Elsevier

Expert Systems with Applications

Volume 59, 15 October 2016, Pages 195-207
Expert Systems with Applications

An adaptive stock index trading decision support system

https://doi.org/10.1016/j.eswa.2016.04.025Get rights and content

Highlights

Abstract

Predicting the direction and movement of stock index prices is difficult, often leading to excessive trading, transaction costs, and missed opportunities. Often traders need a systematic method to not only spot trading opportunities, but to also provide a consistent approach, thereby minimizing trading errors and costs. While mechanical trading systems exist, they are usually designed for a specific stock, stock index, or other financial asset, and are often highly dependent on preselected inputs and model parameters that are expected to continue providing trading information well after the initial training or back-tested model development period. The following research leads to a detailed trading model that provides a more effective and intelligent way for recognizing trading signals and assisting investors with trading decisions by utilizing a system that adapts both the inputs and the prediction model based on the desired output. To illustrate the adaptive approach, multiple inputs and modeling techniques are utilized, including neural networks, particle swarm optimization, and denoising. Simulations with stock indexes illustrate how traders can generate higher returns using the developed adaptive decision support system model. The benefits of adding adaptive and intelligent decision making to forecasts are also discussed.

Introduction

Computer technology is becoming increasingly involved in the area of finance. The high computation capability of computers enables one to do complicated analysis utilizing large amounts of historical data. As previously studied (Balvers et al., 2000, Chaudhuri and Wu, 2003, Fliess and Join, 2009, Lee et al., 2010, Lo and MacKinlay, 1988, Lo and MacKinlay, 2011, Rahman and Saadi, 2008, Semenov, 2008), stock prices do not completely move in a random walk, but are nonlinearly related to historical data, as well as various other fundamental, technical, and macroeconomic factors. In studying these nonlinear relationships, nonlinear regression and neural networks have often been utilized for prediction and clustering (Atsalakis and Valavanis, 2009a, Atsalakis and Valavanis, 2009b, Chavarnakul and Enke, 2009, Enke and Mehdiyev, 2013, Hadavandi et al., 2010, Huang and Tsai, 2009). Systems that use technical analysis, along with expert systems and other forms of computational intelligence for chart pattern recognition, have also been developed (Bogullu et al., 2002, Cervelló-Royo et al., 2015, Gorgulho et al., 2011, Lee and Jo, 1999, Leigh et al., 2002, Lin et al., 2011, Thawornwong et al., 2001, Yamaguchi, 1989). Nonetheless, such systems are often sensitive to the selection of input variables and parameter settings, and typically do not offer the ability to adapt once trained or modeled for a specific stock, stock index, or other financial asset.

Decision support systems (DSS) have been recognized as a useful tool for helping users plan and make decisions. DSS hides all theories, modeling, and algorithms behind a friendly user interface. Research has been conducted to determine the best DSS framework within various industries (Berner, 2007, Hoogenboom et al., 2004, Kawamoto et al., 2005, Klein and Methlie, 2009, Kuo et al., 2001, Martinez et al., 2010). However, less work has been done in the financial industry, in particular, with the stock market. Most application of decision support systems for the stock market focus on a simple screening based on multiple criteria (Huang et al., 2005, Quah, 2008, Sevastjanov and Dymova, 2009), while few models have tried to provide intelligent trading assistance through an investigation and analysis of historical and current data (Albadvi et al., 2007, Atsalakis and Valavanis, 2009b, Kuo et al., 2001, Lai et al., 2009). Before the popularity of neural networks, most research focused on time series analysis, which is used to extract meaningful statistics and forecast future events based on historical data. Models developed from time series theory that are widely used in the financial area include ARIMA, ARCH, and GARCH, among others (Bollerslev, 1986, Box et al., 2013, Engle, 1982, Vejendla and Enke, 2013a, Vejendla and Enke, 2013b). However, interest continues to increase for using various forms of computational intelligence, such as artificial neural networks, to forecast individual stock and stock market prices.

Kuo (1998) proposed a DSS for the stock market through integration of fuzzy neural networks and fuzzy Delphi, in which both quantitative and qualitative data were studied. The neural network used develops a relationship between independent values and the dependent variable, such as the tendency of the stock market to rise or fall. The authors used a fuzzy neural network to initialize the weights to overcome the backpropagation learning algorithms inability to respond efficiently in a reasonable time. There are several other approaches developed that utilize a hybrid model of neural networks and fuzzy logic. For example, Atsalakis and Valavanis (2009a) developed a neuro-fuzzy adaptive control system that uses an adaptive neuro-fuzzy inference system controller to direct the stock market process model. Cheng, Chen, Teoh, and Chiang (2008) proposed a fuzzy time-series model that improves the adaptive expectation model technique, allowing modification of forecasts based on recent prices. Hadavandi et al. (2010) combine genetic fuzzy systems (GFS) and self-organizing map (SOM) neural networks for building a stock price forecasting expert system. The model involves three stages: (1) stepwise regression analysis to choose the key variables that are to be considered in the model; (2) categorize the data set into k clusters using SOM; and (3) feed all clusters into independent GFS models with the ability of rule base extraction and database tuning.

