Handout 2

Statistical Independence

Discrete Random Variables

 Statistical Independence

A and B are statistically independent if

P(A/B)=P(A)
or
P(B/A)=P(B)

Occurrence of B does not affect probability of A and vice

versa

If it is known that two events are independent, then we

get useful simplifications in computing probabilities of
joint events

How do we know that two events are independent?

In certain situations, can assume that events are

independent
 The Random Walk Model for Asset Prices

A simple version of the random walk model is as follows: Assume that the current price
of a stock is $500 and that the daily price evolves in the future according to the
following mechanism based on the result of a coin toss

Tomorrow’s Price = Today’s Price + $1 if a coin tossed comes up Heads

= Today’s Price - $1 if a coin tossed comes up Tails

Thus, the price changes either by +$1 or by -$1, depending on whether the coin comes
up Heads or Tails. For example, if the first 7 tosses were HHHTHTT, the price of the stock
for the first 7 days would be

Day Movement Price

1 1 501(=500+1)
2 1 502(=500+1+1) and so on
3 1 503
4 -1 502
5 1 503
6 -1 502
7 -1 501
. . .
. . .

At least a couple of natural objections arise here immediately; viz.

a) that the model above could let prices become negative, when that is not possible
b) that changes in stock prices aren’t limited to just +/- 1$.

Both of these objections can be resolved easily, the first by working with log(Price) and
the second by modelling the change in (log) prices as draws from a continuous
distribution, such as the normal. One can also add on other modifications, such as an
offset in the model above, which results in the “random walk with drift” model.

The bottom line, though, is that all these fancier models obtained after adding these bells
and whistles have the same fundamental properties as the simple toy model shown above
and it suffices, therefore, to understand how that model works and its implications.

Note that though I stated the model for daily price changes, the same idea works for price
changes over a week, a month, a quarter, etc.
Implications of the random walk model: By construction, the change in daily price
(Tomorrow’s price – Today’s price) is dictated by the toss of a coin and thus implies
statistical independence of price changes over consecutive days/weeks/months. This, in
turn, implies that Nothing useful about future price changes can be said based upon past
price changes and so the past history of the prices is irrelevant and useless in making
predictions about Anything regarding future price changes (and, hence, future prices). In
terms of a probability statement, this would mean, for example that

P(Tomorrow’s price change is positive / past price changes) = P(Tomorrow’s price change
is positive).

One could make similar probability statements about any kind of movement in
tomorrow’s price change.

Some evidence in favour of this model:

a) Patterns in prices simulated using the model show the kinds of patterns, including
seeming momentum, that one sees in actual prices.

Thus, a series of statistically independent price changes can produce a price series that
seemingly has patterns as one sees in actual prices, so our model seems not unreasonable
as a candidate for stick prices. If this model holds, then there is no information in past
prices

Time Series Plot of Price

510

508

506
Price

504

502

500

1 10 20 30 40 50 60 70 80 90 100
Time

Time Series Plot of Price

520

515

510
Price

505

500

1 10 20 30 40 50 60 70 80 90 100
time
b) Estimated lack of dependence in real data to predict future price changes: There are
various statistical measures that allow one to numerically estimate how much
dependence there is in price changes over time from the point of view of predicting
future price changes. All such measures that researchers have estimated from data over
the years are pretty much close to zero, thus not contradicting the RW implication that
past price changes don’t have useful information about predicting future price changes.

Evidence against the RW: Benoit Mandelbrot (1963)

Volatility clustering: This phenomenon has now come to be accepted as a fact that occurs
in all reasonably liquid asset prices (frequently traded stocks, indices, exchange rates,
etc.) in developed markets.

Note that this violates the basic RW model, which implied that the past history of the
prices is irrelevant and useless in making predictions about anything regarding future
price changes.

