MFIN 305: Quantitative

Methods of Finance
A Brief Overview of the Classical Linear Regression Model
1. What is a Regression Model?
It is the relationship between a given variable and one or more variables

Names for 𝑦 Names for 𝑥s

Dependent variable Independent variables
Regressand Regressors
Effect variable Causal variables
Explained variable Explanatory variables
2. Regression Versus Correlation
• Correlation is the degree of linear association between 2 variables
• If we say 𝑦 and x are correlated, it means that we treating 𝑦 and 𝑥 in a
completely symmetrical way
• In regression, we treat the dependent variables (𝑦) and the independent
variables(s) (x’s) very differently.
• The 𝑦 variable is assumed to be random or “stochastic” (i.e., to have a
probability distribution). The x variables are, however, assumed to
have fixed (“non-stochastic”) values in repeated samples
3. First step: Scatter Plot

x
4. Line of Best Fit
• The equation of a straight line is used to get the line of best fit:
𝑦 = 𝛼 + 𝛽𝑥 (Perfect fit)
• 𝛼 represents the value of 𝑦 if 𝑥 is zero
• Interpretation of 𝛽: if 𝑥 increases by 1, 𝑦 will be expected, all else
being equal, to increase by 𝛽
• However, it is unrealistic that 𝑥 alone explains 𝑦
4. Line of Best Fit
• Line of best fit:
Too broad as a definition
ൠ ⇒Imprecise
Room for disagrement about "best" fit
• The values of 𝛼 and 𝛽 that place the line as close as possible to all the
points / data are chosen
4. Line of Best Fit
𝑦𝑡 = 𝛼 + 𝛽𝑥𝑡 + 𝑢𝑡 t = 1,2,3, …, T

Dependent Variable Disturbance term

Parameters Regressor

• 𝛼 and 𝛽 should minimize the distance (vertical) from the data points to
the fitted line
5. Ordinary Least squares
• The most common method of fitting a line to the data is Ordinary
Least Squares (OLS)
• OLS is the workhorse of econometrics/data science!
• OLS entails measuring the vertical distance from a point to the line,
squaring it and then minimizing the sum of squares (hence “least
squares”)
5. Ordinary Least squares
Let:
• 𝑦𝑡 denote the actual data point
• 𝑦ො𝑡 denote the fitted value from a regression line, the value predicted by
the model for a given value x
• ^ denote estimated or fitted values (any value that I calculate)
• 𝑢ො 𝑡 denote the residual: 𝑢ො 𝑡 = 𝑦𝑡 − 𝑦ො𝑡

Note: 𝑥𝑡 is assumed to be non-stochastic (fixed) in repeated samples

5. Ordinary Least squares
• Minimizing the sum of 𝑢ො 𝑡2 by finding 𝛼ො and 𝛽መ that do so (optimization
problem). This is OLS.
• Minimizing the sum of squared distances is given by minimizing:
𝑢ො 12 + 𝑢ො 22 + 𝑢ො 32 + 𝑢ො 42 + 𝑢ො 52 or σ5𝑡=1 𝑢ො 𝑡2
• σ5𝑡=1 𝑢ො 𝑡2 is known as the residual sum of squares (RSS) or sum of
squared residuals
6. Deriving the OLS Estimator
More generally, Min σ𝑇𝑡=1 𝑢ො 𝑡2 = Min σ𝑇𝑡=1 𝑦𝑡 − 𝑦ො𝑡 2
𝛼 ෡
ෝ ,𝛽 ෡
ෝ ,𝛽
𝛼

• Let L denote the RSS, also known as a loss function

መ 𝑡 )2
𝐿 = σ𝑇𝑡=1(𝑦𝑡 −𝑦ො𝑡 )2 = σ𝑇𝑡=1(𝑦𝑡 − 𝛼ො − 𝛽𝑥 (1)

• We will Min 𝐿 to find the values of 𝛼ො and 𝛽መ that give the line that is
෡
ෝ ,𝛽
𝛼
closest to the data

Note: Chain Rule: 𝑓 𝑥 = 𝑓(𝑔 𝑥 ) → 𝑓 ′ 𝑥 = 𝑓 ′ (𝑔 𝑥 )𝑔′ (𝑥)

6. Deriving the OLS Estimator
𝜕𝐿 መ 𝑡 =0
• = −2 σ𝑇𝑡=1 𝑦𝑡 − 𝛼ො − 𝛽𝑥 (2)
𝜕ෝ
𝛼
𝜕𝐿 መ 𝑡 =0
• ෡ = −2 σ𝑇𝑡=1 𝑥𝑡 𝑦𝑡 − 𝛼ො − 𝛽𝑥 (3)
𝜕𝛽

