Microsoft Word - 12_Kliber_P.docx DYNAMIC ECONOMETRIC MODELS Vol. 11 – Nicolaus Copernicus University – Toruń – 2011 Paweł Kliber Poznan University of Economics Jumps Activity and Singularity Spectra for Instruments in the Polish Financial Market† A b s t r a c t. In the paper we try to measure the activity of jumps in returns of some instruments from the Polish financial market. We use Blumenthal-Getoor index β for Lévy processes as a measure of jumps’ activity. This allows us to distinguish between processes with rare and sharp jumps and the processes with infinitely-active jump component. We use three different methods. First we use activity signature plots to estimate the activity patterns of jumps. Then we estimate the Blumenthal-Getoor index with Aït-Sahalia and Jacod threshold estimator.Then we use meth- ods based on singularity spectra of Lévy processes. Finally, we compare the results. K e y w o r d s: Blumenthal-Getoor index, singularity spectrum, Lévy exponential models. Introduction The classical models of assets’ returns are based on the assumption of the nor- mality of returns. This is for example the case of classical portfolio theory, de- veloped by Markowitz (1952) and Sharpe (1963), and the option pricing formu- la, derived by Black and Scholes (1973) and by Merton (1973). However it is well-known that the normality assumption does not hold. Many well-established stylized facts about assets’ returns contradict this. The observed returns reveal characteristics such as heavy tails, high kurtosis or volatility clustering, which is not consistent with Gaussian distribution1. There are several methods to deal with non-normality of returns and to make models better fitted to observed data. The most popular approach is the modeling the conditional volatility with some kind of ARCH/GARCH model. † The research was undertaken in the project sponsored by the Polish Ministry of Science and Higher Education, N N111 436 534. The author would like to thank two anonymous referees for valuable comments. 1 The survey of stylized facts concerning assets’ returns can be found in (Cont, 2001). Paweł Kliber 172 When using non-Gaussian distributions of error terms, such models seem to be well-fitted to the data. In this work we take another approach, which lately becomes more and more popular, namely we use Lévy processes in the modeling. Although this line of modeling is as old as the seminal paper of Maldenbrot (1963), in which models with stable distributions were proposed, it became more popular at the beginning of XXI century. Lévy processes can be represented as a sum of continuous diffusion (Wiener process) and discontinuous jumps. Thus, one of the main questions is if there are jumps in the returns of financial instruments and if there are, how much intensive are the jumps. Formally the intensity of jumps of Lévy process is de- scribed by Blumenthal-Getoor index. In the paper we try to estimate the value of this index for the suitably-chosen sample of instrument from Polish financial market. We use three different methods of estimation and then we compare the results. The paper is organized as follows. In the section 1 we present basic facts about Lévy processes and Lévy exponential models of prices. In this section we also introduce the concept of Blumenthal-Getoor index of jumps. In section 2 we try to analyze the type of processes using activity signature plots. In section 3 we estimate Blumenthal-Getoor indexes using threshold estimator. In section 4 we analyze the type of processes using its singularity spectrum. Section 5 concludes. 1. Lévy Processes and Lévy Exponential Models The Lévy process L is the stochastic process with continuous time that starts at zero (i.e. 0 0L  ) and fulfils the following conditions: 1. for any 1 2 3 40 t t t t    the random variables 12t tL L and 4 3t tL L are independent and the distribution of t h tL L  depends only on h and not on t , 2. the process is stochastically continuous, i.e. for all 0t  and all 0    0 lim 0 h t h tP L L      3. the trajectories of the process are cadlag (i.e. they are right-continuous with left limits). We should stress that the second condition does not imply that the process is continuous. In fact the Lévy processes typically have discontinuities (of “jumps”) and some of them are discontinuous at each point t . The condition 2 means only that the process L does not have the jump of arbitrary size  at any pre-specified moment t . Jumps Activity