Warm Up

Xiao Qiao
City University of Hong Kong
Statistics Preliminaries
Our Basic Toolbox

Probability
Inference
Regressions
Time Series

X. Qiao 1
Random Variable

Random variable (RV)

Takes on one of many possible values
Each value is a realization of the random variable
Discrete vs. continuous
Discrete
Probability mass function (pmf)
Cumulative mass function (cmf)
Continuous
Probability density function (pdf)
Cumulative density function (cdf)

X. Qiao 2
Example 1 - Discrete RV

R can take on exactly one of three values:




 1.1 w.p. 1/3

R = 1.0 w.p. 1/3



0.9 w.p. 1/3

X. Qiao 3
Example 2 - Continuous RV

1 (R−µ)2
f (R) = √ e − 2σ2
2πσ

Normal random variable

Mean µ and standard deviation σ are the parameters of the
distribution

X. Qiao 4
Moments 1

Mean
What value is the random variable centered around?
Discrete
X
µ = E (R) = πi Ri
i

Continuous
Z
E (R) = R f (R)dR

X. Qiao 5
Moments 2

Variance
How “spread out” is the random variable?
X
var (R) = σ 2 (R) = E (R − E (R))2 = πi (Ri − E (R))2
i

Standard deviation
Square root of variance
Same units as the mean

X. Qiao 6
Moments 3

Covariance
How do two random variables move with each other?

cov (R a , R b ) = E [(R a − E (R a ))(R b − E (R b ))]

X
= πi [(Ria − E (R a ))(Rib − E (R b ))]
i

Positive if move together, negative if move in opposite directions

Correlation
cov (R a , R b )
corr (R a , R b ) =
σ(R a )σ(R b )

Correlation is always between -1 and 1

X. Qiao 7
Some Properties of Moments

Mean

E (cR) = cE (R)
E (R a + R b ) = E (R a ) + E (R b )

Variance

var (cR a + dR b ) = c 2 var (R a ) + d 2 var (R b ) + 2cd cov (R a , R b )

Covariance

cov (cR a , R b ) = c cov (R a , R b )

X. Qiao 8
Statistics and Inference

How to get probabilities or distributions of random variables from

observed data?
Analog Principle: Use sample moments as proxies for population
moments

Consistency - convergence in the limit as we get infinite data

Law of Large Numbers
Sampling variation
Confidence intervals and hypothesis testing

X. Qiao 9
Some Sample Moments

Sample mean
T
1 X
R̄ = Rt
T
t=1

Additional notation: If µ is the population mean, the sample mean is

denoted as µ̂

Sample variance
T
1 X
s 2 = σ̂ 2 = (Rt − R̄)2
T −1
t=1

X. Qiao 10
Sampling Variation

Standard error
σ̂(R)
s.e.(R) = σ̂(R̄) = √
T
Confidence interval
Assuming normal distribution, 95% confidence interval is
R̄ ± 1.96 × s.e.(R)
If the true mean were µ, the probability for observing a sample mean
outside of the 95% confidence interval is less than 5%
Hypothesis testing
Assume normal or t-distribution

H0 : µ = C
√ R̄ − C
T
σ̂
X. Qiao 11
Regression Setup

In general, we have

yt = α + β1 x1t + β2 x2t + ... + ϵt ; t = 1, 2, ...T

In finance, often in the form

Rt = α + βRm,t + ϵt
Rt = α + βRm,t + γRp,t + ϵt

X. Qiao 12
Regression Facts

Start up with a simple regression

yt = α + βxt + ϵt

Assumption: E (ϵt ) = 0 and E (xt ϵt ) = 0

cov (x,y )
1 Population value of regression coefficient is β = var (x)
2 True β is recovered only if the error term is uncorrelated with the
right-hand variables, i.e. E (xt ϵt ) = 0
1 sales = α + βadvertisingexpense + ϵ, discounts help sales, if
advertisements is correlated with discounts, then β is biased
3 In a multiple regression, coefficients capture effects on y orthogonal
to all other right-hand variables

X. Qiao 13
Time Series Regression

Rt = α + βRm,t + ϵt ; t = 1, 2, ...T

How do the left-hand and right-hand variables move together

contemporaneously?
R 2 summarizes the variation in returns of the left-hand variable
captured by the right-hand variable(s)
Often use the market portfolio on the right-hand side, hence Rm,t

