arXiv:cond-mat/0209685v1 [cond-mat.stat-mech] 30 Sep 2002 † ∗ eoto h miia rpriso fecahmarket aftercrash of properties empirical the [5]. on portfolio stock report a of Risk as, at measures Value has risk the scale of example, relaxation estimation for typical the by a on characterized of implication lack not direct The is crash scale. market typical financial typical a a a of its after dynamics to just relax- relaxation indicators index the stochastic statistical that the implies market in values the form of functional ation specific this of with decays uyaoadkontdya h mr a 4.The earth- aftershock [4]. time of law at number Omori the measured the that quakes as states today law known cen- Omori a and than ago more geophysics tury in discovered law statistical a exceeds return index of value. value threshold num- absolute given the crash the to times a respect of after with ber just regularity returns statistical that show index a to presents market is of paper this dynamics of of the purpose returns The the of index. market evolution a time non-stationary a implies market the a by crash, decay. characterized the slow evolution After stochastic af- a [3]. shows immediately volatility crash and financial at each increase ter dramatic a error experiences [2]. estimation volatility to constant of due hypothesis be the reject could in strongly observed series size time the tests of fact, empirical fluctuations In volatility [1]. whether is pattern for index non-stationary market a a of showing volatility often the example, for as, market [email protected] [email protected] h tutr fteppri sflos nScin2we 2 Section In follows. as is paper the of structure The h icvrdsaitclrglrt saaoost a to analogous is regularity statistical discovered The directly volatility of evolution time non-stationary A implied, and historical both series, time volatility The financial a of indicators basic of evolution time The siuoNzoaeprl iiadlaMtra ntad Pa Unit`a di Materia, della Fisica la per Nazionale Istituto ao rs sdsrbdb oe-a ucinlcigat a lacking function power-law c a the market by of evolution described real time is in the crash that observed major show a behavior also We nonlinear characteriz model. the numerica GARCH(1,1) This that performing By show threshold. inde we geophysics. given of in a behavior observed over originally non-stationary exceedances observed the The of crash. market t sapwrlwfnto.Teobservation The function. power-law a as edsustesaitclpoete fidxrtrsi fi a in returns index of properties statistical the discuss We .INTRODUCTION I. siuoNzoaeprl iiadlaMtra ntad Pa Unit`a di Materia, della Fisica la per Nazionale Istituto -02,Plro tl n iatmnod iiaeTecnol e Fisica di Dipartimento and Italy Palermo, I-90128, yaiso nnilmre ne fe crash a after index market financial a of Dynamics nvri` iPlro il el cez,I918 Pale I-90128, Scienze, delle Viale Palermo, Universit`a di t fe h anearthquake main the after oai .Mantegna N. Rosario arzoLillo Fabrizio bv ie hehl esrda time at measured threshold given a above fafiaca index The financial a changes crash. of logarithm market one-minute the major is a investigated variable after returns frequency high hrceietetm eiso ne eun yinvesti- times by of returns number index sta- the of a series gating of quantitatively time we use the Specifically, make characterize we method. characterize series robust to order time tistically In return period. aftercrash ”normal” non- is the a time highly to This some back is needs be period market crash. to financial aftercrash market the the a because that stationary after fact dif- the extremely period to is time due process the random in this ficult of evolution time number period. time non-stationary the in [6] value threshold yteOoilw[] h mr law Omori The [4]. law geophysics Omori in the described well by is quantity This earthquake. meitl fe nnilcah h ae nswith period ends time paper remarks. the The concluding during some crash. financial Risk a at after Value immediately the discusses the 4 empirical of Section the estimation section. describe previous the the to of that unable properties shows is 3 Section model GARCH(1,1) crash. the after financial recorded 1-minute Monday