Classical theories of economics were developed in the times then elaborate analysis of empirical data was impossible. Therefore classical economics are unable to give convincing explanation of interesting effects discovered, while analyzing high frequency ($min^{-1}$ – $h^{-1}$) time series. One of those, in classical sense, anomalous, but universal, effects is power law statistical properties observed in various financial markets around the world [1]. Though in scientific literature one usually finds analysis of mature financial markets (such as New York Stock Exchange (further NYSE)), which exhibit high liquidity, while juvenile markets, which exhibit low liquidity, are neglected. Therefore it is of scientific interest to know how statistical properties scale as liquidity drops.
We analyzed tick by tick trades of 4 stocks traded on NASDAQ OMX Vilnius Stock Exchange (further VSE), low liquidity financial market, for 50 months since May 2005, empirical data was provided by VSE. We also extended our previous analysis [2] of NYSE empirical data – 24 stocks traded for 27 months since January 2007. By definition differing market liquidity causes evident difference in mean inter-trade times – 362 s in VSE, 3.02 s in NYSE. But scaling of statistical properties of absolute return, which we define as
$$r_{\Delta t}(t) = [\ln p(t) - \ln p(t - \Delta t)], \tag{1}$$
where $p(t)$ is price function of time and $\Delta t$ is width of return time window (we use 60 s, 600 s and 1800 s values), seems less evident. We obtained perfect match of distributions by ignoring zero return values, probability of which dramatically rises if mean inter-trade times are not significantly smaller than return time window. While spectral density, obviously, can't be matched by using previous technique, though power spectra from both financial markets seem to overlap better with larger absolute return time windows. Obtained matches allows us to conclude that main reason for visible differences in statistical properties is differing financial market liquidity.
We proposed to model return as double stochastic process with long-range memory [2,3]. In order to improve control of model results, we rewrote main model stochastic differential equation as
$$dx = \left[\eta - \frac{\lambda}{2} - \left(\frac{x}{x_{max}}\right)^2\right] \frac{(1+x^2)^{\eta-1}}{\left(1+\epsilon\sqrt{1+x^2}\right)^2} x dt_s + \frac{(1+x^2)^{\frac{\eta}{2}}}{1+\epsilon\sqrt{1+x^2}} dW_s, \tag{2}$$
where $x$ is dimensionless return, $t_s$ is dimensionless time and $W_s$ is accordingly scaled Wiener process. As in [2, 3], we integrate solutions of Eq. (2) within dimensionless time intervals, corresponding to return calculation time window $\Delta t$, and modulate them with $q$-Gaussian [4] noise in order to obtain time series statistically comparable with empirical data. Using same model parameter set we obtained match between statistical properties exhibited by NYSE, VSE and our model in three different time scales ($\Delta t$ values 60 s, 600 s and 1800 s).