Sie sind auf Seite 1von 4

Research on the mixed method of option price

forecasting based on support vector machine


Li Meng
School of Management, Xiamen University, PR China
Email: lmeng@xmu.edu.cn
Abstract In this study, the authors discuss a mixed
method which involves some forecasting techniques like
linear technique (ARMA), BS option pricing function, and
mixed BRF. These forecasting techniques are combined by
support vector machine (SVM) to forecast the option price.
This mixed method can avoid disadvantages which the
single forecasting techniques have, and be more accurate
and stable to forecast the option prices when the maturity
is coming. The rationality of the method has been proved
by experimental result which is based on the foreign
exchange option price data.
Keywords ARMA, BS, BRF, SVM, RMSE.
I. INTRODUCTION
Since the Black and Scholes modeled classic BS
formula in 1973, the option pricing has become much easier
and more accurate than before, so that financial derivative
product developed rapidly. However, the hypothesizes of BS
do not agree with the financial market reality. Lots of deputes
on the rationality of BS are always existing. In recently year,
A new method--the parameter-less methodwas used into
this field. There into, the neural network and genetic
algorithm are the most popular. Its the first time to use
neural network into option pricing by Hutchinson Lo and
Poggio
[1]
in 1994, the arithmetic was based on BP and BRF
neural network to approximate the BS function by using S/X
and T-t as input parameters and C/X as outputs. In 1996, Qi
&Maddala
[2]
used the parameters of BS function(except
volatility) as neural network input, S (option price) as output,
and S&500 as sample to obtain a better forecasting result
than BS. But, there still were limits on the research because
of lack of parameter input and disadvantage of network.
In 1997, Laibcygier P.
[3]
provided a mixed neural
network model. This model used the BP neural network(NN)
to fit the difference between the BS calculated value and the
actual value, and then the fitted difference was added up to
the calculated value to get a better forecasting option price,
and this method had been proved by experimental research.
In 2004, Panayiotis C. Anddreou
[4]
provided a kind of mixed
neural network which do not only used the difference
between the BS calculated value and the actual value as the
input, but also imported implied volatility, and got a result
which was better than normal neural network, BS and CS
Corrado&Su (1996). Wide use of neural network make
a large development on option pricing nonlinear research.
One of the most popular neural network is BP, but there is a
critical limiteasy to fall into local extremum. Recently,
Zhang Yanhong etc.(2008) combined wavelet neural network
with genetic algorithm to get a intelligent nonlinear
arithmetic.
These years, a new intelligent optimization algorithm
support vector machine becomes popular. It has lots of
advantages which the neural network havent. It can be ensured
to get the global optimum solution, avoid the dimension
disaster, and have a strong generalization ability. It has been
widely used into financial forecasting. Zhang Yuchuan(2007)
used SVM to judge the a single stocks fluctuation in the
futurehighs or lows, and the accuracy rate reaches 60%. It
provides a new way for investment decision.
Wang Shouyang etc. proposed a mixed method based on
the foreign exchange rates forcasting model ,which was
composed by a traditional linear econometrical models, an
artificial neural network models and a text mining and rule-
based expert systems techniques. And all of these techniques
were combined by SVM to formulate TEI@I method for
foreign exchange forecasting. As one of financial derivative
products, foreign exchange option market has the similar law as
the stock or foreign exchange market. Based on this, the study
tries to formulate ARMA, RBF and BS by using SVM.
Thereinto, ARMA are used to forecast the main trend of the
option price, the nonlinear components of prices are analyzed
by using RBF, while the impacts of irregular and the infrequent
future factors on the option prices are not considered , because
we consider that the emergent events directly influence foreign
exchange price, and the foreign exchange price influence
foreign exchange option price directly, so we think the
emergent events influence the option price through the foreign
exchange price indirectly. The emergent events impacts on
foreign exchange option can be reflected by the foreign
exchange price which has been used as one of input parameters.
And according to experimental results, one can see that the
mixed method this study proposed are effective.
Financed by the Science Foundation of Educational Program
China (Grant No. DIA080119)
6755
978-1-4577-0536-6/11/$26.00 2011 IEEE
II. THREE SINGLE FORECASTING METHOD
A. Traditional linear econometrical models ARMA
The ARMA is the integration of autoregressive model
AR(p) and moving average model MA(q), it can be seen as the
general form for AR(p) and MA(q). So the ARMA model can
be written in ARMA(p, q).
There is a preconditions to be obeyed for the time series
which will be analyzed by ARMA: the time series must be
stable, namely it should approximate to a discrete stochastic
process. But, the financial time series usually has some trend
which is not exactly stochastic, for instance, it may have a total
increasing or decreasing trend. Especially for the option price
series, there is an obviously decreasing trend when the maturity
date is coming. In this condition, difference method can be
used to make the time series stable.
B.RBF neural network
RBF neural network has been proposed by J.Moody and
C.Darken in 1980s, it can approximate to any continuous
functions with any precision. Compared to BP network, BRF
usually needs more neurons, but it has faster calculating speed.
Its hidden layer adopts radial basis function-which is usually
Gauss Function- as an activation function. The network graph
is shown as following:
C. Support vector machine(SVM)
SVM is a kind of new machine learning method, also a new
method of data mining. It complies to structural risk
minimization principle, and has good learning and
generalization ability. It can deal with small sample problem,
nonlinear problem, dimension disaster, local extremum
problem effectively, and can minimize the empirical error and
maximize geometric marginal division. SVM has classification
and regression abilities. In this paper, the regression ability is
used. Assuming that a separable sample group is ( , )
i i
x y
1, 2, 3,.... i n = ,
d
i i
x R y R e e x
i
is input variable, y
i
is
expectation output variable. Then vector X is mapped to high
dimensional feature space D by nonlinear map | . And by
linear approximation, function f can be find to approximate the
given vector. Namely linear function
( ) ( ) f x x b e| = +
(*) is
used to fit hyper plane. Then there is optimization problem as
following:
2
*
1
1
min ( ),
2
n
i i
i
C e
=
+ +

