Linear Return Prediction Models

Oxford, July-August 2013

Allan Timmermann1

1
UC San Diego, CEPR, CREATES

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 1 / 52

1 Linear Prediction Model - Goyal and Welch, RFS 2008

2 Approximate log-linearized PV model - Campbell and Shiller, RFS 1988

3 Predictive Regressions - van Binsbergen and Koijen, JF 2010

4 Sum of the parts - Fereira and Santa-Clara, JFE 2011

5 Predictive regressions - Stambaugh, JFE 1999

6 Predictive Systems - Pastor and Stambaugh, JF 2009

7 Cenesizoglu and Timmermann (JBF, 2012)

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 2 / 52
Predictability of Stock returns: Empirical Findings

Macro variables appear to forecast mean stock returns: investment/capital

ratio, consumption wealth ratio (CAY)
Predictability linked to few episodes (T-bill rate) and can appear elusive
Predictability sometimes pick up a time-varying business cycle component
(default spread, term spread)
Higher risk premium during recessions (price of risk is countercyclical)
Return predictability is low when it could be exploited (and the risk be
controlled) at high frequency
return prediction models have low signal-to-noise ratios
Time-series predictability appears to have declined during the 1990s
Standard empirical models predict episodes with negative excess returns on
stocks

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 2 / 52

Fundamental questions

Why are returns predictable?

Market ine¢ ciency
Time-varying investment opportunities

1 = Et [mt +1 rt +1 ] )
1 covt (rt +1 , mt +1 )
Et [rt +1 ] =
Et [mt +1 ]

Data-snooping
Learning: parameter estimation unceratinty, model instability and model
uncertainty

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 3 / 52

Goyal-Welch (2008) exercise

Estimate univariate prediction models for the equity premium (excess return),
rt +1 :
rt +1 = β0 + β1 xt + εt +1

Predictors (xt ):
Valuation ratios: dividend price ratio, dividend yield, earnings-price ratio,
10-year earnings-price ratio, book-to-market ratio
Bond yields: three-month T-bill rate, yield on long term government bonds,
term spread, default yield spread, default return spread
Equity risk estimates: long term return, stock variance
Corporate …nance variables: dividend payout ratio, net equity expansion,
percent equity issuing
In‡ation

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 4 / 52

Goyal-Welch (2008): empirical results

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 5 / 52

Goyal-Welch (2008): Cumulated SSE di¤erences

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 6 / 52

Approximate log-linearized PV model (Campbell-Shiller)

From the de…nition of returns

Pt +1 + Dt +1
Rt + 1 =
Pt

Denote logs in lowercase letters,

rt +1 = log(Pt +1 + Dt +1 ) log(Pt )
D
= log(Pt +1 (1 + t +1 )) log(Pt )
Pt + 1
= log(Pt +1 (1 + exp(log(Dt +1 ) log(Pt +1 )) log(Pt )
= pt +1 pt + log(1 + exp(dt +1 pt +1 ))

The last term is a nonlinear function of the log dividend-price ratio

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 7 / 52

Log-linearized PV model: returns

Using a …rst-order Taylor expansion around d p

f ( dt + 1 pt + 1 ) f (d p ) + f 0 (d p )(dt +1 pt + 1 (d p ))

we have

rt +1 pt + 1 pt + log(1 + exp(d p ))
exp(d p )
+ ( dt + 1 pt + 1 (d p ))
1 + exp(d p )
k + ρpt +1 + (1 ρ)dt +1 pt

where
1
ρ= , k= log(ρ) (1 ρ) log(1/ρ 1)
1 + exp(d p)

ρ is a constant slightly smaller than one

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 8 / 52

Log-linearized PV model: prices

Approximate return equation can be rearranged to give a recursive equation

for log-prices:
pt = k + ρpt +1 + (1 ρ)dt +1 rt +1

Iterating forwards under the assumption that limj !∞ ρj pt +j = 0,

∞
k
pt =
1 ρ
+ ∑ ρj [(1 ρ ) dt + 1 + j rt + 1 + j ]
j =0

Ex-post dynamic accounting identity: Holds under the log-linearized

approximation and the transversality condition

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 9 / 52

Prices, expected dividends and expected returns

Taking expectations conditional on current information, gives the ex ante

relationship
" #
∞
k
pt = + Et ∑ ρ [(1 ρ)dt +1 +j rt +1 +j ]
j
1 ρ j =0
∞ ∞
k
=
1 ρ
+ (1 ρ ) Et ∑ ρ j dt + 1 + j Et ∑ ρj rt +1+j ]
j =0 j =0
| {z } | {z }
p dt p rt

