LONG-RANGE MEMORY IN NONLINEAR GARCH(1,1) MODELS

Aleksejus Kononovicius1, Julius Ruseckas1

1 Vilnius University, Institute of Theoretical Physics and Astronomy, Lithuania

Econophysics started out as purely empirical science. Physicists around the world analyzed huge amounts of financial market data and discovered certain statistical regularities, which is now often referred to as stylized facts [1, 2]. One of these regularities is long-range, decaying as power-law, correlations often referred to as long-range memory [2]. Power spectral density (PSD) of the time series with long-range memory behaves as $1/f^{\beta}$ with $\beta$ values being close to 1.

At the same time economists proposed, applied and further developed auto-regressive conditionally heteroskedastic (ARCH) models [3]. The basic idea behind ARCH models lies in an assumption that certain heteroskedastic economical observables, $z_t$, may be modeled as being composed of two parts - stochastic part, $\omega_t$, and time dependent volatility, $\sigma_t$:

$$z_t = \sigma_t \omega_t\tag{1}$$

Stochastic part may be usually assumed to follow Gaussian distribution (though it may depend on actual application), while time dependent part may be assumed to be driven by the following generalized GARCH(p,q) process (though there are numerous other possibilities, once again dependent on actual applications):

$$\sigma_{t}^{2} = a + \sum_{i=1}^{p} b_{i} \sigma_{t-i}^{2} \omega_{t-i}^{2} + \sum_{i=1}^{q} c_{i} \sigma_{t-i}^{2}.\tag{2}$$

In this case if $p = q = 1$, then new $\sigma_t$ values depend only on the immediate prior $\sigma_{t-1}$ value. As $\sigma_t$ does not depend on further history, any $\sigma_{t-i}$ ($i > 2$), the model appears to be memory-less - it exhibits Markov property.

GARCH(1,1) process, Eq. (2) with $p = q = 1$, may be rewritten in continuous time as stochastic differential equation (SDE):

$$dy = (A - Cy)dt + |B|ydW_t,\tag{3}$$

where $W_t$ is Wiener process (standard one dimensional Brownian motion) and with the following relation of parameters: $Ah = a$, $B^2h = 2b_1^2$ and $Ch = 1 - b_1 - c_1$ (in the infinitesimally small time step limit, $h \to 0$). The obtained SDE may be further compared to a general class of SDEs exhibiting power law statistics, including $1/f^{\beta}$ PSD:

$$dx = \sigma^2 \left( \eta - \frac{1}{2} \lambda \right) x^{2\eta - 1} dt + \sigma x^{\eta} dW_t,\tag{4}$$

which was proposed in [4]. From the comparison it is evident that linear GARCH(1,1) will not be able to reproduce $1/f^{\beta}$ noise as $\eta = 1$. Though it is able to reproduce another interesting statistical feature - power-law distribution.

We consider two possible nonlinear modifications of Eq. (2), with $p = q = 1$:

$$\sigma_t^2 = a + b_1 \sigma_{t-1}^{\mu} \omega_{t-1}^{\mu} + c_1 \sigma_{t-1}^2,\tag{5}$$

where $\mu > 2$ is an odd integer, and

$$\sigma_t^2 = a + b_1 \sigma_{t-1}^{\mu} |\omega_{t-1}|^{\mu} + \sigma_{t-1}^2 - c_1 \sigma_{t-1}^{\mu},\tag{6}$$

where $\mu$ may be any positive real number. Both of these nonlinear GARCH(1,1) processes are able to reproduce power-law distributions, with exponent $\lambda = \mu$, as well as PSD of the $1/f^{\beta}$ form, with $\beta = 1 + \frac{\mu - 3}{\mu - 2}$, [5].


[1] A. Chakraborti, I. M. Toke, M. Patriarca and F. Abergel, Econophysics review: I. Empirical facts, Quantitative Finance 7, 991-1012 (2011).

[2] R. Cont, Long range dependence in financial markets (Springer, 2005).

[3] T. Bollerslev, Glossary to ARCH (GARCH) (CREATES, 2008).

[4] B. Kaulakys and J. Ruseckas, Stochastic nonlinear differential equation generating 1/f noise, Physical Review E 70, 020101 (2004).

[5] A. Kononovicius and J. Ruseckas, Nonlinear GARCH model and 1/f noise, to appear in Physica A (2015). arXiv: 1412.6244 [q-fin.ST].