returns Black the index 500 concerns Poor’s discussion and The Standard series. time index httenme fatrhc atqae e nttime unit per time earthquakes at aftershock measured of number the that oe a.I re oaoddvrec at divergence avoid to order In law. power etfruaino h mr a,wihi oesuitable more is which law, Omori the of formulation lent where a sotnrwitnas rewritten often is law em,VaedleSine -02,Plro Italy Palermo, I-90128, Scienze, delle Viale lermo, nti eto,w hrceietetm eiso ultra of series time the characterize we section, this In iia prahi sdi h netgto fthe of investigation the in used is approach similar A ∗ K iuain n hoeia modelling, theoretical and simulations l I MIIA RPRISOF PROPERTIES EMPIRICAL II. † n to saaoost h mr law Omori the to analogous is ation and pclscale. ypical ( eun scaatrzdi terms in characterized is returns x t aca aktjs fe major a after just market nancial FECAHSEQUENCES AFTERCRASH e nttm fatrhc earthquakes aftershock of time unit per ) ahscno edsrbdb a by described be cannot rashes em,VaedleScienze, delle Viale lermo, au tRs bevdjs after just observed Risk at Value τ r w oiiecntns nequiva- An constants. positive two are t n fe h anerhuk easa a as decays earthquake main the after r ( geRelative, ogie m,Italy rmo, ( t = ) t .Adrc hrceiaino the of characterization direct A ). K ( t | + r ( t τ ) ) | − seceigagiven a exceeding is p , n t ( fe h main the after t ) t ,Omori 0, = ∝ t − p says (1) 2 for comparison with real data, can be obtained by inte- 300 grating Eq. (1) between 0 and t. In this way the cumu- lative number of aftershocks after the main earthquake observed until time t is 200 N(t)= K[(t + τ)1−p τ 1−p]/(1 p), (2) − − 10−1 when p = 1 and N(t)= K ln(t/τ +1) for p = 1. When6 the process is stationary the frequency of af- N(t) (t) −2 av 10 tershock n(t) is on average constant in time and there- 100 n fore the cumulative number N(t) increases linearly in time. We have tested that N(t) increases approximately 10−3 linearly in a market period of approximately constant 100 101 102 volatility such as, for example, the 1984 year. For inde- t (trading day) 0 pendent identically distributed random time series it is 0 20 40 60 possible to characterize n(t) in terms of an homogeneous t (trading day) Poisson process [7]. As an example of the behavior of N(t) empirically ob- served after a market crash, we investigate the index re- FIG. 1: Cumulative number N(t) of the number of times turns during the time period just after the Black Monday |r(t)| is exceeding the threshold ℓ = 4σ during the 60 trad- crash (19 October 1987) occurred at New York Stock Ex- ing days immediately after the Black Monday financial crash. change (NYSE). This crash was one of the worst crashes The parameter σ is the standard deviation of the process r(t) occurred in the entire history of NYSE. The Standard computed over the entire investigated period. The dashed line and Poor’s 500 Index (S&P500) went down 20.4% that is the best fit of Eq. (2). The value of the exponent p is 0.85. day. In our investigation, we select a 60 trading day af- In the inset, we show nav(t) which is a moving average of n(t) in a sliding window of 200 trading minutes. The dashed line tercrash time period ranging from 20 October 1987 to 14 is a power-law curve with an exponent equal to −0.85 and is January 1988. For the selected time period, we inves- shown as a guide for the eye tigate the one-minute return time series of the S&P500 Index. The unconditional one-minute volatility is equal to σ = 4.91 10−4. In Fig. 1 we show the cumula- × Quote (TAQ) database issued by the NYSE. tive number of events N(t) detected by considering all Our empirical results show that index return cannot be the occurrences observed when the the absolute value modeled in terms of independent identically distributed of index return exceeds a threshold value ℓ = 4σ. We random process after a big market crash. In ref. [6] we observe a nonlinear behavior in the entire period. Fig. introduce a simple stochastic model which is able to ex- 1 also shows our best nonlinear fit performed with the plain the Omori law in financial time series. Specifically functional form of Eq. (2). The agreement between