Risk Management

Lecture 4-2: Multi-Factor Asset Pricing Model

Chen Tong

SOE & WISE, Xiamen University

October 22, 2024

Chen Tong (SOE&WISE) Risk Management October 22, 2024 1 / 52

Multi-factor explanation for the failure of the CAPM

▶ Merton (1973) provides an inter-temporal capital asset pricing model

(ICAPM), which is an extension of the traditional CAPM, which has
only one factor, the market, into multiple factors.

▶ In the ICAPM, investors are concerned not only with their

end-of-period payoff, but also with the opportunities about how to
consume or invest these payoff in the future.

Chen Tong (SOE&WISE) Risk Management October 22, 2024 2 / 52

ICAPM

▶ When choosing a portfolio today, they consider not only how their
wealth might vary with future state variables tomorrow such as labor
income, the prices of consumption goods, and future stock expected
return, but also expectations about the labor income, consumption,
and investment opportunities available after tomorrow.

Chen Tong (SOE&WISE) Risk Management October 22, 2024 3 / 52

ICAPM

▶ As a result, there are extra-market sources of risk, and the expected

return should be related to all of them.

▶ Therefore, we have the following extension of the CAPM:

S
Erj = rf + βj [E (rm ) − rf ] + ∑ βsj [E (rs ) − rf ]
s=1

where
S = the total number of state variables in the economy;
rs = the return on state s mimicking portfolio (i.e., the portfolio that
has the maximum correlation with state);
βsj = asset j’s beta related to rs

Chen Tong (SOE&WISE) Risk Management October 22, 2024 4 / 52

CAPM and ICAPM

▶ In contrast to the standard CAPM, the expected return on asset j

now has an extra risk premium, which is the sum of its risk
premiums on its exposure to all the state risks. The states S serve
as S additional factors, beyond the market, in determining the
equilibrium asset’s expected return.

Chen Tong (SOE&WISE) Risk Management October 22, 2024 5 / 52

Arbitrage Pricing Theory (APT)

▶ The APT model of Ross(1976) assumes that we can have a general

K -factor model that determine stock returns:

rit − rft = αi + βi1 f1t + ⋯ + βiK fKt + εit

where fjt is the realization of the j-th factor in period t and εit is the
disturbances or random errors.

▶ The K -factor are systematic since they affect all the stock returns.
The εit can be regarded as firm-specific factors or idiosyncratic risks.

Chen Tong (SOE&WISE) Risk Management October 22, 2024 6 / 52

Arbitrage Pricing Theory (APT)

▶ Then, under the assumption of no arbitrage, we have that:

E [˜
ri ] = rf + βi1 λ1 + ⋯ + βiK λK

where βik is the beta or risk exposure on the k-th factor, and λk is
the k-th factor risk premium.

▶ The CAPM can then viewed as a special case of the APT with
K = 1, f1,t = rm,t − rf , and λ1 = E (rm ) − rf .

Chen Tong (SOE&WISE) Risk Management October 22, 2024 7 / 52

Interpretation of factor risk premium

▶ If one asset has only a unit beta risk on the first factor, that is,
βi1 = 1 and zero betas on all other factors, then its expected return
will be
E [ri ] = rf + λ1
This is, the risk-free rate increased by an extra return to compensate
for taking the factor risk on f1 . This is why λ1 is called the factor
risk premium on f1 , or the extra return one earns by taking one unit
beta risk on the factor. Interpretations for other λs follow the same
way.

Chen Tong (SOE&WISE) Risk Management October 22, 2024 8 / 52

Differences between CAPM and APT (1)

▶ The CAPM is based on mean-variance theory and assumes that only

means and variances (whose calculations include covariances) matter
in portfolio choice. This requires restrictions on either the statistical
distribution of the asset returns, such as normality, or on the form of
the utility function, such as the quadratic form.