If stock prices are high today, investors must expect either

high future dividends or
low future returns

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 10 / 52

Log-linearized PV model

Use price equation to derive expression for the stationary log dividend-price
ratio, dt pt :
" #
∞
k
dt pt = + Et ∑ ρ [ ∆dt +1 +j + rt +1 +j ]
j
1 ρ j =0

Dividend-price ratio is high when future dividend growth is expected to be

low and future returns are expected to be high

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 11 / 52

Return innovations

∞ ∞
rt +1 Et [rt +1 ] = Et + 1 ∑ ρj ∆dt +1+j Et ∑ ρj ∆dt +1+j
j =0 j =0
| {z }
η dt +1
!
∞ ∞
Et + 1 ∑ ρj rt +1+j Et ∑ ρ j rt + 1 + j
j =1 j =1
| {z }
η rt +1

Increases to expectations of future dividend growth result in a capital gain

while increasing expected future returns leads to a capital loss
η dt +1 : changes in expected discounted future dividend growth
η rt +1 : changes in expected discounted future returns

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 12 / 52

Empirical …ndings

VAR restrictions are typically strongly rejected

Large unexplained component is present in the log dividend-price ratio
Interest rates and consumption are not good at explaining time-varying
expected returns
Time-varying expected returns explains between one half and two thirds of
the variability in the dividend-price ratio

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 13 / 52

Predictive Regressions - van Binsbergen and Koijen (2010)

Returns, price-dividend ratio, and dividend growth are de…ned as

Pt +1 + Dt +1
rt +1 log
Pt
Pt
PDt
Dt
Dt + 1
∆dt +1 log
Dt

Expected returns, µt = Et [rt +1 ], and expected dividend growth,

gt = Et [∆dt +1 ], follow AR(1) processes:
µ
µ t +1 = δ0 + δ1 ( µt µ 0 ) + ε t +1
gt +1 = γ0 + γ1 (gt g0 ) + εgt+1

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 14 / 52

Predictive regressions: dividend growth and returns

Realized dividend growth:

∆dt +1 = gt + εdt+1

Campbell-Shiller log-linearized returns:

rt +1 κ+ρ pdt +1 + ∆dt +1 pdt

pdt log(PDt )
pd = E [pdt ]
κ = log(1 + exp(pd )) ρ pd
ρ = exp(pd )/(1 + exp(pd ))

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 15 / 52

Predictive regressions: log price-dividend ratio

Iterating on the model, the log price-dividend ratio is linear in expected

returns and expected dividend growth:

pdt = A B1 (µt δ0 ) + B2 (gt γ0 )

κ γ δ0
A = + 0
1 ρ 1 ρ
1 1
B1 = , B2 =
1 ρδ1 1 ργ1

Distribution of three shocks: i.i.d. with covariance matrix

0 g 1 2 2 3
ε t +1 σg σg µ σgd
Σ var @ εut+1 A = 4 σg µ σ2µ σ µd 5
εdt+1 σgd σµd σ2d

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 16 / 52

Predictive regressions

Model simpli…es to a state space system with one transition (state) equation
and two measurement equations:

ĝt +1 = γ1 ĝt + εgt+1

∆dt +1 = γ0 + ĝt + εdt+1
B1 εt +1 + B2 εgt+1
µ
pdt +1 = (1 δ1 )A1 + B2 (γ1 δ1 )ĝt + δ1 pdt
ĝt = gt γ0 : de-meaned dividend growth

Linear model can be estimated by means of a Kalman …lter

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 17 / 52

State space model with market-reinvested dividends

Reduced form model for market-reinvested dividends earning a return of

exp(εM
t +1 )

