IE471

Models and Data

Woo Chang Kim

[email protected]

Industrial & Systems Engineering Department

KAIST

Woo Chang Kim, KAIST

IE471

Factor Models
SECTION 8.2

IE471(KAIST) / IE412(UNIST) AI for Finance 1

Woo Chang Kim, KAIST
IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Inputs of M-V model

• A major obstacle in implementing the mean-variance
theory: estimation of the parameter values (assume 𝒏𝒏
assets)
– Expected returns of assets (𝑛𝑛 parameters)
𝑛𝑛 𝑛𝑛+1
– Covariances of the returns of assets ( parameters)
2
– E.g. a universe of 1,000 stocks => 501,500 parameters

• Therefore, we need to reduce the dimension

– From individual assets to factors

Woo Chang Kim, KAIST 2

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Factor models

• Clearly, it will help to have a simplified approach

• Fortunately, the randomness displayed by the returns of 𝑛𝑛
stocks often can be tracked back to a smaller number of
underlying basic sources of randomness that influence
the individual returns ➔ Factors
• A factor model that represents this connection between
factors and individual returns leads to a simplified
structure for the covariances among assets
• Factors used to explain randomness must be chosen
carefully, and the proper choice depends on the universe of
assets being considered
• For common stocks, the factors might be the size of a
company, gross domestic product (GDP), unemployment
rate, and so forth

Woo Chang Kim, KAIST 3

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Single-factor model
• Single-factor models are the simplest of the factor models

– There is a single factor 𝑓𝑓 which is a random quantity

– We assume that some value 𝑦𝑦 and the factor are related by

𝑦𝑦 = 𝑎𝑎 + 𝑏𝑏𝑏𝑏

Constants
– The 𝑏𝑏’s are termed factor loadings because they measure the
sensitivity of the value 𝑦𝑦 to the factor

Woo Chang Kim, KAIST 4

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Single-factor model

• Suppose there are 𝑛𝑛 stocks, with rates of return 𝑟𝑟𝑖𝑖 , 𝑖𝑖 =

1,2, … , 𝑛𝑛

– We assume that the return of a stock 𝑖𝑖 and the factor are related by

𝑟𝑟𝑖𝑖 = 𝑎𝑎𝑖𝑖 + 𝑏𝑏𝑖𝑖 𝑓𝑓 + 𝑒𝑒𝑖𝑖

Constants Error
(with E 𝑒𝑒𝑖𝑖 = 0 and uncorrelated with 𝑓𝑓)

Woo Chang Kim, KAIST 5

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Single-factor model

• Single-factor model can be viewed graphically as defining a

linear fit
– Imagine that several independent observations are made of both the
return 𝑟𝑟𝑖𝑖 and the factor 𝑓𝑓
– Since both are random quantities, the points are likely to be
scattered
– A straight line is fitted through these points

Woo Chang Kim, KAIST 6

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Single-factor model

• If we know the model (the straight line), we can predict the data
points will fall near the line
• If we first obtain data points, we can then construct the line that
fits the data, and then predict that new data points will fall near
the line
– If a historical record of stock returns and the factor values are
available, the parameter of a single-factor model can be estimated by
actually fitting straight lines

Woo Chang Kim, KAIST 7

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Single-factor model

• If we use a single-factor model, then the standard

parameters for the mean-variance model can be
determined using the factor

𝑟𝑟𝑖𝑖 = 𝑎𝑎𝑖𝑖 + 𝑏𝑏𝑖𝑖 𝑓𝑓 + 𝑒𝑒𝑖𝑖

E 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 𝑖𝑖 → 𝑟𝑟𝑖𝑖̅ = 𝑎𝑎𝑖𝑖 + 𝑏𝑏𝑖𝑖 𝑓𝑓 ̅

