MATH11131 Part 3: Financial Time Series

⊲ MATH11131
Time Series
Part 3: Financial

VARMA models

GARCH models

Stochastic Volatility models

Factor models

Ioannis Papastathopoulos
School of Mathematics, University of Edinburgh

1 / 29
MATH11131 Part 3: Financial Time
Series

⊲ VARMA models
Models for multiple time series
Fitting VARMA models in R–returns
of EU markets
Fitting VARMA models in R

GARCH models

Stochastic Volatility models

Factor models

VARMA models

2 / 29
 Models for multiple time series
 ARMA modelling may be extended to multivariate cases, but is much more complicated.
MATH11131 Part 3: Financial Time
Series  One simple case is the vector autoregressive (VAR), where with Yt a k × 1 vector,
VARMA models
⊲ Models for multiple time series
we have p
Fitting VARMA models in R–returns iid
X
of EU markets
Fitting VARMA models in R
Yt = µk×1 + Φj Yt−j + wt , wt ∼ N (0k×1 , Σk×k ),
GARCH models j=1
Stochastic Volatility models

Factor models
where Φ1 , . . . , Φp are k × k matrices and Φp 6= 0; such models have a lot of parameters.
 The VMA models can be defined likewise.
 This extends to the vector autoregressive moving average (VARMA) model,
p p
iid
X X
Yt = µk×1 + Φj Yt−j + wt + Θj wt−j wt ∼ N (0k×1 , Σk×k )
j=1 j=1

where Θ1 , . . . , Θq are also k × k matrices with Θq 6= 0.

 We define the AR and MA operators Φ(B)k×k = I − Φ1 B − Φ2 B 2 − · · · Φp B p and
Θ(B)k×k = I + Θ1 B + Θ2 B 2 + · · · Θq B q ; the process is causal if the roots of |Φ(z)| lie
outside D, and is invertible if the roots of |Θ(z)| lie outside D
 Complicated conditions are needed to ensure uniqueness and identifiability, but can be
avoided by fitting only VAR models—see Tsay (2014) Multivariate Time Series Analysis
with R and Financial Applications. John Wiley.

3 / 29
 Fitting VARMA models in R–returns of EU markets

MATH11131 Part 3: Financial Time

Series

0.04
VARMA models
Models for multiple time series

0.00
Fitting VARMA models in
⊲

DAX
R–returns of EU markets

−0.04
Fitting VARMA models in R

GARCH models

0.04 −0.08
Stochastic Volatility models

Factor models

0.00
SMI
−0.04
−0.08
0.06
0.02
CAC
−0.02
−0.06
0.04
FTSE
0.00
−0.04

0 500 1000 1500

Time

4 / 29
 Fitting VARMA models in R

MATH11131 Part 3: Financial Time

- base::ar
Series
- MTS::VAR, VMA, VARMA
VARMA models
Models for multiple time series
Fitting VARMA models in R–returns
of EU markets
 If we fit a VAR model to the European stock markets, AIC gives p = 1 and the
⊲ Fitting VARMA models in R
estimated matrix of coefficients Φ̂1 and their standard errors are
GARCH models

Stochastic Volatility models > VAR.model <- VARMA(returns) ## first calculate returns from data
Factor models
Constant term:
Estimates: 0.0007532118 0.000823424 0.0005525078 0.000473503

DAX SMI CAC FTSE

DAX 0.00497 -0.09642 0.03919 0.0483
SMI -0.00738 -0.00682 0.03611 0.0674
CAC -0.02784 -0.11229 0.06117 0.0939
FTSE -0.01084 -0.08922 -0.00395 0.1655
 Here n = 1860 so n−1/2 = 0.023 helping us assess which of these coefficients differ
significantly from zero:
– all react positively to a jump in the FTSE the previous day.
– all react negatively to a jump in the SMI the previous day.
– all are somewhat decoupled from the DAX.