▶ In contrast, the APT does not impose as strong distributional

assumption as the CAPM. It assumes that only the asset returns are
affected by a few factors. Moreover, it does not impose any
restrictions on the form of utility functions except the trivial
assumption that investors prefer more to less.

Chen Tong (SOE&WISE) Risk Management October 22, 2024 9 / 52

Differences between CAPM and APT (2)

▶ In the CAPM, all investors are informed about the true means,
variances, and covariances of the asset returns, and they all use
Markowitz portfolio theory to make their optimal investment
decisions. As a result, all investors hold the same market portfolio of
risky assets and differ only, depending on their risk aversions.

▶ In contrast, the APT assumes that it is sufficient for only some

investors to be able to take advantage of arbitrage opportunities. If
the assets are mispriced so that the expected returns of these assets
deviate from what is accounted for by their beta risk, smart investors
can construct an arbitrage portfolio to make abnormal returns.

Chen Tong (SOE&WISE) Risk Management October 22, 2024 10 / 52

Statistical model & theoretical model

▶ Technically, the APT assumes a K -factor model. That is, the asset
returns are influenced by K factors in the economy via linear
regression equations,

r˜it − rft = αi + βi1 f˜1t + ⋯ + βiK f˜Kt + ε̃it

▶ Theoretically, under the assumption of no arbitrage, the asset pricing

relation of the APT as given by following equation must be true.

E [˜
ri ] = rf + βi1 λ1 + ⋯ + βiK λK

Chen Tong (SOE&WISE) Risk Management October 22, 2024 11 / 52

APT: one-factor model
▶ Consider a one-factor model in which there are two assets and
suppose the factor model has no errors:

r˜1 = µ1 + 0.8f˜
r˜2 = µ2 + 1.6f˜
where E (f˜) = 0.

▶ Consider one portfolio (which we call portfolio z) given by:

z̃ = 2r̃ 1 − r̃ 2 = 2µ1 − µ2

▶ Under the non-arbitrage condition, we must have

2µ1 − µ2 = rf

Chen Tong (SOE&WISE) Risk Management October 22, 2024 12 / 52

APT: one-factor model (cont.)

▶ which can be rewritten as

µ1 − rf µ2 − rf
λ1 = =
0.8 1.6
so we have
µ1 = rf + 0.8λ1
µ2 = rf + 1.6λ1

▶ Similarly, the same beta pricing relation holds for any asset; that is,

µi = rf + βi λ1

Chen Tong (SOE&WISE) Risk Management October 22, 2024 13 / 52

APT: one-factor model

▶ Consider again the previous one-factor example. For any portfolio

weights w1 and w2 ,
w1 + w2 = 1
that satisfy the condition of no risk:

w 1 β1 + w 2 β2 = 0

and the portfolio:z̃ = w1 r˜1 + w2 r˜2 will be riskless. Hence, it has the
risk-free rate of return

w1 µ1 + w2 µ2 = rf

Using this portfolio condition, we get

w1 (µ1 − rf ) + w2 (µ2 − rf ) = 0

Chen Tong (SOE&WISE) Risk Management October 22, 2024 14 / 52

APT: K-factor model

▶ To see why the beta-pricing relation remains true in the N assets

case, consider now the K -factor model:

r˜it − rft = µi + βi1 f˜1t + ⋯ + βiK f˜Kt + ε̃it

▶ consider the portfolio:

N
r˜p = ∑ wi r˜i
i=1
N N N N
= ∑ wi µi + (∑ wi βi1 ) f˜1t + ⋯ (∑ wi βiK ) f˜Kt + ∑ wi ε̃it
i=1 i=1 i=1 i=1

Chen Tong (SOE&WISE) Risk Management October 22, 2024 15 / 52

 ▶ Suppose the portfolio weights are orthogonal to all betas,

N
∑ wi βik = w1 β1k + w2 β2k + ⋯ + wN βNk = 0, k = 1, 2, . . . , K
i=1

▶ Then the portfolio has only unsystematic risk:

N N
r˜p = ∑ wi µi + ∑ wi ε̃it
i=1 i=1

Chen Tong (SOE&WISE) Risk Management October 22, 2024 16 / 52

 ▶ Under certain conditions, this risk will be approximately zero. To see
why, assume all the unsystematic risks are independent across assets
2
and they have the same variance σu . Then the variance of r˜p is:
2 2 2 2 2
σ (˜
rp ) = (w1 + w2 + ⋯ + wN ) σu

When there are a large number of assets, the weights can be roughly
of the same magnitude of 1/N, and hence σ (˜
2
rp ) should be of
magnitude of:
2 2 2
(1/N) × N × σu = σu /N
which approaches zero as N approaches infinity. This implies that r˜p
is almost risk free. Hence, in the absence of arbitrage, the return on
r˜p is approximately rf , or

w1 (µ1 − rf ) + w2 (µ2 − rf ) + ⋯ + wK (µK − rf ) ≈ 0

Chen Tong (SOE&WISE) Risk Management October 22, 2024 17 / 52

 ▶ In terms of linear algebra, equations above says that
- the portfolio vector is orthogonal to the beta vector [that is, their
inner product is zero].
- any vector orthogonal to the beta vector will also be orthogonal to
the expected excess return vector.
- This can only be true if the expected excess return vector can be
generated by the betas, that is, a linear function of the betas.

▶ So, we have the expression of APT:

E [˜
ri ] = µi = rf + βi1 λ1 + ⋯ + βiK λK

Chen Tong (SOE&WISE) Risk Management October 22, 2024 18 / 52

Estimation of factor models ( Known Factors)

▶ The simplest case of factor models is where the K factors are

assumed known or observable, so that we have time-series data on
them. In this case, the K -factor model for the return-generating
process,
r˜it − rft = αi + βi1 f˜1t + ⋯ + βiK f˜Kt + ε̃it
is a multiple regression for each asset, and is a multivariate
regression if all of the individual regressions are pooled together.

Chen Tong (SOE&WISE) Risk Management October 22, 2024 19 / 52

 ▶ For example, if one believes that the gross domestic product (GDP)
is the driving force for a group of stock returns, one would have a
one-factor model,

r˜it − rft = αi + βi1 G̃ DPt + ε̃it

one can run time-series regressions to obtain all the parameters. (it
seems like the first-stage of fama-macbeth regression).

▶ What if we run the second-stage of fama-macbeth regression?

Chen Tong (SOE&WISE) Risk Management October 22, 2024 20 / 52

Review of fama-macbeth regression (∗): Stage 1

▶ The Stage 1 regressions have the following form:

ri,t = αi + βi,1 F1,t + . . . + βi,m Fm,t + ϵi,t , t = 1, . . . , T , ∀i

▶ Regressions of this form are run for each asset (or portfolio) in the
sample, so we have n time-series regressions in total.

▶ The output of this stage is a set of factor sensitivities

β̂i,1 , β̂i,2 , . . . , β̂i,m for each asset.

▶ Note that the "Factors" in Stage 1 could be any time series variable
that is supposed to influence asset/portfolio returns...

Chen Tong (SOE&WISE) Risk Management October 22, 2024 21 / 52

Review of fama-macbeth regression (∗): Stage 2

▶ Stage 2 regressions are cross-sectional for a given period and

produce factor risk premia estimates
▶ For each periods t, run the following cross-sectional regression:

ri,t = λ0,t + λ1,t β̂i,1 + λ2,t β̂i,2 + . . . + λm,t β̂i,m + εi,t , i = 1, . . . , n

▶ The regression produces T estimates λ̂k,t of the premium of factor k

for all period t, therefore FM methodology proposed to use the
T
average of these estimates, T1 ∑t=1 λ̂k,t , as the final estimated value
of factor k risk premium.