DtM+1 = Dt +1 exp(εM
t +1 )
ρM = corr (εM r
t +1 , ε t +1 )

Linear system with two measurement equations and one state equation - can
be estimated by a …lter

∆dtM+1 = γ0 + ĝt + εdt+1 + εM

t +1 εM
t
pdtM+1 = (1 δ1 )A + B2 (γ1 δ1 )ĝt + δ1 pdtM
B1 εt +1 + B2 εgt+1 εM M
µ
t +1 + δ 1 ε t
ĝt +1 = γ1 ĝt + εgt+1

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 18 / 52

van Binsbergen and Koijen (2010): cash/mkt reinvested
dividends

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 19 / 52

van Binsbergen and Koijen (2010): MLE results

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 20 / 52

van Binsbergen and Koijen (2010): Contrast w. OLS

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 21 / 52

van Binsbergen and Koijen: main points

Treat conditional expected returns and conditional expected dividend growth

as latent variables
Use …ltering to obtain estimates of expected returns and expected dividend
growth
Filtered series have strong predictive power over future returns and dividend
growth
Dividend growth and returns contain persistent, predictable components
Expected dividend growth is less persistent than expected returns

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 22 / 52

Fereira and Santa-Clara (2011): Sum of parts
Decompose returns into capital gains plus dividend yield
1 + Rt +1 = 1 + CGt +1 + DYt +1
Pt +1 D
= + t +1
Pt Pt

Capital gains can be written as

Pt + 1 P /E E M E
1 + CGt +1 = = t +1 t +1 t +1 = t +1 t +1
Pt Pt /Et Et Mt Et
= (1 + GMt +1 )(1 + GEt +1 )

Et +1 : earnings per share at t + 1

Mt +1 : price-earnings multiple at t + 1
GEt +1 : earnings growth rate at t + 1
Decompose the dividend yield into
Dt + 1 D P
DYt +1 = = t +1 t +1 = DPt +1 (1 + GMt +1 )(1 + GEt +1 )
Pt Pt +1 Pt

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 23 / 52

Returns as sum of parts

Using these expressions, we can write

1 + Rt +1 = (1 + GMt +1 )(1 + GEt +1 )(1 + DPt +1 )

In logs
rt +1 = log(1 + Rt +1 ) = mt +1 + get +1 + dpt +1

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 24 / 52

Returns forecasts from sum of parts

Fereira and Santa-Clara forecast the return components separately: Sum of

Parts (SoP) method:
µ̂t = µ̂gm ge dp
t + µ̂t + µ̂t

µ̂ge 20
t = ḡt : estimated from 20-year MA of growth in earnings per share
µ̂dp
t = dpt : Dividend price ratio expected to follow a random walk
µ̂gm
t = 0 : no growth in multiples (simplest model)

µ̂t = µ̂ge dp 20
t + µ̂t = ḡt + dpt

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 25 / 52

Alternative approaches

Linear regression for price multiple:

c t +1 = α̂ + β̂xt
gm

Multiple reversion: deduce ‘abnormal’part in multiple, ut , from regression

mt = a + bxt + ut

Expect that mt reverts to its historical mean given current value of xt :

gmt +1 = c + d ( ût ) + vt
µ̂gm
t = ĉ + d̂ ( ût )

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 26 / 52

Ferreira and Santa-Clara (2011): SoP summary stats

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 27 / 52

Ferreira and Santa-Clara (2011): SoP forecasts

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 28 / 52

Ferreira and Santa-Clara (2011): Forecasting performance

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 29 / 52

Ferreira and Santa-Clara (2011): main points

SoP method uses a disaggregated approach to forecasting the dividend yield,

earnings growth, and earnings multiples
Empirically, the approach appears to generate better out-of-sample forecasts
than the prevailing mean of Goyal and Welch (2008) and univariate
prediction models
"In the investment world, we show that there are important gains from
timing the market. To the extent that what we are capturing is excessive
predictability rather than a time-varying risk premium, the success of our
analysis eventually destroys its usefulness. Once enough investors follow our
approach to predict returns, they will impact market prices and again make
returns unpredictable." (page 535)

