Enhancing Stock Trend Prediction using Computational Intelligence Techniques

Abstract

The highly volatile stock market forecasting has been a subject of interest for many individuals, researchers, and traders. Such time-series data can be computationally analyzed with an aid of nancial literacy to derive hidden market patterns and to predict potential future market trends. This thesis primarily aims to integrate computational intelligence techniques for enhancing stock trend prediction. For this purpose, the proposed approaches target speci c modules such as pre-processing, preparation, parameter optimization, and prediction in respective chapters; also, the necessity to address these modules is discussed for the primary goal of stock trend prediction. This thesis introduces the concept of cross-referencing for companies having been listed on two or more stock exchanges. A cross-reference to exchange-based stock trend prediction approach is proposed for companies having been listed on multiple domestic stock exchanges. This approach is further extended for companies having been listed on domestic, as well as international, stock exchanges to demonstrate the impact of considering data from di erent exchanges for the same company. While the performance is evaluated based on limited features, the thesis signi es the importance of feature selection by adopting coe cient of variation for feature selection. This method is supported with an existing k-means algorithm, as well as two proposed techniques, median range and top-M, to further examine the impact of selecting features of speci c variability characteristics. Another way of enhancing the prediction performance is proposed through information retrieval-based term frequency{inverse document frequency approach. This concept of term importance is adapted for numerical data and a feature weight matrix is created to highlight features in a way to enhance forecasting of stock price movement direction. One of the essential parts in handling non-linear stock market data is the identi cation of inherent patterns. The temporal data series requires to undergo di erent operations that can determine meaningful information. For this purpose, this thesis concentrates on various fusion aspects that can transform data into useful presentations. One of the proposed approaches determines how data sampling can be useful and integrated quantization to discretize the continuous temporal stock market data. The other approach is proposed based on representation fusion; here, time-series data vi are presented with triangle shapes so as to concentrate on the major modi cations in data values of consecutive days. While these two methods individually fuse data at di erent levels, the third approach is proposed with multi-level fusion such that data type, representation, and decision are fused within a proposed approach to predict stock trend. With an understanding of the data, features, and models, it can be observed that selection of an appropriate set of features as well as identi cation of an optimal set of parameters can be crucial aspects in forecasting. Therefore, this thesis further considers the signi cance of metaheuristics in stock trend prediction. One of the proposed approaches is based on the concept of genetic memory that is given to mutation operation of the genetic algorithm to preserve the contributing gene positions for parameter optimization and hence, an enhanced prediction performance. Subsequently, the concept of genetic diversity is introduced through inter-intra crossover and adaptive mutation operations within genetic algorithm to address parameter optimization and feature selection tasks in parallel. The prediction results of each of the proposed approaches are elaborated with a step-by-step procedure and evaluated with a series of experimentations. Also, the potential future research directions are discussed based on the proposed approaches. This thesis primarily concentrates on the two major concepts of computation intelligence viz. neural networks and metaheuristics for one of the complex real-world problems of stock trend prediction. The motivation behind various concepts elaborated and proposed in this thesis can provide the rationale behind the development of these approaches. The results indicate the applicability and expansibility of the thesis in addressing complex real-world stock trend prediction problems.