Fractal nature of stock market behavior. Neural networks in Kuwait and Saudi Arabia

The presence of stock market inefficiency and bias of security returns suggests that the application of standard asset pricing models is inappropriate. This is so because of the random walk assumption. The validity of the random walk assumption determines the accuracy of the asset pricing models. In essence, investors need to understand to what extent it applies to the markets of Kuwait and Saudi Arabia. The purpose of this study is to test the random walk assumption in Kuwaiti and Saudi Arabian to assess the implications for sophisticated trading strategies. In other words, we test if these inefficiencies can produce economically significant profits. We use rescaled range analysis to estimate the fractal dimension of price returns and test the Efficient Market Hypothesis and the Fractal Market Hypothesis in the Kuwaiti and Saudi Arabian stock markets. By assessing the fractal dimension of asset returns, investors with recent innovation in financial prediction and computational power, can exploit these signals. And in markets that are less efficient, where no one has used these tools before, the hypothesis is expected to give interesting results. The research problem in hand is, "Are the market inefficiencies that have been identified in Kuwait and Saudi Arabia economically significant for sophisticated investors?";The Hurst Exponent for each time series is measured to access the predictability of a time series. We would expect the stock time series with a higher Hurst Exponent would have better results with a neural network/Genetic algorithm based trading rule. A technical trading system using Artificial Neural Networks and Genetic Algorithms is created for the largest stock in the MSCI index of the Kuwait Stock Exchange and the MSCI Index of the Saudi Arabian Stock Exchange (TADAWUL). After choosing the inputs, a combination of topologies are tested, and the best topology is chosen for each stock model. The output used is an optimal signal and a genetic algorithm is added to optimize the signal threshold. Finally, the best network is chosen for each model. We take into account the transaction costs for each market.;This study provides insights regarding financial trading in the Kuwaiti Stock market and the Saudi Stock Market. The primary objective of this study is to investigate whether the combination of neural networks and genetic algorithms can enhance trading returns, providing realistic profitable trading systems. Our formal testable hypothesis for the models is:;"Can trading strategies created using a combination of neural networks and genetic algorithms with technical/fundamental economic data outperform the buy-and-hold returns for stocks that have a fractal dimension less than 1.5";We find four stocks with large Hurst exponents and two with low ones. This suggests that the markets are not totally random in those periods. Some stocks had a strong trend structure that was learnt from the Artificial Neural Networks and Genetic Algorithm model while the others didn't. Since the Hurst exponent provides a measure of predictability, we can use it as a guide before a model is built for a specific stock or index. We can focus on periods with large Hurst exponents, which would lead to better forecasting. We accept the hypothesis. It is clear that our results do not support the Efficient Market Hypothesis for the four stocks that had high Hurst exponents. Models that outperformed the buy-hold strategy are inconsistent with the weak form of the Efficient Market Hypothesis, but consistent with the fractal market hypothesis. These stocks are Hurst processes.;Kuwaiti and Saudi Arabian stock market returns do not comply with the assumption and properties of a normal distribution in most cases, while in some cases it did. They also do not comply with the weak form of the efficient market hypothesis and the random walk assumption in most cases tested. In the absence of a normal distribution and the random walk assumption, asset-pricing models may not adequately capture the investment risk and probabilities of equity returns. In essence one should test the individual stock to find whether it qualifies for a normal distribution. One could take advantage of these inefficiencies if detected by using nonlinear dynamics, soft computing, or tools from chaos theory; in our case we used Artificial Neural Networks and Genetic Algorithms.