relative perceived volatility. In fact, perceived stock market volatility versus bond mar-
when I directly estimated this relationship, I found ket volatility is, then the higher stock yields must be.
that it fits well for the long term and fits well today. For a long time (before the 1950s), stocks outyielded
The reason the study reported here is a prob- bonds because the realized volatility of stocks ver-
lem for theories like those proposed in Dozv 36,000 sus bonds was much higher than in modem times.
is that I say the rise in stock prices today, rather than Consider the second question: Why did stock
simply beginning as investors start to perceive how and bond yields move relatively independently, or

March/April 2000 109

 Financial Analysts Journal

even perversely, in the 1927-98 period but strongly link expected stock returns to ex ante volatility
move together in the later 40 years of this period? through realized stock returns.^*^ However, realized
Stock and bond yields appear to move indepen- stock retums are very noisy. I hypothesized that D/P
dently or even perversely over long periods (e.g., (or E/P) is a proxy for expected stock retLims and that
1926-1998), but this appearance is an artifact of y is a proxy for expected bond returns and found
missing a part of their structural relationship. If the strong confirmation that the difference between
impact of changing volatility is taken into account, these proxies is a positive function of differences in
stock and bond yields are strongly positively corre- experienced volatility. In other words, unlike many
lated over the entire period for which we have data, other studies, I have documented a strong positive
which many strategists and economists would have intertemporal relationship between expected retum
hypothesized. and perceived risk.
Finally, consider the third question: Why are This article demonstrated that the relative long-
today's stock market yields so low and what does term volatility experienced by investors is a strong
that fact mean for the future? Today's stock market driver of the relative yields they require on stocks
yields are so low simply because bond yields are versus bonds; it did not show that these long-term
low and recent realized stock market volatility has realized volatility figures are accurate forecasts of
been low when compared with bond market vola- future volatility. Thus, I have clearly identified a
tility. I do not need to resort to "the world has behavioral relationship that I believe is important,
changed" types of arguments to explain today's but I offer no verdict on market efficiency.^"^
low yields. The model fully explains them. And the The bottom line is that today's stock market (as
model indicates that they will not go much lower of May 1998) has very low yields (D/P and E/P) for
unless realized stock versus bond volatility or inter- the simple reason that bond yields are low and stock
est rates fall farther. voiatiiity has been low as compared with bond vol-
Although testing a long-term, slowly changing atility. These conditions historically lead investors
relationship has statistical difficulties, the model to accept a low yield (and expected retum) on
easily survived every reasonable robustness check, stocks. If one is a short-term investor, knowing that
including out-of-sample testing of a previously these low yields are not abnormal may be comfort-
untouched period (1871-1945) and the formation of ing. A long-term investor, however, might be very
completely nonoverlapping, nonserially correlated nervous, because raw stock yields (D/P and E/P)
independent and dependent variables for the entire are the best predictors of long-term stock market
1871-1998 period. returns and these raw yields are currently at very
This work li,as strong theoretical implications. low levels.
A link between volatility and expected return is one
of the strongest implications of modern finance.^''
Researchers have found compelling evidence of this
phenomenon in comparing asset classes (i.e., stocks The author would like to thank Jerry Baesel, Peter
versus bonds), but evidence of a link within asset Bernstein, Roger Clarke, Tom Dunn, Eugene Fama,
classes (e.g., testing the capital asset pricing model Ken French, Britt Harris, Brian Hurst, Antti Unianen,
for stocks) or an intertemporal link within one asset Raylwanowski, David Kabiiler, Bob Krail, Tom Philips,
class has been weak. This article addresses the inter- Jim Pieerno, Rex Sinquefield, and especially John Liew
temporal link. Past studies failed to convincingly for helpful comments and editorial guidance.