• From (2): መ 𝑡 =0
σ 𝑦𝑡 − 𝛼ො − 𝛽𝑥 (4)
σ 𝑦𝑡 − 𝑇𝛼ො − 𝛽መ σ 𝑥𝑡 = 0 (5)
Use σ 𝑦𝑡 = 𝑇𝑦 and σ 𝑥𝑡 = 𝑇𝑥 to write:
𝑇𝑦 − 𝑇𝛼ො − 𝑇𝛽෡ 𝑥 = 0 (6)
መ =0
Or 𝑦 − 𝛼ො − 𝛽𝑥 (7)
6. Deriving the OLS Estimator
መ 𝑡 =0
• From (3): σ𝑡 𝑥𝑡 𝑦𝑡 − 𝛼ො − 𝛽𝑥 (8)
መ
• From (7): 𝛼ො = 𝑦 − 𝛽𝑥 (9)
• Substituting into (8) for 𝛼ො from (9) yields:
σ 𝑥𝑡 (𝑦𝑡 − 𝑦 + 𝛽𝑥መ − 𝛽𝑥 መ 𝑡) = 0 (10)
መ σ 𝑥𝑡 − 𝛽መ σ 𝑥𝑡2 = 0
σ 𝑥𝑡 𝑦𝑡 − 𝑦 σ 𝑥𝑡 + 𝛽𝑥 (11)
መ 𝑥ҧ 2 − 𝛽መ σ 𝑥𝑡2 = 0
σ 𝑥𝑡 𝑦𝑡 − 𝑇𝑥 𝑦 + 𝛽𝑇 (12)
6. Deriving the OLS Estimator
መ
• Rearrange for 𝛽:

𝛽መ 𝑇𝑥ҧ 2 − σ𝑡 𝑥𝑡 2 = 𝑇 𝑥 𝑦 − σ 𝑥𝑡 𝑦𝑡 (13)

• Line of best fit:

σ 𝑥𝑡 𝑦𝑡 −𝑇𝑥 𝑦
መ
𝛽= 2 and 𝛼
ො = 𝑦 − ෡𝑥
𝛽
2
σ 𝑥𝑡 −𝑇 𝑥

σ 𝑥𝑡 −𝑥 (𝑦𝑡 −𝑦)
መ
• More intuitively: 𝛽 = Sample cov(x,y)
σ𝑡(𝑥𝑡 −𝑥)ҧ 2
Sample variance of x
6. Deriving the OLS Estimator
σ 𝑥𝑡 −𝑥ҧ (𝑦𝑡 −𝑦)
• Note: Sample covariance 𝜎𝑥𝑦 =
(𝑇−1)

σ 𝑥𝑡 −𝑥 (𝑦𝑡 −𝑦) 𝜎𝑥𝑦

Sample Correlation 𝜌𝑥𝑦 = =
𝑇−1 𝜎𝑥 𝜎𝑦 𝜎𝑥 𝜎𝑦

• From the equation of 𝛼,

ො it is obvious that the regression line will go
through the mean of the observations (i.e. ( 𝑥, 𝑦 ) lies on the regression
line)
7. What do we use 𝛼ො and 𝛽መ For?
Suppose that we have the following data on the excess returns on a fund
manager’s portfolio (“fund XXX”) together with the excess returns on a
market index:
Year, t Excess return = 𝒓𝑿𝑿𝑿,𝒕 − 𝒓𝒇𝒕 Excess return on market index = 𝒓𝒎𝒕 − 𝒓𝒇𝒕

1 17.8 13.7
2 39.0 23.2
3 12.8 6.9
4 24.2 16.8
5 17.2 12.3
7. What do we use 𝛼ො and 𝛽መ For?
45
Excess return on fund XXX

40
35
30
25
20
15
10
5
0
0 5 10 15 20 25
Excess return on market portfolio
7. What do we use 𝛼ො and 𝛽መ For?
• Using the data, calculate 𝛼ො and 𝛽መ
• We will get 𝛼ො = -1.74 and 𝛽መ = 1.64

• The fitted line is:

𝑦ො𝑡 = −1.74 + 1.64𝑥𝑡

(Fitted) Excess returns Excess return on market portfolio

on the fund (𝑟𝑚 − 𝑟𝑡 ) 𝑜𝑟 𝑚𝑎𝑟𝑘𝑒𝑡 𝑟𝑖𝑠𝑘 𝑝𝑟𝑒𝑚𝑖𝑢𝑚
7. What do we use 𝛼ො and 𝛽መ For?
Suppose 𝑟𝑚 − 𝑟𝑓 (or 𝑥𝑡 ) next year (year 6) is 20%. What would the
expected return on the fund be?
𝑦ො𝑡 = −1.74 + 1.64 ⨯ 20 = 31.06

Beta in a CAPM Excess return over the

Risky fund risk free rate of 31.06%

መ if x increases by 1 unit, y will be expected,

• Interpretation of 𝛽:
everything else being equal, to increase by 1.64 units
8. The Population and the Sample
Population Sample
• The total collection of all objects or • Some items from the population
people to be studied
• It may be either finite or infinite • It is usually random
• Ex. All stocks on LSE • Representative of the population of
interest
• Each individual in the population is
equally likely to be drawn
• Size = number of available
observations
9. The DGP, PRF and SRF
Data Generating Process Population Regression Function Sample Regression Function
(DGP) (PRF) (SRF)
Process describing the way the • Description of the model It the relationship estimated
actual observations on y came thought to be generating the using the sample observations
about actual data (information)
• Represents the true
relationship between the መ 𝑡
𝑦ො𝑡 = 𝛼ො + 𝛽𝑥
variables
• Embodies the true values of መ 𝑡 + 𝑢ො 𝑡
𝑦𝑡 = 𝛼ො + 𝛽𝑥
𝛼 and 𝛽 = 𝑦ො𝑡 + 𝑢ො 𝑡
𝑦𝑡 = 𝛼 + 𝛽𝑥𝑡 + 𝑢𝑡 𝑦ො𝑡 : fitted value, 𝑢ො 𝑡 : residual
10. Linearity
• Linearity assumption: in order to use OLS, a linear model is required
(i.e. the relationship between y and x is a straight line)
• The model must be linear in the parameters (𝛼 and 𝛽), but it does not
necessarily have to be linear in the variables (y and x)
• Some nonlinear models can be made to be linear
𝛽
𝑌𝑡 = 𝐴𝑋𝑡 𝑒 𝑢𝑡 (exponential regression model)
ln𝑌𝑡 = ln 𝐴 + β ln 𝑋𝑡 + 𝑢𝑡 ⇒ 𝑦𝑡 = 𝛼 + 𝛽 𝑥𝑡 + 𝑢𝑡

𝑦𝑡 𝛼 𝑥𝑡
10. Linearity
• 𝛽 is an elasticity (strictly, unit changes on a logarithmic scale)
• Say 𝛽መ = 1.2, then a rise in x of 1% will lead, on average, everything else
being equal, to a rise in 𝑦 of 1.2%
• If the theory suggests that should be inversely related to a model of the
form:
𝛽
𝑦𝑡 = 𝛼 + + 𝑢𝑡
𝑥𝑡
You can estimate the regression by OLS by setting:
1
𝑧𝑡 =
𝑥𝑡
• Some models are intrinsically non-linear: 𝑦𝑡 = 𝛼 + 𝛽𝑥𝑡𝛿 + 𝑢𝑡
11. Estimator or Estimate?

Estimator Estimate
Actual numerical values for the
Formulae used to calculate the
coefficients obtained from the
coefficients
sample
12. The Classical Linear Regression Model

𝑦𝑡 = 𝛼 + 𝛽𝑥𝑡 + 𝑢𝑡

• After assumptions are made about 𝑢𝑡 , this becomes the classical linear
regression model (CLRM)

• Assumptions are made about the disturbances 𝑢𝑡 , not the residuals

𝑢ො 𝑡
12. The Classical Linear Regression Model

Assumptions Interpretation
(1) E 𝑢𝑡 = 0 The errors have zero mean
The variance of the errors is constant and finite over all
(2) var 𝑢𝑡 = 𝜎 2 < ∞
values of 𝓍𝑡
(3) cov 𝑢𝑖 , 𝑢𝑗 = 0 The errors are linearly independent of one another
There is no relationship between the error and
(4) cov 𝑢𝑡 , 𝑥𝑡 = 0
corresponding x variate
12. The Classical Linear Regression Model
• If assumption (1) holds, assumption (4) can be written as E 𝑥𝑡 𝑢𝑡 = 0
(i) This implies that the regressor is orthogonal (i.e. unrelated to) the error
term
(ii) An alternative assumption to (4), which is slightly stronger is that the 𝑥𝑡
are non-stochastic or fixed in repeated samples (i.e. no sampling variation
in 𝑥𝑡 and its value is determined outside the model)

