Forecasting issue of financial time series is an important subject
of finance. Nonlinearity is the widely accepted feature for the financial time
series. Discussions about the market efficiency theory are based on the fact
that the return series is a random walk. That means you cannot forecast the
prices. When the hypothesis of prediction of the return is rejected then this
supports the market efficiency theory which is tested by linear model. On the
other hand studies have shown that economic relations are nonlinear. The logic
that economic relations are nonlinear is the basis of the nonlinearity tests.
If we can prove the relation is nonlinear then we can try to forecast the
returns by using the nonlinear model. In this paper three of these tests, which
are BDS, Brock et. al (1996), Hinich Bispectrum Test, Hinich (1982), and
Lyapunov Exponent test, Nychka et. al. (1992), is applied to the
Istanbul Stock Exchange Index. The data is daily closing index between
02/01/1989 and 26/01/2001, which is 2995 observations, obtained by Central Bank
of Turkey.
Özet
Bu makalenin temel amacı
günlük IMKB getirisinin doğrusal olmayan bir yapıya sahip olup olmadığının
ortaya çıkarılmasıdır. Bu amaçla üç tane test yapılmıştır: BDS testi, Hinich
Bispectrum testi, NEGM Lyapunov üssü testi. Bu çalışmada kullanılan veriler
02.01.1989 26.01.2001, tarihleri arasında olup toplam 2995 tane gözlemden oluşmaktadır.
*Department of Business Administration,
Tel:
+ 90 216 345 86 29
Fax:
+ 90 216 345 86 29
E-mail: mcinko@marmara.edu.tr
1. Introduction
The chief corollary of the idea that markets are efficient, that prices
fully reflect all information, that price movements do not follow any patterns
or trends. This means that past price movements cannot be used to predict price
movements. Rather, prices follow what is known as a random walk, an
intrinsically unpredictable pattern. In fact Fama (1970) defined three types of
(informational) capital market efficiency, each of which based on a different
notion of exactly what type of information is understood to be relevant. In
particular, markets are weak form, semistrong-form, and strong form efficient.
The weak form of the Efficient Market Hypothesis (EMH) asserts that all past
market prices and data are fully reflected in asset prices. The implication of
this is that technical analysis cannot be used to beat the market. The
semi-strong form of the EMH asserts that all publicly available information is
fully reflected in asset prices. The implication of this is that neither
technical nor fundamental analysis can be used to beat the market. The strong
form of the EMH asserts that all information - public and private- fully
reflected in asset prices. The implication of this is that not even insider
information can be used to beat the market. Clearly, strong-form efficiency
implies semistrong-form efficiency, which in term implies weak-form efficiency,
but the reverse implications do not follow, since a market easily could be weak-form
efficient but not semistrong-form efficient or semistrong-form efficient but
not strong-form efficient. An immediate and direct implication of an efficient
market is that no group of investors should be able to consistently beat the
market using a common investment strategy. Recently the efficient markets
hypothesis and the notions connected with it have provided the basis for a
great deal of research in financial economics. Briefly stated, the hypothesis
claims that asset prices are rationally related to economic realities and
always incorporate all the information available to the market. This implies
the absence of exploitable excess profit opportunities.
Most of the empirical tests are designed to detect linear structure in
the financial data - that is, linear predictability is the focus. However, as
Campbell, Lo, and MacKinlay (1997) argue many aspects of economic behavior may
not be linear. Experimental evidence and causal introspection suggest that
investor’s attitudes toward risk and expected return are nonlinear. It is for
this reason that interest in deterministic nonlinear chaotic process has in the
recent past experienced a tremendous rate of development. In fact, the possible
existence of chaos could be exploitable and even invaluable. If, for example,
chaos can be shown to exist in asset prices, the implication would be that
profitable, nonlinearity-based trading rules exist (at least in the short run
and provided the actual generating mechanism is known). Prediction, however,
over long periods is all but impossible, due to the sensitive dependence on
initial conditions property of chaos.