Notes
1. A set of assumptions sufficient for this equality to hold for ing tny bond retum series more homoscedastic. The choice
coupon-bearing bonds is that the yield curve be flat and of a duration of 7.0 wa.s arbitrary and had no effect on the
unchanging. results. 1 performed the regression of tliis excess retum
2. The sources aud/or construction of the data for this article series on the excess monthly retum (tf the Ibbotson Associ-
are as follows: For stocks, retum and earnings yield data on ates long- and intermediate-term bond series for January
the S&P SOO came from Datastream and dividend yields ]980toMay 1998. For January 1926 to December 1979,1 used
from lhbolson Associates. Forboiids, return data For |aiiu- the fitted values on the Ibbotson retum series to approxi-
ary 1980 to May 1998 are from the|.P. Morgan Government mate the 7.n-year duration J.P. Morgan govemment bond
Bond Index levered to a constant duration oi 7.0 (i.e., the series. For bond yields, I used the 10-year benchmark yield
monthly return usod is tlie T-bill rale plus 7.0 divided by from Datastream from January I980toM<iy 199!^. For Janu-
the beginning-of-the-month J.P. Morgan duration times the ary 1926 to December 1979, I used the fitted multiple regres-
return on the J.P. Morgan index minus the T-bill return). I sion forecast (fitted from the regression over the January
constructed a constant-duration bond in the hopes of mak- 1980- May 1998 period) of the 10-year yield on the Ibbotson

110 ©2000, Association for Investment Management and Research

 sbort-, intermediate-, and long-term governinent bond
series. The results are not sensitive to precise definitions of
the bond yield or return.
3. The earnings yield I used is prior year's earnings over
current price. All tbe economic results in this article are
robust to using eitber a 3- or 10-year moving average of real
earnings in the numerator.
4. Equation 3 almost assuredly should be augmented witb
variables proxying time-varying expected dividend growth
(see Fama and French 1988). I have tested such proxies and
found them to be statistically significant, but I omitted them
from tbis article because they affect none of the results or
conclusions significantly.
5. Bernstein (1993,1997) examined a related (altbougb sligbtly
different) model and came to some of tbe same conclusions.
6. The results pre.sented here were insensitive to assuming
other reasonable functional forms for this relationship (for
example, assuming linearity in the log of tbe volatilities
rather than the levels).
7. Kane, Marcus, and Nob (1996) examined a related model
for the first difference of market P/Es (a somewhat different It's How Bank One
exercise) and came to some conclusions similar to mine.
8. Tbese studies used forms similar to Equation 5. Encourages Ideas
9. Another logical extension of Equations 3, 4, and 5 is Y - c A big part of the reason that Sank One is the
+ eT-bill + i/a(Bonds). That is, the yield on bonds moves nation's 5'" largest bank holding company is our
(possibly at a multiple) witb the short-term interest rate, commitment to encouraging every employee to
and tbis weighted d i fference between long-term and sbort- express their brightest ideas-and proactiveiy act
term yields is a positive function of perceived bond vola- on the ideas of others. As a member of our team,
tility. Although not the focus of this article (but the focus you'li have the power to approach managers and
of a future paper), empirical tests of tbis equation strongly supervisors with your thoughts on improving
support tbis specification. service to customers, streamlining operations or
