A.

Fischer
SS 2008

Lecture 2: Predictability of Asset Returns

Random Walk Hypotheses, CLM 2.1

One of the biggest questions in finance is: are financial asset prices fore-
castable? Considerable resources and human effort (i.e., mathematicians,
economists, etc) have been invested in trying to “beat the market”. Others
in a less serious manner make a joke out of it. For example, the dart throwing
monkey versus the financial market pros contest reported in the Wall Street
Journal.
Random Walk and Martingale models will provide a benchmark against
other theoretical models. In the end it is not clear, whether the fundamentals
model outperforms the simple random walk model.
• Random Walk Hypothesis
Orthogonality condition of the RW hypothesis
cov[f (rt ), g(rt+k )] = E[(f (rt ) − µt )(g(rt+k ) − µt+k )] (1)
= 0
where f (·) and g(·) are two arbitrary functions. All random walk mod-
els and the martingale hypothesis have this property. If f (·) and g(·)
are linear, then (1) is equivalent to a RW 3. Else if f (·) is unrestricted
and g(·) is linear, then (1) is equivalent to a martingale model. Last,
if both are unrestricted, then we have a RW 1 and a RW 2.
• Martingale Model - “fair game”
The martingale model is defined by the following conditions:
E(Pt+1 |Pt , Pt−1 , · · ·) = Pt
E(Pt+1 − Pt |Pt , Pt−1 , · · ·) = 0
Tomorrow’s price is today’s price: the best forecast based on past prices
is today’s price. Two points: first, this definition says only something
about the mean (i.e., first moment) and nothing about the higher mo-
ments. Second, the orthogonality condition applies to non overlapping
price changes: this implies the ineffectiveness of all linear forecasting
rules for future price changes based on historical prices.
Samuelson defines the martingale model to be equal to weak-form ef-
ficiency. The necessary assumption is market efficiency, namely that

1
 today’s price reflects all the information from past prices. Under this
old definition it says that it is not possible to make profits based on in-
formation from past asset prices. Newer definitions recognize that there
is a tradeoff between risk and expected returns. Hence, the information
from risk is captured in the distribution’s higher moments.
• ME Example #2.1: Hall’s Martingale with Consumption
Let zt be a vector containing a set of macroeconomic variables (such
as money supply, or GDP) including aggregate consumption ct for pe-
riod t. Hall’s (1979) martingale hypothesis is that consumption is a
martingale with respect to zt :
E(ct |zt−1 , zt−2 , · · · , z1 ) = ct−1
This formalizes the notion in consumption theory called “consumption
smoothing”: the consumer, wishing to avoid fluctuations in the stan-
dard of living, adjusts consumption in t − 1 to the level such that no
change in subsequent consumption is anticipated. See Hayashi page
101.
• ME Example #2.2: Naive Inflation Forecasts
Lets us define a simple naive forecast using the martingale model for
inflation. The naive forecast says that the best inflation forecast for
t + h, defined as πt+h|t , is the current observed inflation rate, πt :

E(πt+h |It ) = πt .
The naive forecast is a martingale if the information set It = {pt , pt−1 , · · ·}.
The naive forecast is frequently used as the simplest forecasting bench-
mark against ARIMA models or the Phillips curve model. Most studies
show that they have difficulty in beating the naive inflation forecast,
see Akteson and Ohanian (2000).

Random Walk 1: IID Increments

• RW1 Specification

pt = µ + pt−1 + t , t ∼ IID(0, σ 2 ). (2)

the drift term, µ, is interpreted as the expected change in prices and the
error term, , is independently and identically distributed (i.e., Normal,
Weibull, Exponential). Independence ⇒ cov(t , t−k ) = 0, the reverse

