Market Predictability

Pedro Santa-Clara
Universidade Nova de Lisboa
 Predicting returns
• Importance
– Return predictability is the holy grail for
investment management
– Huge impact in corporate finance decisions
• A small degree of return predictability is
consistent with rational and efficient markets
– Expected returns may change slowly with
economic conditions to reflect changing risk or
changing risk aversion
• However, a lot of return predictability would
indicate market inefficiency

1231 Topics in Finance 2 Pedro Santa-Clara

Predictive Regressions
 Predictive regressions

• When we regress returns on a lagged variable, we model

the conditional expectation

• We use forecasting variables that are related to the risk

premium like bond yields or spreads, valuation multiples,
and volatility
– These variables are typically very persistent (not i.i.d.)
• Typically, the R-square of predictive regressions is quite
low
– Why?

1231 Topics in Finance 4 Pedro Santa-Clara

 Dividend yield through time

1231 Topics in Finance 5 Pedro Santa-Clara

 Fama and French (1988)

1231 Topics in Finance 6 Pedro Santa-Clara

 Summary of empiral evidence

1231 Topics in Finance 7 Pedro Santa-Clara

 Predicting 7-year returns

1231 Topics in Finance Pedro Santa-Clara

8
 Fama and French (1988)

• Dividend yield forecasts stock market returns

• The regression coefficient and the R2 from the
regression increase with the horizon of the
return
• Is this evidence against market efficiency? Or
does this reflect a time-varying risk
premium?
– There has been a huge debate...

1231 Topics in Finance 9 Pedro Santa-Clara

Econometric Problems
 Stambaugh (1999) critique
• Regressor is persistent – near unit root
• Shocks to prices (returns) are negatively correlated
with shocks to dividend yield
– The OLS hypothesis that the error is uncorrelated with the
regressor fails
– The coefficient of the predictive regression is biased upward
and the t-statistic is also biased
• Stambaugh derives a correction to the coefficient
estimator and standard error
• Finds the predictive regressions insignificant at
conventional significance levels
• Is return predictibility dead?

1231 Topics in Finance 11 Pedro Santa-Clara

 Overlapping observations and long horizons

• Valkanov (2003), Boudoukh, Richardson, Whitelaw

(2005) show that using overlapping observations will
spuriously increase the R-square and t-statistics of
long-horizon predictive regressions
– They show that these statistics will grow proportionately
with the forecast horizon

1231 Topics in Finance 13 Pedro Santa-Clara

Out-of-Sample Predictability
 Goyal and Welch (2006)
• Goyal and Welch conduct a horse race between predictive variables
in the literature
• They focus on out-of-sample performance of each predictor
– Avoid in-sample overfitting
– Relevant for investment decisions
• Approach
– Estimate a predictive regression in a training sample
– Apply the estimated coefficients with current variables to predict
returns in the future
– Roll the sample and estimation forward
– Compare the performance of the predictor with the historic sample
mean
• Findings
– No robust predictor, performance erodes over time
– Even for the best variables and best times, market timing only adds
10bp of certainty equivalent return per year

1231 Topics in Finance 20 Pedro Santa-Clara

 Predictability by dividend-price ratio

1231 Topics in Finance 22 Pedro Santa-Clara

The Present Value Relation and
Return Predictability
 Campbell and Shiller (1988)
• Start from definition of returns

• Log-linearize around average P/D (lower case

variables are logs of upper case variables)

• For historic average D/P=4%, ρ=0.96 for yearly data

1231 Topics in Finance 26 Pedro Santa-Clara

 Campbell and Shiller (1988)
• Iterate the return equation forward and take expectations

• High dividend yield currently implies

– High future expected returns
and /or
– Low future expected dividend growth
• Put differently, dividend yield must forecast
– Future returns
and /or
– Future dividend growth
Or the dividend yield would be constant, which it is not...
1231 Topics in Finance 27 Pedro Santa-Clara
 Lacerda and Santa-Clara (2010)

• Assume simple processes for expected returns and dividend

growth

• So that

• Then we have

1231 Topics in Finance 28 Pedro Santa-Clara

 Lacerda and Santa-Clara (2010)

• Solving for expected returns

• So, instead of using the dividend-price ratio as a forecaster of

expected returns in predictive regressions, we should use

1231 Topics in Finance 29 Pedro Santa-Clara

 Cochrane (2006)
• Consider the following VAR

1231 Topics in Finance 30 Pedro Santa-Clara

 Cochrane (2006)

• Weak evidence in favor of small return forecastability

• Overwhelming evidence against forecastability of
dividend growth – “the dog that did not bark...”
• Remember, from the present value model, that the null
hypothesis that dividend yield does not forecast
returns must specify that it forecasts (negatively)
future dividend growth

1231 Topics in Finance 31 Pedro Santa-Clara

 Cochrane (2006)
• The present value relation has implications for the coefficients
and the errors of the system

• We cannot have br=0 without changing the dividend growth

forecast bd or the dividend yield autocorrelation φ
• In particular, we cannot have both br=0 and bd=0 . For a null of
br=0, we need to jointly assume that bd<0
• The proper null hypothesis is then (for demeaned variables)

• Which can be simulated to generate artificial samples to

compare the original estimated coefficients with the coefficients
estimated in these simulated samples