s.t.
*
*
,
, 0, 1,...,
,
i i i
i i
i i i
y x b
i n
x b y
e c

e c
s +

> =

+ s +

where C is used to control the complexity of the model and


approximation error.
D. Introduction for SVM Mixed Method
In this section, a new option price forecasting method is
provided, called SVM Mixed Method (SMM). One know that
foreign exchange option is a kind of financial products, and
has close connection with foreign exchange. But It is harder to
forecast because of its much bigger volatility, especially when
the maturity date is coming. So one may need a more complex
model to forecast foreign exchange option prices . We know
that regression technique can describe the total trend of time
series, neural network can reflect the nonlinear trend of time
series effectively, SVM can minimize the empirical error and
maximize geometric marginal division, and deal with the
small sample problem effectively. The reaching idea of the
mixed model in this study is enlightened by the foreign
exchange forecasting model in paper [9]. The authors integrate
the each single technique with SVM to found a mixed foreign
exchange option forecasting model, and wish the model inherit
the advantages of each technique, and have a better forecasting
effect. The concrete structure is as following:
Fig. 2 Mixed foreign exchange option forecasting model
As the figure shows, P is the real foreign exchange price. is
foreign exchange implied volatility, r
f
and rd are respectively
the foreign and domestic riskless rate. St is the target foreign
exchange price. T is the time segment from right now to the
maturity date (unit: year). P
BS
, P
RBF
and P
ARMA
are respectively
the price forecasted by BS formula, RBF and ARMA. Before
each forecasting data group being input into the SVM for
integrating, they have to access a data adjustment process
make some datas which deviate from the reality seriously
return to zero, such as negative values or too large value. P
f
is
the final forecasting value. Through the figure, we are able to
recognize the concrete calculating flow.
B-S
x
x
x
r
r
r
P
a
a
b
b
b
a
n
ARIMA
P
P
BS
P
RB
F
d
P-
BS
D
a
t
a