5 / 29
MATH11131 Part 3: Financial Time
Series

VARMA models

⊲ GARCH models
Motivation
GARCH models
(G)ARCH theory
FTSE
FTSE returns
Weaknesses
Some extensions

Stochastic Volatility models

Factor models
GARCH models

6 / 29
 Motivation
 Many time series show changes in the variance as well as in the mean; this is particularly
MATH11131 Part 3: Financial Time
prominent in financial time series, but arises in many other contexts also.
Series
 However (stationary, invertible, causal) Gaussian (V)ARMA models satisfy
VARMA models

GARCH models
⊲ Motivation
GARCH models
Yt | Y−t = y−t ∼ N (µ + Σt,−t Σ−1
−t (y−t − µ), Σt,t − Σ Σ−1
Σ
t,−t −t −t,t ), 1

(G)ARCH theory
FTSE
FTSE returns in an obvious notation: the conditional variance of Yt given the preceding observations
Weaknesses
Some extensions Y−t does not depend on their values, and so does not vary with time.
Stochastic Volatility models
 This prompts the search for models that do allow such dependence
Factor models
 A prominent class of such models is the generalised autoreressive conditionally
heteroscedastic (GARCH) class, given by
m r
iid
X X
Y t = σt ε t , σt2 = α0 + 2
αj Yt−j + 2
βj σt−j , εt ∼ N (0, 1).
j=1 j=1

and labelled GARCH(m, r). To avoid negative variances we take αj , βj ≥ 0.

 A GARCH model is not unlike an ARMA model in the variances σt2 , and the fitting
approach is similar. If β1 = · · · = βr = 0, we have an ARCH(m) model.
 Typically in practice m, r are small: often we take m = r = 1
   
1 µX ΣX ΣXY

If (X, Y ) ∼ N , then Y | X = x ∼ N µY − ΣTXY Σ−1
X (x − µX ), ΣY − ΣTXY Σ−1
X ΣXY .
µY ΣTXY ΣY

7 / 29
 GARCH models

MATH11131 Part 3: Financial Time

Series Definition 1. For t ∈ Z, let Ht denote the entire history of the process {Yt } up to time t.
VARMA models

GARCH models
Motivation
⊲ GARCH models Lemma 1. A GARCH model has zero mean, and satisfies E(Yt | Ht−1 ) = 0; thus it is an
(G)ARCH theory
FTSE uncorrelated sequence: cor(Yt+h , Yt ) = 0. The quantity Yt is called a martingale
FTSE returns
Weaknesses difference.
Some extensions

Stochastic Volatility models

iid
Factor models
Lemma 2. An ARCH(1) model with standard Gaussian innovations, {εt } ∼ N (0, 1), may
be written
Yt2 = α0 + α1 Yt−1
2
+ Vt ,
where Vt = σt2 (ε2t − 1), so the {Vt } are non-Gaussian white noise. Thus the process is
stationary and causal if 0 ≤ α1 < 1. In this case we have var(Yt ) = α0 /(1 − α1 ) and if in
addition 3α12 < 1,
E(Yt4 ) 1 − α12
2
=2 2 ≥ 3.
var(Yt ) 1 − 3α1
The implication of this last result is that the marginal distribution of the ARCH(1) model
has heavier tails than does the normal.

8 / 29
 (G)ARCH theory

 For the ARCH(1) model;

MATH11131 Part 3: Financial Time
Series

VARMA models
– If 0 ≤ α1 < 1, then {Yt } is white noise and its unconditional distribution is
GARCH models symmetrically distributed around zero. This unconditional distribution is leptokurtic
Motivation
GARCH models (has heavier tails than the normal).
⊲ (G)ARCH theory
FTSE
FTSE returns
Weaknesses
– If in addition 3α12 < 1, then the process follows {Yt2 } follows a causal AR(1) model
Some extensions |h|
Stochastic Volatility models
with ACF ρh = α1 .
Factor models

– If 3α12 ≥ 1 but α1 < 1, then {Yt2 } is strictly stationary with infinite variance.