Need 𝑎𝑎𝑖𝑖 , 𝑏𝑏𝑖𝑖 , 𝜎𝜎𝑒𝑒2𝑖𝑖
var 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 𝑖𝑖 → 𝜎𝜎𝑖𝑖2 = 𝑏𝑏𝑖𝑖2 𝜎𝜎𝑓𝑓2 + 𝜎𝜎𝑒𝑒2𝑖𝑖 for each asset
plus 𝑓𝑓 ̅ and 𝜎𝜎𝑓𝑓2
cov 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 𝑖𝑖, 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 𝑗𝑗 → 𝜎𝜎𝑖𝑖𝑖𝑖 = 𝑏𝑏𝑖𝑖 𝑏𝑏𝑗𝑗 𝜎𝜎𝑓𝑓2

Woo Chang Kim, KAIST 8

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Single-factor model

slope 𝑏𝑏1
𝑟𝑟
asset 1 𝑒𝑒1
𝑎𝑎1

𝑓𝑓
asset 2 𝑎𝑎2

𝑒𝑒2
slope 𝑏𝑏2

Woo Chang Kim, KAIST 9

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

MV parameters from single factor model

𝑟𝑟𝑖𝑖 = 𝛼𝛼𝑖𝑖 + 𝛽𝛽𝑖𝑖 𝐹𝐹 + 𝜀𝜀𝑖𝑖 , 𝑖𝑖 = 1, … , 𝑛𝑛

• Parameters for mean-variance analysis

– Expected return: E[𝑟𝑟𝑖𝑖 ] = 𝛼𝛼𝑖𝑖 + 𝛽𝛽𝑖𝑖 𝐸𝐸[𝐹𝐹]
– Variance: var 𝑟𝑟𝑖𝑖 = 𝛽𝛽𝑖𝑖2 var 𝐹𝐹 + var[𝜀𝜀𝑖𝑖 ]
– Covariance: cov 𝑟𝑟𝑖𝑖 , 𝑟𝑟𝑗𝑗 = 𝛽𝛽𝑖𝑖 𝛽𝛽𝑗𝑗 var 𝐹𝐹

– Therefore, for 𝑛𝑛 assets, we need to estimate

• 𝛼𝛼𝑖𝑖 for 𝑖𝑖 = 1, … , 𝑛𝑛
• 𝛽𝛽𝑖𝑖 for 𝑖𝑖 = 1, … , 𝑛𝑛
• var[𝜀𝜀𝑖𝑖 ] for 𝑖𝑖 = 1, … , 𝑛𝑛
• 𝐸𝐸[𝐹𝐹] and var 𝐹𝐹

– Now, we need only 3𝑛𝑛 + 2 parameters

𝑛𝑛 𝑛𝑛+3
• cf. parameters when estimated individually
2

Woo Chang Kim, KAIST 10

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Portfolios from single factor model

𝑟𝑟𝑖𝑖 = 𝛼𝛼𝑖𝑖 + 𝛽𝛽𝑖𝑖 𝐹𝐹 + 𝜀𝜀𝑖𝑖 , 𝑖𝑖 = 1, … , 𝑛𝑛

• Consider the following portfolio 𝒑𝒑:

𝑛𝑛
𝑟𝑟𝑝𝑝 = Σ𝑖𝑖=1 𝑤𝑤𝑖𝑖 𝑟𝑟𝑖𝑖
𝑛𝑛 𝑛𝑛 𝑛𝑛
= Σ𝑖𝑖=1 𝑤𝑤𝑖𝑖 𝛼𝛼𝑖𝑖 + Σ𝑖𝑖=1 𝑤𝑤𝑖𝑖 𝛽𝛽𝑖𝑖 𝐹𝐹 + Σ𝑖𝑖=1 𝑤𝑤𝑖𝑖 𝜀𝜀𝑖𝑖
𝑛𝑛 𝑛𝑛 𝑛𝑛
– Let 𝛼𝛼𝑝𝑝 = Σ𝑖𝑖=1 𝑤𝑤𝑖𝑖 𝛼𝛼𝑖𝑖 , 𝛽𝛽𝑝𝑝 = Σ𝑖𝑖=1 𝑤𝑤𝑖𝑖 𝛽𝛽𝑖𝑖 , and 𝜀𝜀𝑝𝑝 = Σ𝑖𝑖=1 𝑤𝑤𝑖𝑖 𝜀𝜀𝑖𝑖
𝑛𝑛
• Note that E 𝜀𝜀𝑝𝑝 = 0 and var 𝜀𝜀𝑝𝑝 = Σ𝑖𝑖=1 𝑤𝑤𝑖𝑖2 var[𝜀𝜀𝑖𝑖 ]
– Then,
• Portfolio return: 𝑟𝑟𝑝𝑝 = 𝛼𝛼𝑝𝑝 + 𝛽𝛽𝑝𝑝 𝐹𝐹 + 𝜀𝜀𝑝𝑝
• Portfolio expected return: E[𝑟𝑟𝑝𝑝 ] = 𝛼𝛼𝑝𝑝 + 𝛽𝛽𝑝𝑝 𝐸𝐸[𝐹𝐹]
• Portfolio variance: var 𝑟𝑟𝑝𝑝 = 𝛽𝛽𝑝𝑝2 var 𝐹𝐹 + var 𝜀𝜀𝑝𝑝

