Extensive academic research was performed for the financial market as it is closely connected to practical economy. Research in traditional financial economics resulted in economic indicators and the econometrics was instrumental for quantitative rese...
Extensive academic research was performed for the financial market as it is closely connected to practical economy. Research in traditional financial economics resulted in economic indicators and the econometrics was instrumental for quantitative research in financial market. However, it proved to be difficult to predict the market behavior as it is a result of complex interaction among many agents with their own agenda. An effective tool to predict a change in market would be beneficial for policy makers and market participants to assist them with rational and consistent decision making. On the other hand, inconsistent prediction would lead to a suboptimal and inconsistent market activity which sometimes result in sudden collapse in the market as it did in 2008 Financial crisis and 1997 Asian financial crisis. The purpose of this dissertation is to develop approach based on econophysics and machine learning to systematically analyze the financial market.
The main focus of this dissertation involves the network structure of stock market. To predict the change in market behavior, it is critical to understand the relationship or correlation among the market participants beforehand, and complex network analysis is one of the most prominent methods for such study. The fractal theory was employed as the primary approach to analyze the network structure of financial market. The empirical study shows that the network of financial market exhibits fractal properties. Also, analysis of fractal dimension and network topology led to two key discoveries. First, the fractal dimension and the Strong effective repulsion between distinct network nodes known as the hub are closely related. Second, the fractal dimension reveals the shortcut of network structure. Through further analysis, these two properties were proved to be useful for risk management in financial market. Three fractal measures were proposed to specify network structure for ease of implementation in future studies.
In the second step, the fractal measures were implemented in a financial market to assess its ability to predict the market movement. Recently, studies were conducted to determine if a new measure or index improves the prediction accuracy for financial time series. These studies are advantageous for future studies as it proposes new indices for other implementation and further analysis rather than studying the precision of their own method. In this paper, machine learning algorithms were employed to assess the predictive properties of fractal measures. Empirical experiments were performed to predict direction of market movement, which is effectively a classification task, and prediction for returns, a regression task. The studies concludes that the fractal measure proposed were effective in prediction for long-term stock returns of more than three months period.
Finally, a model to improve trading strategy based on learning-to-rank algorithm and the fractal measures was introduced. Previous studies are often based on the modern portfolio theory(MPT), but it is insufficient for real-world application as it doesn’t provide any implication for rebalancing period of portfolio. The optimal rebalancing model proposed in this study allows its application with traditional portfolio methods. The experiments were carried out in two steps. The model learns to predict the better time period to perform rebalancing between two time periods in the future, followed by the empirical simulation to apply the model in real world trading scenario. Two traditional portfolio methods, equal weighted and maximized Sharpe ratio, were taken for experiment. The result affirms that the optimal rebalancing model was able to capture the better time period of rebalancing portfolio. In addition, the model outperformed a simple rebalancing method of fixed time period. When the fractal measures were employed as an input variable, the model performance was further improved. The primary contribution achieved through this model is that it allows application and expansion into all traditional portfolio models. Also, the fractal measures observed in the network structure grants insight regarding the market behavior and empirically proved that the measure provides benefit in prediction for the real-world stock market.