Return Predictability

Giacomo Morelli
Sapienza University
[email protected]

February 25, 2022

Giacomo Morelli Financial Risk Management Financial Risk Management 1 / 25

Return predictability

We know from Asset Pricing that if expected returns are constant, then the EMH
implies that actual returns (net of transaction costs) should be unpredictable.
Empirical studies of predictability:
1 Seasonal patterns in returns
2 Autocorrelation in short-horizon (daily, weekly, monthly) returns:
Q-statistics, ARMA models
3 Autocorrelation in long-horizon (multi-year) returns:
AR models, variance ratio tests
4 Predictability based on valuation ratios (dividend yield, price-earnings ratio)
and firm characteristics (size, book-to-market ratio)
5 Predictability based on macroeconomic variables (consumption, GDP, interest
rates)

Giacomo Morelli Financial Risk Management Financial Risk Management 2 / 25

Seasonal patterns in returns

Research in the 1980s documented seasonal effects in stock returns

(‘week-end’ and ‘January’ effects, etc).

But when investors discovered these effects, they often disappeared!

And if you incorporate transactions costs, there are no real gains left.

Giacomo Morelli Financial Risk Management Financial Risk Management 3 / 25

Autocorrelation in short-horizon returns

The simplest first check is to have a look at autocorrelation coefficients of

(one-period) returns. Define the autocovariance of order s as
γ(s) = Cov(Rt , Rt−s )
The autocorrelation of order s is defined as
Cov(Rt , Rt−s ) γ(s)
ρ(s) = p p =
Var(Rt ) Var(Rt−s ) γ(0)

The empirical counterparts are given by

T −s T
1 X 1 X
γ̂(s) = (Rt − R)(Rt−s − R), R= Rt
T t=1 T t=1
γ̂(s)
ρ̂(s) =
γ̂(0)

Giacomo Morelli Financial Risk Management Financial Risk Management 4 / 25

Autocorrelation in returns (one-period-ahead)

If Rt are normally distributed, the following basic result holds approximately:

√ d
T ρ̂(s) → N(0, 1)

To test whether ρ(1) = ρ(2) = . . . = ρ(s) = 0 we typically rely on the Q-statistics

(Box-Pierce).
s
d
X
Qm = T ρ̂(i)2 → χ2 (s)
i=1

We often use the modified Box-Pierce statistic (Ljung-Box) due to better finite
sample properties:
s
X ρ̂(i)2 d 2
Q̃m = T (T + 2) → χ (s) (1)
T −i
i=1

Giacomo Morelli Financial Risk Management Financial Risk Management 5 / 25

Example: Autocorrelation in returns

Size portfolios (10 deciles), U.S. stock market value-weighted returns:

1963/01-2004/12, monthly frequency, T =504, from Fama and French
(http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_
library.html)

Portfolio ρ̂(1) ρ̂(2) ρ̂(3) Q̃( 3) p-value

1 (Small) 0.244 -0.013 -0.049 31.497 0.000
2 0.173 -0.030 -0.072 18.310 0.000
3 0.150 -0.062 -0.074 16.177 0.001
4 0.155 -0.042 -0.064 15.207 0.002
5 0.132 -0.038 -0.053 10.920 0.012
6 0.135 -0.072 -0.039 12.579 0.006
7 0.111 -0.048 -0.030 7.862 0.049
8 0.070 -0.055 -0.045 5.048 0.168
9 0.066 -0.051 -0.036 4.228 0.238
10 (Big) -0.014 -0.027 0.030 0.922 0.820

Giacomo Morelli Financial Risk Management Financial Risk Management 6 / 25

Autocorrelation in returns (one-period-ahead)

Consider the regression:

Rt+1 = α + γ 0 It + ωt+1 (2)
where It contains lagged returns (Rt , Rt−1 , etc.) and possibly lagged forecast
errors (ωt , ωt−1 , etc.)
Example: Consider an ARMA model for daily returns with seasonal effects
Statistically significant predictability based on past forecast errors, and
statistically significant week-end effect
R 2 very low, and the effects are probably not economically significant!

