MODELING AND FORECASTING STOCK RETURN VOLATILITY USING A RANDOM LEVEL SHIFT MODEL Lu Hongda Yu Lu Meng Shuangshuang
Feb 25, 2016
MODELING AND FORECASTING STOCK RETURN VOLATILITY USING A RANDOM LEVEL SHIFT MODEL
Lu HongdaYu LuMeng Shuangshuang
Introduction Main goal : to forecast volatility proxied
by daily squared returns Our approach: extends Starica and
Granger (2005)’s work Apply log-absolute returns and daily
returns rather than intra-daily data
Three main parts
1• The model, the specification
adopted and the estimation procedure
2 • Estimation results and diagnostics
3 • The forecasting comparisons
1. THE MODEL AND THE ESTIMATION METHOD
Lu Hongda
1.1 Model
The model we apply to log-absolute returns is given by
is a constant is the random level shift components is a short-memory process, included to model the remaining noise. The level shift component is specified by:
Where ,is a binomial variable that takes value 1 with probability α and value 0 with probability (1 − α). If it takes value 1, then a random level shift occurs, specified by ∼ i.i.d. N(0,).
1.2 Assumptions
For simplicity, we assume:
a) Short memory: = , ()b) The components and are mutually independentc) Normality assumption for , and (needed to
construct the likelihood function)
1.3 Model Estimation Specify level shift component as a random walk process with innovations distributed according to a mixture of two normally distributed processes: where , and is a Bernoulli random variable that
takes value one with probability α and value 0 with probability 1 − α. By specifying and
Estimation of :
The probability of level shifts is very small in all cases considered. Thus… Given that the shifts occur so infrequently,
the noise component accounts for the bulk of total variation.
Then…
Apply the method of Bai and Perron (2003) to obtain the estimates of the break dates that globally minimize the following sum of squared residuals:
where m is the number of breaks, (i = 1,… , m) are the break dates with and and (i = 1, ..., m + 1) are the means within each regime which can easily be estimated once the break dates are.
1.3 Model Estimation
Continue…
We next specify the model in terms of first-differences of the data:
where We then have the following state space form:
Or more generally
Where, in the case of an AR(p) process
And…
and is a p-dimensional normally distributed random vector with mean zero and covariance matrix
1.4 The estimation method
Similar to Markov regime switching models in estimation methodology
The log likelihood function is:
where
1.4 The estimation method
where
where 1 represents a (4*1 )vector of ones denotes element-by-element multiplication with element, with the element of
11( )ˆ
t tt tvec
),( ji th
( )t tvec ),( ji th
11( , , ; )tt t ttf i jy s s Y
2 2
-1 -1-11 1
( ; ) ( ; ) Pr( ; )(3)-1 -1, , , |tt t tti j
f f t t tti s j Yy y i j YsY s s
1ˆ1'( )tt t
-1Pr( ; )-1 , |
tt ti j Ys s
1.4 The estimation method
For when We first focus on the evolution of
-1-2 -1-1 -1
-2 -2-2 -1 -2 -1-1
-2-1
Pr( , ; )
( , , ; ) Pr( , , ; )
( ; )
ki
tt tt t
t tt t t tt
tt
k i
f k i k i
f
s s Yy s s s sY Y
y Y
-1-1-12
-1-2 -11
Pr( , ; )
Pr( ) Pr( , ; ),
ij
tt tt t
tt t tk
i j
j k i
s s Ys s s Y
1.4 The estimation method
Therefore, the evolution of is given by:
Or more compactly by:
1ˆt t
~~~~
~~)(1(
~~)(1(
~~(
~~(
~~~~
22
12
21
11
2212
2111
2212
2111
22
1
12
1
21
1
11
1
)1()1(0000)1()1(
0000
)
)
)
)
tt
tt
tt
tt
tttt
tttt
tttt
tttt
tt
tt
tt
tt
)4(ˆˆ1 tttt
1.4 The estimation method
With The conditional likelihood for is the
following Normal density:
: the prediction error : the prediction error
variance
1
1' )
ˆˆ
ˆ(1
tt t
t ttt t
)5(,2
exp21);,,(
)(~1'
2/1
11
vfvfYssyij
t
tttt
ij
t
ijt
ijtij
tjif
yyvi
ttt
ij
t
1
)('
vvf ijtE ij
t
ij
t
1.4 The estimation method
The best forecast for the state variable and its associated variance conditional on past information and are
We have the measurement equation
ist 1
1 1 1
'
1 1 1
, (6)
.
