ebook img

Bridging the ARCH model for finance and nonextensive entropy PDF

0.62 MB·English
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Bridging the ARCH model for finance and nonextensive entropy

Bridging the ARCH model for finance and nonextensive entropy S´ılvio M. Duarte Queir´os and Constantino Tsallis ∗ Centro Brasileiro de Pesquisas F´ısicas, Rua Dr. Xavier Sigaud 150, 22290-180, Rio de Janeiro-RJ, Brazil 4 (February 2, 2008) 0 Engle’sARCHalgorithmisageneratorofstochastictimeseriesforfinancialreturns(andsimilar 0 quantities) characterized by a time-dependent variance. It involves a memory parameter b (b = 0 2 corresponds to no memory), and the noise is currently chosen to be Gaussian. We assume here n a generalized noise, namely qn-Gaussian, characterized by an index qn ∈ R (qn = 1 recovers the a J Gaussian case, and qn > 1 corresponds to tailed distributions). We then match the second and fourth momentaof theARCHreturndistribution with those associated with theq-Gaussian distri- 14 butionobtainedthroughoptimizationoftheentropySq = 1−q−i1piq,basisofnonextensivestatistical mechanics. The outcome is an analytic distribution for the rPeturns, where an unique q ≥qn corre- ] sponds to each pair (b,qn) (q =qn if b =0). This distribution is compared with numerical results h andappearstoberemarkablyprecise. Thissystemconstitutesasimple,low-dimensional,dynamical c mechanism which accommodates well within the current nonextensiveframework. e m PACS numbers: 05.40.-a, 05.90.+m, 89.65.Gh - t a t Time series are ubiquitous in nature. They appear in zt =σt ωt, (1) t.s geoseismic phenomena, El Nin˜o, finance, electrocardio- (cid:26)σt2 =a+ si=1bi zt2−i , a and electroencephalographic profiles, among many oth- P m ers. Someoftheseseriescanbeconstructedbyusing,for or, equivalently, - successive values ot time t, random variables associated d with the same distribution for all times: they are called n homoskedastic. SuchisthecaseoftheordinaryBrownian zt = a+ si=1bi zt2 i ωt (2) co motion. But many phenomena exist in nature which do ( σt2 =aq+ siP=1bi σt2 i−ωt2 i , − − [ not accomodate with such a property, i.e., the distribu- P tion associated with each value of t depends on t. Such wherea,b ,σ 0,andω representsanindependentand 2 i t ≥ t v randomvariables are then called heteroskedastic. A sim- identicallydistributedstochasticprocesswithmeanvalue 1 ple illustration would be to use say a centered Gaussian nullandunitvariance,(i.e.,hωti=0and ωt2 =1),cur- 8 at all steps, but with a randomly varying width. rentlychosentofollowaGaussiandistribution,butother 1 (cid:10) (cid:11) Nowadays, time series that are intensively studied are choices are possible. In this manuscript we will discuss 1 the financial ones, where nonstationary volatility (tech- in detail the usual case s = 1, ARCH(1), which will be 0 4 nical name for second-order moment of say returns) is a from now on simply designated by ARCH. In general, 0 common feature [1–3]. In order to mimic explicitly and the distribution, Pn(ω), together with parametersa and t/ analyze this type of time series, R.F. Engle introduced {bi}, specify the particular ARCH process (n stands for a in 1982 the autoregressive conditional heteroskedasticity noise). m (ARCH) process[4]. Its prominencecanbe measuredby As canbe seenfromEqs. (2), parameters b charac- i - { } its wide use, by the amount of its extensions introduced terize the memory mechanism. For b = 0 ( i), there is d i ∀ n later [5,6], and — last but not least — by the fact that no memory effect, and consequently the ARCH process o the 2003 Nobel Prize for Economics was awardedto En- forreturnsreducestogeneratingthenoiseω (multiplied t c gle “for methods of analyzing economic time series with by√a). Wecanverifythat,ingeneral, z = σ ω = : h ti h tih ti v time-varying volatility (ARCH)”. This and similar pro- 0, and ztzt′ = σtσt′ ωtωt′ = σtσt′ δtt′. Therefore, h i h ih i h i i cessesarecommonlyusedinthe implementationofasset even for nonvanishing b ’s, the returns z remain uncor- X i t pricingtheories,marketmicrostructuremodels,andpric- related, whereas correlations do exist in the variance σ2 r t a ing derivative assets [6,7] (see also [8]). Specifically, the [5]. We observe that this stochastic process captures the ARCH(s) [4] process generates, for the returns z , a dis- tendency for the so called volatility clustering, i.e., large t crete time series whose variance σ2, at each time step, (small) values of z are followed by other large (small) t t depends linearly on the s previous values of [z ]2. It is values of z , but of arbitrary sign. In other words, t t+1 defined as follows: zt zt′ by no means is proportional to δtt′. h| || |i ∗[email protected], [email protected] 1 Hereon we focus on ARCH(1), i.e., b b and b = FIG.2. Time dependence of the volatility σ for the same 1 i ≡ two ARCH processes shown in Fig. 1. The largest value for 0 (i = 2, 3,...). From Eqs. (1) it is simple to obtain σ occurs at t=238. the n-th moment for the P(z) stationary distributions, particularly the second moment Weshallnowestablishapossibleanalyticalconnection a σ2 ≡ zt2 =hσt2i= 1 b (b<1), (3) betweenthememoryparameterb,P (z)andPn(ω)under the framework of the nonextensive statistical mechanics − (cid:10) (cid:11) and the fourth moment, introduced by one of us [13]. Consider a probability dis- tributionP (z)suchthatitmaximizestheentropicform 1+b ′ z4 =a2 ω4 . (4) t (1 b)(1 b2 ω4 ) Withou(cid:10)t (cid:11)loss of(cid:10)gen(cid:11)eral−ity, we−canhatssiume that the Sq = 1− −+∞∞[P′(z)]qdz (q ), (6) q 1 ∈ ℜ ARCH procedure generates a time series with unit vari- R − ance,i.e.,σ2 =1,hencea=1 b. Now,forz ,thefourth t + m(aopmoesnsitbilse mnuemaseurriceaollfyneoqnu-gaalu−tsositahneitkyuorrtopsiesakkexdn≡eshshxx2f4oii2r wsthtiaetnhcdolsinmfsotqrr→aBi1noStlsqtzm=+a∞n−nPR-G−′(∞i∞zb)bdPsz)′.(=zM)1lanxaPnim′d(izz)idn+zg∞E≡zq2P.SB′((6Gz))d(wzBitG=h probability distributions), and we can easily get, 1 we have −∞ −∞ R R k 1 z4 =k =k 1+b2 ω − (k b2 <1), (5) t z ωh 1−kωb2i ω P′(z)= [1+ (qA1)z2]q−11 (q <5/3), (7) Itis(cid:10)str(cid:11)aightforwardlyverifiedthat,whatevertheformof B − Pn(ω),theARCHprocessgeneratesprobabilitydistribu- with tions P (z) with a slower decay and consequently with a kurtosis kz > kw [9–12]. See in Figs. 1 and 2, typical Γ 25q−32q runs for a Gaussian noise. − , (8) B ≡ 2(q h1)Γ i3−q − 2q 2 − h i and Γ 1 q 1 − (2q 1) . (9) A≡ √2πΓh 3−iq − B 2q−2 p h i In the q 1 limit, the normal distribution is recovered. → The fourth