METHODS IN FINANCE
FORECASTING
2
MODEL
IDENTIFICATION
ACF,
PACF
plot
4
MODEL
ADEQUACY
Box
Test
for
serial
correlaEon.
closing
Last 14.14
Mean Zero
[20070103/20120706]
80
T test shows mean not zero of the closing prices. p-value <2.2e-16 alternaEve hypothesis: true mean is greater than 0 Since the mean is not zero, we dierence the data.
60
40
20
Jan 03 2007
Jan 02 2008
Jan 02 2009
Jan 04 2010
Jan 03 2011
Jan 03 2012
The augmented DickeyFuller (ADF) staEsEc, used in the test, is a negaEve number. The more negaEve it is, the stronger the rejecEon of the hypothesis that there is a unit root at some level of condence. Dickey-Fuller = -12.2629, Lag order = 11, p-value = 0.01 alternaEve hypothesis: staEonary Conclusion: Since p value<0.05, reject null of non staEonarity, therefore data is staEonary A model that includes a constant and a Eme trend is esEmated using sample of 50 observaEons and yields the staEsEc of 4.57. This is more negaEve than the tabulated criEcal value of 3.50, so at the 95 per cent level the null hypothesis of a unit root will be rejected.
[20070104/20120706]
10
Jan 04 2007
Jan 02 2008
Jan 02 2009
Jan 04 2010
Jan 03 2011
Jan 03 2012
NORMALITY TEST
Jarque - Bera Normalality Test Test Results: X-squared: 58437.3676, P VALUE: < 2.2e-16 the JarqueBera test is a goodness-of-t test of whether sample data have the skewness and kurtosis matching a normal distribuEon. The null hypothesis is a joint hypothesis of the skewness being zero and the excess kurtosis being zero. Hence, we conclude that the data does not have a normal distribu@on.
Basic STATS MS.Close nobs 1385.000000 NAs 0.000000 Mean -0.001226 Variance 0.002144 Stdev 0.046308 Skewness 1.393276 Kurtosis 31.649301 Gaussianity refers to the probability distribu@on with respect to the value, in this context the probability of the signal reaching an amplitude, while the term 'white' refers to the way the signal power is distributed over @me or among frequencies.
MODEL IDENTIFICATION
0.2
There is no clear indicaEon of a AR(p) or MA(q) process here, as neither the ACF nor the PACF has a gradual decline at consecuEve lags nor do they have spikes at certain lags. Second method of p,q es@ma@on. best.order<-c(0,0,0) best.aic<-Inf for(i in 0:4)for(j in 0:3){ t.aic<-AIC(arima(resid(arma) ,order=c(i,0,j))) if(t.aic<best.aic){ best.order<-c(i,0,j) best.arma<-arima(resid(arma1) ,order=best.order) best.aic<-t.aic}}
10
20
LAG
30
40
0.0
0.2
PACF 0.4
0.6
0.8
1.0
10
20
LAG
30
40
STATISTICAL
METHODS
IN
FINANCE
ARIMA(1,0,1),
AIC=
-5.141276
ARIMA(3,0,1),
AIC=
-5.165905
Box-Ljung
test
(H 0 :
The
data
are
independently
distributed)
X-squared
=
41.8899,
df
=
12,
p-value
=
3.476e-05
Since
p<0.05,
we
reject
the
null
and
say
data
is
dependent.
Here
an
ARIMA
p=1,3
and
q=1
with
d=0(as
data
is
staEonary)
has
been
med
to
the
data
without
indicaEons
of
removal
of
serial
correlaEons
between
the
square
of
residuals.
We
nd
a
relaEon
between
the
variance
of
the
residuals
signifying
that
the
data
was
not
totally
characterized
by
the
ARIMA
models
used
here
and
more
analysis
is
needed
to
get
a
white
noise
in
the
end.
1.0
Fi\ng
ARMA
Series: armaresi4^2
0.0
0
0.2
0.8
10
20
LAG
30
40
0.0
0
0.2
0.8
1.0
10
20
LAG
30
40
STATISTICAL METHODS IN FINANCE G6-ARMA(3,1)+GARCH(1,1) AIC=-4.134697 G8-ARMA(1,1)+GARCH(1,1) AIC=-4.132427 Box-Ljung test X-squared = 30.7012, df = 12, p-value = 0.002188 Box-Ljung test X-squared = 47.8836, df = 12, p-value = 3.275e-06 The Error analysis for G6 showed insignicact presence of AR 2 and 3 and the model could only be characterised by ARMA(1,1) + Garch(1,1), but G6 had a bemer AIC value and both G6 and G8 did not remove the serial correlaEon between the residuals. The QQ plot shows that the residuals are sEll not normal and the ACF and Ljung-Box test of residual squared shows that there is serial correlaEon between them. Therefore, the models are not complete.
200
400
600
800
1000
1200
1400
ACF of Residuals
0.4
Sample Quantiles
10
0.2
ACF
0.0
0.2
10
20 LAG
30
40
Theoretical Quantiles
0.2
0.0
10 lag
15
20
GARCH(1,1), ARMA(3,1)+GARCH(1,1), ARMA(3,1)+GARCH(1,1), ARMA(3,1)+GARCH(1,1), t normal t skewed t AIC Ar1 ar2 ar3 ma1 alpha beta -4.134258 0 0 0 0 0.13389 0.85514 -4.040692 0.55364 0.01263 -0.069487 -0.58672 0.13272 0.85570 -4.134697 0.7038 0.01662 -0.04825 -0.7336 0.1189 0.8843 -4.133303 0.70338 0.16594 -0.48111 -0.73270 0.12188 0.88402
Rt t ht
Where, should be WN(0,1). But since our GARCH does not capture all the characteris@cs of the model it is not in this case.