• A fifth assumption is required to make valid inferences about the population

መ
parameters (𝛼 and 𝛽) from the sample parameters (𝛼ො and 𝛽):

(5) 𝑢𝑡 ~ 𝑁(0, 𝜎 2 ) ⇒ 𝑢𝑡 is normally distributed

13. Properties of the OLS Estimator
If assumptions (1)-(4) hold, then the estimators 𝛼ො and 𝛽መ determined by
OLS will have a number of desirable properties
Under assumptions (1)-(4), the OLS estimator is the best linear unbiased
estimator (BLUE)
• Best: means that the OLS estimator has minimum variance among the
class of linear unbiased estimators. The Gauss-Markov theorem proves
that the OLS estimator is best by examining an arbitrage alternative
linear unbiased estimator and showing in all cases that it must have a
variance no smaller than the OLS estimator
• Linear: 𝛼ො and 𝛽መ are linear estimators (linear combinations) of y
13. Properties of the OLS Estimator
• Unbiased: on average, the actual vales of 𝛼ො and 𝛽መ will be equal to
their true values

• Estimator: 𝛼ො and 𝛽መ are estimates of the true values of 𝛼 and 𝛽

13. Properties of the OLS Estimator
OLS estimator is consistent, unbiased and efficient
Consistency
• 𝛽መ is consistent if: lim 𝑃𝑟 𝛽መ − 𝛽 < 𝛿 = 0 ∀ 𝛿 > 0
𝑇→∞
• As the sample size increases (tends to infinity), the estimates will
converge to their true values
• Consistency is a large sample or asymptotic, property
• The assumptions that E 𝑥𝑡 𝑢𝑡 = 0 and E 𝑢𝑡 = 0 are sufficient to
derive consistency
13. Properties of the OLS Estimator

T=300

0 T=100 T=200
13. Properties of the OLS Estimator
Unbiasedness
OLS estimates are unbiased
𝐸 𝛼ො = 𝛼 On average, the estimated values for the
መ ቋ coefficients will be equal to their true
𝐸 𝛽 =𝛽
values
• No systematic over or under-estimation of true coefficients
• Proving unbiasedness requires that cov(𝑥𝑡 , 𝑢𝑡 ) = 0
• It is a stronger condition than consistency since it holds for small as
well larger samples
13. Properties of the OLS Estimator
Efficiency
• An estimator of parameter 𝛽 is said to be efficient if no other
estimators has a smaller variance
• Minimizing the probability that it will be far away from the true value
• Estimator’s probability distribution is Efficient / Best
narrowly dispersed around the true value

0
14. Precision and Standard Errors
• Set of estimates 𝛼ො and 𝛽መ are specific to the sample used in their
estimation
• Different sample → different (𝑥𝑡 , 𝑦𝑡 ) → different 𝛼ො and 𝛽መ
መ How good these estimates are?
• Measure the reliability of 𝛼ො and 𝛽.
• Precision (confidence in the estimates) → standard error
• Would estimates vary from one sample to another (out of the
population)?
• Precision can be calculated using available data
14. Precision and Standard Errors
Given assumptions 1 to 4, valid estimators of the standard errors are
given by:
σ 𝑥𝑡2 σ 𝑥𝑡2
𝑠𝑒 𝛼ො = 𝑠 = 𝑠
𝑇 σ(𝑥𝑡 − 𝑥) 2 𝑇 σ 𝑥𝑡2 − 𝑇 2 𝑥ҧ 2

1 1
𝑠𝑒 𝛽መ = 𝑠 2
=𝑠
σ(𝑥𝑡 − 𝑥) σ 𝑥𝑡 2 − 𝑇𝑥ҧ 2

Where s is the estimated standard deviation of the residuals

14. Precision and Standard Errors
• Standard errors give a measure of the degree of uncertainty in the
estimated values for the coefficients
• 𝑠 2 : estimate of the variance of the disturbance term
• 𝜎 2 : actual variance of the disturbance term
 2
15. Estimating the variance of the 𝑢𝑡 ; 𝜎
0
2
𝜎2 = var 𝑢𝑡 = 𝐸 𝑢𝑡 − 𝐸(𝑢𝑡 )