10 This work is not sensitive tci the definition of generation as enhancing any other aspect of our business. Even
precisely 20 years. better, your efforts wili be both recognized and
11 Note that I am not attempting to use tbe best short-term rewarded. Join us in Columbus for the encourage-
conditional estimate of volatility. Sbort-term changes in ment you need to achieve great things.
volatility may be mostly transitory. If so, they would have
little impact on stock prices and required stock yields (see, Mutual Funds
for instance, Poterba and Summers 1986).
12. All R values were adjusted for degrees of freedom. Marketing Manager
13. Granger and Newfcold (1974) found that in regressions of
You'ii manage the planning and execution of
one random walk on another, rejection of tbe null hypoth- marketing and communication programs for One
esis is more tbe rule than the exception. Also see Kirby Group Mutual Funds. Success will invo[ve oversee-
(1997) or Goetzmann and jorion (1993). ingthe creation of dynamic product marketing and
14. As mentioned previously, the results of this article are not communication materials, developing introduction
very sensitive to the choice of a 20-year window for volatil- programs for new products and presenting a
ity. For instance, using a 10-year window for volatility esti- unified brand strategy to support sales efforts for
mation greatly reduced (but did not eliminate) tbe degree a variety of distribution channels.
of autocorrelation in the rigbt-band variables. Wben 1 rees-
timate Equation 9 using 10-year rolling volatility (wbich We require 4-7+ years of experience managing
also added 10 more years, 1936-1945, to the regression), the marketing projects, preferably for investments or
(-statistics did not materially change; the f-statistics on Y, mutual funds. A background developing and
a(Stocks), and a(Bonds) were, respectively, +10.00, +14.45, evaluating creative production is essential. The
and -14.75. Using a 7-year window (now adding data from ability to effectively interface with investment
1933-1945 to tbe regression), the (-statistics were +5.21, management professionals, sales managers and
+11.54, and -10.93. A later section addresses this issue more wholesalers is also expected. The strategic thinker
directly by using longer-term data and analyzing nonover- we seek must be an excellent communicator with a
lapping 20-year periods. keen understandingofthe creative process.
15. The sources for these data are Robert J. Sbiller's Web page
We'll reward your efforts with competitive compen-
(an update of tbe data in Chapter 26 of Sbilier 1989) and the
sation, comprehensive benefits and serious
company Global Financial Data.
advancement potential. Send your resume,
16. Tbese ratios are somewhat higher than reported in Figure indicating iob code C00039, to: Bank One, P.O.
5because the duration of the bond used in these annual tests Box 540277, WalUiam, MA 02454-0277; e-mail;
was, on average, somewbat shorter than the duration of 7.0 [email protected]; Fax; 1-800-424-
years used in the monthly tests. Thus, bond volatility is 3 1 8 8 . Although resumes are processed in
somewhat lower in these annual tests. This cbange is only Wattham-position is based in Columbus, OH. Visit
a matter of scale and has no economic effect on the tests. us on the Internet at www.bankone.com/careemav.
17. For instance, Fama and Frencb (1988) found that the param- BankOneisan equal opportunity employer and we
eters of the Lintner (1956) model for explaining dividend support diversity in the workforce m/f/d/v.
changes changed radically during the 1927-86 period.
18. As a final cbeck, 1 reestimated Equation 24 using the
Gochrane-Orcutt procedure to adjust for first-order auto- PoweT