Many studies show the nonlinearity of financial time series. Hsieh
(1993) found that the following financial time series are nonlinear. The data
set is consist of daily prices for currency futures contracts traded on the
Chicago Merchantile Exchange (CME), British Pound, Deutsche Mark, Japanese Yen,
Swiss Frank, from February 22 1985 to March 9 1990 (1275 observation). Abhyankar
et al. (1995) found the nonlinearity for the data set
consisting of 60.000 minute by minute real time returns on the UK-FTSE 100
index for the first 6 month of 1993. Hsieh (1989) found that the data consist
of daily closing bid prices of foreign currencies in terms of US dollars are
nonlinear. He used the following financial time series British Pound, Canadian
Dollar, Deutsche Mark, Japanese Yen, Swiss Frank. There are a total of 2510
daily observations from January 2 1974 to December 30 1983. Scheinkman et al
(1989) found the nonlinearity for the data set consisted of 5200 daily returns
(including dividends) on the value-weighted portfolio of the center for
research security prices at the University of Chicago. Abhyankar et
al. (1997) found the nonlinearity evidence for the data
set consists of real-time observations for the period September 1 to November
31 1991 at 1-minute frequency in the case of the FTSE-100, the Deutscher Aktein
Index (DAX), and Nikkei-225 and 15-second frequency for the S&P 500
futures. Totally six series are used, four published cash indexes (the
FTSE-100, the S&P 500, the DAX, and Nikkei) and two series constructed for
futures on the FTSE-100 and S&P 500. Barnett et al. (1997)
generate 5 different models of random number series in two categories one is
small sample size (380 observation) and large sample size (2000 observation).
Model I was fully deterministic, chaotic series. Model II was GARCH process.
Model III was a nonlinear moving average process. Model IV was ARCH process.
Model V was ARMA process. BDS test was successful 2 out of 5 model in the small
sample case and 3 out of 5 is ambiguous. In large sample case, BDS test, was
successful for all of the data series. Hinich bispectrum test was successful 3
out of 5 model in the small sample cases and it failed 2 out of 5 cases. In
large sample case test is successful in 3 out of five, ambiguous in one case
and failed in one case. Liapunov exponent test is successful in all cases.
Frank
and Stengos (1989) found low dimensional chaos. They used Correlation dimension
test to see whether the data series is chaotic or not. They used daily price
for gold beginning of 1975 to June 1986. They also used daily price for silver
beginning of 1974 to June 1986.
Brockett, Hinich, and Patterson (1988) use the test
for the following data set and found that: first series is both linear and
Gaussian but second and third series are both non-linear and non-Gaussian. The
first series examined was a sequence of values generated by a random number
generator that simulates Gaussian white noise. The series divided into 100
records, each containing 1024 samples. Each record was tested for linearity and
Gaussianity. Since the test statistics are one-sided only large positive values
are significant. Hence, for this series it is found that accept both linearity
and Gaussianity. The results of applying the bispectral tests to ten different
common stock series are: daily returns are realizations of nonlinear random
process. Although stock-return time series closely resemble white noisy, this evidence
in daily suggests a much higher degree of dependence in daily stock returns.
The series of stock-prices relatives is decidedly non-Gaussian and nonlinear.
Third example, In applying the preceding statistical tests to the analysis of
the forward and spot time series for foreign exchange rates, both the original
price quotes and the log of the price are examined. For the analysis U.S. dolar
and Japanese yen exchange rate are chosen. Two time period are examined, from
January 1,1981, to mid-1982 and from December 12,1981, to mid-1983. Daily
quotes for rates from The Wall Street Journal are used and it is found that
time series is nonlinear and non-Gaussian.
In the next section test statistics are going to be
introduced and in section three findings are going to be presented. In the last
section conclusion is going to be given.
2.
Nonlinearity Tests
In this
section each of the test statistics is going to be introduced. There are many
tests in the literature for nonlinearity. They are collected under the name of
Lagrange Multiplier tests and Portmanteau tests. BDS, Hinich and Lyapunov
exponent test are under the second group. BDS test is the most popular one
since the test has a power against all type of nonlinearity. Hinich bispectrum
test is a frequency domain test and it is based on the fact that linear process
has a constant skewness function. If the process is Gaussian then skewness
function is constant and also equal to the zero. Last test is based on the estimation of
lyapunov exponent. Posite lyapunov exponent is the operational definition
of chaos.
2.1
BDS test
This test for independence based on estimation of correlation integrals
at various dimensions. It has power against virtually all types of linear and
non-linear departure. While the estimation
of the BDS statistic is non-parametric, the test statistic asymptotically
follows a normal distribution with zero mean and unit variance and therefore
lends itself for easy hypothesis testing. BDS test is the most widely used
nonlinearity test may be because of no distributional assumption, or easily
applied to time series, or theory behind it.