𝜎 2 = 𝑣𝑎𝑟 𝑢𝑡 = 𝐸(𝑢𝑡2 )
• 𝑢𝑡 is unobservable (population disturbances), thus:
1
𝑠2 = σ 𝑢ො 𝑡2 Biased estimator of 𝜎 2 , but consistent
𝑇
ෝ𝑡2
σ𝑢 ෝ𝑡2
σ𝑢
𝑠2 = ⇒𝑠= standard error of the regression or
𝑇−2 𝑇−2
standard error of the estimate
• s is a broad measure of the fit of the regression equation.
• The smaller s is, the closer is the fit of the line to the actual data
16. Comments on the Standard Error
Estimators
σ 𝑥𝑡2 σ 𝑥𝑡2
𝑠𝑒 𝛼ො = 𝑠 2
=𝑠 2
𝑇 σ(𝑥𝑡 − 𝑥) σ 2
𝑇(( 𝑥𝑡 ) − 𝑇 𝑥 )
መ , se(𝛼)
1. As the sample size T , se(𝛽) ො
2. As 𝑠 2 , se(𝛽)መ , se(𝛼)
ො → The larger the residuals are, the worse is
the fit of the line
3. The larger σ(𝑥t − 𝑥ҧ )2 is, the smaller the coefficient variances
→ More variation in the x’s around their mean is better
The importance of the deviation of x from its
mean
y
y

y y

x 0 x x
0 x
Aside: Accuracy of the intercept estimate
y

0 x
17. How to Calculate the Parameters and
Standard Errors
Assume the following data have been calculated from a regression of y
on a single variable 𝑥 and a constant over 22 observations:
σ 𝑥𝑡 𝑦𝑡 = 830102, 𝑇 = 22, 𝑥 = 416.5, 𝑦 = 86.65
σ 𝑥𝑡2 = 3919654, 𝑅𝑆𝑆 = 130.6
σ 𝑥𝑡 𝑦𝑡 −𝑇 𝑥ҧ 𝑦ത 830102−(22⨯416.5⨯86.65)
መ
•𝛽= 2 = = 0.35
2
σ 𝑥𝑡 −𝑇 𝑥 3919654−22(416.5) 2

• 𝛼ො = 𝑦 − 𝛽෡ 𝑥 = 86.65 − 0.35 ⨯ 416.5 = −59.12

መ 𝑡 = −59.12 + 0.35𝓍𝑡
• Sample regression function: 𝑦ො = 𝛼ො + 𝛽𝑥
17. How to Calculate the Parameters and
Standard Errors
ෝ𝑡 2
σ𝑢 130.6
• 𝑠𝑒 regression : 𝑠 = = = 2.55
𝑇−2 20

σ 𝑥𝑡2 3919654
• 𝑠𝑒 𝛼ො = 𝑠 2 = 2.55 ⨯ = 3.35
𝑇 σ 𝑥𝑡2 −𝑇 2 𝑥 22⨯(3919654−22⨯416.52 )

1 1
• 𝑠𝑒 𝛽መ = 𝑠 σ 𝑥𝑡2 −𝑇𝑥 2
= 2.55 ⨯ = 0.0079
3919654−22⨯416.52

• We write: 𝑦ො𝑡 = −54.12 + 0.35𝑥𝑡

(3.35) (0.0079)
18. An Introduction to Statistical Inference
• Financial theory will often suggest that certain coefficients should take on
particular values, or values within a given range
• Sample point estimates 𝛼ො and 𝛽መ of true but unknown population parameters
𝛼, 𝛽
• How reliable are these estimates? 𝑠𝑒 𝛼ො , 𝑠𝑒(𝛽)መ
• 𝛼 and 𝛽 describe the true relationships between the variables, but they are
never available
• Inferences are made concerning the likely population values from the
regression parameters that have been estimated from the sample of data to
hand
• The aim is to determine the differences between the coefficient estimates
actually obtained and expectations arising from financial theory
19. Hypothesis Testing
• Null Hypothesis (𝐻0 ): statement or statistical hypothesis being tested, it is
always an equality
• Alternative hypothesis (𝐻1 ): Remaining outcomes of interest
• Hypothesis testing is always conducted on population values
𝐻0 : 𝛽 = 0.5
• Two-sided test: ቊ
𝐻1 : 𝛽 ≠ 0.5
𝐻0 : 𝛽 = 0.5 𝐻0 : 𝛽 = 0.5
• One-sided test: ൜ or ቊ
𝐻1 : 𝛽 > 0.5 𝐻1 : 𝛽 < 0.5

