Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
The first author would like to thank University College London (UCL), for awarding Overseas Research Scholarship (ORS) and Graduate Research Scholarship (GRS) funding.
FTSE 100 Returns and Volatility estimation using Higher Order Neural NetworksJanti Shawash, Dr David R. Selviah
Department of Electrical and Electronic Engineering
Conclusion
Simulations
Results
CHONN-EGARCH
CHONN-GARCH
Linear-EGARCH
CHONN
1.55 1.60
CHONN-EGARCH
CHONN-GARCH
Linear-EGARCH
CHONN
1.0 1.2 1.4
Data
Kurtosis = 1.7157Skewness = -0.0941
! = 882.5572
Year Frequency
FTSE
100
daily
pri
ce se
ries
3000
4000
5000
6000
7000
2002 2004 2006 2008 0 50 100 150 200 250
(a)
Kurtosis = 9.1589Skewness = -0.0976
! = 1.3487
Year Frequency
FTSE
100
daily
retu
rn se
ries
(%)
-10
-5
0
5
10
2002 2004 2006 2008 0 50 100 150 200 250
(b)
Auto
Corr
elat
ion
Coef
ficie
nts
-0.4
-0.2
0
0.2
0.4
Lag (n)0 5 10 15 20 25 30
Lag (n)0 5 10 15 20 25 30
(c) (d)
Daily returns Daily returns2
# inputs
# Hiddenneurons
#parameters
0
500
1000
0
100
11
NN CHONN HorizHONN CPHONN FXPHONN
Returns Volatility
Kurtosis = 7.9794Skewness = -0.3904
! = 1.335
Year Frequency
Residual
-10
-5
0
5
10
2002 2004 2006 2008 0 50 100 150 200
(a)
Auto
Cor
rela
tion
Coe
ffici
ents
-0.4
-0.2
0
0.2
0.4
Lag (n)0 5 10 15 20 25 30
Lag (n)0 5 10 15 20 25 30
(c) (d)
Residual Residual2
Year Frequency
Dai
ly v
olat
ility
0
1
2
3
4
5
2002 2004 2006 2008 0 20 40 60 80 100 120 140
(a)
Kurtosis = 3.2290Skewness = -0.2720
! = 1.0007
Year Frequency
Stan
dard
ized
Resi
dual
-10
-5
0
5
10
2002 2004 2006 2008 0 20 40 60 80 100
(b)
Auto
Cor
rela
tion
Coe
ffici
ents
-0.4
-0.2
0
0.2
0.4
Lag (n)0 5 10 15 20 25 30
Lag (n)0 5 10 15 20 25 30
(c) (d)
Standardized Residuals Standardized Residuals2
J =δLayer
δWLayer
ModelMSE = min
W(1
N
N�
t=1
�2t )
Wnew = Wold − [∇2J − µI]−1J�
Time
0
5
Epochs
510
HRo 50
52
HRi 55
60
RMSEo
1.41.61.8
RMSEi
1.01.2
AIC
1.8
2.0
2.2
NN CHONN HorizHONN CPHONN FXPHONN
MSE MAE
Acknowledgements
(a) FTSE 100 daily price series(b) FTSE 100 daily returns series(c) Autocorrelation coefficients or returns and returns-squared
A graph showing the simulations of networks with input size range of [1-11] input lags. Hidden neurons in the range of [0-10] in models with a single output neuron. The stripped pattern in this graph matches the (input-hidden) parameter count of the graph on the far left. We observe that the in-sample (RMSE, HR) performance increases with the number of parameters in the estimation model. Out-of-sample (RMSE, HR) performance is better in smaller models. The training time was also correlated to the number of parameters. However, the epochs required for convergence were independent of the number of parameter with a mean value of 12 epochs. Two minimisation criteria were used for training as indicated in the legend.
(a) FTSE 100 Volatility estimate. (c,d) Autocorrelation coefficients or returns and returns-squared. (b) FTSE 100 daily returns residual after volatility compensation.Results show that most of the volatility and returns information was extracted leaving a residual close to a normal-Gaussian distribution.
(a) Residual of FTSE 100 returns estimation using the Correlation Higher Order Neural Network.
Even after returns estimation Kurtosis and ! levels remained the same.
HorizHONN
CHONN
CPHONN
FPXHONN
Linear
NN
1.32 1.34 1.36
CHONN
HorizHONN
NN
CPHONN
FPXHONN
Linear
1.65 1.70
Linear
HorizHONN
CHONN
CPHONN
FPXHONN
NN
58 60
AIC RMSEo Hit Rate RMSEo
AIC
(c) Autocorrelation of Residuals(d) Autocorrelation residuals-squared.
These results show that the autocorrelation levels were reduced in the residuals domain, however; Residuals-squared stayed at the same level indicating no change in volatility.
AIC performance improvement* when using HONNs over standard linear models.
Out of sample Root Mean square error slightly improved as well.
Hit Rate performance was worse for Linear model andthe best was NN.* AIC variation was in the 3rd significant figure indicating consistent performance.
Both AIC and RMSEo indicators were improved significantly after extracting the volatility information from the residual of the FTSE-100 index daily returns.
The FTSE 100 series was converted into a returns series with properties closer to a stationary Gaussian distributed data set using the following equation.
W11
W21
W22
W23
W2M
W12
W1N
yt!n
yt!2
yt!1
yt
NeuralNetworkHigher
Order Function
!!
t
2
!1
!2
!q
!t"1
2
!t"2
2
!t"q
2
!t"1
2
!t"2
2
!t" p
2
!1
!2
!p
Feedback
GARCH
Number of parameters of a Neural Network and 4 Higher Order Neural Networks.
Higher Order Neural Network output y(t) as a function of Weights (W,b) and inputs Xi.
Optimisation criterion for network weight update estimation.
Error gradient estimation represented as the Jacobian matrix
Weight update using Levenberg-Marquardt algorithm.
Root Mean Squared Error, evaluation criteria used to benchmark performance of model estimation.
Hit Rate (HR), is an estimate of the rate of correct direction of the estimated output.
Generalised AutoRegressive Conditional Heteroskedasticity (GARCH)
Volatility estimation from previous error values along with the previous volatility estimate.
Log-Likely-hood minimisation criterion that incorporate changes in volatility along with the error.
Akaike information criterion, used to select the best models based on the error and the number of parameters used for estimation.
This work compared Higher Order Neural Networks (HONN) with Neural Networks, and linear regression for short term forecasting of stock market index daily returns.
Two new HONNs, the Correlation HONN (CHONN) and the Horizontal HONN (HorizHONN) outperform all other models tested in terms of the Akaike Information Criterion, out-of-sample root mean square error, of FTSE100 and NASDAQ giving out-of-sample Hit Rates of up to 60% with AIC improvement up to 6.2%. New hybrid models for volatility estimation are formed by combining CHONN with E/GARCH are compared with conventional EGARCH, providing up to 2.1% and 2.7% AIC improvement for FTSE100 and NASDAQ.