X. Qiao 14
Forecasting Regression

Rt+1 = α + βRp,t + ϵt+1 ; t = 1, 2, ...T − 1

How do the right-hand variable(s) predict the movement of the

left-hand variable?
For annual data, R 2 of 0.1 is very good for stock returns
Right-hand variable does not have to be portfolio returns, could use
other characteristics such as dividend-price ratio or book-to-market

X. Qiao 15
Cross-Sectional Regression

Ri = α + βi λ + ϵi ; i = 1, 2, ...N

How do variation in betas correspond to differences in returns at one

point in time across assets?
If α is small, the asset pricing model “prices” the left-hand assets
R 2 is not super important

X. Qiao 16
Definition of Time Series

Set of repeated observations of a random variable over time

Subscript denotes time
Example: y1 , y2 , y3 , ...yt , ...
If the series were iid, then knowing the past does not give us
information about the future
If the series were not iid, then we can model its dynamics

X. Qiao 17
Conditional vs. Unconditional

Unconditional moments are simply the population moments we have

seen earlier
E (y ), σ 2 (y ), etc.
Conditional moments use information from the past, exploiting
dynamics of the series
E (yt+1 |yt , yt−1 , ...) or E (yt+1 |Ft )
Ft denotes information known at time t

X. Qiao 18
Autocovariance and Autocorrelation

Apply definitions of covariance and correlation to time series

observations
Autocovariance

cov (yt , yt−j ) = E [(yt − E (yt ))(yt−j − E (yt−j ))]

Autocorrelation
cov (yt , yt−j )
corr (yt , yt−j ) =
σ 2 (y )
Tell us if we see today’s observation, where tomorrow’s observation is
likely to be
If autocorrelation is positive, a high value today is likely followed by a
high value tomorrow
If autocorrelation is negative, a high value today is likely followed by a
low value tomorrow
X. Qiao 19
Stationarity

A series is stationary if its unconditional moments are not changing

through time
Strong stationarity

f (yt , yt−1 , ...yt−J ) = f (yt−i , yt−i−1 , ...yt−i−J )

Weak stationarity (also called covariance stationary)

E (yt ) = E (yt−j ) ∀j
σ 2 (yt ) = σ 2 (yt−j ) ∀j
cov (yt , yt−j ) = γj does not depend on t

X. Qiao 20
White Noise Process

Let’s start from the simplest process, the white noise process

yt = ϵt

ϵt has zero mean and a constant variance σ 2 (ϵt ) = σ 2

ϵt s are uncorrelated through time, corr (ϵt , ϵt−j ) = 0, ∀j ̸= 0
Autocovariances and autocorrelations are all zero
If a time series y is a white noise process, any past observations of
this series do not tell us anything about where the series will go next

X. Qiao 21
MA(1)

Let’s inject the white noise process with some dependence

yt = θϵt−1 + ϵt

This is called an MA(1) process, or moving average of order one

Moments

E (yt ) = 0
σ 2 (yt ) = (1 + θ2 )σ 2
θ
corr (yt , yt−1 ) =
1 + θ2
corr (yt , yt−j ) = 0, ∀j ≥ 2

X. Qiao 22
MA(1) Continued

MA(1) has a one-period dependence

We can also see this by looking at the conditional moments

Et (yt+1 ) = θϵt
Et (yt+j ) = 0, ∀j ≥ 2
σt2 (yt+1 ) = σ 2
σt2 (yt+j ) = (1 + θ2 )σ 2 , ∀j ≥ 2

Et is the conditional expectation using all information at time t

The conditional moments revert to unconditional moments after one
period
Could add a level to the unconditional mean, i.e.

yt = µ + θϵt−1 + ϵt
X. Qiao 23
MA(q)

MA(1) has one-period dependence; MA(q) has q-period dependence

Autocorrelation is non-zero for the first q lags, zero for any further
lags
Conditional mean reverts back to unconditional mean after q + 1
periods
Conditional variance reverts back to unconditional variance after
q + 1 periods

X. Qiao 24
AR(1)

MA(q) captures dependence for q periods, but what if we want to

capture long-term dependence past q periods?
Use autoregressive process of order one, AR(1)

yt = ρyt−1 + ϵt

Assuming stationarity

E (yt ) = 0
σ2
σ 2 (yt ) =
1 − ρ2
corr (yt , yt−j ) = ρj

Autocorrelations of AR(1) drops to zero slowly, unlike the MA(q)

process for which the values cut off at q + 1
X. Qiao 25
AR(1) Continued

Let’s look at the conditional moments of the AR(1)