em- we assume that the stochastic variable r(t) is the product pirical data and the functional form of the Omori law is of a deterministic time dependent scale γ(t) times a sta- pretty good. With a single time series it is not possible tionary stochastic process rs(t) during the time period to have a good description of n(t) due to the discreteness after a big crash. Under these assumptions, the num- of the data. In order to solve this problem we compute ber of events of r(t) larger than ℓ observed at time t is nav(t) which is the running average of n(t) in a sliding proportional to | | window of 200 trading minutes. In the inset of Fig. 1 we show this quantity as a function of time in a bilogarith- ℓ n(t) 1 F , (3) mic plot. A power-law behavior is observed for almost ∝ − s γ(t) two decades and the exponent of the best power-law fit is in agreement with the exponent p obtained by fitting where Fs(rs) is the cumulative distribution function of N(t). the random variable rs(t). We observe a similar behavior when we set the thresh- By assuming that the stationary return probability old value ℓ equal to 5σ, 6σ and 7σ. The best fit of the density function behaves asymptotically as a power-law exponent p slightly increases when ℓ increases and even- tually converges to a constant value. 1 This paradigmatic behavior is not specific of the Black fs(rs) α+1 , (4) ∼ rs Monday crash of the S&P 500 index. In fact, we observe similar results also for a stock price index weighted by and that the scale of the stochastic process decays as a market capitalization for the time periods occurring af- power-law γ(t) t−β, the number of events above thresh- ter the 27 October 1997 and the 31 August 1998 stock old is power-law∼ decaying as n(t) (γ(t)/ℓ)α 1/tp. market crashes [6]. This index has been computed select- The exponent p is given by ∼ ∼ ing the 30 most capitalized stocks traded in the NYSE and by using the high-frequency data of the Trade and p = α β. (5) 3

300 with the G@RCH 2.3 package [12]. The best estima- tion of the parameters on the time series of one minute 250 logarithmic price change during 60 trading days after −8 Black Monday crash are α0 = 2.87 10 , α1 = 0.38 and 200 4 β1 = 0.54. With these parameters we generate 10 sur- rogate time series according to Eq. (6) and we compute 150 the average behavior of N(t). To mimic the dynamics af- N(t) ter the crash, we set as an initial condition a large value 100 of return in each realization. Fig. 2 shows N(t) ver- sus time for real data and GARCH(1,1) model (dotted 50 line) with a threshold ℓ = 4σ (the same as in Fig. 1). The GARCH(1,1) time series converges to its stationary 0 phase very quickly and it is unable to show a significant 0 10 20 30 40 50 60 nonlinear behavior. time (trading day) In order to take into account the non-stationary be- havior of the return time series after a crash we have FIG. 2: The dotted line is the mean value of the cumula- performed a different analysis. Specifically we divide the tive number N(t) of the number of times |r(t)| is exceeding 4 60 trading day period in 6 nonoverlapping time intervals the threshold ℓ = 4σ over 10 simulations obtained with a of 10 trading days and we estimate the GARCH(1,1) pa- GARCH(1,1) model with parameters estimated in a period of 4 60 trading days immediately after the Black Monday financial rameters specific for each interval. We then generate 10 crash. The dashed line is the mean of N(t) over 104 simula- GARCH(1,1) surrogate time series using the estimated tions obtained with a GARCH(1,1) model with parameters parameters for each subinterval. The mean value of N(t) estimated in 6 nonoverlapping windows of 10 days each. As a as a function of time obtained with this procedure is also reference we show also N(t) for the real data already shown shown in Fig. 2 as a dashed line. In each subinterval in Fig. 1 (solid line). N(t) increases linearly for almost all the period. This is due to the fact that the non-stationary part is extremely small compared with the time scale of the figure. The previous relation links the exponent p governing the Below we theoretically describe the behavior of N(t) number of events exceeding a given threshold