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 30 / 52

Predictive regressions (Stambaugh, 1999)

Return regressions

yt = α + βxt 1 + ut
E (ut jxs , xw ) 6= 0, s < t w

Residuals are correlated with past or future values of the regressor

xt is assumed to be persistent:

xt = θ + ρxt 1 + vt , j ρ j < 1
ut σ2u σuv
Σ = cov , ut vt =
vt σuv σ2v

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 31 / 52

Distributional results, Stambaugh (1999)

Distribution and moments of OLS estimator β̂ can be derived from

Proposition 1 in Stambaugh
Assuming Gaussian innovations, Marriott and Pope (1954) and Kendall
(1954) show that
(1 + 3ρ)
E [ρ̂ ρ]
T

Bias in ρ̂ transmits to a bias in β̂. Stambaugh shows that

σuv
E [ β̂ β] = E [ρ̂ ρ]
σ2v
σuv (1 + 3ρ)
= + O (T 2 )
σ2v T

Bias can be substantial: Stambaugh estimates a bias around 0.40 in a

regression of returns on the dividend yield from 1977-1996 (T = 240)

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 32 / 52

Stambaugh (1999): Table 1 - …nite sample properties of
beta estimate

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 33 / 52

Stambaugh (1999): main points

"When the innovation in a lagged stochastic regressor is correlated with the

regression disturbance, the OLS estimator can exhibit …nite-sample properties
that deviate sharply from those in the standard regression setting." (p. 408)
highly relevant to return regressions that use the dividend yield
Finite-sample p-values substantially higher than when computed in the
traditional way
Bayesian inference can yield sharper results, although they depend on the
prior, assumptions about the initial observation (…xed or stochastic), and
stationarity assumptions for the predictor: "In the 1952-1996 period, for
example, the p-value equals 0.15, so a classical test would accept the
zero-slope hypothesis at conventional signi…cance levels. In contrast, the
posterior probability that the slope is less than or equal to zero ranges
between 0.01 and 0.05, depending on the speci…cation of the likelihood and
prior." (Stambaugh, p. 408)

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 34 / 52

Predictive system, Pastor and Stambaugh, JF 2009

Linear regression
rt +1 = a + b 0 xt + et +1

rt +1 : returns
xt : predictors
et +1 : innovation (MDS)
This model is too simple if the predictors are imperfectly correlated with
expected returns, µt

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 35 / 52

Predictive system

State space system

rt +1 = µ t + ut + 1
xt + 1= (I A)Ex + Axt + vt +1
µt +1 = (1 β)Er + βµt + wt +1
2 3 0 1
ut 0 σ2u σuv σuw
4 vt 5 N @ 0 , σvu Σvv σvw A
wt 0 σwu σwv σ2w

µt : expected return
ut +1 : unexpected returns

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 36 / 52

Conditional expected returns

∞
E [rt +1 jDt ] = Er + ∑ ω s εU 0
t s + δ s vt s
s =0
εU
t = rt Er
ωs = m ( β m )s
δs = n ( β m )s
1
m = ( βQ + Cov (u, w jv )) (Q + Var (u jv ))
n = (σwv mσuv )Σvv1

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 37 / 52

Return equation
Plugging µt = Er + ∑i∞=0 βi wt i into the return equation,
∞
rt +h = Er + ∑ β i wt + h 1 i + ut + h
i =0
h 1
= (1 β h 1 ) Er + β h 1 µ t + ∑ βh 1 i wt +i + ut +h
i =1

Hence
βσ2w
Cov (rt , rt h ) = βh 1 ( + σuw )
1 β2

Persistence in µt induces positive serial correlation in returns. σuw can be

negative so serial correlation in returns can be of either sign
Alternatively, the return process can be written as an ARMA(1,1):

rt +1 = (1 β)Er + βrt + εt +1 γεt

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 38 / 52

Predictive regressions: perfect predictors
Suppose we regress returns directly on the predictors xt :

rt +1 = a + b 0 xt + et +1

Perfect predictors require that there is a b such that

wt = b 0 vt
0
Ab = βb

In univariate models, we must have ρvw = 1, and A = β

When this holds, m ! 0, n ! b, and so
∞
E [rt +1 jDt ] = Er + b 0 ∑ As vt s = Er + b 0 (xt Ex ) = a + b 0 xt
s =0