Samaras and Matsatsinis (2004) describe a multi-criteria decision support system that aims at presenting an evaluation of stocks on the Athens Stock Exchange on the basis of fundamental analysis. Tsaih, Hsu, and Lai (1998) developed a hybrid AI system for forecasting the S&P 500 stock index. They first applied a rule-based system for providing training samples. A reasoning neural network was then employed to generate trading triggers. Tsai and Hsiao (2010) combine multiple feature selection methods to find more representative variables for better prediction. In their research, principle component analysis, genetic algorithms, and a decision tree (CART) were used. Using representative variables that had been filtered, a backpropagation neural network was then developed to predict the stock price. Amornwattana, Enke, and Dagli (2007) illustrate the use of a hybrid options pricing model that uses multiple neural network models, along with a decision support system structure, to identify the correct stock options trading strategy based on direction and volatility forecast, as well as return forecasting (Enke & Amornwattana, 2008). The use of technical analysis, along with employing a neural network as a decision maker, has also been used for stock trading (Chavarnakul and Enke, 2009, Thawornwong et al., 2003)

Previous research illustrated that neural networks can be an effective tool for stock market prediction. Additional studies have also been done to consider the variables that are selected as inputs to the neural network. Chang, Liu, Lin, Fan, and Ng (2009) developed a three-stage system, CBDWNN, for stock trading prediction. These stages include: (1) screening out the potential stocks and the important influential factors; (2) using a backpropagation neural network to predict the buy/sell points of stock price; and (3) adopting case-based dynamic windows to further improve the forecasting results. Thawornwong and Enke (2003) performed an adaptive selection of financial and economic variables for artificial neural networks, utilizing methods for finding input data with the most information content.

However, most of the previous research tried to find one model that could be applied to all stocks. When it comes to the entire stock market, it is questionable whether we can find a general model to fit every stock or index. A model fitting one security well may perform poorly for another security. Such a finding will be confirmed later in this paper. To solve this problem, an adaptive stock index trading support system has been developed which utilizes an artificial neural network for predicting the future stock index price movement direction, adaptively creating a new model for every single stock index, rather than applying a generalized model. This information is then used to help users make better trading decisions. A particle swarm algorithm has been integrated into the system to overcome the disadvantages of the backpropagation algorithm, a common learning algorithm used for training neural networks. Furthermore, most researchers have focused on predicting the stock index price. However, it is believed that if one can predict the future direction of stock index price movement, the prediction result could be more helpful for making profit. Therefore, a system has been developed for forecasting the price movement direction instead of price level itself. Research has shown that such a forecast can often result in more accurate trading results (Enke and Thawornwong, 2005, Thawornwong and Enke, 2003).

The remainder of the paper begins with a review of previous research in the areas of neural networks and particle swarm optimization. This is followed by an overview of the architecture and methodology of the proposed trading system. Finally, numerous simulation results will be demonstrated, both illustrating how the system works and evaluating how well the system performs.

Section snippets

Artificial neural network

Artificial neural networks (ANN) are inspired by the functional and structural aspects of biological neural networks. ANNs have become an important tool in statistical data modeling, especially for discovering the non-linear relationship between an input and output dataset. An ANN consists of groups of neurons that are interconnected. The architecture usually includes one input layer, one or more hidden layers, and an output layer. The ANN updates its interconnection weights toward optimization

System overview

An adaptive stock direction prediction system has been developed in this paper. The system utilizes an artificial neural network to forecast the movement of future stock prices. Direction information is then used as trading signal in order to assist users in making trading decisions. The output from the neural network will be standardized between zero and one: output from 0.5 to 1 will be categorized to one, indicating that the stock index price is going up; otherwise, output from 0 to less

Adaptive methodology overview

The accuracy of the prediction and network performance are greatly impacted by the selection of the inputs for the neural network. Initially, a general model was developed to describe the relationship among stock index variables. Principal component analysis, genetic algorithms, and decision trees are widely used for filtering representative variables. However, for this research it will be very difficult to determine the best fit model when considering multiple stock indices, such that there

Independent variables

Analysis starts from the selection of potential variables, which will be used in the model as inputs. Economic variables are usually released monthly and quarterly and typically stay the same for a specific period. As a result, economic variables are excluded from our daily forecast. For daily prediction, the most important parameters are the closing price and volume of the stock index for the last trading day. These two pieces of information contain the most recent stock index trading

Forecast evaluation

The proposed system trained the neural network with two years of daily data from January 2008 through December 2009. A twelve-month sample from January 2010 through December 2010 was retained to access the prediction performance of the neural network.