Woo Chang Kim, KAIST 11

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Single-factor model for stocks

• Example: Four stocks and one index
– Assume a single-factor model using the index as the factor

Woo Chang Kim, KAIST 12

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Single-factor model for stocks

• Example: Four stocks and one index (cont’d)
– Let us first compute the average (expected) return of the stocks and
index from the sample
10
1
𝑟𝑟̂𝑖𝑖̅ = � 𝑟𝑟𝑖𝑖𝑘𝑘
10
𝑘𝑘=1

Woo Chang Kim, KAIST 13

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Single-factor model for stocks

• Example: Four stocks and one index (cont’d)
– Then, variance
10
1 2
var(𝑟𝑟𝑖𝑖 ) = � 𝑟𝑟𝑖𝑖𝑘𝑘 − 𝑟𝑟̂𝑖𝑖̅
9
𝑘𝑘=1

– For each stock, also compute the covariance between the index
10
1
cov(𝑟𝑟𝑖𝑖 , 𝑓𝑓) = � 𝑟𝑟𝑖𝑖𝑘𝑘 − 𝑟𝑟̂𝑖𝑖̅ 𝑓𝑓 𝑘𝑘 − 𝑓𝑓̂̅
9
𝑘𝑘=1

Woo Chang Kim, KAIST 14

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Single-factor model for stocks

• Example: Four stocks and one index (cont’d)

Woo Chang Kim, KAIST 15

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Single-factor model for stocks

• Example: Four stocks and one index (cont’d)
– Then, find 𝑎𝑎𝑖𝑖 and 𝑏𝑏𝑖𝑖 for the single-factor model for each stock

– Once the covariances are estimated, we find the values of 𝑏𝑏𝑖𝑖 and 𝑎𝑎𝑖𝑖
from the formulas

cov 𝑟𝑟𝑖𝑖 , 𝑓𝑓
𝑏𝑏𝑖𝑖 =
var 𝑓𝑓
�
�� 𝑖𝑖 − 𝑓𝑓�
𝑎𝑎𝑖𝑖 = 𝑟𝑟

Woo Chang Kim, KAIST 16

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Single-factor model for stocks

• Example: Four stocks and one index (cont’d)

𝑎𝑎 + 𝑏𝑏𝑓𝑓 ̅ 14.96 14.32 10.91 15.08

Woo Chang Kim, KAIST 17

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Single-factor model for stocks

• Example: Four stocks and one index (cont’d)
– We can also estimate the variance of the error
var 𝑒𝑒𝑖𝑖 = var 𝑟𝑟𝑖𝑖 − 𝑏𝑏𝑖𝑖2 var(𝑓𝑓)