and Singularity Spectra for Instruments in the Polish Financial Market 173 The Lévy processes can be seen as the extension of Wiener process. In fact if we change the condition 2 and require the process L to be continuous, then the only processes that fulfill the definition are Wiener processes with drift. There is also another connection between Wiener process and Lévy processes. The increments of a Wiener process are normally-distributed, while the incre- ments of Lévy process belong to the family of infinitely-divisible distributions. It is the biggest family of distributions that can serve as limits for the sums of independent variables (see Feller, 1971). Thus, if we believe that the returns of the assets in financial markets result from many independent shock, then the Lévy processes are natural choice as an engine in the model. There are two fundamental theorems that reveal the structure of the Lévy processes. The first one is the Lévy-Itô decomposition, which states that any Lévy process L can be uniquely represented as the sum of Wiener process with drift, Poisson process and purely discontinuous martingale: l s t t t tL XW Xt    , (1) where  and 0  are constants, W is a standardized Wiener process, lX is a compound Poisson process of large jumps and sX is purely discontinuous martingale (i.e. it is discontinuous at each point t ). The process sX represents small jumps – in every finite interval there are infinitely many of them, but they are very small, so that the process sX does not explode. The second theorem characterizes the characteristic function of the Lévy processes. According to Lévy-Khintchine representation the characteristic func- tion of the Lévy process is: [ ] exp[ ( )]tiuLE e t u , (2) where the characteristic exponent  equals:   2 2 11 1( ) ( )2 iux x R e iuxu i u u xd         . (3) The first two terms in the sum (3) are the same as in the characteristic function for the Gaussian distribution. These terms describes continuous part of the pro- cess (diffusion). The measure  (called Lévy measure) describes the jumps of the process L . The value of ( )R is the intensity of jumps. If finite, it is the average number of jumps in the unit of time – the process is then said to be “finitely active”. If ( )v R   , then the process is infinitely active – in any in- terval the number of jumps is infinite. The values (( , ))v a b give the relative intensity of jumps with sizes between a and b (i.e. jumps such that ( , )t t tLL aL b   ). Thus the Lévy measure contains information of both the Paweł Kliber 174 intensity of jumps and distribution of jumps’ sizes. The triple ( , , )   , called “characteristic triple”, gives unique characterization of the process. The measure  , which contains all information of jumps and its structure, can vary, depending on the type of the process or on the probability distribu- tions of increments of the process. There exists however one synthetic index that divides measures  into certain categories and gives characteristics of jump behavior. Blumenthal-Getoor index is defined as 1 1 inf 0 : ( ) b db x x               . (4) The index  takes values in the interval [0, 2) . Higher values of  mean that jumps are more intensive and smaller and the process L resemble continuous process. If 0  the process is finitely active. All other values mean that the activity of the process is infinite. In the case 0  the discontinuous part of the process L is the compound Poisson process, i.e. 0sX  in the decompo- sition (1). The models of assets prices driven by Lévy process usually take the form of exponential Lévy models: 0 tL tS S e , (5) where tS denotes the asset price at time t . The models can be also formulated as stochastic differential equation (as in classis Black-Scholes model): t t tdS S dV , (6) where V is a Lévy process, whose characteristics can be derived from L 2. The logarithms of the prices are described by Lévy process L and thus the logarith- mic returns are increments of Lévy process. According to (1) the logarithmic price is given by: ln l st t t t tts XWS X     . (7) In the literature one considers also the generalization of (7), assuming that the volatility of continuous part is not constant. Such a model can be specified as follows: d t t t ts Wt L   , (8) where t is a process or a function representing volatility and dL is discontinu- ous part of the process L and represents jumps. 