In principle no distributional
assumptions need to be made about the data under the null hypothesis other than
that it is independent identically distributed (i.i.d). It can be interpreted
as a test for nonlinearity, if the appropriately used in conjunction with
Autoregressive Moving Average (ARMA) modeling. In the first step, the
best-fitted ARMA (p,q) is determined and fitted to the data, thus eliminating all
linearity from the data. Only in the second step is the test applied by running
it on the residuals of that ARMA model so that any dependence found in the
residuals must be non-linear in nature. Because of this reason, BDS test can be
used to produce indirect evidence for non-linearity. The order of the ARMA
model is decided by looking at the BIC (Bayesian Information Criteria).
Choosing the best ARMA means choosing the lowest BIC value.
Null hypothesis for the BDS test
is data-generating process is IID. The alternative hypothesis is not specified
that means the data generating process may be linear nonlinear or chaotic. In
practice the test has been used to examine whether fitted model’s errors are
IID. Test statistics is given by Brock et.
al. (1996)
~ N(0,1).
(1)
where Vm,n(e)
is the variance of Tm,n(e)
Tm,n (e) = C (N,m, e) - C
(N,1, e)m
(2)
H(z)
is the Heavside function which maps the positive arguments into one, and
nonpositive arguments into zero.
2.2 Hinich Bispectrum Test
Hinich (1982) developed a test to detect Linearity and Gaussianity of
the time series. His approach provides a direct test for Linearity and
Gaussianity. Tests statistic has a known asymptotic sampling distribution under
both Linearity and Gaussianity.
If
a time series is Gaussian and linear then its third moment is zero. Hinich test is a test in the frequency domain
of flatness of the bispectrum. Bispectrum is a double Fourier transformation of
skewness function. If linearity is tested then the null hypothesis will be
‘skewness function is flat’ or ‘lack of third order nonlinear dependence’. Z
will denote test statistics for linearity. If Gaussianity will be tested then
the null hypothesis will be ‘time series is Gaussian’ or ‘skewness function is
flat and equal to zero’. H will denote test statistics for Gaussianity.
Flatness of the skewness function is necessary but not sufficient condition for
general linearity and Gaussianity. However, flatness of the skewness function
is necessary and sufficient condition for third order nonlinearity.
Autoregressive (AR) or
Autoregressive Moving average (ARMA) generation process is the implicit
assumption of most of the time series. Let {xt} be of the form:
(3)
where
e t
are independent identically distributed random innovations with E(e (t))=0. If the input is Gaussian
then the output is Gaussian and its covariance function completely determines
the joint distributions of the process.
The Bispectrum is double
Fourier transformation of third order cumulant. The bispectrum, B(w1,w2), gives a
measure of the multiplicative nonlinear interaction of frequency components in
{xt}. For a real stationary time series with E x(t) =0, the
bispectrum is defined as follows:
Bx(w1,w2)=
(4)
Assuming that | Cxxx (m,n) | is summable. Given the symmetries
of Bx(w1,w2),
its principle domain is the triangular set
(5)
(6)
for all w1 and w2 in W
whenever {xt} is linear. If u(t) is Gaussian, m3 = 0 and thus
B (w1,w2)=0.
Nonlinear
dynamical systems can behave in ways that are hard to distinguish from a random
process. Nychka et al. (1992) introduce
a test statistic for chaotic time series. If a series is bounded and it has a
positive Lyapunov exponent then this series is called chaotic. Lyapunov
exponents are generalizations to nonlinear systems of the eigenvalues or roots
of linear system. There are three possible value for the Lyapunov exponent
where Lyapunov exponent denoted by l:
(i) l < 0,
Lyapunov exponent is less than zero system is convergent (ii) l = 0, Lyapunov exponent is equal to
zero indicate that the system is in some sort of steady state mode. (iii) l > 0, Lyapunov exponent is greater
than zero indicate a sensitive dependence on initial conditions, i.e. system is
chaotic. NEGM is a procedure for
testing for chaos by estimating the dominant Lyapunov exponent. As mentioned
before, positive Lyapunov exponent for a bounded system is the operational
definition of chaos. As is mentioned before Jacobian method is used to
calculate Lyapunov exponent where the neural network method has been used to
estimate this exponent. In the earlier studies BIC (Bayesian
Information Criteria) values were used, but in the more recent version of NEGM
approach GCV (Generalized Cross Validation) criteria has been using to determine the order of linear time series process
(i.e. the order to find the best combination of L, d, q). As
appropriate value of L, d, and q are unknown where q is the number of units in
the hidden layer of the neural network, d is the embedding dimension and L is
the time delay.