– We have already seen how to estimate the parameters α0 , α1 using—the Markov

property, and—least squares.
 Likewise the GARCH(1,1) has an ARMA(1,1) representation

Yt2 = α0 + (α1 + β1 ) Yt−1

2
+ Vt − β1 Vt−1 ,

where Vt is defined as before.

 Estimation for GARCH(m, r) is performed by supposing that σ12 = · · · = σr2 = 0, and
building the likelihood for Yr+1 , . . . , Yt .

9 / 29
 FTSE

MATH11131 Part 3: Financial Time

Series The Financial Times Stock Exchange Index, 1991–1998.
VARMA models

GARCH models
Motivation
GARCH models
(G)ARCH theory
⊲ FTSE
FTSE returns

6000
Weaknesses
Some extensions

Stochastic Volatility models

Factor models

5000
FTSE Index

4000
3000

1992 1993 1994 1995 1996 1997 1998

Time

10 / 29
 FTSE returns

MATH11131 Part 3: Financial Time

Series [Demo]
VARMA models

GARCH models
Motivation
GARCH models
(G)ARCH theory
FTSE
⊲ FTSE returns
Weaknesses
Some extensions

Stochastic Volatility models

4
Factor models

2
FTSE returns (%)

0
−2
−4

1991 1992 1993 1994 1995 1996 1997 1998

Time

11 / 29
 Weaknesses

MATH11131 Part 3: Financial Time

Series  (G)ARCH models have the following weaknesses
VARMA models

GARCH models
– financial markets react more strongly to negative shocks, but in a GARCH model σt
Motivation
GARCH models
reacts identically to positive and negative shocks;
(G)ARCH theory
FTSE
FTSE returns
⊲ Weaknesses
– they react slowly to large shocks in the return, so may over-predict volatility;
Some extensions

Stochastic Volatility models

– strong restrictions on the parameters are needed for stationarity and finite variance;
Factor models

– they may describe the second-order properties, but they are not based on a
financial/economic theory so give no insight into why a series behaves as it does;

12 / 29
 Some extensions

MATH11131 Part 3: Financial Time

Series  The exponential GARCH (EGARCH) model allows the variance to depend on the
VARMA models
sign of the series, for example giving
GARCH models
Motivation

|Yt−1 |
GARCH models
  
2 2α
(G)ARCH theory
σt = σt−1 exp α∗ + γ + θ sign(Yt−1 )
FTSE
FTSE returns σt−1
Weaknesses
⊲ Some extensions

Stochastic Volatility models

for suitable constants α, α∗, γ, θ; we expect that θ < 0 if negative shocks have higher
Factor models impacts.
 GARCH models have been extended to integrated GARCH (IGARCH) models, e.g.,

σt2 = α0 + β1 σt−1
2 2
+ (1 − β1 ) Yt−1 ,

which is analogous to an ARIMA model, in that past volatility shocks persist.

13 / 29
MATH11131 Part 3: Financial Time
Series

VARMA models

GARCH models

⊲ Stochastic Volatility models

Parameter- and observation-driven
models
Parameter driven models
Simple stochastic volatility model
Simple stochastic volatility model
Bayesian inference
Markov chain Monte Carlo
Example: AR(1) with measurement

Stochastic Volatility models

error
Full conditional densities
AR(1) with measurement error
Comments
MCMC for SV
Discussion

Factor models

14 / 29
 Parameter- and observation-driven models

MATH11131 Part 3: Financial Time

Series  A useful division of time series models is into observation-driven models and
VARMA models
parameter-driven models.
GARCH models

Stochastic Volatility models  Easiest to explain in a very simple context: suppose