This led researchers to construct a new RW model in which the price changes are what
are called martingale differences with volatility clustering (the latter phenomenon is also
called conditional heteroscedasticity). The interesting thing about this revised RW model
is that it manages to exhibit the volatility clustering phenomenon (so past price changes
have useful information about the magnitude of future price changes) while, at the same
time, implying that past price changes contain no useful information in predicting the
future price change (i.e. whether the future price change goes up, goes down, etc).
For adherents of the RW model, this model fulfills 3 things:

a) Prices simulated from it exhibit the same momentum style patterns seen in the
original RW
b) By construction, the revised RW is such that it implies that future price changes
cannot be predicted by past price changes
c) By construction, it takes into account the volatility clustering seen in real data

The interesting thing is that though Mandelbrot noticed the volatility clustering effect in
the early 60’s it wasn’t till the early 80’s, 20 years later, that attention turned to this
phenomenon and models were developed for it.
The first statistical model for the price changes possessing volatility clustering was the
ARCH (AutoRegressive Conditional Heteroscedasticity) model, proposed in the early
1980’s by an econometrician named Rob Engle, who has been on the Finance faculty at
Stern for the last 18 years. This fundamental model resulted in an explosion of more
flexible models with the same fundamental properties, such as the GARCH (Generalized
ARCH) and Stochastic Volatility models etc.

Practitioners realized that such models were crucial in assessing risk, which would vary
depending on the most recent volatility in the asset and these models became
workhorses in such risk assessments. This work had such impact that in 2003, Rob Engle
was awarded the Nobel Prize in Economics.
Simplification in computing joint probabilities when two
events are independent

If A and B independent, then

P(A and B)=P(A) P(B)

This extends to more than two events

Group/pooled testing for a disease (covid)

Assume that the test has sensitivity=specificity=100%. A

random sample of, say, 5 people is selected, a nasal swab
taken from each and then the material from the 5 swabs is
pooled/combined and one single test is run on the pooled
material. If this test comes back negative, all 5 people are
declared negative. If this test comes back positive, then all
the 5 individuals are tested separately (i.e. 5 tests)

Suppose that 10% of the population in a city is infected. A

pooled test is carried out on 5 randomly selected people.
What is the probability that the test comes back positive?
Random Variable: Variable which takes on numerical
values determined by the outcome of the phenomenon
under study

Ex: Number on the top face of a die

Number of heads in 8 tosses of a coin
Number of phone calls in an hour at the Service desk
Salary of an individual

Discrete random variable: a variable which takes values

arising from a counting process or a finite number of
values

Probability distribution for a discrete r.v. : Table or formula

which associates a probability with each of the values of
the random variable

Ex: Roll a dice. If {1,2,3} come up, you lose $1. If {4} comes
up, you lose nothing. If {5,6} come up, you win $1

Y = earning in the game

Y P(Y=y)
-1 3/6
0 1/6
1 2/6
 Expected value of a discrete r.v.

Computation:

Suppose Y takes N different values y1 , y 2 ,... y N

N
E (Y ) = ∑ y i P (Y = y i )
i =1

= y1P(Y = y1 ) + y 2 P(Y = y 2 )+...+ y N P(Y = y N )

Ex: Y=earning in the game

E(Y)= -1(3/6) + 0(1/6) + 1(2/6) = - 1/6 $

Variance of a discrete r.v.

Computation:

N 2

Var (Y) = ∑ ( yi − E (Y) ) P(Y = yi )

i =1

= ( y1 − E (Y)) 2 P(Y = y1 ) + ... + ( y N − E (Y)) 2 P(Y = y N )

Ex: Y=earning in the game

Var (Y) = (−1 − (−1 / 6)) 2 (3 / 6) + (0 − (−1 / 6)) 2 (1 / 6)

+ (1 − (−1 / 6)) 2 (2 / 6)

= 174 / 216
Standard Deviation of a discrete r.v.

The standard deviation is the positive square root of the

variance of the random variable

Ex: earning in a game

Std. deviation is 0.8975