A bank requires traders to keep their portfolio’s beta below 0.5

(financial theory)
19. Hypothesis Testing
Hypothesis testing Level of significance
approach
Test of significance approach Confidence
interval approach
• Intuition: if 𝛽መ = 0.5091 and the hypothesized value is 5, the null
hypothesis is likely to be rejected
20. Probability Distribution of Least Squares
Estimators
• CLRM assumption: 𝑢𝑡 ~ 𝑁(0, 𝜎 2 )
• Since 𝑢𝑡 is normally distributed, 𝑦𝑡 will be normally distributed
• Since OLS estimators are linear combinations of the random variables
𝛽መ = σ 𝑤𝑡 𝑦𝑡 ; 𝑤𝑡 :weights
and since the weighted sum of normal random variables is also
normally distributed, the OLS estimates will be normally distributed
𝛼~𝑁
ො 𝛼, v𝑎𝑟 𝛼ො መ
𝑎𝑛𝑑 𝛽~𝑁(𝛽, v𝑎𝑟 𝛽መ )
• Even if the errors do not follow a normal distribution, the coefficient
estimates still follow a normal distribution CLT
20. Probability Distribution of Least Squares
Estimators
• Standard normal variables can be constructed from 𝛼ො and 𝛽መ by subtracting
the mean and dividing by the square root of the variance:
𝛼ො − 𝛼 𝛽መ − 𝛽
~𝑁 0,1 and ~𝑁(0,1)
v𝑎𝑟(𝛼)
ො መ
v𝑎𝑟(𝛽)

• v𝑎𝑟(𝛼)
ො and v𝑎𝑟(𝛽) መ are unknown. Thus, their counterparts, the calculated
standard errors of the coefficient estimates 𝑠𝑒(𝛼) መ are used
ො and 𝑠𝑒(𝛽)
𝛼ො − 𝛼 𝛽መ − 𝛽
~𝑡𝑇−2 and ~𝑡𝑇−2
𝑠𝑒(𝛼)
ො መ
𝑠𝑒(𝛽)
21. A Note on the t and the Normal
Distribution
• The normal distribution is “bell shaped”, it
is symmetric around the mean 𝑓(𝓍)

• We can scale a normal variate to have a Normal

distribution
zero mean and unit variance by subtracting
its mean and dividing by its standard
deviation
• There is a specific relationship between the t-distribution
t-distribution and the standard normal
distribution, and the t-distribution has
𝜇 𝓍
another parameter, its degrees of freedom
21. A Note on the t and the Normal
Distribution
Normal Distribution t-distribution
• Normal distribution is fully • Depends on its degree of freedom
characterized by its first two moments parameter
(mean and variance)
• symmetric • Looks like the normal distribution
(symmetric)
• Bell shaped • Has fatter tails than the normal
distribution
• Can be scaled to have zero mean and • Smaller peak at the mean
unit variance
21. A Note on the t and the Normal
Distribution
In the limit, a t-distribution with an infinite number of degrees of
freedom is a standard normal, i.e. 𝑡∞ = 𝑁(0,1)

Significance level 𝑁(0,1) t(40) t(4)

50% 0 0 0
5% 1.64 1.68 2.13
2.5% 1.96 2.02 2.78
0.5% 2.57 2.70 4.60
22. The Test of Significance Approach
Assume the regression equation is given by: 𝑦𝑡 = 𝛼 + 𝛽𝑥𝑡 + 𝑢𝑡

ො 𝛽መ and 𝑠𝑒 𝛼ො , 𝑠𝑒(𝛽)
(1) Estimate 𝛼, መ
𝐻0 : 𝛽 = 𝛽 ∗
(2) State the null and alternative hypothesis: ቊ
𝐻1 : 𝛽 ≠ 𝛽 ∗
and complete the test statistic:
෡
𝛽−𝛽 ∗
test statistic= ෡
𝑠𝑒(𝛽)

where 𝛽 ∗ is the value of 𝛽 under the null hypothesis

22. The Test of Significance Approach
(3) Choose a significance level 𝛼, usually 𝛼 = 5%
𝛼: Rejecting 𝐻0 by chance, probability of rejecting 𝐻0 when 𝐻0 is
true
(4) Obtain from the tabulated distribution a critical value. Compare the
estimated test statistic to the critical value
(5) Rejection and non-rejection regions can be determined
Decision rule: if the t-statistic lies in the rejection region, then reject
𝐻0 . Otherwise, do not reject 𝐻0 .
22. The Test of Significance Approach
𝐻0 : 𝛽 = 𝛽 ∗
Rejection region for a two-sided 5% (level of significance) test: ቊ
𝐻1 : 𝛽 ≠ 𝛽 ∗
𝛽መ