Parameter- and
⊲ observation-driven models
2
Parameter driven models


Simple stochastic volatility model Yt | Zt ∼ N µt , σt , t ∈ Z.
Simple stochastic volatility model
Bayesian inference
Markov chain Monte Carlo
Example: AR(1) with measurement
The distribution of Yt | Zt is normal with mean µt and variance σt2 both functions of
error
Full conditional densities Zt .
AR(1) with measurement error
Comments
MCMC for SV
 Observation-driven models take Zt to be a function of past observations. Examples:
AR(p) model with µt = φ1 yt−1 + · · · + φp yt−p , σ 2 = σ 2 ;
Discussion

Factor models
–
– ARCH(m) model with µt = 0, σt2 = α0 + α1 yt−12 2
+ · · · + αp yt−m ;
– AR(1)-ARCH(1) model, with µt = µ + φ(yt−1 − µ), σt2 = α0 + α1 (yt−1 − µ)2 .
 Three reasons to like observation-driven models for volatility:
– the likelihood is easily computed, so estimation, testing, model-checking are easy;
– finance theory is often specified through one-step-ahead movements, defined with
respect to economic agents’ information;
– they parallel the very successful ARMA-type models.

15 / 29
 Parameter driven models

MATH11131 Part 3: Financial Time

Series  Parameter-driven models take Zt to be a function of some unobserved or latent
VARMA models
component, e.g., with
GARCH models

Stochastic Volatility models

iid
Parameter- and observation-driven
models µt = 0, log σt2 = γ0 + 2
γ1 log σt−1 + ηt ηt ∼ N (0, ση2 ),
⊲ Parameter driven models
Simple stochastic volatility model
Simple stochastic volatility model
Bayesian inference
so that the unobserved log-volatility process log σt2 satisfies an AR(1) model
Markov chain Monte Carlo
Example: AR(1) with measurement
error
Full conditional densities
 Reasons to like these models:
AR(1) with measurement error
Comments – properties are easier to find, manipulate and generalize to the vector case;
MCMC for SV
Discussion

Factor models – have simpler representations in continuous time (e.g., as diffusions that satisfy
suitable SDEs), which makes them natural in finance where much theory is based on
stochastic differential equations;
but parameter-driven models don’t have simple forms for their density f (yt | Ht ), so
can’t easily write down likelihood . . . so inference becomes more difficult;

 Consider simplest such model, based on treatment in Kim, Shephard and Chib (1998),
Review of Economic Studies, 351–393.

16 / 29
 Simple stochastic volatility model

MATH11131 Part 3: Financial Time

Series  Consider
VARMA models

GARCH models

Stochastic Volatility models

Y t = σt ε t , ht = log σt2 , ht+1 = γ0 + γ1 (ht − γ0 ) + ση ηt
Parameter- and observation-driven
models
iid iid
Parameter driven models
⊲ Simple stochastic volatility model
where εt ∼ N (0, 1) independent of ηt ∼ N (0, 1).
Simple stochastic volatility model
Bayesian inference
Markov chain Monte Carlo
Example: AR(1) with measurement
error
 Thus {ht } ∼ AR(1), and if |γ1 |< 1, then we have stationary distribution
Full conditional densities
AR(1) with measurement error
!
Comments ση2
MCMC for SV
ht ∼ N γ0 , 2 ,
Discussion

Factor models
1 − γ1

with long excursions above or below γ0 if γ1 ≈ 1. (why?)

17 / 29
 Simple stochastic volatility model

MATH11131 Part 3: Financial Time

Series  Inference for this model has close link to state space models, but in brief:
VARMA models

GARCH models
– likelihood inference can be replaced by quasi-likelihood inference using the relation
Stochastic Volatility models
Parameter- and observation-driven
log Yt2 = ht + ε2t , where (with εt normal), we have E(log ε2t ) = −1.27 and
models
Parameter driven models
var(log ε2t ) = 4.93, but this doesn’t work well;
Simple stochastic volatility model
⊲ Simple stochastic volatility model – otherwise we can use Bayesian inference and Markov chain Monte Carlo
Bayesian inference
Markov chain Monte Carlo
Example: AR(1) with measurement (MCMC) to estimate the time series {ht}.
error
Full conditional densities
AR(1) with measurement error
Comments – Next few slides explain MCMC and Bayesian inference in the context of an AR(1)
MCMC for SV
Discussion model with measurement error, before returning to stochastic volatility modelling.
Factor models

