Article Preview
Top1. Introduction
Currently, in the Chinese stock market, there are more than 100 million individual investors, among which 95% are non-professional. Institutional investment only accounts for 10% of total market capitalization, and the majority of total outstanding shares are owned by individuals (Wang & Xu, 2004). These individual investors are in great need of mobile decision support system (apps) for trading analysis and decision making.
Although numerous commercial software and freeware, such as Dazhihui, Qianlong, Tonghuashun (from China), ProfitSource, eSingal, and VectorVest (from other countries, see http://stock-software-review.toptenreviews.com) are available to common investors, none of them incorporates data mining functions; they only provide information retrieval, statistical analysis, trade ordering, and technical analysis-based program trading. Their build-in trading rules/models are empirical and static, and thus they cannot adapt to the varying market in time. In addition, they cannot process large-scale financial data due to the limited computing ability of mobile devices (Goh & Taniar, 2004).
Data mining technologies are suitable to address these limitations because they can handle large-scale, non-linear, noisy data. These techniques are used to study data and discover new, hidden, unexpected, valuable trends or patterns from existing databases, and have gained increasing attention in science and business areas (Daly & Taniar, 2004). Client-Server (C/S) architecture can be deployed to partition workloads and share computing resources. Therefore, this paper is to develop a C/S-based Mobile Data-Driven Stock Trading System (iTrade), which can provide intelligent decision support for non-professional investors on mobile devices. The unique characteristics of iTrade can be outlined in three aspects.
First, a data-driven intelligent learning model is constructed for accurate stock (trend) prediction. Compared to the empirical technical trading rule-based stock analysis software, the proposed model is based on a well-known data mining algorithm, Support Vector Machine (SVM) (Vapnik, 1995), which is gaining more and more popularity in stock prediction (Hu et al., 2013; Huang, 2012; Lee, 2009).
Second, a concept drift adaptation process is proposed to identify market structure changes and adapt the learning model to these changes. This is about how to identify the most informative and up-to-date predictors (e.g., fundamental/technical factors) that can explain future excess returns. This process can be realized by combining feature selection and sliding window method (Tsai & Hsiao, 2010; Zhang et al., 2014). In this paper, feature selection is based on the Least Absolute Shrinkage and Selection Operator (Lasso) algorithm because of its effectiveness in sparse and consistent model selection (Zhao & Yu, 2006). Moreover, sliding window method is combined with Lasso to perform adaptive feature selection, which can handle the phenomenon of concept drift that the most informative predictors are ever-changing from time to time, especially from bull to bear market, vice versus. Concept drifts generally exists in stock market and might be derived from mass psychology, macroeconomic, the development of technology, and so on, which makes adaptive feature selection imperative.
Third, a rigorous benchmark analysis is provided to evaluate iTrade, including performance comparisons with 1) the quantified strategies of three world-famous master investors (Warren E. Buffett, William J. O’Neil and Richard Driehaus), and 2) the Buy-and-Hold strategy. This analysis indicates whether the data-driven intelligent model can defeat the human experts.