Woo Chang Kim, KAIST 18

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

CAPM as a single factor model

• Let us look at a single-factor model for stock returns
𝑟𝑟𝑖𝑖 = 𝑎𝑎𝑖𝑖 + 𝑏𝑏𝑖𝑖 𝑓𝑓 + 𝑒𝑒𝑖𝑖
• Assume that the single factor is the market rate of return 𝑟𝑟𝑀𝑀
𝑟𝑟𝑖𝑖 = 𝑎𝑎𝑖𝑖 + 𝑏𝑏𝑖𝑖 𝑟𝑟𝑀𝑀 + 𝑒𝑒𝑖𝑖
• For convenience, we can subtract the constant 𝑟𝑟𝑓𝑓 from the
two returns
𝑟𝑟𝑖𝑖 − 𝑟𝑟𝑓𝑓 = 𝑎𝑎𝑖𝑖 + 𝑏𝑏𝑖𝑖 𝑟𝑟𝑀𝑀 − 𝑟𝑟𝑓𝑓 + 𝑒𝑒𝑖𝑖
or express it as
𝑟𝑟𝑖𝑖 − 𝑟𝑟𝑓𝑓 = 𝛼𝛼𝑖𝑖 + 𝛽𝛽𝑖𝑖 𝑟𝑟𝑀𝑀 − 𝑟𝑟𝑓𝑓 + 𝑒𝑒𝑖𝑖

