Financial predictions using intelligent systems

Abstract

This thesis presents a collection of practical techniques for analysing various market properties in order to design advanced self-evolving trading systems based on neural networks combined with a genetic algorithm optimisation approach. Nonlinear multivariate statistical models have gained increasing importance in financial time series analysis, as it is very hard to fmd statistically significant market inefficiencies using standard linear modes. Nonlinear models capture more of the underlying dynamics of these high dimensional noisy systems than traditional models, whilst at the same time making fewer restrictive assumptions about them. These adaptive trading systems can extract information about associated time varying processes that may not be readily captured by traditional models. In order to characterise the fmancial time series in terms of its dynamic nature, this research employs various methods such as fractal analysis, chaos theory and dynamical recurrence analysis. These techniques are used for evaluating whether markets are stochastic and deterministic or nonlinear and chaotic, and to discover regularities that are completely hidden in these time series and not detectable using conventional analysis. Particular emphasis is placed on examining the feasibility of prediction in fmancial time series and the analysis of extreme market events. The market's fractal structure and log-periodic oscillations, typical of periods before extreme events occur, are revealed through recurrence plots. Recurrence qualification analysis indicated a strong presence of structure, recurrence and determinism in the fmancial time series studied. Crucial fmancial time series transition periods were also detected. This research performs several tests on a large number of US and European stocks using methodologies inspired by both fundamental analysis and technical trading rules. Results from the tests show that profitable trading models utilising advanced nonlinear trading systems can be created after accounting for realistic transaction costs. The return achieved by applying the trading model to a portfolio of real price series differs significantly from that achieved by applying it to a randomly generated price series. In some cases, these models are compared against simpler alternative approaches to ensure that there is an added value in the use of these more complex models. The superior performance of multivariate nonlinear models is also demonstrated. The long-short trading strategies performed well in both bull and bear markets, as well as in a sideways market, showing a great degree of flexibility and adjustability to changing market conditions. Empirical evidence shows that information is not instantly incorporated into market pnces and supports the claim that the fmancial time series studied, for the periods analysed, are not entirely random. This research clearly shows that equity markets are partially inefficient and do not behave along lines dictated by the efficient market hypothesis.