18 / 29
 Bayesian inference

MATH11131 Part 3: Financial Time

Series  Bayesian inference requires that a prior density be specified for every unknown
VARMA models
parameter in a statistical problem, and then just applies the laws of probability to obtain
GARCH models

Stochastic Volatility models

the posterior density for all unknowns, conditional on the observed data.
Parameter- and observation-driven
models
Parameter driven models
 Given the data model f (y | θ) ≡ f (y; θ) and prior density π(θ) for a parameter θm×1 ,
Simple stochastic volatility model
Simple stochastic volatility model
Bayes’ theorem posterior density
⊲ Bayesian inference
Markov chain Monte Carlo
Example: AR(1) with measurement π(θ) f (y | θ)
error π(θ | y) = R
Full conditional densities π(θ) f (y | θ) dθ
AR(1) with measurement error
Comments
MCMC for SV
Discussion
which contains all information about θ, conditional on the observed data.
Factor models
 We often use the Markov chain Monte Carlo simulation to generate samples from
π(θ | θ−i )

19 / 29
 Markov chain Monte Carlo

MATH11131 Part 3: Financial Time

Series  Gibbs sampler: starting with an (arbitrary) initial θ(0) , for r = 1, . . . , R, iterate
VARMA models

GARCH models – set θ(r) = θ(r−1) ,

Stochastic Volatility models
Parameter- and observation-driven (r) (r)
models
Parameter driven models
– for i = 1, . . . , m, generate θi∗ ∼ π(θi | θ−i ) and replace θi by θi∗
Simple stochastic volatility model
Simple stochastic volatility model
Bayesian inference – Let θ−i denote the elements of θ without θi , for i = 1, . . . , m. Suppose that it is
⊲Example:
Markov chain Monte Carlo
AR(1) with measurement possible to sample from the distributions π(θ | θ−i ), for i = 1, . . . , m.
error
Full conditional densities
AR(1) with measurement error
Comments – Under suitable conditions and for large R, the distribution of the sequence
MCMC for SV
Discussion
θ(1) , . . . , θ(R) approximates π(θ | y), so we can use the θ(r) to estimate properties of
Factor models the posterior density (often dropping the initial transient part of the
sequence—knowns as burn-in. )

20 / 29
 Example: AR(1) with measurement error

Suppose {ηt} follows an AR(1) process with parameter α and innovation variance τ 2,
MATH11131 Part 3: Financial Time
Series 
VARMA models ind
GARCH models
and Yt | η1 , . . . , ηn ∼ N (ηt , σ 2 ), giving an AR(1) with measurement error.
Stochastic Volatility models

Here θ = (α, τ 2, σ 2 , η) has dimension m = n + 3.

Parameter- and observation-driven
models 
Parameter driven models
Simple stochastic volatility model
Simple stochastic volatility model
Bayesian inference  For simplicity we take independent prior densities
Markov chain Monte Carlo
Example: AR(1) with
⊲ measurement error
η1 ∼ N (0, α2 ), α ∼ N (0, b2 ), τ 2 ∼ IG(c, d), σ 2 ∼ IG(e, f )
Full conditional densities
AR(1) with measurement error
Comments
MCMC for SV
Discussion
where IG(c, d) denotes the inverse gamma density dc x−c−1 e−d/x /Γ(c), for x > 0 and
Factor models c, d > 0.