i i
t t t t
i i
t t t tQ
X FXP FP F
ttt HXy
1.4 The estimation method
Where Hence, the prediction error and variance
are:
Applying standard updating formulae, we have (given ),
2,( ) 0, var( )
0, (1 )t t
pmean
p
RHHPfHXyv
j
i
tt
ij
t
i
ttt
ij
t
'
1
1
iandj ss tt
1,,
PRHPHHPPP
HXyRHPHHPXXi
tt
i
tt
i
tt
ii
tt
i
ttt
i
tt
i
tt
ij
tt
Hjitt
jitt
1
1'
11
1
1'
11
)(
)('
1
)7(),('1
1.4 The estimation method
To reduce the dimension of the estimation problem, we adopt the re-collapsing procedure suggested by Harrison and Stevens (1976), given by
2
1
'2
1
'2
1 1
2
1
2
12
1 1
~)(~
)(
~
~
])([
);Pr(
])([);,Pr(
)8(,);Pr(
);,Pr(
i
ij
tt
ij
tt
j
tt
ij
tti
ij
tt
tt
ij
tt
j
tt
ij
tti tttj
tt
i
ij
tt
ij
tti
ij
tt
tt
ij
tti tttj
tt
XXXXP
YsXXXXPYss
P
XYs
XYssX
ijtt
jtt
j
ijtt
jttji
j
ji
1.4 The estimation method
By doing so, we make unaffected by the history of states before time t − 1
If we define we then have four possible states corresponding to
with the transition matrix Π as defined in (4)
ij
t
),(1ssS ttt
)2,2(,4
)1,2(,3
)2,1(,2
)1,1(,1
1
1
1
1
ssSssSssSssS
ttt
ttt
ttt
ttt
when
when
when
when
1.4 The estimation method
The vector of conditional densities
thus have a more compact representation given by:
Where are as defined in (5)
)9(,2
exp21);,(
)(1'
2/1
1
vfvfYSyl
t
ttt
l
t
lt
ltl
tlf
fandvltl
t,,
),...( 41 tt
l
t
1.5. Apply model and estimation method Four major market indices: the S&P 500,
AMEX, Dow Jones and NASDAQ
Daily returns:
Offset parameter/small constant:
1.5. Apply model and estimation method Specification of short-memory component
a) , ,by Starica and Granger(2005)
b) , process, as a robustness check
1.5. Apply model and estimation method Initialize the state vector and its
covariance matrix by their unconditional expected values(all components of states are stationary), i.e., and
We obtain estimates by directly maximizing the likelihood function:
)2(,);(ln)ln(1 1
T
t ttYyfL
2. EMPIRICAL RESULTS FOR RETURNS ON STOCK MARKET INDICES
Yu Lu
2.1. Estimation results The cases in which the short-memory component
is specified to be white noise and an AR (1) process
the standard deviation of the original seriesa) : the standard deviation of level shift
component, b) α: the probability of a shiftc) :the standard deviation of the stationary
componentd) φ :the autoregressive coefficient when
considering the AR (1) specification for .
Estimation results?