to the α expected for a GARCH(1,1) time series. First of all we 2 exponent of the stationary power-law return cumulative note that the expectation value of σt conditioned to the 2 distribution and to the β exponent of the power-law de- value σ0 is given by [13] caying scale. t The hypotheses of our model are consistent with re- 2 2 2 t 1 (α1 + β1) E(σt σ0)= σ0 (α1 + β1) + α0 − . (7) cent empirical results. In fact, a return probability den- | 1 (α1 + β1) − sity function characterized by a power-law asymptotic This equation shows that, under the assumption of fi- behavior has been observed in the price dynamics of nite unconditional variance α + β < 1, the mean value several stocks [8, 9] and a power-law or power-law log- 1 1 of the scale of the process decays exponentially to the periodic decay of implied volatility has been observed in unconditional value in a GARCH(1,1) time series. The the S&P500 after the 1987 financial crash [10]. In [6] we decaying time is given by τ = 1/ ln(α + β ). With the have tested the hypotheses of our model by measuring 1 1 parameters estimated we have−τ = 12 trading minutes in independently the exponents α and β from real data and the case considered in Fig. 2. we have shown that the relation between exponents de- Since the variance of the GARCH(1,1) model depends scribed by Eq. (5) is satisfied in all the three investigated only on α and on α + β , we perform a set of numeri- crashes. 0 1 1 cal simulations of GARCH(1,1) time series with the same value of α0 and by keeping constant α1 +β1 =0.92, which are the values observed in the real data. Different time III. COMPARISON WITH GARCH(1,1) MODEL series of simulations are obtained by changing the value of α1 which controls the leptokurtosis of the unconditional Here we compare the empirical behavior of a market density. Fig. 3 shows the average of n(t) over 105 realiza- index after a major crash with the predictions of a sim- tions of a GARCH(1,1) model with α1 =0.02, α1 =0.18 ple autoregressive model, specifically the GARCH(1,1) and α1 = 0.38. This last value corresponds to the one model [11] described by the equation estimated from the real data. In the first two cases the unconditional kurtosis is finite, whereas for α1 =0.38 the 2 2 2 σt = α0 + α1rt−1 + β1σt−1, (6) fourth moment diverges. The value of the threshold is set to ℓ =4σ in all cases. The unconditional value of n(t) is where rt is price return and σt is the volatility. To this larger for larger values of α1. This is expected because end we estimate the best value of the parameters of the the unconditional densities have the same variance and GARCH(1,1) model. The estimation has been performed the tail are fatter for larger values of α1. For all values 4

0 10 TABLE I: Simulated and theoretical values of n(t = 2) in a GARCH(1,1,) model for different values of α1 and for the -1 10 two regimes described in the text. Case (i) corresponds to −3 −3 r0 = 3.4 10 and case (ii) corresponds to r0 = 1 10 . The −8 −3 -2 other parameters are: α0 = 2.87 10 , ℓ = 4σ = 2.4 10 . 10 case (i) case (ii) α1 β1 nsim nteo nsim nteo n(t) (2) (2) (2) (2) -3 −1 −1 −2 −2 10 0.02 0.90 4.41 10 4.41 10 1.11 10 1.15 10 0.18 0.74 4.34 10−1 4.31 10−1 1.26 10−2 1.40 10−2 0.38 0.54 4.11 10−1 3.95 10−1 1.75 10−2 2.26 10−2 -4 10 0.58 0.34 3.73 10−1 3.33 10−1 2.27 10−2 3.75 10−2

-5 10 0 50 100 150 200 Under the assumption that the density is not very dif- time (trading minute) ferent from Gaussian at time t = 2, we can approximate the density with a first order Edgeworth expansion [14] FIG. 3: Mean of n(t), the number per unit of time that |r(t)| is 5 ℓ σ 2 2 exceeding the threshold = 4 , for 10 numerical simulation −r2 /2s2 of three different GARCH(1,1) models with the same value of e r2 κ4(t = 2) p(r2)= 1+ He4 4 , (11) α0 and of α1 + β1 = 0.92. The values of α1 are 0.02 (dotted √2π  s2  4! s2  line), 0.18 (solid line) and 0.38 (dashed line). This last value is the one estimated from the real time series of Black Monday where He4(x) is the fourth Chebyshev-Hermite polyno- aftercrash period. The gray line is the prediction based on the mial and κ4(t) is the fourth cumulant of rt. By integrat- approximation of Gaussian density. ing