Given
a time series Nychka propose to test the Lyapunov exponent by using
nonparametric regression. Assume that xt are generated by nonlinear
autoregressive model.
(7)
The
estimation technique of the functional form of the equation (7) is
(8)
where
G(u) is the logistic distribution function. This technique finds dominant lyapunov
exponent by using nearal network.
3.
Data Analysis
The
daily closing index of Istanbul Stock Exchange is obtained from the Central
Bank of Turkey. The data,
which is 2995 observations, is between 02/01/1989 and 26/01/2001. Returns of
the ISE composite index are calculated as rt = ln (ISEt /
ISEt-1). Before the nonlinearity tests stationarity test should be
done and giving the descriptive statistics should be helpful to understand the
structure of the data. In table 1 descriptive statistics and normality test
Jarque-Bera statistic, which rejects the normality, are given.
Table 1 Descriptive Statistics
|
Observations 2995 |
|
Mean 0.002664 |
|
Median 0.002168 |
|
Std. Dev. 0.032416 |
|
Skewness 0.139312 |
|
Kurtosis 7.976217 |
|
Jarque-Bera
3099.870 |
|
Probability 0.000000 |
In
table 2 ADF (Augmented –Dickey Fuller test (1979)) and P-P (Phillips-Peron test
(1988)) test statistics are given. Both tests reject the unit root
hypothesis.
Table
2 Unit Root Tests Result
|
ADF Test Statistic |
-24.14516 |
1% Critical Value* |
-3.4356 |
|
|
|
|
5%
Critical Value |
-2.8630 |
|
|
|
|
10% Critical Value |
-2.5676 |
|
|
PP
Test Statistic |
-48.76425 |
1%
Critical Value* |
-3.4356 |
|
|
|
|
5%
Critical Value |
-2.8630 |
|
|
|
|
10% Critical Value |
-2.5676 |
|
|
*MacKinnon
critical values for rejection of hypothesis of a unit root. |
||||
3.1
BDS Test Result
By
using the minimum BIC criteria most appropriate ARMA model is found and it is
ARMA(6,6). Filtering the original return series by ARMA(6,6) residuals are
obtained. BDS test is done to these residuals and result shows that residuals
are not independently identically distributed. It is known that dependence is
not linear then it is nonlinear. BDS test statistics are given in the table 3. Critical value (for a =0.01) is 2.33. For all of the embedding dimensions BDS test reject the
null hypothesis of iidness that means returns are nonlinearly dependent.
Table
3 BDS Test Statistics
|
|
Embedding Dimension
|
||||||
|
e |
2 |
3 |
4 |
5 |
6 |
7 |
8 |
|
0.75 |
14.16 |
19.81 |
24.1 |
28.67 |
34.56 |
42.41 |
52.99 |
|
1 |
14.53 |
19.55 |
22.76 |
25.97 |
29.76 |
34.25 |
39.62 |
|
1.25 |
14.68 |
19.12 |
21.54 |
23.77 |
26.24 |
28.91 |
31.82 |
|
1.5 |
14.84 |
18.78 |
20.57 |
22.15 |
23.81 |
25.46 |
27.08 |
|
2 |
14.81 |
17.83 |
18.85 |
19.63 |
20.46 |
21.23 |
21.77 |
|
2.5 |
15.22 |
17.24 |
17.76 |
18.07 |
18.54 |
18.97 |
19.15 |
3.2
Hinich Bispectrum Test Result
As it is said that Hinich Bispectrum test is a
frequency domain approach and two hypotheses can be tested by this test:
Linearity and Gaussianity.
Linearity test:
This test is a test of whether there exist third
order nonlinear dependence or not (i.e tests the flatness of skewness function
or lack of third order nonlinear dependence). If we reject the null of
linearity then the series can be accepted to be nonlinear. However fail to
reject the null hypotheses does not mean that series is linear it only says
that series is not third order nonlinear. Test statistics is given in the table
4 shows that null hypothesis of linearity cannot be rejected.
Gaussianity test:
In this test we test whether the
daily ISE returns are Gaussian or not. In this case, the test statistic, is
1.00 is again smaller than the critical value which implies that the null
hypothesis of Gaussianity can not be rejected. That does not mean that ISE
returns are Gaussian, this test can reject Gaussainity but Gaussianity cannot
be accepted by this test since all Gaussian series have a flat skewness
function but inverse is not always correct.