March/Aphl 2000 111

 Finaiiciat Analysts Journal

correlation in the residuals. Each coefficient was essentially 23. Two good sources for a scholarly but readable review of
the same and remained statistically significant, whereas the these issues are Siegel (1994) and Cornell (1999).
first-order annual residual autocorrelation was highly sta- 24. Glassman and Hassett did offer some reasons why stock
tistically significant at 0,55, volatility might be lower in the future than in the past, but
19. This low autocorrelation matches the results of Poterba and their central argument does not need this farther drop to
Summers, who found only very short-term persi.stence in happen because their argument is that stocks are no more
market volatility. Interestingly, I found that long-term roll- risky than bonds right now.
ing estimates of volatility seem to be crucial in determining 25. In all fairness, the actual practical investment advice in the
the required expected return on the market but do not book Dozi' 36,000 appears quite reasonable, although it is
forecast the next period of long-term volatility itself. Thus, still easy lo see how an investor who believes in the authors'
although investor perceptions of volatility drive market premise will not act so reasonably.
expected returns, those perceptions have not necessarily 26. When this article was written. May 1998 data were the latest
been accurate. My model might correctly describe investor used. As of November ] 999, the model's short-term forecast
behavior, but reconciling this behavior with market effi- for stocks had joined the long-term forecast of stocks as
ciency may be difficult (although not necessarily impossi- below average, although not nearly as severely below aver-
ble). I leave this endeavor to future work. age as the long-term forecast. I would be happy to provide
20. In fact, the failure of the one regression (1895-1995) was a more up-to-date forecast and can be contacted at
driven by the 1975 observation (without this observation, cliff,asness@aqrca pital.com. Of course, trade on such a fore-
the regression had an R" of 89,5 percent and a (-statistic of cast at your own risk!
+5.93). Furthermore, by the luck of the draw, this regression 27. This link does not need to hold precisely for inefficient
did not include values for either the ,v or y variable as portfolios.
extreme as in Figures 7 and 8, which lowered the power of 28. An exception is Kane, Marcus, and Noh, who correctly
this test. pointed out that this relationship is much clearer in ex ante
21. These regression results should not be considered an accu- measures than in ex post returns.
rate test of a short- or long-term trailing strategy. First, the 29. Unfortunately, I also could not determine the rationality of
regressions used D/P, which because it has price in the the predictive power of D/ P(Error) over short horizons and
denominator, is known to induce a small bias toward find- D/P(Fit) over long horizons. Modigliani and Cohn (1979)
ing a positive coefficient. The regressions also used the full- argued that when inflation (and presumably bond yields)
period data to form D/P(Error), which would not have been is low, investors mistakenly (i.e., irrationally or ineffi-
known prior to the end of the period. Finally, of course, the ciently) overprice equities (and vice versa). The empirical
regressions do not account for trading costs. These regres- results of this study support their hypothesis in one way:
sions are meant to be indicative of the forecasting power of When volatility is held constant, investors do price stocks
the model versus traditional models. Formal tests of a trad- at higher P/Es and P/Ds when interest rates are low (and
ing strategy based on these methods are not available from vice versa). This empirical finding is an important contribu-
the author; trying to profit from such strategies is what I do tion, because more-naive tests (which fail to account for
fora day job, relative volatility changes) do not pick up this relationship.
22. These tests were carried out on in-sample regression resid- However, distinguishing whether the short-term predictive
uals tu retain the full 1946-98 period. Analogous signifi- power of D/P(Error) or the long-term predictive power of
cant results (although a bit weaker) were found for 1966- D/P(Fit) is coming from such mispricing or rational vari-
1998 when rolling out-of-.sample versions of D/P(Fit) and ance in expected return (perhaps caused by changing risk
D/P(Error) vi'ei^ used. aversion) is beyond the scope of this article.

References
Bagwell, Laurie Simon, and John B. Shoven, 1989, "Cash . 1988. "Stock Prices, Earnings, and Expected Dividends."
Distributions to Shareholders," journal of Economic Perspectives, journal of Finance, vol. 43, no. 3 (July): 661-676.
vol. 3, no. 3 (Summer):129-140, Cole, Kevin, Jean Helwege, and David LasEer. 1996. "Stock
Bernstein, Peter L, 1993, Volatility: Two Views, Two Signals. Market Valuation Indicators: Is This Time Different?" Financial
Newsletter. (November 15). Analysts Journal, vol. 52, no. 3 (May/June):56-64.
, 1996. "Are Stocks the Best Place to Be in the Long Run? Cornell, Bradford. 1999. The Ecfuity Risk Premium. New York:
A Contrary Opinion." journal of Investing, vol, 5, no, 2 John Wiley & Sons.
(Summer) :6-9. Eama, Eugene F,, and Kenneth R. French, 1988. "Dividend
. 1997. Stock/Band Risk Perceptions and Expected Returns. Yields and Expected Stock Retums." Journal of Financial
Newsletter. (February 1). Economics, vol. 22, no. 1 (October) :3-26,
Bogle, John C. 1991. "Investing in the 1990s: Occam's Razor . 1989. "Business Conditions and Expected Returns on
Revisited," journal of Portfolio Management, vol. lH, no. 1 Stocks and Bonds," journal of Financial Economies, vol, 25, no. 1
(Fall):88-91, (July): 23-50,
. 1995. "The 1990s at the Halfway Mark." journal of French, Kenneth R., G, William Schwert, and Robert F.
Portfolio Management, vol. 21, no. 4 (Summer):21-31. Stambaugh. 1987. "Expected Stock Retums and Volatility,"
Boudoukh, Jacob, and Matthew Richardson. 1994, "The Journal of Financial Econoinics, vol. 19, no. 1 (January) :3-29.
Statistics of Long-Horizon Regressions Revisited," Mathematical Glassman, James K., and Kevin A. Hassett, 1999, Dow 36,000.
Finance, vol. 4, no. 2 (Aprii):103-n9, New York: Random House,
Campbell, John Y., and Robert J. Shiller. 1998, "Valuation Ratios Goetzmann, William N., and Philippe Jorion. 1993. "Testing the
and the Long-Run Stock Market Outlook." journal of Portfolio Predictive Power of Dividend Yields." fournal of Finance, vol. 48,
Management, vol. 24, no, 2 (Winter):ll-26. no. 2 aune):663-679.