moment (which for this case coincides with the kurtosis) is given by, FIG.1. Two examplesofARCHtimeseries obtained fora z4 =3Γ 72−q−52q Γ 23q−−q2 (1<q < 7). (10) tGhaeupssrioabnabnioliitsye (foi.re.l,aqrnge=va1l)u.esMoefm|zotr|y, i(.ie.e.,.,fabt>ta0il)sinincrPea(sze)s. (cid:10) (cid:11) h Γ 52i−q32qh 2 i 5 − The large value at t = 237 is virtually never observed for n h io b=0. Letus now makethe ansatz P (z) P (z). Consistently ′ ≃ we impose the matching of Eqs. (5) and (10) . This yields q as a function of b and k . Assuming that noise ω ω follows a generalizeddistribution, Eq. (7), defined by t 3.0 b = 0.0 an entropic index qn (such that kw(qn)b2 <1) we are able b = 0.5 to establish a relation between the memory parameter t 2.5 b and the entropic indexes q and qn, which character- izerespectivelythe distributionsP (z)andP (w). This ′ n 2.0 connection is straightforwardly given by 1.5 2 2 1.0 b= rG(q)nΓh52−qn3−qn2io −G(qn)nΓh25q−−32qio , (11) 2 0 150 t 300 sG(qn)(cid:18)3G(q)−nΓh52−q−32qio (cid:19) 2 where G(x) Γ 7 5x Γ 3 x . This connection, de- comparing the sixth-order moment. It is simple to ver- ≡ 2−x 2 2x−2 − − ify that these momenta for P(z) and P (z) exhibit quite pictedinFig. 3,chonstituitesh,tothiebestofourknowledge, ′ smalldiscrepancies. Forexample,forthecasesillustrated the first ever found which analytically expresses the re- in Figs. 4 and 5, we have obtained discrepancies never turn distribution in terms of the noise distribution and largerthan 2.9%(occuring in factfor b 0.5)for q =1 the memory parameterb. In what followswe shallverify n ≃ andthan2.7%(occuringinfactforb 0.3)forq =1.3. thattheaboveansatzisindeedsatisfiedwithinaremark- n ≃ Such minor discrepancies are, in financial practice, com- able precision. To do this for typical values of b and q , n pletely inocuous (for example, we may check in [14,15] we first generate, through a standard algorithm based that realistic return distributions are larger than 10 6, on Eq.(1) (with s = 1), a set of ARCH time series and − whereas in our present Figs. 4 and 5 we have simulated theircorrespondingprobability densityfunctions(PDF’s) down to 10 8). P(z). The results are indicated in Figs. 4 and 5. Then − we compare with a histogram (with a conveniently cho- sen unit interval δ) associated with the distribution (7), with q satisfying Eq. (11). In other words, we compare 1 with 1 (1 q)x2 1−q, where is given by Eq. δ A −B − B (8), and (cid:2) (cid:3) = +δ/2p(z)dz = δ Γ[q−11] Γ[25q−−32q] Aδ −δ/2 √2πΓ[23q−−q2]rΓ[23q−−q2] (12) R F 1, 1 ;3, δ2Γ[25q−−32q] ×2 1(cid:18)2 q−1 2 − 8Γ[23q−−q2] (cid:19) FIG. 4. PDFs for a qn = 1 noise and typical values of b. a) b = 0, q = 1 (χ2 = 1.14 × 10−11); b) b = 0.1, q = 1.01976 (χ2 = 2.34×10−10); c) b = 0.4, q = 1.242424 (χ2=1.82×10−10);d) b=0.5, q= 4 (χ2 =2.98×10−10). 