2 Kallsen (2000) has shown that specifications (5) and (6) are equivalent and gave the formu- lae how to express V in terms of L and vice versa. Jumps Activity and Singularity Spectra for Instruments in the Polish Financial Market 175 Let us select some time-scale, i.e. choose some frequency in which we sam- ple the process. Suppose that we sample every h units of time (in practice it can be a period from several seconds to one day). The increments of the process at the specified frequency are logarithmic returns of the assets. We denote them by ( )ix h : ( 1) ( 1)( )i i h ih i h ihh s sx L L    . (9) If there is no ambiguity about frequency we omit brackets and denote returns simply by ix . In the next three sections we will use data about returns to calcu- late index  , using three different methods. 2. Estimating Jumps Activity with Activity Signatures The first method is based on power variation of stochastic process, defined as: 1 ( , ) ( ) p i i N V p h x h    . (10) In the case 2p  , the (2, )V h is well-known realized volatility, which tends to the quadratic variation of the process, as the sampling frequency tends to infini- ty. The behavior of ( , )V p h in other cases depends on the type of the process. As Barndorff-Nielsen and Shephard (2002) have shown, if the process L con- tains no jumps (i.e. 0dL  in (8)), then: 1 / 2 0 0 plim ( , ) T pp s h pph shV d      , (11) where p is appropriate constant. If the process L contains no diffusion part ( 0t  ), then as 0h  the sum (10) diverge for p  . On the other hand for p  the sum is convergent: 0 0 (plim , ) T p s h s p h LV     . (12) If the process contains both diffusion and jump parts, then the following equa- tions holds: Paweł Kliber 176 1 2 / 2 0 0 2 0 00 0 0 plim fo( , ) , (0, 2)r pli , (2, ) , ,m for 2 pli (m , for 2, .) T pp p s T T s s s T p s h h s h p h h h V ds p L p h L V ds p V p                               (13) Based on these limit behavior of the power variation, Todorov and Tauchen (2009) proposed a qualitative test of the type of the process. They have defined activity signature function ˆ ( ; , )p k h as: lnˆ ( ; , ) ln ln ( , ) ln ( , ) p k p k h k V p kh V p h     . (14) The graph of ˆ ( ; , )p k h with respect to p are called “activity signature plot”. If the process is continuous, then for all 0p  : 0 ˆplim ( ; , ) 2 h p k h   . (15) For processes of pure jumps we have: 0 , for (0, ),ˆplim ( ; , ) , for .h p p k h p p         (16) In the case, when the process contains both jumps and diffusion: 0 2, for (0, 2),ˆplim ( ; , ) , for .h p p k h p p       (17) We apply method of activity signatures to several instruments from Polish financial markets. The sample contains of four stocks: two liquid ones (PKN Orlen and KGHM) and two less liquid (Agora and BRE Bank), three stock market indexes (WIG, WIG20 and MWIG40), one future contract (FWIG) and two currencies (euro and US dollar). The sample was chosen as to contain pos- sibly wide range of different instruments. We have used intraday data for the period from the beginning of 2009 to the March of 2011. In the computation we used 10-minutes returns. We have tried some other frequencies and decided that this frequency is high enough to justify the usage of the limits (15)-(17). On the other hand it is not so high, that the market microstructure noise affects the re- sults3. In the computations of ˆ ( ; , )p k h we took 2k  , as in the original work 3 We have used methods of Zhang, Mykland, Aït-Sahalia (2005) to control the microstructure noise. For sampling period of 10 minutes the difference between “two time scales” estimator of Jumps Activity and Singularity Spectra for Instruments in the Polish Financial Market 177 of Todorov and Tauchen (2009). Thus we have worked with two time scales: 10 minutes and 20 minutes returns. We do not present results for each instrument, but only show three typical cases. Figure 1 shows activity signature plot for Agora. Such a graph is typical for less-liquid stocks. It represents the case (16) of pure jump process with 0  , which means that the prices are driven by compound Poisson processes. Figure 1. The activity signature plot for Agora (vertical axis – exponents p , horizon- tal axis – the values of ˆ ( ; 2, )p h ) Figure 2. The activity signature plot for PKN Orlen (vertical axis – exponents p , horizontal axis – the values of ˆ ( ; 2, )p h ) quadratic variance and realized variance was small, so we decide that microstructure effect for this frequency is negligible. Paweł Kliber 178 The Figure 2 represents the activity signature