S&P 500 0.7524 0.00198 0.73995(SD:) 0.74290 0.00082 0.73920 0.00030AMEX 0.84579 0.00157 0.72346(SD:) 0.85127 0.00097 0.70671 -0.0146Dow Jones
0.97270 0.00120 0.73291
(SD:) 0.94899 0.00121 0.73196 -0.00087NASDAQ 1.53121 0.00081 0.74280(SD:) 1.53283 0.00079 0.74286 -0.00029
Table 1: Maximum Likelihood Estimates
2.2. Noteworthy features of results:
Firstly, for short-memory componenta) When considering process, estimate of φ is very
small and close to zerob) Adopting either an AR(1) or white noise
specification for yields very similar resultsThus…In subsequent sections we shall only consider results based on the white noise specification for the short-memory component. These results are in agreement with those of Starica and Granger (2005)
2.2. Noteworthy features of results:
Secondly, for level shift componenta) The probability of level shifts is very
small in all cases considered. Thus… Given that the shifts occur so infrequently,
the noise component accounts for the bulk of total variation.
2.3. The effect of level shifts on long-memory and conditional heteroskedasticity Investigating whether the shifts can
explain:
a) the well-documented feature of long-memory
b) the presence of conditional heteroskedasticity
Plot:a) the ACF of the original series b) the ACF of its short-memory
component(obtained by subtracting the fitted trend from )
Whether the level shift can account for the long-memory feature of the series?
2.3.1. Effects of level shift on long-memory
a)the ACF of the original series
The log-absolute returns clearly display an autocorrelation function that resembles that of a long-memory process: it decays very slowly and the values remain important even at lag 300.
b) the ACF of its short-memory component For all practical purposes, we can view the
short-memory component as being nearly white noise.
2.3.2 Effects of level shift on conditional heteroskedasticity GARCH (1, 1) model with Student-t errors
given by, for the demeaned returns process
① i.i.d. Student-t distributed with mean 0 and variance 1
② the parameters of interest and measure the extent of conditional heteroskedasticity present in the data.
CGARCH model
where if t is in regime i, i.e., and 0 otherwise, with (i = 1, ..., m) being the break dates documented in Figure 1 (again and ). The coefficients , which index the magnitude of the shifts, are treated as unknown and are estimated with the remaining parameters, while the number of breaks is obtained from the point estimate of α.
Table2. Parameter EstimatesS&P 500
Coefficient
Estimate Std. Error
t-Statistic
p-value
No level shifts in GARCH
0.057 0.004 13.55 00.938 0.004 214.75 0
No level shiftsin CGARCH
0.058 0.007 7.98 00.939 0.009 95.34 00.9996 0.0003 3250.19 00.027 0.005 5.36 0
Level shiftsin CGARCH
-0.085 0.069 4.73 0.0820.525 0.217 1.54 0.0880.622 0.012 106.98 00.138 0.089 2.47 0.031
2.3.2 Effects of level shift on conditional heteroskedasticity
Both estimates are highly significant for all series. In particular, the value of is quite high.
when we none of the estimates ofare significant.
Use the standard GARCH (1,1) model
allow for level shifts(CGARCH)
Summary:
The level shifts model with white noise errors appears to provide an accurate description of the data.
The level shift component is an important feature that explains both the long-memory and conditional heteroskedasticity features generally perceived as stylized facts.