Table 4 Result of
Hinich Bispectrum Test
|
Test Statistic Critical value Conclusion |
|
Linearity test 1.07 2.55 (a =0.05)
Fail to reject Linearity |
|
Gaussianity
test 1.00 Fail to reject Gaussianity |
3.3 Lyapunov Exponent Test Results
GCV rule shows that
best combination of (d,L,q) is (3,1,3). The null and the alternative of this
test can be written as
H0: Time series is
chaotic
H1: Time series is not
chaotic
By looking at the
estimated Lyapunov exponent for the time series appears to be -0.7480689. Since calculated Lyapunov exponent is less than zero
null hypothesis of positive Lyapunov exponent or chaos is rejected. Therefore,
based on the NEGM test it can be concluded that the daily ISE return is not
chaotic.
4. Conclusion
In this
paper three popular tests for nonlinearity are applied on the daily ISE return.
We have rejected the presence of linearity for ISE return by the BDS. This can
be considered to be significant evidence against linearity since this test is
accepted to be powerful tests against any departure from linearity. By NEGM
test we could not find evidence on chaos. Furthermore, by Hinich’s bispectrum
test we could not find any conclusion that supports a third order nonlinear
dependence. It can be concluded that we have found sufficient evidence for
nonlinearity in the ISE daily returns but the form of nonlinearity can be any
type except choatic or third order nonlinear.
REFERENCES
Abhyankar, A.,Copeland, L.S, Wong,W. (1995)
Nonlinear dynamics in real time equity
market indeces: evidence from the United Kingdom, Royal Economic
Society, 105 (July), 864-880
Abhyankar, A.,Copeland, L.S, Wong,W. (1997) Uncovering nonlinear
structure in real time
stock market indexes: The S&P 500, the DAX, the Nikkei
225, and the FTSE-100, Journal
of
business and Econnomic Statistic, January1997, vol15 no:1, 1-14
Barnett, W.A.,Gallant, R., Hinich, M.J., Jungeilges, J.A., Kaplan, D.,
Jensen, M. (1997) A single blind controlled competition among tests for
nonlinearity and chaos, Journal of
Econometrics 82, 157-192
Brock, W.A., D.,Dechert, B. Lebaron, and J. Scheinkman (1996) A test for
independence based on a correlation dimension, Econometric Review 15, 197-235
Brockett, P.L., Hinich, M., Patterson, D. (1988) Bispectral-based tests
for the detection of gaussianity and linearity in time series, Journal of
American Statistical Association September, vol 83 no:403, 657-664
Campbell, Y.J., Lo, A.W., MacKinlay, A.C. (1997) The Econometric of
Financial Markets, Princeton University Press
Dickey, D.A. and W.A. Fuller (1979) “Distribution of the Estimators for
Autoregressive Time Series with a Unit Root,” Journal of the American
Statistical Association, 74, 427–431
Fama, E. (1970) Efficient
capital markets: A review of theory and empirical works, Journal of Finance 25, 34-105
Frank,M.,Stengos,T. (1989) Measuring the stangeness of gold and silver
rates of return, Rewiev of Economic Studies, 56, 553-567
Hinich, M.J (1982) Testing for Gaussianity and linearity of a stationary
time series, Time Series Journal,3(3),169-176
Hsieh, D.A. (1989) Testing for nonlinear dependence in daily foreign
exchange rates, Journal of Business vol
62 no:3 339-369
Hsieh, D.A. (1993) Implication of nonlinear dynamics for financial risk
management, Journal of Financial and
Quantitative Analysis, vol 28 no:1, 41-64
Jarque, C. and A. Bera (1980) "Efficient Tests for Normality,
Homoskedasticity, and Serial Independence of Regression Residuals,"
Economics Letters, 6, 255–259.
Nychka, D., Ellner, S., Gallant, R., and McCaffrey,D. (1992) Finding chaos
in noisy systems, Journal of The Royal Statistical Society B, 54, no:2, 399-426
Phillips, P.C.B. and P. Perron (1988) “Testing for a Unit Root in Time
Series Regression,” Biometrika, 75, 335–346
Scheinkman J.A, B. LeBaron (1989) Nonlinear dynamics and stock returns,
Journal of Business vol 62, no:3, 311-337