3 FIG. 3. Diagram (q,qn,b) for the ARCH process with z2 =1 (for b=0 we havethe straight line q=qn). FIG. 5. PDFs for a qn = 1.3 noise and typical val- (cid:10) (cid:11) ues of b. a) b = 0, q = 1.3 (χ2 = 3.93 × 10−10); b) We verify that the agreement is quite satisfactory. b = 0.1, q = 1.30762 (χ2 = 8.86 × 10−10); c) b = 0.2, In order to quantify the small discrepancy between the q = 1.32950 (χ2 = 9.72×10−10); d) b = 0.3, q = 1.36306 ARCH PDF’s P(z) and the analytical ones P′(z), we (χ2=2.68×10−10). have indicated in the captions of Figs. 4 and 5 the val- N Let us conclude by saying that the fact that a close ues of χ2 1 [P(z ) P (z )]2, N being the number ≡ N i − ′ i connectionhas beenshownto existbetweenthe possibly i=1 of points. An aPlternative evaluation of the discrepancy ubiquitous ARCH stochastic processes and the nonex- of the ARCH and analyticaldistribution can be done by tensive entropy Eq.(6) strongly suggests further possible 3 connections. Werefertophenomenawhichmightpresent [7] T.H. McCurdy and T. Stengos, J. Econometrics 52, 225 some kind of scale invariant geometry, e.g., hierarchi- (1992). calormultifractalstructures,low-dimensionaldissipative [8] I. Zovko and J. Doyne Farmer, Quantitative Finance 2, and conservative maps [16], fractional and/or nonlinear 387 (2002). [9] H. White, Econometrica 50, 1 (1982). Fokker-Planck equations [17], Langevin dynamics with [10] A.A. Weiss, Journal of Time Series Analysis 5, 129 fluctuating temperature [18], possibly scale-free network (1984). growth [19], long-range many-body classical Hamiltoni- [11] T. Bollerslev, Rev.Econ. Stat. 69, 542 (1987). ans[20–22],amongothers. Amoredetailedstudyofthis [12] B. Podobnik, P.Ch. Ivanov, Y. Lee, A. Cheesa and H.E. aswellasofotherheteroskedasticmodels(e.g.,GARCH) Stanley , Europhys.Lett. 50, 711 (2000) is in progress. [13] C. Tsallis, J. Stat. Phys. 52, 479 (1988). A regu- We acknowledge useful remarks from J.D. Farmer, larly updated bibliography on the subject is avaible at as well as partial support from Faperj, CNPq, http://tsallis.cat.cbpf.br/biblio.htm. PRONEX/MCT (Brazilian agencies) and FCT/MCES [14] R. Osorio, L. Borland and C. Tsallis, in Nonextensive (contract SFRH/BD/6127/2001)(Portuguese agency). Entropy - Interdisciplinary Applications, eds. M. Gell- MannandC.Tsallis(OxfordUniversityPress,NewYork, 2004), in press. [15] L. Borland, Phys. Rev. Lett. 89, 098701 (2002); L. Bor- land, Quantitative Finance 2, 415 (2002). [16] M.L.LyraandC.Tsallis,Phys.Rev.Lett.80,53(1998); F. Baldovin and A. Robledo, Europhys. Lett. 60, 518 (2000). [1] B.M. Mandelbrot, J. Business 36, 394 (1963). [17] L. Borland, Phys. Rev.E 57, 6634 (1998). [2] E.F. Fama, J. Business 38, 34 (1965). [18] C. Beck and E.G. Cohen, Physica A 322, 267 (2003). [3] J.-Ph. Bouchaud and M. Potters, Theory of Financial [19] R. Albert and A.L. Barabasi, Phys. Rev. Lett. 85, 5234 Risks: From Statistical Physics to Risk Management (2000). (Cambridge University Press, Cambridge, 2000); R.N. [20] V.Latora,A.RapisardaandC.Tsallis,Phys.Rev.E64, Mantegna and H.E. Stanley, An Introduction to Econo- 056134 (2001). physics: Correlations and Complexity in Finance (Cam- [21] S. Abe and Y. Okamoto, eds., Nonextensive Statistical bridge UniversityPress, Cambridge, 1999). Mechanics and Its Applications(Springer-Verlag,Berlin, [4] R.F. Engle, Econometrica 50, 987 (1982). 2001) [5] T.Bollerslev R.Y.Chou and K.F.Kroner,J. Economet- [22] M. Gell-Mann and C. Tsallis, eds., Nonextensive En- rics 52, 5 (1992). tropy–InterdisciplinaryApplications(OxfordUniversity [6] S.H. Poon and C.W.J. Granger, J. Econ. Lit. 41, 478 Press, New York,2004), in press. (2003). 4

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.