Stochastic Volatility in a Quantitative Model of Stock Market Returns

Abstract

Standard quantitative models of the stock market predict a log-normal distribution for stock returns (Bachelier 1900, Osborne 1959), but it is recognised (Fama 1965) that empirical data, in comparison with a Gaussian, exhibit leptokurtosis (it has more probability mass in its tails and centre) and fat tails (probabilities of extreme events are underestimated). Different attempts to explain this departure from normality have coexisted. In particular, since one of the strong assumptions of the Gaussian model concerns the volatility, considered finite and constant, the new models were built on a non finite (Mandelbrot 1963) or non constant (Cox, Ingersoll and Ross 1985) volatility. We investigate in this thesis a very recent model (Dragulescu et al. 2002) based on a Brownian motion process for the returns, and a stochastic mean-reverting process for the volatility. In this model, the forward Kolmogorov equation that governs the time evolution of returns is solved analytically. We test this new theory against different stock indexes (Dow Jones Industrial Average, Standard and Poor s and Footsie), over different periods (from 20 to 105 years). Our aim is to compare this model with the classical Gaussian and with a simple Neural Network, used as a benchmark. We perform the usual statistical tests on the kurtosis and tails of the expected distributions, paying particular attention to the outliers. As claimed by the authors, the new model outperforms the Gaussian for any time lag, but is artificially too complex for medium and low frequencies, where the Gaussian is preferable. Moreover this model is still rejected for high frequencies, at a 0.05 level of significance, due to the kurtosis, incorrectly handled.