plot for PKN Orlen, but the graph is similar to the plots for liquid stock, future contract and currency prices. The plot resemble the case (17), when the process contains both jump and diffusion parts. The plot for index WIG, shown on the Figure 3, is typical for all indexes. In this case the process is continuous. Figure 3. The activity signature plot for WIG (vertical axis – exponents p , horizontal axis – the values of ˆ ( ; 2, )p h ) 3. Estimating Blumenthal-Getoor Index Using Threshold Estimator The method of activity signatures allows us to identify the type of process that underlines prices, but usually does not allow estimating the value of Blumen- thal-Getoor index. To estimate the values of this index we use threshold estima- tor proposed by Aït-Sahalia and Jacod (2009). The estimator is given by the formula: ln ( , ) ln ( , )ˆ ( , , ) ln U h U k h k h k       , (18) where ,( )U h is counting function which counts the exceedances of the threshold:  ( ) 1 ( ), 1 i N x h h i hU      , (19) with (0,1 / 2) . Estimator (18) uses two time scales, as activity signature method, but the latter method is based on limit properties of power variation, while the threshold estimator is based on different exceedance rates in different time scales. It allows us to estimate Blumenthal-Getoor index in the case when the process contains diffusion (continuous) part. It also allows for testing the accuracy of the estimation. The asymptotic standard error of the estimator equals: Jumps Activity and Singularity Spectra for Instruments in the Polish Financial Market 179 1 1 1 ln ( ,) (, )k hk U U h   . (20) We have calculated the estimator ˆ ( , , )k h  for all instruments in the sam- ple. As in the original work of Aït-Sahalia and Jacod (2009) we have taken 1 / 5  and 2k  . As for the threshold level  it was taken as seven times the estimated standard error of the continuous part of the process. The results are shown in Table 1. Table 1. Estimators of Blumenthal-Getoor index Instrument ˆ ( , , )k h  Std. dev. AGO 0.7382 0.0087 BRE 1.0888 0.0074 KGH 1.5486 0.0052 PKN 1.5405 0.0085 WIG 2.6208 0.0285 WIG20 1.9684 0.0097 MWIG40 2.2870 0.0185 FWIG20 1.9280 0.0024 EUR 1.9835 0.0097 USD 2.0491 0.0069 The results are only partially consistent with the ones obtained with activity signature method in the previous section. For the less-liquid stocks (AGO and BRE) the estimators of  are significantly lower than for the other instruments. They are however greater than 0. In case of stock indexes (WIG, WIG20 and MWIG40) the estimators are high (close to 2 or even greater than 2), what is consistent with previous results, that trajectories of these instruments are con- tinuous. Liquid stocks have  between 1 and 2. As for the currencies (EUR, USD) and futures contract (FWIG20), the obtained values are close to 2, what stands in contradiction with the results from the previous section. The last col- umn in the Table 1 contains estimators’ errors. However they were calculated with asymptotic formula (20) and it is dubious if they give the true errors of estimators, especially if some values are greater than 2. 4. Singularity Spectra and Jump Activity The third method of estimating Blumenthal-Getoor index is based on singularity spectra of observed trajectories of prices. It is non-statistical methods. It is based on the fact that the trajectories of Lévy processes with different  almost surely (i.e. with probability 1) reveals different types of continuity. Let us first introduce the concept of singularity spectrum. Paweł Kliber 180 Take any function :f R R . We say that the function is a -Hölder contin- uous in the point 0t if there exists a polynomial ( )P t of order lower then a such that in some neighborhood of 0t : 0( ) ( ) a f t P t K t t   (21) for some 0K  . Let  be a number such that for a  function f is a -Hölder continuous at t and for all a  f is not a -Hölder continuous at t . The  is called local Hölder exponent of the function f at the point t and is denoted by ( )fh t . It is the measure of “regularity” of f in the neighbor- hood of t . The higher ( )fh t , the more regular the function is. For example the trajectories of Wiener motion almost surely have local Hölder exponent equal to 0.5 at each point. Let ( )f  be the set of all points in which the function f has local Hölder exponent equal to  :  ) :( ( )f ft h t    . (22) Singularity spectrum it is the mapping which for all  returns the Hausdorff dimension (see Mallat, 