Chen Tong (SOE&WISE) Risk Management October 22, 2024 22 / 52

something about the covariance matrix

▶ We assume the return vector Yt adopts following factor structure:

Yt = BFt + Zt

where Yt is a d-dimensional vector process, Ft is a r-dimensional

common factor process with r < d, Zt is the idiosyncratic
component, and B is a constant factor loading matrix of size d × r .

▶ Then we have the decomposition for the covariance matrix of Yt :

′
ΣY = BΣF B + ΣZ

Chen Tong (SOE&WISE) Risk Management October 22, 2024 23 / 52

Multifactor Models with Known Factors

▶ The Fama-French three-factor model is a special case of factor

model with K = 3,

r˜it − rft = αi + βim (˜

rmt − rft ) + βis S̃MBt + βib H̃MLt + ε̃it

where S̃MBt and H̃MLt are the book-market and size factors,
respectively. These factors are supported by studies that suggest
that market capitalization and stocks classified as either growth or
value impact excess returns.

▶ Note: it’s just a statistical model!!!

Chen Tong (SOE&WISE) Risk Management October 22, 2024 24 / 52

Latent Factors

▶ While some applications use observed factors, some use entirely

latent factors, that is, they take the view that the factors ft in the
K -factor model:
R̃t = α + β f˜t + ε̃t
are not directly observable.

▶ Without imposing what ft are from our likely incorrect belief, we can
statistically estimate the factors based on the factor model and data
(will be detailed in the next part).

Chen Tong (SOE&WISE) Risk Management October 22, 2024 25 / 52

Both Types of Factors

▶ Rather than taking the view that there are only observable factors or
only latent factors, we can consider a more general factor model
with both views,
R̃t = α + β f˜t + βg g̃t + ε̃t
where f˜t is a K -vector of latent factors, g̃t is an L-vector of
observable factors, and βg are the betas associated with g̃t .

▶ This model makes intuitive sense. If we believe a few fundamental

and macroeconomic factors are the driving forces, we will use them
to create the g̃t vector. Since we may not account for all the
possible factors, we need to add additional K unknown factors,
which are to be estimated from the data.

Chen Tong (SOE&WISE) Risk Management October 22, 2024 26 / 52

Both Types of Factors (cont.)
▶ The estimation of the above factor model usually involves two steps.
In the first step, a regression of the asset returns on the known
factors is run in order to obtain β̂g , an estimate of βg . This allows
us to compute the residuals,

ût = Rt − β̂g gt

that is, the difference of the asset returns from their fitted values by
using the observed factors for all the time periods.

▶ Then, in the second step, a factor estimation approach is used to

estimate the latent factors for ût ,

ũt = α + β̃ft + ṽt

where ũt is the random differences whose realized values are ût . The
estimation method for this model is the same as estimating a latent
factor model.

Chen Tong (SOE&WISE) Risk Management October 22, 2024 27 / 52

Predictive Factor Models

▶ An important feature of factor models is that they use time t factors

to explain time t returns. This is to estimate the long-run risk
exposures of the assets, which are useful for both risk control and
portfolio construction. On the other hand, portfolio managers are
also very concerned about time-varying expected returns. In this
case, they often use a predictive factor model such as the following
to forecast the returns:

R̃t+1 = α + β f˜t + ε̃t

Chen Tong (SOE&WISE) Risk Management October 22, 2024 28 / 52

cont.

▶ However, it should be emphasized that the R 2 , a measure of model

2
fitting, is usually very good in explanatory factor models (most R s
are larger than 50%).

▶ In contrast, if a predictive factor model is used to forecast the

2
expected returns of various assets, the R rarely exceeds 2%. This
simply reflects the fact that assets returns are extremely difficult to
predict in the real world.

▶ More discussion in Section of time-series Analysis.

Chen Tong (SOE&WISE) Risk Management October 22, 2024 29 / 52

Some of the properties of the (latent) factor model

▶ First, the factors are not uniquely defined in the model, but all sets
of factors are linear combinations of each other. This is because if f˜t
is a set of factors, then, for any K × K invertible matrix A, we have:

R̃t = α + β f˜t + ε̃t = α + (βA ) (Af˜t ) + ε̃t

−1

Chen Tong (SOE&WISE) Risk Management October 22, 2024 30 / 52

cont.