3 FORECASTINGPerformance of the level shift model &GARCH(1,1)in forecasting volatility
Meng Shuangshuang
3.1 Design of forecasting experiment
Following Starica and Granger(2005) Start forecasting at observations 2,000 Re-estimate every 20 days, up to 200
days Proxy: realized squared returns( )
22)( rE
3.2 Evaluation of forecasting performance
evaluated by the ratio of their MSE:
MSE(p)
where n : no. of forecasting produced
)(/)( PP MSEMSE GARCHLS
= 1n σ (rഥt,p2 −σt,p2 )2 n t=1
rҧt,p2 =σ rt+k2pk=1 : the realized volatility over [t+1,t+p]
σt,p2 = σ σ t+k2pk=1 : σ t2 be a p-step ahead forecast of σt+p2 the variance
of returns rtat time t+p
3.3 construction of the forecasts The level shift model(1) The level shift model an appropriate transformation
where C : a positive constant used to bound returns away from zero
yields
𝐥𝐨𝐠ሺa(𝐫𝐭a(+ 𝐂ሻ= 𝛕𝐭+ 𝐂𝐭
a(𝐫𝐭a(+𝐂= 𝐡𝐭𝟏/𝟐𝛜𝐭 where, ht = e2τt E(e2ct) and 𝛜𝐭 = ect /[E(e2ct)]1/2
3.3 construction of the forecasts The level shift model(2) continue
ignore level shifts when forecasting rare uncertain about timing and
magnitudes
E𝛜𝐭2=1 𝛜𝐭 is independent of ht and i.i.d
2ct~ i.i.d N(0, 4σe2) Eሺe2ctሻ= e0+4σe2/2 = e2σe 2
3.3 construction of the forecasts The level shift model(3)
Continue
the K -period ahead forecast of the squared returns is :
Et(a(rta(+C)2 = Etht+k = exp( 2τt +2σe2) Etrt+k2 = expሺ2τt +2σe2ሻ−2CEta(rt+ka(−C2
Where Et |rt+k | = Etexp (𝛕𝐭 +𝐂𝐭+𝐤) - C=exp (τt +0.5σe2) - C
3.3 construction of the forecasts The GARCH model(1) Student-t GARCH(1,1) model
Transformation:
Rt෪ = rt −u = σtεt σt2 = α1 +α2𝐫𝐭−𝟏 +α3σt−12
εt : i.i.d~N(0,1)
rt : demeaned returns
rt2=α1 + (α2+α3)rt−12 +ωt−α3ωt−1 * where ωt = rt2-σt2 : the forecast error associated with the forecast of rt2 , white noise
3.3 construction of the forecasts The GARCH model(2) assuming < 1, recursive form for the squared
demeaned returns
So for k > 1, the k-period ahead forecast
Where the forecasts for the squared returns
E(rt2)=σ2=α1/(1− α2−α3)
rt2-σt2=(α2+α3)( rt−12 -σ2)+ ωt−α3ωt−1
Etrt+k2 =σ2 + ( α2+α3 )k−1(Etrt+12 -σ2)= σ2 +ሺ α2+α3 ሻk−1(σt+12 −σ2) σt+12 = α1 +α2rt2+α3σt2.
Etrt+k2 = Etrt+k2 + (Etrt+k)2 ≈ Etrt+k2 + μ2
3.4 Forecasting Comparisons
use the fitted means obtained from the full sample to get the mean
re-estimate every 20 observations to get the mean
the random level shifts model the GARCH(1,1)
The comparison indicates that with a precise estimate of the mean of log-absolute returns at a given date, we can obtain much better forecasts from the level shift model than the other models.
3.4 Forecasting ComparisonsHorizon Horizon
P days S&P 500 P days S&P 500
20 0.5 120 0.51
40 0.44 140 0.54
60 0.44 160 0.56
80 0.46 180 0.59
100 0.48 200 0.62
)(/)( PP MSEMSE GARCHLS)(/)( PP MSEMSE GARCHLS
3.4 Forecasting Comparisons Table: volatility forecasting performance
of the LS model & GARCH(1,1) model ratio < 1 : volatility forecast of the LS
model is more precise than the GARCH(1,1) model
The figure shows an over-all better longer-horizon volatility forecast performance of the LS model
Conclusion
a simple estimation method for a model with a random level shift & a short-memory
component The level shift component : explains the
presence of both the long-memory and conditional heteroskedasticity
a short-memory component: assume to be white noise
impressive results : when model applied to log-absolute returns of many indices
Conclusion
Volatility forecasting with the model: improvement with squared returns as a
proxy Noticeable when mean is well estimated Difficult to obtain precise estimates of the
mean
ENDThank you