2.5% rejection region 2.5% rejection region

95% non-rejection region

- t-critical value 𝛽∗ = 0 t-critical value
22. The Test of Significance Approach
𝐻0 : 𝛽 = 𝛽 ∗
Rejection region for a one-sided hypothesis test: ቊ ∗
መ
𝐻1 : 𝛽 < 𝛽
𝛽

5% rejection region
95% non-rejection region
- t-critical value 𝛽∗ = 0
22. The Test of Significance Approach
𝐻0 : 𝛽 = 𝛽 ∗
Rejection region for a one-sided hypothesis test: ቊ
𝐻1 : 𝛽 > 𝛽 ∗
𝛽መ

5% rejection region
95% non-rejection region

𝛽∗ = 0 t-critical value
22. The Test of Significance Approach
𝛽መ − 𝛽∗
𝑡𝑒𝑠𝑡 𝑠𝑡𝑎𝑡𝑖𝑠𝑡𝑖𝑐 =
መ
𝑠𝑒(𝛽)
• If: (1) 𝑠𝑒(𝛽) መ , test statistic (we are confident in the estimate)
(2) 𝛽መ is far away from 𝛽∗ , test statistic
• Degrees of freedom: number of pieces of information (additional) beyond
minimum
• If two parameters are estimated, a minimum of two data points are needed
• If the degrees of freedom , critical value , one is more confident of the
results of the hypothesis test
• 𝛼 = significance level = size of the test
23. The Confidence Interval Approach to
Hypothesis Testing
• Say 𝛽መ = 0.93 and the 95% confidence interval for 𝛽 is (0.77, 1.09)
• This means that in many repeated samples, 95% of the time, the true
but unknown value of 𝛽 will be contained within this interval
• Confidence intervals are always for true but unknown population
parameters. Given a level of significance, we have a 1 − 𝛼 %
confidence interval. 1 − 𝛼 is the level of confidence.
𝛼 = 1% 99% CI
𝛼 = 5% 95% CI
𝛼 = 10% 90% CI
23. The Confidence Interval Approach to
Hypothesis Testing
• Confidence intervals and test of significance approaches always give
the same conclusion for a given 𝛼

𝐻0 : 𝛽 = 𝛽 ∗
• ∗ ቋ CIs are always used for two-sided hypotheses
𝐻1 : 𝛽 ≠ 𝛽

• The null hypothesis will not be rejected if the test-statistic lies in the
non-rejection region
23. The Confidence Interval Approach to
Hypothesis Testing
• The confidence interval contains the same information (but
rearranged) as the test of significance approach:
𝛽መ − 𝛽 ∗
−𝑡𝑐𝑟𝑖𝑡 ≤ ≤ +𝑡𝑐𝑟𝑖𝑡
𝑠𝑒 𝛽መ
𝛽መ − 𝑡𝑐𝑟𝑖𝑡 . 𝑠𝑒 𝛽መ ≤ 𝛽 ∗ ≤ 𝛽መ + 𝑡𝑐𝑟𝑖𝑡 . 𝑠𝑒(𝛽)
መ

• If the hypothesized value lies within the confidence interval, do not

reject 𝐻0
• If the hypothesized value lies outside the confidence interval, reject 𝐻0
24. Carrying out a hypothesis test using
confidence intervals
(1) Compute 𝛼, መ 𝑠𝑒 𝛼ො , 𝑠𝑒 𝛽መ
ො 𝛽,
(2) Choose 𝛼. This is significant to choosing a 1 − 𝛼 ⨯ 100% CI
(3) Find the appropriate critical value, which will again have T-2
degrees of freedom
(4) Construct the confidence interval as
𝛽መ − 𝑡𝑐𝑟𝑖𝑡 . 𝑠𝑒 𝛽መ ≤ 𝛽 ∗ ≤ 𝛽መ + 𝑡𝑐𝑟𝑖𝑡 . 𝑠𝑒(𝛽)
መ

(5) Perform the test

26. Hypothesis Testing: Example
Using the regression:
𝑦ො𝑡 = 20.3 + 0.5091𝑥𝑡
(14.38) (0.2561)

and 𝑇 = 22 , test the hypothesis that 𝛽 = 1 against a two-sided

alternative using both the test of significance and confidence interval
approaches
𝐻0 : 𝛽 = 1
ቊ
𝐻1 : 𝛽 ≠ 1
26. Hypothesis Testing: Example
Test of significance approach
෡
𝛽−𝛽 ∗ 0.5091−1
• Test-statistic = ෡ = = −1.917
𝑠𝑒(𝛽) 0.2561