Et (yt+1 ) = ρyt
Et (yt+j ) = ρj yt
σt2 (yt+1 ) = σ 2
σt2 (yt+2 ) = (1 + ρ2 )σ 2

X. Qiao 26
AR(p)

yt = ρ1 yt−1 + ρ2 yt−2 + ... + ρp yt−p + ϵt

More complex dynamics than AR(1), but same idea

Captures time series dependence past the first lagged observation

X. Qiao 27
ARMA

We could combine AR and MA processes into ARMA processes

ARMA(1,1)

yt = ρyt−1 + θϵt−1 + ϵt

Richer dynamics than AR or MA alone

To capture longer dependence, could also have ARMA(p,q)
No need to work out moments by hand, use software

X. Qiao 28
Finance Warm Up
Finance in a Nutshell

Finance is all about assets

Buy assets to defer consumption
How do you save/invest your money if you don’t want to spend it all
today?
What does finance study?
Financial market participants
Types of financial assets

X. Qiao 29
Types of Assets

Financial Assets
Liquid, frequently traded
Claims on real assets
Real Assets
Used to produce goods and services, e.g. factories and machines
Illiquid, used for specific purposes, not traded frequently

X. Qiao 30
What does Finance Study?

General issues in finance

Asset valuation
Portfolio choice
Corporate finance
Investments
Discounting is central to asset valuation
Prices at different dates are related
The value of a stock today depends on its future cash flows
Time and risk determine the relationship
How much would you pay to get $100 in one year?
How much would you pay to get $200 with probability 0.5 and $0 with
probability 0.5?
X. Qiao 31
Financial Market Participants

Households
Want to invest funds, save to defer consumption today (e.g.
retirement)
Corporations
Obtain external funds for its projects (e.g. develop a new product)
Invest accumulated profits
Financial Sector
Role of an intermediary (e.g. banks, brokers)
Government
Obtain funds for its projects (e.g. build a highway)

X. Qiao 32
Types of Financial Assets

Equity
Common stock
Preferred stock (promises a dividend)
Warrants (e.g. executive stock options)
Debt
Government vs. corporate
Bank debt
Money market instruments (short-term lending and borrowing)

X. Qiao 33
Types of Financial Assets

Commodities
Soybeans, crude oil, etc.
Physical inputs to production
Foreign Exchange
Relation between different currencies

X. Qiao 34
Types of Financial Assets

Derivatives
An asset whose value is derived from some underlying asset
The underlying could be anything: stocks, bonds, stock index (e.g.
Hang Seng)
Examples: options, futures, credit-default swaps (CDS)
Underlying could also be a derivative itself

Some assets fall into multiple categories

Mortgage-backed securities (debt, derivative)

X. Qiao 35
Holding Period Return

We use returns to measure how well an asset performs

Gross return Rt+1 is the payoff tomorrow compared to the price you
paid today Pt
Pt+1 + Dt+1
Rt+1 =
Pt
Pt+1 is the price you get when you sell tomorrow, and Dt+1 is the
dividend
Dt+1 can be viewed more generally as some cash flow for assets other
than a stock, such as coupon payments (bonds) or rent (real estate)
Net return rt+1 = Rt+1 − 1

X. Qiao 36
Calculating Returns

Duolingo’s initial public offering (IPO) on July 28, 2021

Offering price: $102/share
Closed at $136.4 after two days of trading
Two-day return to IPO investors?
136.4
Rt+2 = = 1.337
102
rt+2 = 1.337 − 1 = 33.7%

Note the different representation of gross and net returns

We commonly quote net returns when speaking about investments

X. Qiao 37
Expected vs. Realized Returns

We do not know how much the stock market will go up or down next
year
We represent an unknown value with a random variable r˜t+1
If we are at time t,

P̃t+1 + D̃t+1 − Pt
r˜t+1 =
Pt

P̃t+1 , D̃t+1 , and r˜t+1 are unknown today, Pt is known so no tilde˜

X. Qiao 38
Expected vs. Realized Returns

Expected return

E [P̃t+1 ] + E [D̃t+1 ] − Pt
E [˜
rt+1 ] =
Pt
Realized return, after observing Pt+1 , Dt+1 , is
Pt+1 + Dt+1 − Pt
rt+1 =
Pt
I’ve removed the tildes to indicate Pt+1 and Dt+1 are now known
E [˜
rt+1 ] and rt+1 can be very different