t
o

A
d
j
u
s
t
S
V
R
6756
. EXPERIMENT AND ANALYSIS
The sample data is from China Merchants Bank (CMB).
Foreign exchange option: EURUSD, GBPUSD maturity date:
4/8/2009, from 1/13/2009 to 4/8/2009; Foreign exchange:
EURUSD, GBPUSD, from 1/13/2009 to 4/8/2009; each
countrys libor data is used to be the freeless rate. And
implied volatility is used to calculate the option price in this
study. (http://www.econstats.com/r/rlib__d11.htm)Table 1 is
the particial sample data of EURUSD.
Tab. 1 EURUSD
P St rd
3.2498,3.3645,
3.009,3.2251,
2.6472,2.2581,
2.5127,2.3326,
2.2293,2.8656,
2.5925,1.9287,
1.6619,1.5999,
1.9981,1.4999,
1.2808,1.4763,
1.4147,1.4464,
1.3222,1.3224,
1.1968,0.9704,
0.7718,0.697,
0.7029,0.964,
...
1.3198,1.3216,
1.3155,1.3264,
1.309,1.2882,
1.3001,1.3004,
1.2965,1.3211,
1.3146,1.2932,
1.282,1.2812,
1.3009,1.2851,
1.2788,1.2942,
1.2958,1.291,
1.2864,1.2881,
1.2861,1.2797,
1.2582,1.2565,
1.2656,1.2827,
...
0.0114938,0.0115688
,0.0116094,0.011659
4,0.0117688,0.01191
88,0.012075,0.0122,0
.0123188,0.012275,0.
0122625,0.0122219,0
.0122281,0.0122688,
0.012875,0.0129938,
0.0130875,0.0131563
,0.0132,0.0132594,0.
0133125,0.013125,0.
012925,0.0128375,0.
0127663,0.0127125,0
.0126625,0.0126438,
...
rf T
0.026075,0.025612
5,0.025,0.0245,0.0
24075,0.0236875,0
.023075,0.0225313
,0.0219,0.021425,0
.021325,0.021175,
0.0210375,0.02090
63,0.0207625,0.02
06125,0.0205125,0
.0203875,0.020212
5,0.0200438,0.019
925,0.019775,0.01
95625,0.01935,0.0
192313,0.0191125,
...
0.235616438,0.232876712,0.230136986
,0.22739726,0.219178082,0.216438356,
0.21369863,0.210958904,0.208219178,
0.197260274,0.194520548,0.191780822
,0.189041096,0.180821918,0.17808219
2,0.175342466,0.17260274,0.16986301
4,0.161643836,0.15890411,0.15616438
4,0.153424658,0.150684932,0.1424657
53,0.139726027,0.136986301,0.134246
575, ...
A. Founding ARMA
The software Eviews is utilized to found and test
ARMA model. First, the time series P should be stable. We
draw the scatter diagrams of P, and find the obviously
decreasing trend. So the time series P is considered as instable
series. It need to be treated by second-order difference. The
treated series is shown as Fig. 4.
Calculating the autocorrelation function and partial
autocorrelation function of the treated series, which are
tending to zero from the second order and the third order
respectively, so both of them are trailing. Through trail
method, we find that ARMA1 2 has a preferable
goodness fit and less AIC. The model is as following:
1 2
0.8491 0.9804
t t t
P P c

=
the intuitive fitting condition can be seen as Fig.5. The
forecasting values are shown in Tab.2.
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
2009M01 2009M02 2009M03
P
Fig.3 The sample data of EURUSD(P)
-1.2
-0.8
-0.4
0.0
0.4
0.8
1.2
2009M01 2009M02 2009M03
DDP
Fig.4 The treated series by second-order difference
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
2009M01 2009M02 2009M03
PF P
Fig.5 regression result by ARMA
B. The forecasting results calculated by SVM
Seeing in Fig.2, we can realize that the input parameters
of SVM are P
BS
, P
RBF
, d
P-BS
, P
ARMA
. Each forecasting results
are shown in Tab.2:In Table 2, P is the sample option price
from 3/19/09 to 4/8/09; BRF is the forecasting price
forecasted by BRF during the same period. The input
parameters for BRF are , r
f
, r
d
, St, T, whose time series are
Tab.2 The forecasting results for each method(EURUSD)

P

BRF

ARMA

BS
2.2717 2.3004 1.7583 2.3115
1.8683 1.909 2.3712 1.7531
1.7734 1.7465 1.8050 1.776
1.1786 1.2016 1.7309 1.0304
1.4427 1.4473 1.1377 1.4087
1.1983 1.2009 1.4857 1.1127
0.5717 0.52333 1.1107 0.3351
0.3011 0.32534 0.5238 0.095579
0.3199 0.32897 0.5238 0.12761
0.2402 0.22412 0.2296 0.085632
0.4414 0.51764 0.1751 0.40188
6757
0.2536 0.12555 0.4177 0.35146
0.107 0.46182 0.1755 0.012793
0.1096 -1.4459 0.0695 0.00065139
0.1066 -6.3381 0.033 1.20E-06
RMSE(7) 1.4835 0.18048 0.14999
RMSE(8) 2.7954 0.16939 0.14459