21 / 29
 Full conditional densities

The full conditional densities π(θi | θ−i ) required to run a Gibbs sampler are:
MATH11131 Part 3: Financial Time
Series
n−1 n
VARMA models 1 1 X
2 1 X
GARCH models α | rest ∼ N (Bα /Aα , 1/Aα ), Aα = 2 + 2 η t , Bα = 2 ηt ηt−1 ,
Stochastic Volatility models
b τ t=1 τ t=2
Parameter- and observation-driven
n
models
( )
Parameter driven models n − 1 1 X
Simple stochastic volatility model τ 2 | rest ∼ IG c + ,d + (ηt − α ηt−1)2 ,
Simple stochastic volatility model
Bayesian inference
2 2 t=2
Markov chain Monte Carlo n
( )
Example: AR(1) with measurement
n 1 X
error
⊲ Full conditional densities σ 2 | rest ∼ IG e + , f + (yt − ηt)2 ,
AR(1) with measurement error 2 2 t=2
Comments
MCMC for SV
2 n−1
Discussion 1 α 1 αη2 y1 X
Factor models η1 | rest ∼ N (B1 /A1, 1/A1 ), A1 = 2 + 2 + 2 , B1 = 2 + 2 , ηt2,
a τ σ τ σ t=1
1 1 + α2 α(ηt−1 + ηt+1 )η2 y1
ηt | rest ∼ N (Bt /At, 1/At ), At = 2 + 2
, Bt = 2
+ 2, t 6= 1, n,
σ τ τ σ
1 1 αηn−1 yn
ηn | rest ∼ N (Bt /At, 1/At ), A n = 2 + 2 , Bn = 2
+ 2.
τ σ τ σ
The next slide shows the results when this is run for R = 10, 000 iterations, preceded by a
‘burn-in’ of length 200. I took a = 100, b = 10, c = d = e = f = 0.01, to give very vague
prior information. The R code for the entire examples, including pictures, is 60 lines.

22 / 29
 AR(1) with measurement error

MATH11131 Part 3: Financial Time

Series

VARMA models

GARCH models

Stochastic Volatility models

Parameter- and observation-driven
models
Parameter driven models
Simple stochastic volatility model
Simple stochastic volatility model
Bayesian inference
Markov chain Monte Carlo
Example: AR(1) with measurement
error
Full conditional densities
⊲ AR(1) with measurement error
Comments
MCMC for SV
Discussion

Factor models

Figure 1: Top left: the red line is the posterior mean for η, the grey band are 95% pointwise
credible intervals for η, the blue line is the true η. In the other panels the true value used is
shown by a vertical line.

23 / 29
 Comments

MATH11131 Part 3: Financial Time

Series  Often the code can be accelerated by updating blocks of (conditionally) independent
VARMA models
variables. In the example above, the most time-consuming part is the iteration over
GARCH models

Stochastic Volatility models

η1 , . . . , ηn . These have a multivariate normal distribution conditional on θ, and updating
Parameter- and observation-driven
models them simulateneously would involve inverting an n × n matrix, but we could avoid this
Parameter driven models
Simple stochastic volatility model by updating η1 , η3, . . ., conditional on η2, η4 , . . . , and then vice verse (both conditional
Simple stochastic volatility model
Bayesian inference on θ), noting that the odd ηs are conditionally independent given the even ones, and
Markov chain Monte Carlo
Example: AR(1) with measurement vice versa. This replaces a loop over η1 , . . . , ηn with two steps, making the algorithm
error
Full conditional densities much faster.
AR(1) with measurement error
⊲ Comments
MCMC for SV
Discussion
 MCMC algorithms can be developed for many models, though usually some full
Factor models conditional densities π(θi | θ−i ) will be unavailable in a simple form and
Metropolis-Hastings steps must be used.

 Reversible jump MCMC algorithms (Green, 1995. Biometrika, 82, 711–732) can be
used when the number of underlying states is unknown.

 Huge and widely applied ideas with very large literature.