i 112 ©2000, Association for Investment Management and Research

 stocks versus Bonds

Goyal, Amit, and Ivo Welch. 1999. "Predicting the Equity Modigliani, F., and R. Cohn. 1979. "Inflation, Rational
Premium." Working paper. University of California at Los Valuation., and the Market." Financial Analysts journal, vol. 35,
Angeles. no. 2 (March/April):24-44.
Granger, CJ.VV., and P. Newbold. 1974. "Spurious Regressions Phillips, Thomas K. 1998. "Why Do Valuation Ratios Forecast
in Econometrics." journal of Econometrics, vol. 2:111-120. Long Run Equity Returns?" Working paper. Paradigm Asset
Management.
Ilmanen, Antti. 1995. "Time-Varying Expected Retums in
International Bond Markets." journal of Finance, vol. 50, no. 2 Poterba, James M., and Lav^^rence H. Summers. 1986. "The
(June) :481-506. Persistence of Volatility and Stock Market Fluctuations."
American Economic Review, vol. 76, no. 5 {December):1142-51.
Kane, Alex, Alan J. Marcus, and Jacsun Noh. 1996. "The P/E
Samuelson, Paul A. 1994. "The Long-Term Case for Equities and
Multiple and Market Volatility." Financial Analysts Journal, vol.
How It Can 'BeOversold." journal of Portfolio Management, vol. 21,
52, no. 4 (July/August):16-24.
no. 1 (Fall):15-26.
Kirby, Chris. 1997. "Testing the Predictable Variation in Stock Schwert, G. WiUiam. 1989. "Why Does Stock Market VolatUity
and Bond Retums." Review of Financial Studies, vol. 10, no. 3 Change over Time?" journal of Finance, vol. 44, no. 5
(FaU):579-630. (December):in5-54.
Lander, Joel, Athanasios Orphanides, and Martha Shiller, Robert J. 1989. Market Volatility. Boston, MA:
Douvogiannis. 1997. "Earnings Forecasts and the Predictability Massachusetts Institute of Technology Press.
of Stock Returns: Evidence from Trading the S&P." journal of Siegei, Jeremy. 1994. Stocks for fhe Long Run. Burr Ridge, IL:
Portfolio Management, vol. 23, no. 4 (Summer):24-35. Irwin.
Lintner, John. 1956. "Distribution of Incomes of Corporations Sorensen, Eric H., and Robert D. Arnott. 1988. "The Risk
among Dividends, Retained Earnings, and Taxes." American Premium and Stock Market Performance." Journal of Portfolio
Economic Review, vol. 46:97-113. Management, vol. 14, no. 4 (Si.m-imer):50-55.
Merton, R.C. 1980. "On Estimating the Expected Retum on the Stambaugh, Robert F. 1984. "Bias in Regressions with Lagged
ol. 8,no.4(June):323-361. Stochastic Regressors." Manuscript.

March/April 2000 113