▶ The second property is that we can assume all the factors have zero
mean, that is, E [f˜t ] = 0. This is because if µf = E [ft ], then the
factor model can be written as:

R̃t = α + β f˜t + ε̃t = (α − βµf ) + β (f˜t − µf ) + ε̃t

▶ Hence, without loss of generality, we will assume that the mean of

the factors are zeros in our estimation in the next section.

Chen Tong (SOE&WISE) Risk Management October 22, 2024 31 / 52

cont.

▶ Recall that we have the decomposition for the covariance matrix of

Yt :
′
ΣY = BΣF B + ΣZ
regardless of whether the factors are observable or latent.

▶ However, through factor rotation, we can make a new set of factors

so as to have the identity covariance matrix. In this case with
Σf = IK , we say that the factor model is standardized, and the
covariance equation then simply becomes:
′
ΣY = BB + ΣZ

Chen Tong (SOE&WISE) Risk Management October 22, 2024 32 / 52

Factor Model Estimation (Main objective)

▶ We need to consider only how to estimate the latent factors f˜t from
the K -factor model,
Ỹt = β f˜t + ε̃t
where both the dependent variables and the factors have zero means,

E (f˜t ) = 0, E [Ỹt ] = 0

▶ This version of the factor model is obtained in two steps. First, we

substract the factor ft from its mean so that the alphas are the
expected returns of the assets. Second, we substract the asset
returns from their means. In other words, we let Ỹt = R̃t − α.

Chen Tong (SOE&WISE) Risk Management October 22, 2024 33 / 52

cont.
▶ In practice, suppose that we have return data on N risky assets over
T time periods. Then the realizations of the random variable Ỹt can
be summarized by a matrix:

⎛ Y11 Y21 ⋯ YN1 ⎞

Y =⎜
⎜ ⋮ ⋮ ⋮ ⋮ ⎟
⎟
⎝ Y1T Y2T ... YNT ⎠

where each row is the N asset returns after subtracting their sample
means at time t for t = 1, 2, . . . , T .

▶ Our task is to estimate the realizations (unobserved) on the K

factors, f˜t , over the T periods:

⎛ F11 F21 ... FK 1 ⎞

F =⎜
⎜ ⋮ ⋮ ⋮ ⋮ ⎟ ⎟
⎝ F1T F2T ... FKT ⎠

Chen Tong (SOE&WISE) Risk Management October 22, 2024 34 / 52

Factor Model Estimation (PCA)

▶ In this section, we provide first a step-by-step procedure for

estimating the factor model based on the popular and implementable
approach, the principal components analysis (PCA). But note that
we need T ≫ N.

▶ PCA is a statistical tool that is used by statisticians to determine

factors with statistical learning techniques when factors are not
observable. That is, given a variance-covariance matrix, a
statistician can determine factors using the technique of PCA.

Chen Tong (SOE&WISE) Risk Management October 22, 2024 35 / 52

What is inner product?
▶ For instance, we have

A = (x1 , y1 ) , B = (x2 , y2 )
and the inner produce is defined by
A ⋅ B = x1 x2 + y1 y2 = ∣A∣∣B∣ cos(α)

Chen Tong (SOE&WISE) Risk Management October 22, 2024 36 / 52

What is a projection vector

▶ For instance, we could define a vector ξ satisfying

⎛ ζ1 ⎞
⎜ ζ2 ⎟
ξ=⎜
⎜ ⎟
⎟
′
⎜ ⎟ , ξξ=1
⎜ ⋮ ⎟
⎝ ζp ⎠
[p×1]

▶ Then, for any vector A, we have

A ⋅ ξ = ξ A = ∣A∣ cos(α)
′

So, A ⋅ ξ is the projection of A on the direction of ξ.