X. Qiao 39
Compounding

Compounding is about earning interest on interest

An investment yields rt+1 at time t + 1 and rt+2 at time t + 2. Let’s
ignore dividends for now
Pt+1 − Pt
rt+1 = → Pt+1 = Pt (1 + rt+1 )
Pt
Similarly

Pt+2 = Pt+1 (1 + rt+2 )

We can now relate Pt+2 with Pt

Pt+2 = Pt (1 + rt+1 )(1 + rt+2 )

Pt+2 − Pt
rt,t+2 = = (1 + rt+1 )(1 + rt+2 ) − 1
Pt
X. Qiao 40
Multi-Period Returns

If you earn 20% in year 1 and 10% in year 2, what’s your two-year
return?

rt,t+2 = (1 + rt+1 )(1 + rt+2 ) − 1 = (1 + 0.2)(1 + 0.1) − 1 = 32%

What is the return per year?

Simple average gives (20% + 10%)/2 = 15%
But if we compound 15% for two years, we get something different:

(1 + 0.15)(1 + 0.15) − 1 = 32.25%

We need to distinguish between arithmetic and geometric returns

X. Qiao 41
Multi-Period Returns

Geometric average is the one number that gives us the same outcome
that we get from using actual rates

(1 + r )(1 + r ) = (1 + 0.2)(1 + 0.1)

Solve for r

(1 + r )2 = (1 + 0.2)(1 + 0.1)
p
1 + r = (1 + 0.2)(1 + 0.1)
p
r = (1 + 0.2)(1 + 0.1) − 1 = 14.9%

Arithmetic return: 15%, geometric return: 14.9%

X. Qiao 42
Arithmetic vs. Geometric Returns

The difference can be large if returns are highly variable or the

horizon is long
-20%, 35%, -25%, 40%, and 10%
Arithmetic: 8%, geometric: 4.5%
Geometric returns is what you make if you hold the asset
A stock trades $10 today, $20 one year from now, and $10 two years
from now
The two one-year returns are 100% and -50%
Arithmetic: 25%, geometric: 0%

We will often use arithmetic returns, because when we estimate the

mean, we use arithmetic returns

X. Qiao 43
Historical Returns

X. Qiao 44
Historical Returns

Total returns = capital gains (Pt+1 /Pt ) + income yield (Dt+1 /Pt )

X. Qiao 45
Nominal vs. Real Returns

We just looked at nominal returns

Real returns tell us how the purchasing power of wealth changed over
time
The price of one unit of consumption good is $1 and annual inflation is
5%
The amount of goods consumers can buy with $1 in one year is
1
1+0.05 = 0.952
If π is inflation rate, then
Rnominal
Rreal =
1+π
1 + rnominal
1 + rreal =
1+π

X. Qiao 46
Nominal vs. Real Returns

Nominal returns

Real returns

X. Qiao 47
Risk

Risk is our lack of knowledge of a random variable

What will Tesla’s stock price be next year?
How do we measure risk?
Standard deviation σ, also called volatility
Range of the variable: maximum and minimum, 5th and 95th
percentiles
Value-at-risk, drawdown, ...

X. Qiao 48
Risk

We use sample estimates as approximations of true population means

and variances

X. Qiao 49
Risk and Return

Poll: Flip a coin. If it lands heads, you receive $40,000. If it lands tails,
you lose $20,000. What would you pay to play this game?
1 Nothing
2 $8,000
3 $10,000
4 $15,000
5 $20,000

X. Qiao 50
Risk and Return

Most people would take a sure $10,000 over playing an uncertain

game. Both choices offer expected reward of $10,000, but one is sure
and one is uncertain
You might choose to pay $8,000 to play a game with expected reward
of $10,000. Or you may not
Takeaway
People are generally risk averse (they dislike risk)
For people to do something risky, they have to be rewarded for it

X. Qiao 51
Risk and Return

Stocks offer higher average returns than bonds because stocks are
riskier
For risk-averse people to hold riskier stocks, they have to receive
higher average returns
The difference in average returns between stocks and bonds should
reflect the higher level of risk of stocks

X. Qiao 52