Article Preview
TopIntroduction
In financial trading, two major approaches are used for forecasting the price trend: technical analysis and fundamental analysis. In fundamental analysis, the most used indicators are market policy, financial statements, earnings, and sales on the asset price as well as macroeconomic and industry factors (Frankel & Froot, 1990). Technical analysis is based on analysing the behaviour of the asset price and defining indicators which facilitate identifying the direction of the current price trend (Murphy, 1999). Combinations of the technical and fundamental indicators can create potential trading rules. Identifying potential patterns in the price time series is a nontrivial task due to the uncertainties involved in the price values. The financial forecasting techniques used in the literature can be categorized as statistical models and artificial intelligence (AI) models. The statistical models, such as ARIMA and GARCH, are affected by the assumption of linear correlation structure for time-series values. AI techniques are effectively used in the development of algorithmic trading strategies (algo-trading). An algorithmic trading strategy is software that places automated orders to make a profit (Hsu, Lessmann, Sung, Ma, & Johnson, 2016). Algo-trading is designed to optimize pre-determined trading rules or generate new ones and optimize them afterwards. Several AI techniques have been effectively used on financial applications, including artificial neural networks (ANNs) (Grothmann, 2002; Zimmermann, Neuneier, & Grothmann, 2001; Chang, Liu, Lin, Fan, & Ng, 2009; Kamruzzaman & Sarker, 2003), evolutionary algorithms (EAs) (Aguilar-Rivera, Valenzuela-Rendón, & Rodríguez-Ortiz, 2015; Bauer, 1994; Chen & Yeh, 2001; Cliff, 1998; Dunis, Laws, & Karathanasopoulos, 2013; Hsu, 2011; Karatahansopoulos, Sermpinis, Laws, & Dunis, 2014; Leigh, Purvis, & Ragusa, 2002; Aloud, 2016; Aloud, 2017), and support vector machines (SVMs) (Tay & Cao, 2001; Kim, 2003; Lu & Wu, 2009). The agent-based approach is an effective technique to develop algo-trading designed for decision support systems. These AI techniques use a set of data points to identify patterns to forecast price movement. Although these techniques for forecasting trading signals are effective, they focus on constant values that restrict the time of trading, resulting in validity only for short-term trading.
The key to successful financial market forecasting is achieving the finest results with the minimum required input data and least complex algo-trading strategy. One of the key issues of AI learning processes designed for forecasting financial time series is selecting representative input data for the forecasting system. Data input selection is one of the critical steps in any machine-learning process, which is facilitated by selecting the most applicable data inputs and then reducing the dimensionality of the data. This will assure an improvement in the accuracy and efficiency of the designed algo-trading strategy.
The EAs and particularly genetic algorithms (GAs) for data input optimization are widely used as searching processes for an optimal subset of data inputs designed for a decision tree algorithm (Aguilar-Rivera, et al. 2015). The main issues with GAs are the time utilization and the random selection of the initial population of candidates for the decision tree algorithm. For these reasons, improvement of GAs for the selection of input data must be considered. For instance, all individuals of a population have the same stop criteria. As a result, the algorithm fails to provide a fitting solution. Different machine-learning approaches have been used to improve GAs using the results of a filter algorithm to optimize the set of input data (Hayward, 2006; Versace, Bhatt, Hinds, & Shiffer, 2004).