An evolutionary trend reversion model for stock trading rule discovery
Introduction
Quantitative investment (QI) is an important research topic for stock trading in the big data era [25]. Big data in stock market are of high volume, high velocity, high variety and high variability (4V), including heterogeneous data sources—real-time (or online) and offline transaction data of different time granularities, as well as (textual) news, financial statements and macroeconomic statistics released by governments, institutions, social media and listed companies. Great challenges exist for knowledge discovery in stock trading/transaction data because they are generally known to be non-linear, noisy, and of high dimension [4], [5]. Different from the traditional expert/individual experience-based trading strategies, QI models adopt systemic and quantitative methods, such as statistics, artificial intelligence, machine learning (ML) and data mining techniques. Owing to these techniques of big data analysis, QI models are effective in discovering knowledge from abundant homogeneous and/or heterogeneous stock market data, and thus have attracted increasing interest.
However, “black-box” QI models, such as those based on neural networks [22], lack interpretability and cannot provide understandable knowledge about the inter-actions between trading strategy indicators (i.e., model inputs) and price movement (i.e., model output). Investors often feel uncomfortable and insecure about making stock trading decisions with a “black-box” model [10]. To transform this negative perception, researchers have invested a lot of effort in finding comprehensible stock trading rules, especially those correspond with classical investment strategies, such as trend following and reversion [11], [13]. Therefore, both “black-box” and “comprehensibility” issues must be considered.
To address the “black-box” issue, the eXtended Classifier Systems (XCS) is quite suitable. XCS has excellent learning and explicit expression abilities, because it delicately combines three representative ML technologies, including classification rule mining (i.e., non-black-box), evolutionary learning and reinforcement learning. In stock market prediction/trading, XCS-based QI models were demonstrated to be more profitable compared to the random and buy-and-hold strategies [9], [18]. Advantages of XCS include: (1) XCS can properly learn explicit rules from noisy, complex and non-linear environment that continuously changes, (2) XCS can make real-time and accurate learning and response, and (3) XCS can adjust itself to strengthen its inward knowledge step by step [33]. However, the original pure explore mode of XCS is not optimal for stock trading because this random-exploring strategy cannot adequately use historical information. For instance, when previous information regarding stock price movement is available, it is unnecessary to “randomly guess” the best action (i.e., buy or sell a stock). Thus, it is imperative to design a new strategy/mode for “learning from history”.
To address the “comprehensibility” issue, it is necessary to integrate classical investment strategies into ML-based QI model. Trend-following QI models have been widely examined as simple but effective investment tools [10], [12], [11]). Trend-reversion ML-based QI model, however, has attracted less attention. Actually, the premise of trend-reversion strategy is the mean reversion/reversal theory of stock price movement, which has been statistically verified in numerous studies [3], [11]. Our paper explores the profitability of using sole trend-reversion indicators in a ML-based QI model.
This paper proposes a novel XCS with learn mode (XCSL) and establishes an Evolutionary Trend Reversion Model (eTrendRev) that integrates trend-reversion strategy with XCSL. The eTrendRev is highlighted in three aspects: (1) XCSL is able to generate explicit rules that are more understandable than black-box models, such as neural networks, and can provide justifiable knowledge to guide stock trading; (2) the original pure explore mode of XCS is substituted by the proposed learn mode, which can make full use of historical information and lead to more stable and better performance; (3) a variety of trend-reversion strategies are integrated and made dynamic through the evolutionary learning framework of XCSL.
To evaluate eTrendRev, experiments were carried out on the historical data of the Shanghai Composite (SH) Index of the Chinese stock market and the NASDAQ Composite (NASDAQ) Index of the US stock market. In addition, a brief analysis of buy/sell signals on market turning points is provided.
The rest of the paper is organized as follows. A brief review of related works is presented in Section 2. The architecture and internal logic of eTrendRev are described in Section 3. Section 4 presents experiment results and analysis. Finally, a summary and conclusion is provided in Section 5.
Section snippets
Quantitative stock trading strategy
With respect to stock trading, a quantitative strategy differs from a discretionary strategy in that quantitative trading can reduce the arbitrariness of discretionary trading by utilizing computerized and systematic methods to eliminate subjective decisions driven by emotion, indiscipline, passion, greed, and fear. Generally, quantitative models for stock selection and/or timing can be grouped into two categories: (1) theory-driven and (2) data-driven [26]. They have the same objective of
Description
An overall picture of eTrendRev is given in Fig. 1. The eTrendRev takes XCSL as its core, and is designed for selecting the optimal actions in a stock market environment. Following the representative presentation in [35], Fig. 1 mainly depicts the XCS mechanism, including its data sources, data flow and process flow.
With respect to data sources, input data of eTrendRev can be offline and online daily stock trading data, including the open, high, low, closing prices, etc. These data are first
Experiment
In this section, eTrendRev was evaluated in a back-testing experiment. As a reaction to the current stock market states, eTrendRev selects and performs the proper buy/sell action, and then evolves itself according to the subsequent market reward. In addition, eTrendRev is periodically re-trained based on the available historical data, i.e., the previous data before the time of training.
Conclusion
To cope with the challenge of big data analysis in QI, this paper proposes a model called eTrendRev, which integrates the proposed XCSL with trend-reversion strategy. In the context of 4V in big data era, eTrendRev can adapt itself to the varying stock market environment owing to the evolutionary and reinforcement learning capabilities. Experimental results showed that eTrendRev can generate effective buy/sell signals and outperform the buy-and-hold strategy with high Sortino ratio after
Acknowledgments
This research was partly supported by the National Natural Science Foundation of China (71271061, 70801020), Major Project of National Social Science Foundation of China (14ZDA074), Science and Technology Planning Project of Guangdong Province, China (2010B010600034, 2012B091100192), and The Ministry of Education Innovation Team Development Plan, Guangdong Natural Science Foundation Research Team, and Business Intelligence Key Team of Guangdong University of Foreign Studies (S2013030015737,
References (38)
- et al.
Using genetic algorithms to find technical trading rules
J. Financ. Econ.
(1999) - et al.
Momentum and mean reversion across national equity markets
J. Empir. Financ.
(2006) - et al.
An adaptive network-based fuzzy inference system (ANFIS) for the prediction of stock market return: the case of the Istanbul stock exchange
Expert Syst. Appl.
(2010) - et al.
A comparison between Fama and French’s model and artificial neural networks in predicting the Chinese stock market
Comput. Oper. Res.
(2005) - et al.
Mining associative classification rules with stock trading data – a GA-based method
Knowl.-Based Syst.
(2010) - et al.
Integrating extended classifier system and knowledge extraction model for financial investment prediction: an empirical study
Expert Syst. Appl.
(2006) - et al.
Mean reversion in stock market prices: New evidence based on bull and bear markets
Res. Int. Bus. Financ.
(2010) - et al.
Trend following algorithms in automated derivatives market trading
Expert Syst. Appl.
(2012) - et al.
An examination of evolved behavior in two reinforcement learning systems
Decis. Support Syst.
(2013) - et al.
Applying a GA kernel on optimizing technical analysis rules for stock picking and portfolio composition
Expert Syst. Appl.
(2011)