1231 Topics in Finance 32 Pedro Santa-Clara

 Sampling distribution of statistics

1231 Topics in Finance 33 Pedro Santa-Clara

 Economic significance
• Using the joint null, the hypothesis of return forecastability is
accepted
– Indeed, all the variation in the dividend yield corresponds to
changes in expected returns and none to news about future
dividend growth
• Expected returns vary through time with a standard deviation
of 5.5% which is of the same magnitude as the average equity
premium (7.7%)
• Long-horizon forecasts (implied from one-year regressions) are
more powerful to assess the significance of the predictor

• Despite this, out-of-sample forecasts of returns are unreliable --

be careful timing the market based on this variable
– Small data samples and persistent regressor makes OLS
estimates unreliable
1231 Topics in Finance 34 Pedro Santa-Clara
The Sum-of-the-Parts Approach

Ferreira and Santa-Clara (2011)

 A simple decomposition of returns
• Capital gains

• Dividend yield

• Total returns

• In logs

1231 Topics in Finance 36 Pedro Santa-Clara

 Decomposition of cumulative returns

7.0
gm dp
ge r
6.0

5.0
Yearly cumulative return

4.0

3.0

2.0

1.0

0.0

-1.0
1948 1952 1956 1960 1964 1968 1972 19 76 1980 1984 1988 1992 1996 2000 2004

1231 Topics in Finance 37 Pedro Santa-Clara

 History of return components

Mean Median Sdt Dev Min Max Skew Kurt AR(1)

gm 0.44 -1.44 26.33 -62.26 78.83 0.27 3.12 -0.17
ge 5.09 9.64 21.49 -70.56 56.90 -1.02 5.42 0.17
dp 3.90 3.60 1.64 1.13 9.62 0.75 3.99 0.79
rm 9.69 13.51 19.42 -60.97 43.60 -0.97 4.50 0.09

Correlations
gm ge dp rm
gm 1.00
ge -0.66 1.00
dp -0.21 -0.16 1.00
rm 0.60 0.19 -0.38 1.00

1231 Topics in Finance 38 Pedro Santa-Clara

 The sum-of-the-parts approach

• We forecast each component of returns

separately

• Expected dividend price estimated by the

current dividend-price ratio
– Assumes this ratio follows a random walk
• Expected earnings growth estimated with a
20-year past moving average
– Earnings growth nearly impossible to forecast
– Tried analyst consensus forecasts with worse
results
1231 Topics in Finance 39 Pedro Santa-Clara
 The sum-of-the-parts approach
• 3 alternatives to estimate expected multiple
growth
– No multiple growth

– Multiple growth regression (with shrinkage)

– Multiple reversion (with shrinkage)

1231 Topics in Finance 40 Pedro Santa-Clara

 Out-of-sample performance
Out-of-Sample R-square
Variable Predictor SOP SOP SOP
no multiple multiple growth multiple
multiple regression reversion
13.43
SVAR Stock variance 12.74 13.65
DFR Default return spread 14.40 12.98
LTY Long term bond yield 10.92 7.61
LTR Long term bond return 12.62 16.94
INFL Inflation 12.91 14.05
TMS Term spread 11.28 15.57
TBL T-bill rate 11.51 11.67
DFY Default yield spread 12.57 14.46
NTIS Net equity expansion 13.31 14.21
ROE Return on equity 13.66 9.02
DE Dividend payout 12.60 9.72
EP Earnings price 14.31
SEP Smooth earning price 11.07
DP Dividend price 8.99
DY Dividend yield 12.51
BM Book-to-market 10.20
Constant 14.40

1231 Topics in Finance 41 Pedro Santa-Clara

 Return forecast (no multiple growth)

1231 Topics in Finance 42 Pedro Santa-Clara

 Return forecast vs T-bill rate

1231 Topics in Finance 43 Pedro Santa-Clara

 Forecast vs realized returns

1231 Topics in Finance 44 Pedro Santa-Clara

 SOP vs predictive regressions vs mean

1231 Topics in Finance 45 Pedro Santa-Clara

 International evidence
Out-of-Sample R-square
Predictor Predictive Predictive SOP SOP SOP
regression regression no multiple multiple growth multiple
(shrinkage) multiple regression reversion
U.K.
10.73
LTY Long term bond yield -47.54 -5.61 4.16 11.27
TMS Term spread -14.71 -1.13 9.26 11.60
TBL T-bill rate -20.87 -3.07 6.39 11.51
DY Dividend yield -9.19 5.07 13.28 10.78
Constant 11.75
Japan
12.14
LTY Long term bond yield -11.01 -1.86 12.11 11.87
TMS Term spread -5.46 -0.89 5.75 5.82
TBL T-bill rate -7.57 -0.62 5.14 5.62
DY Dividend yield 3.12 6.63 10.25 11.99
Constant 11.91
U.S.
7.75
LTY Long term bond yield -20.73 -1.51 4.47 3.12
TMS Term spread -12.05 -0.99 8.24 5.50
TBL T-bill rate -21.18 -2.00 5.06 3.40
DY Dividend yield 0.96 2.68 6.64 5.73
Constant 5.92

1231 Topics in Finance 46 Pedro Santa-Clara

 Monte Carlo simulation
• Simulate expected returns and expected dividend
growth from Binsbergen and Koijen (2010)

• Use Campbell-Shiller present value approximation

• Simulate 10,000 samples of 80 years of returns,

dividend growth, and the dividend-price ratio
1231 Topics in Finance 47 Pedro Santa-Clara
 Monte Carlo simulation

1231 Topics in Finance 48 Pedro Santa-Clara