• t-critical= 𝑡20,5% = ±2.086

Since test − statistic < 2.086

1.917 < 2.086
⇒ Do not reject H0
26. Hypothesis Testing: Example
Confidence interval approach
𝛽መ ± 𝑡𝑐𝑟𝑖𝑡 ⨯ 𝑠𝑒 𝛽መ = 0.5091 ±2.086⨯(0.2561)

95% CI = (−0.0251; 1.0433)

⇒ Do not Reject H0 since the hypothesized value 1 lies within the

interval
26. Hypothesis Testing: Example

2.5% rejection region 2.5% rejection region

95% non-rejection region

-2.086 2.086
26. Hypothesis Testing: Example
What if in the example before: 𝛼 = 10%?
• 𝑡20,10% = ±1.725
• Given that the test-statistic = 1.917, the null is rejected

Note:
• Test of significance approach → better if the size of test 𝛼 changes
• CI approach → useful for testing many hypothesis at a given 𝛼
26. The Exact Level of Significance Approach
• Commonly known as the p-value approach
• Gives the marginal significance level where one would be indifferent
between rejecting and not rejecting the null hypothesis
• If the test statistic is large in absolute value, the p-value will be small
and vice-versa
26. The Exact Level of Significance
Approach
• Informally, the p-value is also often referred to as the probability of
being wrong when the null hypothesis is rejected
• A p-value of 0.05 or less leads the researcher to reject the null
hypothesis
• The p-value is also termed the plausibility of the null hypothesis
• Decision rule: if p-value < 𝛼, reject H0
27. Terminology: Significance and t-ratio
• Reject 𝐻0 at 𝛼 = 5% → The result of the test or the estimate of the
parameter is statistically significant

• Reject 𝐻0 at 𝛼 = 1% → Highly statistically significant

• The t-ratio is the test statistic designed to test the null:

𝐻0 : 𝛽 = 0
𝐻1 : 𝛽 ≠ 0
෡
𝛽
• The t-ratio is ෡
𝑠𝑒(𝛽)
• The ratio of the coefficient to its standard error is known as the t-ratio
28. The Errors That we can make using
Hypothesis Tests
• We usually reject 𝐻0 if the test statistic is statistically significant at a
chosen significance level

• There are two possible errors we could make:

1. Rejecting 𝐻0 when it was really true: type I error
2. Not rejecting 𝐻0 when it was in fact false: type II error
28. The Errors That we can make using
Hypothesis Tests
Reality
𝐻0 is true 𝐻0 is false
Reject 𝐻0
Type I error = 𝛼 √
(significant)
Result of Test
Do not reject 𝐻0
√ Type II error = 𝛽
(insignificant)
28. The Errors That we can make using
Hypothesis Tests
• There is always a trade-off between type I and type II errors when
choosing a significance level.

Lower chance of type I error

• Reduce 𝛼
Higher chance of type II error
• One way to reduce both 𝛼 and 𝛽 is to increase the sample size
• Power of a test = 1 − 𝛽
29. Do Actively Managed Mutual Funds
“Beat the Market”? Jensen’s Alpha
• Testing for the presence and significance of abnormal returns
(“Jensen’s alpha” - Jensen, 1968).
• The Data:
Annual Returns on the portfolios of 115 mutual funds from 1945-1964
(T = 20, df = T-2 = 18).
• The model is: for j = 1, …, 115
• We are interested in the significance of j.
• The null hypothesis is H0: j = 0 . (Two sided: t0.025, 18 = 2.10, One
Sided: t0.05, 18 = 1.7341).
29. Do Actively Managed Mutual Funds
“Beat the Market”? Jensen’s Alpha
• Using α = 5%, Two sided: 115×0.025 = 2.87 then 2 to 3 funds
outperform or “beat the market” by pure chance! That is, because of
Type I error and a wrong decision we make about H0.
• Using α = 5%, One Sided: 115×0.05 = 5.75 then 5 to 6 funds “beat
the market” chance!
29. Jensen (1968)’s Findings




̂
29. Frequency Distribution of t-ratios of Mutual
Fund Alphas (gross of transactions costs)
29. Frequency Distribution of t-ratios of Mutual Fund
Alphas (net of transactions costs)
30. Workspace Application
• Obtaining alpha and beta for a mutual fund/ETF
• Obtaining other quantitative measures for a mutual fund/ETF