Woo Chang Kim, KAIST 19

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

CAPM as a single factor model

• So the single-factor model is
𝑟𝑟𝑖𝑖 − 𝑟𝑟𝑓𝑓 = 𝛼𝛼𝑖𝑖 + 𝛽𝛽𝑖𝑖 𝑟𝑟𝑀𝑀 − 𝑟𝑟𝑓𝑓 + 𝑒𝑒𝑖𝑖

• Then, take the expectation of both sides

𝐸𝐸 𝑟𝑟𝑖𝑖 − 𝑟𝑟𝑓𝑓 = 𝐸𝐸 𝛼𝛼𝑖𝑖 + 𝛽𝛽𝑖𝑖 𝑟𝑟𝑀𝑀 − 𝑟𝑟𝑓𝑓 + 𝑒𝑒𝑖𝑖

𝐸𝐸 𝑟𝑟𝑖𝑖 − 𝐸𝐸 𝑟𝑟𝑓𝑓 = 𝐸𝐸 𝛼𝛼𝑖𝑖 + 𝐸𝐸 𝛽𝛽𝑖𝑖 𝑟𝑟𝑀𝑀 − 𝑟𝑟𝑓𝑓 + 𝐸𝐸 𝑒𝑒𝑖𝑖

𝐸𝐸 𝑟𝑟𝑖𝑖 − 𝑟𝑟𝑓𝑓 = 𝛼𝛼𝑖𝑖 + 𝛽𝛽𝑖𝑖 𝐸𝐸 𝑟𝑟𝑀𝑀 − 𝑟𝑟𝑓𝑓 + 0

𝑟𝑟𝑖𝑖̅ − 𝑟𝑟𝑓𝑓 = 𝛼𝛼𝑖𝑖 + 𝛽𝛽𝑖𝑖 𝑟𝑟𝑀𝑀

̅ − 𝑟𝑟𝑓𝑓

Similar to CAPM, except CAPM predicts that 𝛼𝛼𝑖𝑖 = 0

Woo Chang Kim, KAIST 20

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

CAPM as a single factor model

• So the single-factor model is

𝑟𝑟𝑖𝑖 − 𝑟𝑟𝑓𝑓 = 𝛼𝛼𝑖𝑖 + 𝛽𝛽𝑖𝑖 𝑟𝑟𝑀𝑀 − 𝑟𝑟𝑓𝑓 + 𝑒𝑒𝑖𝑖

• Next, take the covariance of both sides with 𝑟𝑟𝑀𝑀

𝑐𝑐𝑐𝑐𝑐𝑐 𝑟𝑟𝑖𝑖 − 𝑟𝑟𝑓𝑓 , 𝑟𝑟𝑀𝑀 = 𝑐𝑐𝑐𝑐𝑐𝑐 𝛼𝛼𝑖𝑖 + 𝛽𝛽𝑖𝑖 𝑟𝑟𝑀𝑀 − 𝑟𝑟𝑓𝑓 + 𝑒𝑒𝑖𝑖 , 𝑟𝑟𝑀𝑀

𝑐𝑐𝑐𝑐𝑐𝑐 𝑟𝑟𝑖𝑖 , 𝑟𝑟𝑀𝑀 = 𝑐𝑐𝑐𝑐𝑐𝑐 𝛽𝛽𝑖𝑖 𝑟𝑟𝑀𝑀 − 𝑟𝑟𝑓𝑓 , 𝑟𝑟𝑀𝑀

𝑐𝑐𝑐𝑐𝑐𝑐 𝑟𝑟𝑖𝑖 , 𝑟𝑟𝑀𝑀 = 𝛽𝛽𝑖𝑖 × 𝑐𝑐𝑐𝑐𝑐𝑐 𝑟𝑟𝑀𝑀 − 𝑟𝑟𝑓𝑓 , 𝑟𝑟𝑀𝑀

𝑐𝑐𝑐𝑐𝑐𝑐 𝑟𝑟𝑖𝑖 , 𝑟𝑟𝑀𝑀 = 𝛽𝛽𝑖𝑖 × 𝑐𝑐𝑐𝑐𝑐𝑐 𝑟𝑟𝑀𝑀 , 𝑟𝑟𝑀𝑀

𝑐𝑐𝑐𝑐𝑐𝑐 𝑟𝑟𝑖𝑖 , 𝑟𝑟𝑀𝑀 𝑐𝑐𝑐𝑐𝑐𝑐 𝑟𝑟𝑖𝑖 , 𝑟𝑟𝑀𝑀

= 𝛽𝛽𝑖𝑖 𝛽𝛽𝑖𝑖 =
𝑐𝑐𝑐𝑐𝑐𝑐 𝑟𝑟𝑀𝑀 , 𝑟𝑟𝑀𝑀 𝑣𝑣𝑣𝑣𝑣𝑣 𝑟𝑟𝑀𝑀

Expression for 𝛽𝛽𝑖𝑖 is exactly the same as the 𝛽𝛽𝑖𝑖 in CAPM

Woo Chang Kim, KAIST 21

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

CAPM as a single factor model

𝑟𝑟𝑖𝑖 − 𝑟𝑟𝑓𝑓 = 𝛼𝛼𝑖𝑖 + 𝛽𝛽𝑖𝑖 (𝒓𝒓𝑴𝑴 − 𝒓𝒓𝒇𝒇 ) + 𝜀𝜀𝑖𝑖 , 𝑖𝑖 = 1, … , 𝑛𝑛

• The market portfolio as a factor

– If 𝛼𝛼𝑖𝑖 = 0, the CAPM holds
– Otherwise, the asset is mispriced under the CAPM and 𝛼𝛼𝑖𝑖 represents
the amount of mispricing

• Note that CAPM is not derived as a factor model

– Factor models start by assuming that asset returns have linear
relationship with factors
– However, the CAPM has no assumption on the dynamics of asset
returns
• Instead, it just assumes that investors are mean-variance investors, the
market portfolio is efficient, etc.

Woo Chang Kim, KAIST 22

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Multifactor model

• Proceeding development can be extended to more than one

factor

• Consider a model with two factors 𝑓𝑓1 and 𝑓𝑓2

𝑟𝑟𝑖𝑖 = 𝑎𝑎𝑖𝑖 + 𝑏𝑏1,𝑖𝑖 𝒇𝒇𝟏𝟏 + 𝑏𝑏2,𝑖𝑖 𝒇𝒇𝟐𝟐 + 𝑒𝑒𝑖𝑖
– Constant 𝑎𝑎𝑖𝑖 is the intercept, and 𝑏𝑏1,𝑖𝑖 and 𝑏𝑏2,𝑖𝑖 are factor loadings