=-

=
mBRF
=+
svr
-0.0398 0.0030589 2.314559
0.1152 0.023488 1.776588
-0.0026 0.006204 1.782204
0.1482 0.13181 1.16221
0.034 0.032183 1.440883
0.0856 0.085642 1.198342
0.2366 0.25335 0.58845
0.205521 0.13891 0.234489 0.28716
0.19229 0.19513 0.32274 0.33702
0.154568 0.16111 0.246742 0.24965
0.03952 0.023277 0.425157 0.4786
-0.09786 -0.11856 0.2329 0.23061
0.094207 0.23291 0.245703 0.29962
0.108949 0.19441 0.195061 0.12208
0.106599 -6.2414 -6.2414 0.020446
0.072712 0.081393
2.4003 0.08209
from some day to 40 days ago counted reversely; ARMA is
the forecasting price forecasted by ARMA from 3/19/09 to
4/8/09; BS is the price calculated by BS from 3/19/09 to
4/8/09; =- is the difference of P and BS, used to be
training group for ; is the forecasting values of
forecasted by BRF; mBRF is the sum of and , which is
the simplest mixed method; svr is the time series forecasted
by SVM, the input datas are from during the first
data to the 13
th
one, and the training data is from during the
first data to the 13
th
one.
Fig 6 The forecasting results of six kind of methods(EURUSD)
Where P is the real sample data. This figure can show us a
more intuitionistic comparison among each method. It prove
SMM is more stable than the other methods once again.
Although at any time, the ARMA do not depart from P very
far, but its RMSE is much higher than SMM. Based on an
overall consideration, The SMM is better than the other
method mentioned in this study.
. CONCLUSION
In this paper, more than one forecasting techniques are
integrated by SVM, then a new mixed forecasting method
SMM is produced. This method can keep the result in a
reasonable scope, because it can adjust unreasonable input
parameter values to be zero, so as to control the final results.
It can integrate each methods advantage, so as to keep the
forecasting results stable. It would be a ideal method for
forecasting the financial products which are seriously risky
and fluctuant.
REFERENCES
[1]Hutchinson et al. Anonparametric approach to pricing and hedging
derivative securities via learning networks [J]. Journal of Finance, 1994. 851-
889.
[2]Qi &Maddala and S Maddala. Option Pricing Using ANN: The Case of
S&P 500 Index Call Options. Proceeding of 3rd International Conference on
Neural Networks in Capital Markets, London, 1996, 78-91.
[3]Laibcygier P., Flitman A., Swan A. and Hyndman R. The Pricing and
Trading of Options Using aHybrid Neural Networks Models with Historical
Volatility [J]. Neurovest Journal, 1997vol. 5, No 1, 27-41.
[4]Panayiotis C. Andreou et al. Martzoukous. Pricing and Trading European
Options by Combining Artificial Neural Networks, Single Factor Parametric
Models, and Implied Parameters. 2004 FMA European Conference, 2004.
June 2-5.
[5] Xie Guangjun, Zhuang Zhenquan. A Step-wise Prediction Model in
Options Pricing [J]. system engineering. 2000, 7(100): 28-31
[6] LIU Feng-qin, MA Jun-ha. Review on Artificial Neural Network
Valuation Methods for Pricing Financial Derivative Securities[J]. Collected
Essays on Finance and Economics. 2008, 5 (137): 47-53.
[7] ZHANG Hong-yan, LIN Hui. Study on Hong Kong Derivative Market by
Applying Hybrid Wavelet neural Network and Genetic Algorithm [J]. Journal
of Systems & Managemen. 2008, 2 (1): 25-31.
[8] SHI Yan-jie. Stock Price Forecasting based on Support Vector Machine[J].
STATISTICS AND DECISION. 2005,2:123-125.
[9] ZHOU Wan-long, YAO Yan. Application of Support Vector Machine on
stock market short-term forecasting[J]. COMERCIAL RESEACH.
2006,6(338): 160-162.
[10] WU Meng, XU Quan-zhi. Application of Support Vector Machines in
Financial Time Series Forecasting[J]. Journal of University of Electronic
Science and Technology of China. 2007,4(2): 442-444.
[11] ZHANG Yu-chuan, ZHANG Zuo-quan. Application of Support Vector
Machines in Stock Price Predicting [J]. JOURNAL OF BEIJING JIAOTONG
UNIVERSITY. 2007,12(6): 73-76.
[12] WANG Shouyang, YU Lean, LAI Kinkeung. TEI@I Methodology and
Its Application to Exchange Rates Prediction [J]. Chinese Journal of
Management. 2007,1(1):21-27.
[13] FEISI Technology Research and Development Center. Neural network
theory and the realization of MATLAB7[M]. Publishing House of Electronics
Industry.2005,3: 116-120.
[14]LE Li-hua, WENn Rong-sheng, ZHU Hui. Stock Market Forecast Based
on RBF Neural Network and MATLAB Realization. SCI-TECH
INFORMATION DEVELOPMENT & ECONOMY. 2008(30):151-152.
1 2 3 4 5 6 7 8
-7
-6
-5
-4
-3
-2
-1
0
1
P
BRF
ARMA
BS
mBRF
svr
6758

Das könnte Ihnen auch gefallen