2003; Falconer, 2003) of the set ( )f  : H( ) dim ( )f fD    . (23) As was shown by Jaffard (1999) the Lévy processes with different Blumen- thal-Getoor index almost surely have different singularity spectra. If the Lévy process L does not contain diffusion part, then: 1 (dim ) for H L        (24) and ( )L    for 1 /  . The shape of singularity spectrum for such a process is show in the Figure 4. If the process L contains diffusion part, then: 1 ( ) for dim 2 H L      , (25) (1di / 2)m 1H L  and ( )L    for 1 / 2  . For the Wiener process (with- out jumps) (1di / 2)m 1H L  and ( )L    for 1 / 2  . Jumps Activity and Singularity Spectra for Instruments in the Polish Financial Market 181 Figure 4. Singularity spectrum for Lévy process with Blumenthal-Getoor index  To calculate singularity spectra from discretely sampled data one uses so called “multifractal formalism”, introduced by Frisch and Parsi (1985) and de- veloped by Jaffard (1997a, 1997b). It can be shown that the function *fD , de- fined as: 0 * 1( ) lim inf ln ( ) ( ) ln q f h D q f x h f x dx h    (26) is the Legendre transform4 of the singularity spectrum D , i.e.:  * ( ) inf ( )f f R D q q D      . (27) If the function )(fD  is convex (as it is generally assumed in the literature), then one can obtain fD by performing Legendre transform on * fD . The func- tion *fD can be estimated from sample moments. One has to calculate power variation (10) for multiple of sampling times h and then to use the regression: *ln ( , ) ( ) lnfV q h C qD h  . (28) This method however is not stable numerically and some better methods were proposed. Most of them use wavelet transform. The review of them can be found in Turiel, Pérez-Vincente, Grazzini (2006) or Oświęcimek (2005). In our research we have used method based on “modulus maxima” of wavelet coefficients, implemented in Matlab package Fraclab. The method was pro- 4 Lagendre transform is well-established method of convex analysis. See for example Rocke- faller (1970), ch. 26. α Df(α) 1/β 1 Paweł Kliber 182 posed by Jaffard (1997a) and very good description can be found in Mallat (2003, ch. 6). Figure 5. Estimated singularity spectrum for PKN Orlen The Figure 5 shows the results for one stock (PKN Orlen). The estimated singularity spectra for all other instrument revealed very similar pattern, so we omit them. One should also stress that the results bears a great resemblance to the results of Cont and Tankov (2004) for the stock from US market. The graph does not star at the origin, which is inconsistent with typical shape depicted on the Figure 4. This however can be due to numerical inaccuracy. The main fea- ture of the graph is the value at which the function reaches its peak, as it is the reciprocal of Blumenthal-Getoor index. For all instrument in the sample the peak values lie in the interval [0.6, 0.8], which means that the indexes  are between 1.2 and 1.8. Conclusions The estimation of the Blumenthal-Getoor index is a complicated problem. We have used three different methods and as one can see the results are in many cases inconsistent. As for the less-liquid stocks, according to the activity signature method the price process is driven by compound Poisson process, while the two other methods reveal positive value of Blumenthal-Getoor index. This inconsistency may be due to the fact of low liquidity of these stocks. There were days when the prices did not change for several hours. The estimators we have used take advantage of limit properties of price changes as time scale tends to 0. It is thus dubious if they gave correct estimators for such illiquid stocks. 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.2 0.4 0.6 0.8 1 Jumps Activity and Singularity Spectra for Instruments in the Polish Financial Market 183 As for the liquid stocks all three estimators gave similar results. Moreover the results are consistent with similar results for other markets (  at the level about 1.5). According to the first two methods (activity signature and threshold estimator) the indexes are continuous processes, while the third method (singu- larity spectrum) revealed 2  . Probably in fact the former result holds, as the method of singularity spectrum is the most prone to numerical errors. The results for currencies and futures contract are ambiguous. This can be a result of active process of jumps with Blumenthal-Getoor index close to 2. As it was pointed out by Zhang (2007) the jump process is then hardly distin- guishable