• When there are 𝒏𝒏 factors,

𝑟𝑟𝑖𝑖 = 𝑎𝑎𝑖𝑖 + 𝑏𝑏1,𝑖𝑖 𝒇𝒇𝟏𝟏 + 𝑏𝑏2,𝑖𝑖 𝒇𝒇𝟐𝟐 + ⋯ + 𝑏𝑏𝑛𝑛,𝑖𝑖 𝒇𝒇𝒏𝒏 + 𝑒𝑒𝑖𝑖

Woo Chang Kim, KAIST 23

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

MV parameters from multi-factor model

• Consider a two-factor model

𝑟𝑟𝑖𝑖 = 𝛼𝛼𝑖𝑖 + 𝛽𝛽1𝑖𝑖 𝑓𝑓1 + 𝛽𝛽2𝑖𝑖 𝑓𝑓2 + 𝜀𝜀𝑖𝑖 , 𝑖𝑖 = 1, … , 𝑛𝑛
• Parameters for mean-variance analysis
– Expected return: E 𝑟𝑟𝑖𝑖 = 𝛼𝛼𝑖𝑖 + 𝛽𝛽1𝑖𝑖 𝐸𝐸 𝑓𝑓1 + 𝛽𝛽2𝑖𝑖 𝐸𝐸[𝑓𝑓2 ]
2 2
– Variance: var 𝑟𝑟𝑖𝑖 = 𝛽𝛽1𝑖𝑖 var 𝑓𝑓1 + 𝟐𝟐𝜷𝜷𝟏𝟏𝟏𝟏 𝜷𝜷𝟐𝟐𝒊𝒊 𝐜𝐜𝐜𝐜𝐜𝐜 𝒇𝒇𝟏𝟏 , 𝒇𝒇𝟐𝟐 + 𝛽𝛽2𝑖𝑖 var 𝑓𝑓2 + var[𝜀𝜀𝑖𝑖 ]
– Covariance: cov 𝑟𝑟𝑖𝑖 , 𝑟𝑟𝑗𝑗 = 𝛽𝛽1𝑖𝑖 𝛽𝛽1𝑗𝑗 var 𝑓𝑓1 + (𝜷𝜷𝟏𝟏𝟏𝟏 𝜷𝜷𝟐𝟐𝒋𝒋 + 𝜷𝜷𝟏𝟏𝒋𝒋 𝜷𝜷𝟐𝟐𝒊𝒊 )𝐜𝐜𝐜𝐜𝐜𝐜 𝒇𝒇𝟏𝟏 , 𝒇𝒇𝟐𝟐 +
𝛽𝛽2𝑖𝑖 𝛽𝛽2𝑗𝑗 var 𝑓𝑓2
• Note that
– cov 𝑟𝑟𝑖𝑖 , 𝑓𝑓1 = 𝛽𝛽1𝑖𝑖 var 𝑓𝑓1 + 𝜷𝜷𝟐𝟐𝒊𝒊 𝐜𝐜𝐜𝐜𝐜𝐜[𝒇𝒇𝟏𝟏 , 𝒇𝒇𝟐𝟐 ]
– cov 𝑟𝑟𝑖𝑖 , 𝑓𝑓2 = 𝛽𝛽2𝑖𝑖 var 𝑓𝑓2 + 𝜷𝜷𝟏𝟏𝒊𝒊 𝒄𝒄𝒄𝒄𝒄𝒄[𝒇𝒇𝟏𝟏 , 𝒇𝒇𝟐𝟐 ]
– Therefore, for 𝑛𝑛 assets, we need to estimate
• 𝛼𝛼𝑖𝑖 for 𝑖𝑖 = 1, … , 𝑛𝑛
• 𝛽𝛽1𝑖𝑖 for 𝑖𝑖 = 1, … , 𝑛𝑛
• 𝛽𝛽2𝑖𝑖 for 𝑖𝑖 = 1, … , 𝑛𝑛
• var[𝜀𝜀𝑖𝑖 ] for 𝑖𝑖 = 1, … , 𝑛𝑛
• 𝐸𝐸[𝑓𝑓1 ], 𝐸𝐸[𝑓𝑓2 ], var 𝑓𝑓1 , var 𝑓𝑓2 , and cov 𝑓𝑓1 , 𝑓𝑓2

– Now, we need only 4𝑛𝑛 + 5 parameters

𝑛𝑛 𝑛𝑛+3
• cf. parameters when estimated individually
2

Woo Chang Kim, KAIST 24

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Fama-French three factor model