24 / 29
 MCMC for SV

 The Gibbs sampler in this case might be:2

MATH11131 Part 3: Financial Time
Series – Initialise h = (h1 , . . . , hn ) and θ = (γ0 , γ1 , ση2 ).
VARMA models

GARCH models – For r = 1, . . . , R, repeat

Stochastic Volatility models
Parameter- and observation-driven
models
1. For t = 1, . . . , n, sample ht ∼ π(ht | h−t , γ0 , γ1 , ση2 , y)
Parameter driven models
Simple stochastic volatility model
Simple stochastic volatility model 2. Sample ση2 ∼ π(h, γ0 , γ1 , , y)
Bayesian inference
Markov chain Monte Carlo
Example: AR(1) with measurement
error 3. Sample γ0 ∼ π(h, γ1 , ση2 , y)
Full conditional densities
AR(1) with measurement error

4. Sample γ1 ∼ π(h, γ0 , ση2 , y)

Comments
⊲ MCMC for SV
Discussion

Factor models  A complete sweep involves cycling once through steps 1-4, and we must take
R = O(103 ) at least, perhaps R = O(106 ), so efficient coding and fast algorithms are
essential.
 Computing the conditional densities here may be hard, or even impossible, but then we
can use Metropolis-Hastings steps, require only the densities up to a constant of
proportionality.
 Once we have the output, we have to check convergence of the algorithm, and if
satisfied we then use the output to estimate the posterior density π(θ | y).
2
Many variant algorithms and procedures exist; this is just the simplest—see Kim et al. (1998).

25 / 29
 Discussion

MATH11131 Part 3: Financial Time

Series  Stochastic volatility modelling is an attractive approach to accounting for changing
VARMA models
variance, because we can allow the hidden process σ 2 to vary according to
GARCH models

Stochastic Volatility models

economic/financial theory, including the role of exogenous variables.
Parameter- and observation-driven
models
Parameter driven models
Simple stochastic volatility model
 But it involves more sophisticated approaches than we have used so far, based on
Simple stochastic volatility model
Bayesian inference
Markov chain Monte Carlo, and these may be hard to implement.
Markov chain Monte Carlo
Example: AR(1) with measurement
error
Full conditional densities
 See Kim et al. (1998) and subsequent papers for more details, developments, etc.
AR(1) with measurement error
Comments
MCMC for SV
⊲ Discussion

Factor models

26 / 29
MATH11131 Part 3: Financial Time
Series

VARMA models

GARCH models

Stochastic Volatility models

⊲ Factor models
Factor models
Factor models

Factor models

27 / 29
 Factor models

MATH11131 Part 3: Financial Time

Series  Popular models for financial time series posit that returns for many series react to
VARMA models
changes in a few underlying factors, which themselves evolve according to stationary
GARCH models

Stochastic Volatility models

time series
Factor models
⊲ Factor models
 Suppose we have k assets and n time periods, and let the return on asset i in time
Factor models

period t be

rit = αi + βi1 f1t + · · · + βim fmt + εit , t = 1, . . . , n, i = 1, . . . , k,

where αi is a constant expressing the intercept, fjt for j = 1, . . . , m are m common

factors, βij is the factor loading for asset i on factor j, and εit is the specific factor
of asset i. We hope that m ≪ k; but in some cases have k ≫ n.

 Assume that (ft )m×1 is a stationary process satisfying E(ft ) = µf , var(ft ) = Σf , and
that the {εit } are uncorrelated white noise series, but with var(εit ) = σi2 .

 We can write
rt = αk×1 + βk×m ft + εt , t = 1, . . . , n,
in an obvious notation, and then have var(rt ) = β Σf β T + diag(σ12 , . . . , σk2 ).

28 / 29
 Factor models

MATH11131 Part 3: Financial Time

Series  Common to assume that the factors are serially uncorrelated.
VARMA models

GARCH models

Stochastic Volatility models

 These models are widely used but suffer from the curse of dimensionality: the
Factor models number of parameters quickly becomes very large, so symmetries must be exploited to
Factor models
⊲ Factor models reduce the number of parameters—see Tsay (2005), Analysis of Financial Time
Series, Second edition, Wiley.

29 / 29