p(x) in Eq. (11) between ℓ and infinity we can find an expression for n(t = 2). By performing the integration, 2 we observe two different regimes: (i) when ℓ < 3(α0(1 + of α , n(t) decays to the stationary value in less than a 2 2 1 α1 + β1) + (α1 + β1) r0), n(t = 2) decreases when α1 in- 2 2 2 trading day. The decay is essentially exponential and a creases; (ii) when ℓ > 3(α0(1 + α1 + β1) + (α1 + β1) r ), −ct 0 best fit with a function n(t) = a + b e gives a good n(t = 2) increases when α1 increases. Table 1 shows two agreement. numerical examples corresponding to the two regimes. Fig. 3 also shows that the larger is α1 (and therefore The results of Table 1 show that the agreement between the unconditional kurtosis), the faster is the decay of n(t) simulated values and predicted values of n(t) is quite for very small values of t. This empirical behavior can be good especially for small values of α1. The fact that modelled and hereafter we show how it is possible to find for large values of α1 the agreement is less precise is due an estimate of n(t) for small values of t. Let us consider to the significant contribution coming from the moments that at time t = 0 the crash occurs and the return is r0. higher than the fourth that have been neglected in the At time t = 1 the density is Gaussian with zero mean Edgeworth expansion of Eq. (11). One could improve 2 and variance given by σ1 = α0 + (α1 + β1)r0 . However the forecast of n(t) by extending the Edgeworth expan- at subsequent times the density is no more Gaussian. As sion to higher order and by using the approximate density a first approximation we can assume that the density is so obtained to calculate a better estimation of n(t). Gaussian with time-dependent variance described by (7). The technique of Edgeworth expansion can also be This is a very rough approximation, which is valid only used to forecast the value of n(t) for value of t larger when α1 is very small. The gray line in Fig. 3 is ngau(t)= than 2. Our numerical simulations show that for moder- 2 erfc(ℓ/ 2σt ) which is the function obtained by using ate values of α1 and/or short forecast horizons the predic- this approximation.p For large values of t, n(t) tends to tions work quite well. For larger value of α1 (especially a constant value as e−t/2τ . The agreement is very good when the fourth moment of the unconditional density is in the case α1 = 0.02. However, for larger values of α1 infinite) the prediction is valid only for small values of t. a better approximation is needed to explain the results These theoretical and numerical observations confirm of numerical simulations. A better approximation can be that the GARCH(1,1) is unable to model the power law obtained by taking into account higher moments. Even decay of n(t) observed in real time series after a major if at time t = 1 the density is Gaussian, at t = 2 the market crash. density is no more Gaussian. Nevertheless we can find the exact expression for the expected value of the first four moments at t = 2 conditioned to the information at IV. VALUE AT RISK AFTER A CRASH time t = 0. One finds The empirical evidence and the theoretical model E (r )= E (r3)=0, (8) 0 2 0 2 shown in the previous sections have direct relevance for 2 2 2 E0(r2)= α0 + (α1 + β1)σ1 s2, (9) risk management. One of the most widely used measure 4 2 4 ≡ 2 E0(r2)=6α1σ1 + E0(r2). (10) of risk is the so called Value at Risk (VaR) [5]. The VaR 5 is the most probable worst expected loss at a given level V. CONCLUSIONS of confidence over a given time horizon. Here we indi- cate the probability density for index return r over the time horizon τ as fτ (r). The VaR ΛV aR associated to a Our empirical observations show that the statistical certain probability of loss PV aR is defined implicitly by properties of index return time series after a major fi- nancial crash are essentially different from the ones ob- −ΛV aR served far from the crash. Other examples of statistical fτ (r) dr = PV aR. (12) Z−∞ properties of market which are specific of the aftercrash period have been observed in the investigation of cross- In stationary market conditions, the VaR is determined sectional quantities computed for