Chen Tong (SOE&WISE) Risk Management October 22, 2024 37 / 52

The main idea

▶ Suppose X is a p × 1 random vector with zero-mean and covariance

matrix ΣX
⎛ x1 ⎞
⎜ x2 ⎟
X =⎜
⎜
⎜
⎟
⎟
⎜ ⋮ ⎟ ⎟
⎝ xp ⎠
[p×1]

▶ Note: X is a zero-mean random vector!!!

Chen Tong (SOE&WISE) Risk Management October 22, 2024 38 / 52

PCA

▶ The goal of PCA is to find a vector ξ to maximize the variance of

′
the random variable ξ X , i.e.

Var (ξ X ) = ξ ΣX ξ
′ ′
max
ξ
′
s.t. ξξ=1

Chen Tong (SOE&WISE) Risk Management October 22, 2024 39 / 52

 ▶ How to solve this problem?

L(ξ) = ξ ΣX ξ + λ (1 − ξ ξ)
′ ′

FOC
ΣX ξ = λξ
So, the λ is the eigenvalue of ΣX , and ξ is the associated
eigenvector!!!

▶ And the λ is also the maximized variance:

′ ′
ξ ΣX ξ = λξ ξ = λ

▶ We have p different λk and ξk for k = 1, 2, 3..., p

Chen Tong (SOE&WISE) Risk Management October 22, 2024 40 / 52

 ▶ Recall the we have a unique spectral decomposition for a nonsingular
covariance matrix:
′
⎛ λ1 0 ⋯ 0 ⎞⎛ ξ1 ⎞
⎜ 0 λ2 0 ⎟ ⎜
⎜
′
⎟
⎟
ξp ) ⎜ ⎟
⋮ ⎜ ξ2 ⎟
ΣX = ( ξ 1 ⎜ ⎟
⎟⎜ ⎟
ξ2 ⋯ ⎜ ⎟ ⎜ ⎟
⎜ ⋮ 0 ⋱ 0 ⎜ ⋮ ⎟
⎝ 0 ⋯ 0 λp ⎠⎝ ξp
′ ⎠

that is
p
′
ΣX = ∑ λk ξk ξk
k=1

Chen Tong (SOE&WISE) Risk Management October 22, 2024 41 / 52

Find the factors?

▶ Define the factor as fk = ξk′ X with variance var (fk ) = λk , and define
′
⎛ ξ1 ⎞ ⎛ f1 ⎞
⎜ ′
⎟ ⎜ ⎟
B=⎜ ⎟ f2
F =⎜ ⎟
⎜ ξ2 ⎟
⎜
⎜ ⎟
⎟ , ⎜
⎜ ⎟
⎟
⎜ ⋮ ⎟ ⎜ ⋮ ⎟
⎝ ξp
′ ⎠ ⎝ fp ⎠
[p×p] [p×p]

then we have
′
⎛ λ1 0 ⋯ 0 ⎞⎛ ξ1 ⎞
⎜ ⎟ ⎜ ′
⎟
⎟⎜ ⎟
0 λ2 0
ξp ) ⎜ ⎟
⋮ ⎜ ξ2 ⎟
ΣX = B Σ F B = ( ξ 1 ⎜
′
⎟⎜ ⎟
ξ2 ⋯ ⎜ ⎟ ⎜ ⎟
⎜ ⋮ 0 ⋱ 0 ⎜ ⋮ ⎟
⎝ 0 ⋯ 0 λp ⎠⎝ ξp
′ ⎠

Chen Tong (SOE&WISE) Risk Management October 22, 2024 42 / 52

The properties the factors:

▶ That means we have

k = {1, . . . , p}
′
fk = ξk X ,
Var (fk ) = λk
Cov (fi , fj ) = 0, i ≠j

▶ We also have
p
′
X = B F = ∑ ξk fk
k=1

and
p p
∑ ξk var (fk )ξk
′ ′ ′
ΣX = B ΣF B = = ∑ λk ξk ξk
k=1 k=1

Chen Tong (SOE&WISE) Risk Management October 22, 2024 43 / 52

How to determine the number of factors?