Denoising approach

Given that excessive trading could eliminate potential profit opportunities, it is worth considering better ways to deal with stock market volatility. While the prices of stock indices fluctuate daily, the price movements do not always reflect the value of the indices or their associated long-term trends. This is due in part to various noises in the stock market, such as those caused by speculation and program trading, among others. In addition, selling stocks at every downturn will result in

Impact of transaction costs

As demonstrated above, the average number of transactions is 14 and 50 for the denoising approach and the model without denoising, respectively. The average returns are 42%, 28%, and 13% for denoising approach, trading following signals, and buy-and-hold strategy, respectively. Assuming that $100 K is invested through a typical online discount broker, that the broker is charging $9.99 for each transaction, and that an investible ETF is available for each index, the results in Table 11 can be

Short selling

The previous study has shown that the system helps significantly increase returns. To further improve the performance, short selling is considered. In general, short sellers hope to profit from a decline in the price of stocks - they still use a buy low and sell high approach, but initially sell at a higher price after borrowing a security, hoping to buy the security back later at a lower price before returning the borrowed security. By predicting the movement direction of a stock's price, it

Conclusions

For this research, an adaptive intelligent stock trading decision support system has been proposed that utilizes particle swarm optimization and an artificial neural network to predict a stock index's future movement direction. While technical analysis has been used in other trading systems, researchers have often focused on price sequences and patterns, with some form of computational intelligence being used to help identify such patterns (Cervelló-Royo et al., 2015, Lee and Jo, 1999, Leigh et

References (60)

  • A. Gorgulho et al.

    Applying a GA Kernel on optimizing technical analysis rules for stock pricing and portfolio composition

    Expert Systems with Applications

    (2011)
  • E. Hadavandi et al.

    Integration of genetic fuzzy systems and artificial neural networks for stock price forecasting

    Knowledge-Based Systems

    (2010)
  • HuangC.L. et al.

    A hybrid SOFM-SVR with a filter-based feature selection for stock market forecasting

    Expert Systems with Applications

    (2009)
  • HuangS.M. et al.

    A study on decision factors in adopting an online stock trading system by brokers in Taiwan

    Decision Support Systems

    (2005)
  • KuoR.J. et al.

    An intelligent stock trading decision support system through integration of genetic algorithm based fuzzy neural network and artificial neural network

    Fuzzy Sets and Systems

    (2001)
  • LaiR.K. et al.

    Evolving and clustering fuzzy decision tree for financial time series data forecasting

    Expert Systems with Applications

    (2009)
  • LeeC.C. et al.

    Stock prices and the efficient market hypothesis: Evidence from a panel stationary test with structural breaks

    Japan and the World Economy

    (2010)
  • LeeK.H. et al.

    Expert system for predicting stock market timing using a candlestick chart

    Expert Systems with Applications

    (1999)
  • W. Leigh et al.

    Stock market trading rule discovery using technical charting heuristics

    Expert Systems with Applications

    (2002)
  • LinX. et al.

    Intelligent stock trading system based on improved technical analysis and echo state network

    Expert Systems with Applications

    (2011)
  • QuahT.S.

    DJIA stock selection assisted by neural network

    Expert Systems with Applications

    (2008)
  • A. Rahman et al.

    Random walk and breaking trend in financial series: An econometric critique of unit root tests

    Review of Financial Economics

    (2008)
  • A. Semenov

    Testing the random walk hypothesis through robust estimation of correlation

    Computational Statistics & Data Analysis

    (2008)
  • P. Sevastjanov et al.

    Stock screening with use of multiple criteria decision making and optimization

    Omega

    (2009)
  • TsaiC. et al.

    Combining multiple feature selection methods for stock prediction: Union, intersection, and multi-intersection approaches

    Decision Support Systems

    (2010)
  • TsaihR. et al.

    Forecasting S&P 500 stock index futures with a hybrid AI system

    Decision Support Systems

    (1998)
  • T. Yamaguchi

    A technical analysis expert system in the stock market

    Future Generation Computer Systems

    (1989)
  • ZhangJ.R. et al.

    A hybrid particle swarm optimization back-propagation algorithm for feedforward neural network training

    Applied Mathematics and Computation

    (2007)
  • S. Amornwattana et al.

    A hybrid options pricing model using a neural network for estimating volatility

    International Journal of General Systems

    (2007)
  • R. Balvers et al.

    Mean reversion across national stock markets and parametric contrarian investment strategies

    The Journal of Finance

    (2000)
  • Cited by (124)

    View all citing articles on Scopus
    View full text