a set of stocks before, by the the density fτ (r). In Section 2 we have discussed a at and after financial crashes [15, 16]. The time period simple market model able to describe quantitatively the just after a crash is characterized by a relaxation to the non-stationary evolution of index return time series after typical market phase that can be modelled in terms of a market crash. In this model the shape of the return a power-law decay of the typical scale of index returns. density remains constant but its scale changes in time. We have shown that this observation rules out the possi- By using the relation r(t)= γ(t)rs(t), it is direct to show bility that the statistical properties of the time evolution that the (instantaneous) VaR after a market crash is not of index returns can be efficiently modelled in terms of constant but changes in time as simple autoregressive models, such as the GARCH(1,1) model, in the non-stationary period after a crash. Our (s) ΛV aR(t)= γ(t)ΛV aR, (13) modelling of the aftercrash period also suggests that the Value at Risk of a financial portfolio measured just after (s) where ΛV aR is the constant VaR obtained by using the a financial crash evolves in a non trivial way character- stationary probability density function fs(rs). We have ized by a power-law evolution lacking a typical scale. In seen [6] that the occurrence of the Omori law after a summary we conclude that the time period after a major market crash is consistent with a time evolution of the market crash is characterized by statistical regularities −β scale γ(t) = c1t + c2. Eq. (13) indicates that after a which are specific to such a time period and not well de- big market crash the VaR decreases in time as a power- scribed by models which presents a typical time scale in law with exponent β (close to 0.2 0.3 in the investigated their stochastic evolution. The authors thank INFM ASI crashes) and eventually relaxes to÷ a constant value. and MURST for financial support.

[1] Pagan, A., 1996. The econometrics of financial markets. H.E., 1998. Inverse Cubic Law for the Distribution of Journal of Empirical Finance. 3, 15-102. Stock Price Variations. European Physical Journal B 3, [2] Schwert, G.W., 1989. Why Does Stock Market Volatility 139-140. Change Over Time? The Journal of Finance. 44, 1115- [10] Sornette, D., Johansen, A., Bouchaud, J.-P., 1996. Stock 1153. Market Crashes, Precursors and Replicas. Journal de [3] Chen, N.-F., Cuny, C.J., Haugen, R.A., 1995. Stock Physique France 6, 167-175. Volatility and the Levels of the Basis and Open Inter- [11] Bollerslev, T., 1986. Generalized Autoregressive Condi- est in Futures Contracts. The Journal of Finance. 50, tional Heteroskedasticity. Journal of Econometrics 31, 281-300. 307-327. [4] Omori, F., 1894. On the after-shocks of earthquakes. J. [12] Information and the package can be dowloaded at the Coll. Sci. Imp. Univ. Tokyo 7, 111-200. web site http://www.egss.ulg.ac.be/garch/. [5] Duffie, D., Pan, J., 1997. An overview of Value at Risk. [13] Baillie, R.T., Bollerslev, T., 1992. Prediction in dynamic Journal of Derivatives. 4, 7-49. models with time-dependent coditional variances. Jour- [6] Lillo, F., Mantegna, R.N., 2001. Power-law relaxation nal of Econometrics. 52, 91-113. in a complex system: The fluctuation decay after a [14] Feller, W. 1971. An Introduction to Probability and Its financial market crash. http://xxx.lanl.gov/abs/cond- Application, Vol.2. J. Wiley & Sons, New York. mat/0111257. [15] Lillo, F., Mantegna, R.N., 2000. Symmetry alteration of [7] Embrechts, P., Kl¨uppelberg, C., Mikosch, T., 1997. ensemble return distribution in crash and rally days of Modelling Extremal Events for Insurance and Finance. financial markets. European Physical Journal B 15, 603- Springer-Verlag, Berlin Heidelberg. 606. [8] Lux, T., 1996. The stable Paretian hypothesis and the [16] Lillo, F., Mantegna, R.N., 2001. Empirical properties of frequency of large returns: an examination of major Ger- the variety of a financial portfolio and the single-index man stocks. Applied Financial Economics 6, 463-475. model. European Physical Journal B 20, 503-509. [9] Gopikrishnan, P., Meyer, M., Amaral, L.A.N., Stanley,