𝑟𝑟𝑖𝑖 − 𝑟𝑟𝑓𝑓 = 𝛼𝛼 + 𝛽𝛽𝑖𝑖 𝑟𝑟𝑀𝑀 − 𝑟𝑟𝑓𝑓 + 𝛽𝛽𝑠𝑠𝑖𝑖 � 𝒇𝒇𝑺𝑺𝑺𝑺𝑺𝑺 + 𝛽𝛽𝑣𝑣𝑖𝑖 � 𝒇𝒇𝑯𝑯𝑯𝑯𝑯𝑯 + 𝜀𝜀𝑖𝑖

• Fama & French (1992) constructed a three-factor model

with two additional factors
– SMB (small minus big): size factor
• Small cap stocks – large cap stocks
• Basu (1981), Reinganum (1981), Rosenberg et al. (1985): small stocks tend
to perform better than large stocks
• If 𝛽𝛽𝑠𝑠𝑖𝑖 is positive, then asset 𝑖𝑖 is close to small cap stocks
• If 𝛽𝛽𝑠𝑠𝑠𝑠 is negative, then asset 𝑖𝑖 is close to large cap stocks
– HML (high minus low): value factor
• High book-to-market stocks – low book-to-market stocks
• Graham (1949): one can find undervalued stocks by looking at P/E ratio or
price-to-book ratio
• If 𝛽𝛽𝑣𝑣𝑖𝑖 is positive, then asset 𝑖𝑖 is close to value stocks
• If 𝛽𝛽𝑣𝑣𝑖𝑖 is negative, then asset 𝑖𝑖 is close to growth stocks

Woo Chang Kim, KAIST 25

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Carhart four-factor model

𝑟𝑟𝑖𝑖 − 𝑟𝑟𝑓𝑓 = 𝛼𝛼 + 𝛽𝛽𝑖𝑖 𝑟𝑟𝑀𝑀 − 𝑟𝑟𝑓𝑓 + 𝛽𝛽𝑠𝑠𝑠𝑠 � 𝑓𝑓𝑆𝑆𝑆𝑆𝑆𝑆 + 𝛽𝛽𝑣𝑣𝑣𝑣 � 𝑓𝑓𝐻𝐻𝐻𝐻𝐻𝐻 + 𝛽𝛽𝑚𝑚𝑖𝑖 � 𝒇𝒇𝑴𝑴𝑴𝑴𝑴𝑴 + 𝜀𝜀𝑖𝑖
• Carhart (1997) added one more factor to Fama-French three
factor model
– MOM: momentum factor
• Positive momentum stocks – negative momentum stocks
• Jegadeesh & Titman (1993) found the tendency for the stock
price to continue rising if it is going up and to continue declining
if it is going down
• If 𝛽𝛽𝑚𝑚𝑖𝑖 is positive, then asset 𝑖𝑖 has positive momentum
• If 𝛽𝛽𝑚𝑚𝑖𝑖 is negative, then asset 𝑖𝑖 has negative momentum

Woo Chang Kim, KAIST 26

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Selection of factors
• Factors should account for some common variability of a
large number of stocks
– External factors: gross national product (GNP), consumer price index
(CPI), unemployment rate, interest rates, etc.
– Extracted factors: market portfolio, industry average, size factor,
value factor, momentum factor, etc.

• How many factors?

– Obviously, a large number of factors can fit data better
– However, if there exist factors that are highly correlated each other,
the model is inefficiently constructed
– Also, if the number of factors is close to the number of assets, there is
no dimension reduction effect
– Generally agreed that for models of U.S. stocks, it is appropriate to
use between 3 to 15 factors

Woo Chang Kim, KAIST 27

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Selection of factors

• Finding factors through dimensionality reduction

– As for factor models, we first have factors (like macroeconomic
variables and market anomalies), then we try to see if these factors
explain the variability of stock returns well or not

– However, some may search for factors that explain the variability of
stock returns with smallest error using dimensionality reduction
techniques in machine learning
• E.g. principal component analysis or autoencoder

– Problem?
• Usually, such factors cannot be intuitively explained
• Therefore, we may not be able to replicate such factors for future
analysis

Woo Chang Kim, KAIST 28

IE471 IE471(KAIST) / IE412(UNIST) AI for Finance

Summary of Chapter 8
• Single-factor models

• Multifactor models

Woo Chang Kim, KAIST 29