2
 however is not true: Independence ⇐ cov(t , t−k ) = 0. If X1 and
X2 are independent random variables for every Borel function (i.e.,
function g(·) guarantee’s that r.v. xt is also a r.v. for yt = g(xt ))
h1 (·) and h2 (·) then E(h1 (X1 )h2 (X2 )) = E(h1 (X1 )) · E(h2 (X2 )). This
is both a necessary as well as a sufficient condition for independence.
One particular case of interest is when h1 (X1 ) = X1 and h2 (X2 ) = X2 ,
then under linear independence we have E(X1 X2 ) = E(X1 ) · E(X2 ).
From this we have cov(X1 , X2 ) = E(X1 X2 ) - E(X1 ) · E(X2 ). Here,
linear independence is equivalent to uncorrelatedness, since it implies
cov(X1 , X2 ) = 0
• RW1 model is stronger than the Martingale Model

Random Walk 2: Independent but not identically distributed

• RW2 Specification: Independent Increments
Over time the assumption of identically distributed increments is im-
plausible and therefore RW 2 relaxes this assumption. Clearly, RW 2
contains RW 1 as a special case.
pt = µ + pt−1 + t , t ∼ IN ID(0, σi2 ).
The unconditional variance is not constant: i.e., a Markov/threshold
model, which takes on two variances, i = 2, and σ12 6= σ22 .

Random Walk 3: Not Ident. and not indep. distributed

• RW3 Specification: Uncorrelated Increments
pt = µ + pt−1 + t , t ∼ N IN ID(0, σt2 ),

where cov(t , t−k ) = 0 and cov(2t , 2t−k ) 6= 0 for k 6= 0. RW 3 contains

RW 2 and RW 1 as special cases.
• An example of a RW 3 process is an ARCH(1) model
rt = pt − pt−1 = t , µ = 0,
2 2
t = α0 + α1 t−1 + νt .
We say this process has uncorrelated increments, but is clearly not
independent since its squared increments are correlated.

3
Random Walk, Time Dependency and Stationarity

pt = µ + pt−1 + t , t ∼ IID(0, σ 2 ).
Can be rewritten as E(pt |p0 ) = p0 + µt and V ar(pt |p0 ) = σ 2 t. To show
this, consider the following steps:

pt = µ + pt−1 + t
pt−1 = µ + pt−2 + t−1
..
.
p1 = µ + p0 + 1
Through repeated substitution
pt = µt + p0 + t + t−1 + · · · + 0 ,
with E(t ) = 0, one obtains
E(pt |p0 ) = µt + p0 .
Next, let us consider var(pt |p0 ):
var(pt |p0 ) = E(pt |p0 − E(pt |p0 ))E(pt |p0 − E(pt |p1 ))
= E(t + · · · + 1 )E(t + · · · + 1 )
with E(t )E(t−k ) = cov(t , t−k ) = 0 for k 6= 0 and E(t , t−k ) = σ 2 . This
gives
t
σ2,
X
var(pt |p0 ) =
i=1
2
= σ t.
The simple Random walk model has a time dependent mean and variance.
The time dependency of the moments makes it a non stationary process. To
obtain stationarity, or in other words non time dependence, we can first-
difference the RW process:
(1 − L)pt = ∆pt = t , with µ = 0,
This yields E(∆pt |∆p0 ) = 0 and var(∆pt |∆p0 ) = σ 2 . Note, there are other
forms of non stationarity that have nothing to do with Random Walk models;
i.e.
pt = µ + αt + t , t ∼ IID(0, σ 2 ).

4
In this simple model, we have a process that is non-stationary in mean,
but stationary in the variance. Note: Random walk models are are non
stationary, but not all forms of non stationary are random walk models.