In this case, forecasts from the predictive system and the linear return
regression are identical
Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 39 / 52
Predictive regressions: imperfect predictors

With imperfect predictors

full history of the predictors matters through a weighted sum of past
innovations
full history of returns also a¤ect expected returns
Diagnostic for imperfect predictors:

cov (et , et +1 ) = βVar (µt jxt ) + cov (ut , wt b 0 vt )

Var (µt jxt ) = 0 and ω t = b 0 vt only if the predictors are perfect

Otherwise, these terms will be nonzero and so lead to autocorrelation in the
residuals from the return equation

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 40 / 52

Negative correlation between expected, unexpected returns
Strong prior of negative correlation between shocks to expected and
unexpected returns: Asset prices fall when discount rates rise:

ρuw < 0

Weight on past return, κ s , in forecast is a function of ρuw

t 1
E [rt +1 jDt ] = ∑ κs rt s
s =0

Low recent returns suggests that (i) lower expected returns since the sample
mean has gone down; (ii) higher expected returns since rising return
expectations lead to lower realized returns
If ρu ω is su¢ ciently negative, the "risk premium" e¤ect dominates the
"sample mean" e¤ect, so recent returns get a negative weight in estimates of
currrently expected returns, while older return data get a positive weight
Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 41 / 52
Pastor and Stambaugh (2009): e¤ect of past returns

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 42 / 52

Pastor and Stambaugh (2009): E¤ect of correlation priors

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 43 / 52

Pastor and Stambaugh (2009): predictive regressions

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 44 / 52

Pastor and Stambaugh (2009): return forecasts

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 45 / 52

Pastor and Stambaugh (2009): main points

Conditional return forecasts will depend on lagged returns and lagged

predictors when the predictor variables are imperfect
If a large part of the variation in unexpected returns comes from time-varying
expected returns, recent returns will get negative weights
Tests for serial correlation in predictive return regression indicates the presence
of imperfect predictors
Diagnostic for imperfect predictor: sign of correlation between residuals of
expected, unexpected return equations
Predictive system approach can incorporate priors about correlations between
shocks to expected and unexpected returns

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 46 / 52

Cenesizoglu and Timmermann (JBF, 2012)

Estimate a range of time-varying mean and volatility models for monthly

returns of the form

rt +1 = β 0 + β 1 xt + ε t + 1
ε t +1 N (0, σ2t +1 )
εt εt
log(σ2t +1 ) = δ0 + δ1 xt + δ2 log(σ2t ) + δ3 j j + δ4
σt σt

Restricted versions of the model impose constant mean (β1 = 0) and/or

constant conditional volatility (δ1 = δ2 = δ3 = δ4 = 0)
Forecasts are used to select the optimal portfolio weights under
mean-variance preferences and under power utility

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 47 / 52

Cenesizoglu and Timmermann: Empirical …ndings

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 48 / 52

Cenesizoglu and Timmermann: Empirical …ndings

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 49 / 52

Cenesizoglu and Timmermann: Empirical …ndings

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 50 / 52

Cenesizoglu and Timmermann: Empirical …ndings

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 51 / 52

Cenesizoglu and Timmermann (2012): main points

Very common for return prediction models to produce higher out-of-sample

mean squared forecast errors than a model assuming a constant equity
premium, yet simultaneously add economic value when their forecasts are
used to guide portfolio decisions
Although there is generally a positive correlation between a return prediction
model’s out-of-sample statistical performance and its ability to add economic
value, the relation tends to be weak and only explains a small part of the
cross-sectional variation in di¤erent models’economic value
Underperformance along conventional measures of forecasting performance
such as root mean squared forecast errors contain little information on
whether return prediction models that allow for a time-varying mean or
variance help or hurt investors

Timmermann (UCSD) Linear prediction models July 29 - August 2, 2013 52 / 52