Giacomo Morelli Financial Risk Management Financial Risk Management 7 / 25

Example: Autocorrelation in returns

Decile 1 portfolio (small stocks): After some specification search based on

information criteria: Model achieving the lowest AIC is the ARMA(1,1) with a
January dummy

Variable Coeffcient Std. Error t-stat. p-value

α 0.786 0.342 2.295 0.022
‘dummy’ 6.286 0.913 6.884 0.000
AR(1) -0.033 0.161 -0.205 0.838
MA(1) 0.307 0.154 1.999 0.046
AIC 6.405
adj. R 2 0.141

Giacomo Morelli Financial Risk Management Financial Risk Management 8 / 25

Example: Autocorrelation in returns

Figure: Seasonal means, small stocks. Decile 1 (small stocks) by season

How about decile 10 (big stocks)? After some specification search based on
information criteria things look quite different there
→ Model with lowest AIC just contains a constant, January dummy is insignificant
Giacomo Morelli Financial Risk Management Financial Risk Management 9 / 25
Autocorrelation in long-horizon returns

Two common approaches:

1 Multi-period time-overlapping AR model

2 Variance ratio test

Giacomo Morelli Financial Risk Management Financial Risk Management 10 / 25

Multi-period AR model

The model is given for k ≥ 1

Rt→t+k = αk + βk Rt−k→t + t+k (3)

Or, for logarithmic returns:

rt→t+k = αk + βk rt−k→t + t+k (4)

Giacomo Morelli Financial Risk Management Financial Risk Management 11 / 25

Remarks on multi-period AR model

The log return version preferable, because multi-period time-overlapping log

returns, rt→t+k , have better statistical properties than multi-period
time-overlapping ‘simple’ returns, Rt→t+k
t+k ∼ MA(k − 1)
⇒ OLS estimate of βk consistent, but SE(β̂k ) inconsistent
⇒ Use Newey-West (HAC-SE) robust standard errors
Empirical studies show:
β̂k ≈ 0 and R 2 ≈ 0 for very low and very large values of k
β̂k < 0 for intermediate values of k. Strongest predictability for k equal to
around 5 years (R 2 ≈ 0.35). β̂k < 0 implies mean-reversion in returns

Giacomo Morelli Financial Risk Management Financial Risk Management 12 / 25

Variance ratio (VR) test - Intuition

If log returns are iid distributed with zero mean and constant variance, σ 2 , we
have the following relationship

Var(rt→t+2 ) Var(rt+1 + rt+2 ) 2σ 2

VR2 = = = =1
2 Var(rt+1 ) 2 Var(rt+1 ) 2σ 2

If the log return follows a weakly stationary process we have

Var(rt→t+2 ) Var(rt+1 ) + Var(rt+2 ) + 2 Cov(rt+1 , rt+2 )

VR2 = =
2 Var(rt+1 ) 2 Var(rt+1 )
2
2σ + 2γ(1)
= = 1 + ρ(1)
2σ 2

Giacomo Morelli Financial Risk Management Financial Risk Management 13 / 25

VR Test

ρ(1) = 0, VR2 = 1, ρ(1) > 0, VR2 > 1 , ρ(1) < 0, VR2 < 1
| {z } | {z } | {z }
No Autocorrelation Pos. Autocorrelation Neg. Autocorrelation
General case:
iid log returns: Var(rt→t+k ) = k × Var(rt+1 )
Positive autocorrelation: Var(rt→t+k ) > k × Var(rt+1 )
Negative autocorrelation: Var(rt→t+k ) < k × Var(rt+1 )
1 Var(rt→t+k )
VRk ≡ (5)
k Var(rt+1 )
 
 VRk < 1 : mean reverting 
VRk = 1 : iid log returns
VRk > 1 : mean averting
 

Giacomo Morelli Financial Risk Management Financial Risk Management 14 / 25

Remarks to VR test

ˆ k is distributed as
Under the iid null hypothesis, VR
√
 
ˆ d 2(2k − 1)(k − 1)
Tk(VR k − 1) → N 0,
3k
r
ˆ k) = 2(2k − 1)(k − 1)
SE(VR
3Tk 2

Empirical studies show:

ˆ k ≈ 1 for small k
VR
ˆ k < 1 for large k → long-term mean reversion
VR

Giacomo Morelli Financial Risk Management Financial Risk Management 15 / 25

Example: VR test

Decile 1 portfolio (small stocks), 1963/01-2004/12, k = 12, 24, 36, 48

k Var(rt→t+k ) k Var(rt+1 ) ˆ k
VR Std. Error VR-stat.
1 0.0041 0.0041 1.00
12 0.0628 0.0492 1.28 0.048 5.75
24 0.1155 0.0984 1.17 0.050 3.48
36 0.1601 0.1476 1.08 0.050 1.68
48 0.2202 0.1968 1.12 0.051 2.35