▶ A simple index:
K
∑i=1 λi
R= p ∈ [0, 1]
∑j=1 λj
where K ≤ p.

▶ In parctice, R > 0.8 is enough!

Chen Tong (SOE&WISE) Risk Management October 22, 2024 44 / 52

A formal criteria of Bai and Ng (2002)

From an econometrics perspective, there is a simple solution in Bai and

Ng (2002) provide a statistical criterion:

N +T NT
IC (K ) = log(V (K )) + K ( ) log ( )
NT N +T

where
N T
2
V (K ) = ∑ ∑ (Yit − β̂i1 fˆ1t − β̂i2 fˆ2t − ⋯ − β̂iK fˆKt )
i=1 t=1

Note that V (K ) is the sum of the fitted squared residual errors of the
factor model.

Chen Tong (SOE&WISE) Risk Management October 22, 2024 45 / 52

Cont.

▶ The smaller the V (K ), the better the K -factor model in explaining

the asset returns. So we want to choose such a K that minimizes
V (K ). However, the more the factors, the smaller the V (K ), but
this improvement is obtained at a cost of estimating more factors
with greater estimation errors. Hence, we want to penalize too many
factors. The second term plays this role. It is an increasing function
of K .

Chen Tong (SOE&WISE) Risk Management October 22, 2024 46 / 52

Matlab code

Chen Tong (SOE&WISE) Risk Management October 22, 2024 47 / 52

An Application to Bond Returns

▶ Consider an application of the PCA factor analysis to the excess

returns on Treasury bonds with maturities 12, 18, 24, 30, 36, 42, 48,
54, 60, 120, and beyond 120 months. Hence, there are N = 11
assets. With monthly data from January 1980 to December 2008.
We have a sample size (i.e., number of data points) of T = 348.

Chen Tong (SOE&WISE) Risk Management October 22, 2024 48 / 52

An Application to Bond Returns

▶ We can easily compute the eigenvalues and eigenvectors from

covariance matrix. The largest three eigenvalues are:

(λ1 , λ2 , λ3 ) = 10 (0.2403, 0133, 0012)

−2

whose sum is more than 99% of the sum of all the eigenvalues.

▶ Thus, it is enough to consider K = 3 factors and use the first three

eigenvectors, PCAs, as proxies for the factors. Denote them as
F1 , F2 , and F3 .

Chen Tong (SOE&WISE) Risk Management October 22, 2024 49 / 52

Factor loadings and explanatory power
▶ Consider now the regression of the 11 excess bond returns on the
three factors:

Rit = αi + βi1 F1t + βi2 F2t + βi3 F3t + εit

where i = 1, 2, . . . , 11.

The PCA factors are effective in explaining the asset returns.

Chen Tong (SOE&WISE) Risk Management October 22, 2024 50 / 52

An Application to Bond Returns

▶ It is interesting that the loadings on the first factor are all positive.
This implies that a positive realization of F1 will have a positive
effect on the returns of all the bonds.

▶ It is clear, however, that F1 affects long-term bonds more than

short-term bonds. As an approximation, F1 is usually interpreted as
a level effect that roughly shifts the returns on bonds plotted against
their maturity (i.e., the yield curve).

Chen Tong (SOE&WISE) Risk Management October 22, 2024 51 / 52

An Application to Bond Returns

▶ The second factor has a different pattern from the first. A positive
realization of F2 will have a negative effect on short-term bonds, and
a positive effect on the long-term ones. This is equivalent to an
increase in the slope of the yield curve. So, F2 is commonly
identified as a steepness factor.

▶ Finally, a positive realization of F3 will have a positive effect on both

short- and long-term bonds, but a negative effect on the
intermediate ones. Hence, F 3 is usually interpreted as a curvature
factor.

Chen Tong (SOE&WISE) Risk Management October 22, 2024 52 / 52