Tests of Random Walk 1 with t ∼ IID, CLM 2.1

• Sequences and Reversals

Consider equation (2) without drift, i.e., µ = 0 and the indicator vari-
able, It , defined as follows:


 1 if rt = pt − pt−1 > 0


It = 
 0 if rt = pt − pt−1 ≤ 0
 

As in the Bernoulli “coin-toss” game, It , indicates whether the com-

pounding continuously return, rt is positive or negative. The Cowles
Jones Test is a comparison of the frequency of sequences and reversals
in asset returns. Sequences are defined as pairs of consecutive returns
with the same sign (i.e., It = 1,1,1 or It = 0,0,0), whereas reversals are
pairs of returns with the opposite returns (i.e. It = 1,0,1,0). Given a
sample of n+1 returns r1 , r2 , · · ·, rn+1 , the number of sequences Ns
and reversals Nr may be expressed as simple functions of the It0 s

n
X
Ns ≡ Yt , Yt ≡ It It+1 + (1 − It )(1 − It+1 )
t=1
Nr ≡ n − Ns

Note: we are generating n transformations with It for a sample of

n+1 returns.

5
 Example 1: Defining the Number Sequences Ns

Return rt Indicator It Sequence Sum Yt

r1 0.1 I1 1 Y1 0

r2 -0.5 I2 0 Y2 1

r3 -0.7 I3 0 Y3 0

r4 0.3 I4 1 Y4 0

r5 ?

• Cowles Jones Statistic without drift

If log prices follow a driftless IID RW1 process and if the distribution
of t is symmetric, then whether rt is positive or negative should be
equally likely. This is similar to a fair coin-toss with probability one-
half of either outcome. This implies that for any pair of consecutive
returns, a sequence and reversal are equally probable; hence the Cowles-
Jones ratio CJ ˆ ≡ Ns /Nr should be equal to one. More formally, this
ratio may be interpreted as a consistent estimator of the ratio CJ of
the probability πs of a sequence to the probability of a reversal 1-πs
since:

ˆ ≡ Ns = Ns /N = πˆs →
CJ
pr πs
= CJ =
1/2
=1 (3)
Nr Nr /N 1 − πˆs 1 − πs 1/2

Some Statistical Concepts

Before we can figure out the case of CJ with drift we need some statistics.
• More Distribution Theory
Bernoulli Distribution:
E(It ) = p = π,
var(It ) = pq = π(1 − π)

6
 Binomial Distribution:
E(Ns ) = np = nπ
var(Ns ) = npq = nπ(1 − π)

where Ns = N Y and Yt = It It+1 + (1 − It )(1 − It+1 ), thus we have

P
PNi=1 t
Binomial = i=1 Bernoulli.
• Variance Transformations with T
This transformation is used frequently in econometrics. It says

(θ̂ − θ0 ) ∼ N (0, V )
1/T (θ̂ − θ0 ) ∼ N (0/T, V /T 2 ) ∼ N (0, V /T 2 )

• Cumulative Normal Distribution

Let X be a random variable. The point function F (·), a function from
a point to a point, R → [0, 1] defined by F (·) = P r(X ≤ x) for all x
∈ R is called the distribution function of X and satisfies the following
properties:
(i) F (·) is non decreasing
(ii) F (−∞) = 0 and F (∞) = 1
(iii) F (·) is continuous from the right. Next, let F (·) be the distribution
function of the random variable X. The non negative function f (x)
defined by
Z x
F (x) = f (u)du, ∀ ∈ R continuous
−∞

or
X
F (x) = f (u), ∀ ∈ R discrete
u≤x

is said to be the probability density function of X.

Example: coin tossing f (0) = 1/4, f (1) = 1/2, and f (2) = 1/4. Draw
for F (·) and f (·).
Example: consider the case where X takes the values in the interval
[a, b] and all values of X are attributed the same probability; we express
this by saying that X is uniformly distributed in the interval [a, b] and
we write X ∼ U (a, b).
The distribution of X takes the form

7
 

0 x<a








F (x) = x−a
a≤x≤b

 b−a




 1 otherwise



The corresponding density function of X is given by


1

a≤x≤b


b−a

f (x) = 

 0

otherwise

Draw for F (·) and f (·).

Next, define the normal distribution and density function as
Z x 1 Z x
−t2 /2
Φ(x) = e dt = φ(t)dt
−∞ (2π)0.5 −∞

2
where x = µ/σ and the normal density function is φ(x) = 1
(2π)0.5
e−t /2 .