Giacomo Morelli Financial Risk Management Financial Risk Management 16 / 25

Figure: Returns and Dividend Yield. Malkiel (2003): The efficient market hypothesis and
its critics, CEPS Working Paper No. 91

Giacomo Morelli Financial Risk Management Financial Risk Management 17 / 25

Multi-period return regression

Rt→t+k = a + bXt + t+k (6)

Xt is a univariate or multivariate series of standard ‘valuation ratios’, such as the
dividend to price ratio (Dt /Pt ) or the earnings to price ratio (Et /Pt )
Again, because multi-period log returns have better statistical properties, the
forecasting regression is typically formulated in logs:

rt→t+k = a + bxt + t+k , (7)

where xt is ln(Dt /Pt ) = dt − pt or ln(Et /Pt ) = et − pt

Giacomo Morelli Financial Risk Management Financial Risk Management 18 / 25

Example: Multi-period return regression

Figure: Multi-period excess return prediction. Cochrane (2008): Financial Markets and
the Real Economy, In: HANDBOOK OF THE EQUITY RISK PREMIUM, Elsevier BV.
(Table 1)

Giacomo Morelli Financial Risk Management Financial Risk Management 19 / 25

Remarks on multi-period return regressions

t+k ∼ MA(k − 1)
⇒ OLS estimate of b consistent, but SE(b) inconsistent
⇒ use Newey-West (HAC-SE) robust standard errors
Empirical studies show:
b̂ ≈ 0 and R 2 ≈ 0 for small k
b̂ 6= 0 and R 2 very high for large k
A high (low) price, relative to dividends or earnings, signals lower (higher)
future long-term stock returns
We return to multi-period forecasting regressions later in the course

Giacomo Morelli Financial Risk Management Financial Risk Management 20 / 25

Example: Multi-period return regression

Figure: Recursive estimates of b of log excess return prediction with log dividend-to-price
ratio. Goyal & Welch (2003): Predicting the Equity Premium With Dividend Ratios,
Management Science 49 (5), pp. 639-654 (Figure 2)

Giacomo Morelli Financial Risk Management Financial Risk Management 21 / 25

Predictability by macro variables

Predictable returns and the state of the business cycle

Several macro variables have been shown to be predictors of returns:

Variables from bond markets: interest rates, term spreads (e.g. Fama and
French, 1989)

Output related measures: e.g. Rangvid (2006, JFE), Cooper/Priestley (2008,

RFS)

Consumption related measures: e.g. Lettau/Ludvigson (2001, JF),

Rangvid/Vinther-Møller (2010)

etc

Giacomo Morelli Financial Risk Management Financial Risk Management 22 / 25

Example: Predictability returns by ‘cay’

Figure: Recursive estimates of b of log excess return prediction with log dividend-to-price
ratio. Goyal & Welch (2003): Predicting the Equity Premium With Dividend Ratios,
Management Science 49 (5), pp. 639-654 (Figure 2)

Giacomo Morelli Financial Risk Management Financial Risk Management 23 / 25

Example: Predictability returns by ‘cay’ (cont’d)

Figure: CAY. Lettau & Ludvigson (2001): Consumption, Aggregate Wealth, and
Expected Stock Returns, Journal of Finance 56 (3), pp. 815-849 (Table III)

Giacomo Morelli Financial Risk Management Financial Risk Management 24 / 25

Caveats

The regressions with multi-period returns are plagued with small-sample

problems when the degree of time-overlap is large relative to the sample size
(large k/T ), which makes the HAC-SEs are unreliable
The regressions are plagued with small-sample bias when the error terms
(ωt+1 and t+k ) are correlated with innovations to the predictor variables (It ,
Xt , and xt ), which will typically be the case when the predictor variable is a
valuation ratio
The VR statistic is biased downwards in small samples
The statistical size of the autocorrelation test has been shown to be poor if
the number of lags relative to the sample size is large
We return to some of these problems later in the course.

Giacomo Morelli Financial Risk Management Financial Risk Management 25 / 25