from continuous diffusion. References Aït-Sahalia, Y., Jacod, J. (2009), Estimating The Degree of Activity of Jumps in High Frequency Data, The Annals of Statistics 37, 2202–2244. Barndorff-Nielsen, O.E., Shephard, N. (2002), Power Variation and Time Change, working pa- per, available in RePEc database. Becry, E., Muzy, J. F., Arnédo, A. (1993), Singularity Spectrum of Fractal Signals from Wavelet Analysis: Exact Results, Journal of Statistical Physics 70, 635–674. Black, F., Scholes, M. (1973), The Pricing of Options and Corporate Liabilities, Journal of Politi- cal Economy 81, 637–654. Cont, R. (2001), Empirical Properties of Assets Returns: Stylized Facts and Statistical Issues, Qualitative Finance 1, 223–236. Cont, R., Tankov, P. (2004), Financial Modelling with Jump Processes, Chapman&Hall, London, New York. Falconer, K. (2003) Fractal Geometry, Wiley. Feller, W. (1971), An Introduction to Probability Theory and Its Applications, vol. 2, Wiley. Frisch, U., Parsi, G. (1985) Fully Developed Turbulence and Intermittency, in: Giil, M. (ed.) Turbulence and Predictability in Geophysical Fluid Dynamics and Climate Dynamics, North Holland, Amsterdam, 84–88. Jaffard, S. (1997a), Multifractal Formalism for Functions Part I: Results Valid for All Functions, SIAM Journal of Mathematical Analysis 28, 944–970. Jaffard, S. (1997b), Multifractal Formalism for Functions Part II: Self-Similar Functions, SIAM Journal of Mathematical Analysis 28, 971–998. Jaffard, S. (1999), The Multifractal Nature of Lévy Processes, Probability Theory and Related Fields 114, 207–227 Jondeau, E., Poon, S.-H., Rockinger, M. (2007), Financial Modeling Under Non-Gaussian Dis- tributions, Springer. Kallsen, J. (2000), Optimal Portfolios for Exponential Lévy Processes, Mathematical Methods of Operational Research 51, 357–374. Maldenbrot, B. B. (1963) The Variation of Certain Speculative Prices, Journal of Business 36, 394–419. Malevergne, Y., Sornette, D. (2006), Extreme Financial Risks, Springer, Berlin, New York. Mallat, S. (2003), A Wavelet Tour of Signal Processing, Elsevier, Singapore. Markowitz, H.M (1952) Portfolio Selection, Journal of Finance 7, 77–91. Merton, R. C. (1973) Theory of Rational Option Pricing, Bell Journal of Economics and Man- agement Science 4, 141–183. Oświęcimek, P. (2005), Multifraktalne charakterystyki finansowych szeregów czasowych (Mul- tifractal Characteristics of Financial Time Series), doctor thesis, Instytut Fizyki Jądrowej Polskiej Akademii Nauk (Nuclear Physics Institute of The Polish Academy of Science). Paweł Kliber 184 Rockafellar, R. T. (1970), Convex Analysis, Princeton University Press, Princeton. Sharpe, W. F. (1963), A Simplified Model for Portfolio Analysis, Management Science 9, 277–293. Todorov, V., Tauchen, G. (2009), Activity Signature Function for High-Frequency Data Analysis, Journal of Econometrics, preprint at http://ideas.repec.org (7.09.2011). Turiel, A., Pérez-Vincente, C., Grazzini, J. (2006), Numerical Methods for The Estimation of Multifractal Singularity Spectra on Sampled Data: A Comparative Study, Journal of Com- putational Physics 216, 362–390. Zhang, L., Mykland, P. A., Aït-Sahalia Y. (2005), A Tale of Two Time Scales: Determining Integrated Volatility with Noisy High-Frequency Data, Journal of the American Statistical Association 100, 1394–1411. Zhang, L. (2007), What You Don’t Know Cannot Hurt You: On the Detection of Small Jumps, working paper at http://tigger.uic.edu/~lanzhang/ (7.09.2011). Aktywność skoków i spectrum osobliwości dla instrumentów z polskiego rynku finansowego Z a r y s t r e ś c i. W artykule podejmujemy próbę oszacowania aktywności skoków w procesach cen kilku instrumentów z polskiego rynku finansowego. Jako miarę aktywności skoków przyjmu- jemy indeks β Blumenthala-Getoora dla procesów Lévy’ego. Pozwala nam to na rozróżnienie procesów charakteryzujących się rzadkimi i dużymi skokami i procesów o nieskończonej aktyw- ności procesu skoków. Aktywność skoków szacujemy trzema różnymi metodami. Wykorzystu- jemy wykresy podpisu aktywności (activity signature plots) do zbadania typu procesu. Następnie korzystamy z estymatora Aït-Sahalii i Jacod, opartego na liczbie przekroczeń, do oszacowania wartości indeksu  . Wreszcie korzystamy ze spektrum ciągłości oraz z odpowiednich twierdzeń na temat przebiegu tej funkcji dla procesów Lévy’ego z różnymi wartościami indeksu  . S ł o w a k l u c z o w e: wykładnicze modele Lévy’ego, indeks Blumenthala-Getoora, spektrum osobliwości.