An intelligent quantitative trading system based on intuitionistic-GRU fuzzy neural networks
Introduction
Quantitative trading refers to substituting mathematical models and statistical methods for human subjective judgments. By using computer technology, a variety of “high probability” events can be selected from huge historical data, where they bring excess returns to make strategies, greatly reduce the impact of investor sentiment fluctuations and avoid irrational investment decisions in the extremely fanatical or pessimistic market. Through analyzing market observations, quantitative finance utilizes mathematical models to search for subtle patterns and inefficiencies in financial markets to improve prospective profits [1]. Meanwhile, accompanied with the development of machine learning, in recent years, abundant of results related to quantitative trading based on intelligent algorithms have been proposed. For instance, in 2018, Tsai et al. [2] used the characteristics of deep learning to imitate the intuitive conclusions of human in the context of trading charts. The implementation of intelligent algorithms would provide an extensive perspective on the study of financial market and quantitative trading [3].
The appropriate feature extraction of financial data is essential for quantitative trading [4], where it is based on the process of analyzing the signals of financial time series. In 1998, Huang et al. [5] proposed the Hilbert–Huang transformation, whose core lies in empirical mode decomposition (EMD) and Hilbert transformation with the characteristics of both time domain and frequency domain. EMD decomposition can decompose complex signals into a finite number of intrinsic mode functions (IMFs), which have good Hilbert transformation and thus can well reflect the local characteristics of the original data. Among them, EMD is a data-driven adaptive algorithm with local feature expressiveness, which can effectively reflect the physical characteristics of the system itself. It is a suitable method for dealing with nonlinear and non-stationary time series. Through decomposition, the embedded structure can be clearly identified with high efficiency. Many studies also show that the financial data is a nonlinear and non-stationary time series with complex behavior [6], so EMD technology has been widely applied in the field of financial research. Specifically, EMD technology has been applied to oil price analysis [7], foreign exchange rate analysis [8] and stock market analysis [9]in financial fields.
Fuzzy Sets, introduced by Zadeh [10], can be used to model nonlinear and uncertain information, so as to capture human experience or knowledge. Fuzzy logic is a conceptual system of reasoning and computation in which objects of discourse and analysis are allowed to be associated with uncertain information. Uncertain information refers to inaccurate, incomplete, or fuzziness in one or more respects. Fuzzy sets theory lays a foundation for dealing with such kinds of data, and has been widely applied in many fields, such as fuzzy reasoning [11], fuzzy logic [12] and fuzzy control [13]. Moreover, in order to solve uncertainty inherent in practical applications, Zadeh further introduced the concept of type-2 fuzzy sets in 1975 [14]. As a representation of information particles, type-2 fuzzy sets can solve the problem of language ambiguity and data noise well. Compared with the type-1 fuzzy sets, it can greatly reduce the uncertainty influence of rule-based fuzzy logic system and deal with fuzzy problems to a greater extent. The fuzzy membership level enables the type-2 fuzzy sets to adapt flexibly to the uncertain environment. This flexibility provides a decision boundary and is very similar to human decision making, so that class objects can have a gradual transition from member to non-member rather than a sudden transition. Therefore, type-2 fuzzy sets have attracted extensive attentions [15], [16], [17].
In 1986, Atanassov extended the concept of fuzzy sets to intuitionistic fuzzy sets (IFS) [18], dealing with uncertainty by considering the degree of membership and non-membership of element into a fuzzy set, as well as additional uncertainty (hesitation degree). By involving the hesitation degree to describe the fuzzy property of neutral state, more ability can be achieved to capture sufficient knowledge in uncertain facts and imprecise data than ordinary fuzzy sets. Furthermore, in 1989, Atanassov and Gargov proposed the concept of interval-valued intuitionistic fuzzy sets [19]. Intuitionistic fuzzy gradually developed into a systematic research field, and began to be widely applied [20], [21], [22]. In recent years, intuitionistic fuzzy sets are still widely used in various fields of research. For example, in 2018, Yazdi used the intuitionistic fuzzy hybrid TOPSIS method to evaluate risk factors, severity and likelihood parameters [23]. In the same year, Vishraj et al. used an intuitionistic fuzzy domain level set method to automatically identify near-pleural lung nodules in chest CT images [24]. In 2019, Zhan et al. combined an intuitionistic fuzzy set and a rough set to construct an intuitionistic fuzzy rough framework for solving decision problems [25]. It has been shown that intuitionistic fuzzy sets has a good ability to express uncertainty and achieve reasoning.
Compared with feedforward neural networks, recurrent neural networks (RNNs) are more effective for sequential learning tasks. However, due to the problem of gradient disappearance, RNNs still suffer for long-term time-dependence. In 1997, Hochreiter and Schmidhuber jointly proposed long short-term memory (LSTM) neural networks, which applied the gate mechanism to learn the long-term information in the sequential data [26]. Due to the good results of LSTM for sequential learning tasks, it is widely used in the fields of language processing [27], prediction system [28], pattern recognition [29] and anomaly detection [30]. Furthermore, in 2014, Cho et al. proposed the gated recurrent unit (GRU), which is a variant of LSTM and can simplify the LSTM structure while maintaining the similar function [31]. In 2018, Zhang et al. proposed a deep neural network combining convolutional GRU to identify hate speech in social media [32]. In 2019, Song et al. designed a GRU-based recurrent neural network 2D LiDAR map prediction to achieve the purpose of robot navigation and path planning [33]. In the same year, Zhou et al. applied GRU technology to Air Pollutant Concentration Prediction [34].
Meanwhile, from the perspective of the financial applications, various prediction models based on neural networks have been studied. For instance, in 2019, Gu et al. [35] proposed a LSTM-RNN model for the price forecasts and financial trading, where the empirical analysis of futures trading verified the practice of the proposed model in terms of risked return. In 2020, Yu et al. [36] proposed a prediction model based on deep neural networks using LSTMs for deep learning, by using which stock prices were predicted. Subsequently, Sermpinis et al. [37] carried out the comparative studies by using multi-layer perceptions, radial basis functions, higher order neural networks and recurrent neural networks for the forecasting and trading in stock markets. And, they noted that RNNs presented the stronger profitability compared to other neural network counterparts. Therefore, RNN-based (LSTM-based) models with their advantages of sequential learning ability are most commonly used in the prediction of financial time series. However, the existing works have not effectively dealt with the characteristics of high noise and large fluctuation of financial time series, which would weaken the practicability of prediction models.
Motivated by the above discussion, this paper proposes an interval type-2 intuitionistic fuzzy neural network with GRU mechanism. In the data preprocessing stage, EMD is used to decompose financial data into a series of intrinsic mode functions containing different time scales, based on which small fluctuations are smoothed out and the main trends of data are extracted. Subsequently, a reasoning system by using interval intuitionistic type-2 fuzzy neural network is constructed, where intuitionistic fuzzy set is involved to enhance the ability to deal with uncertainty and fuzzy information. On this basis, in order to establish the temporal relations in financial data, GRU mechanism is further introduced into reasoning system, which is helpful to solve the problem of long-range timing dependence. The learning process of the reasoning system is regulated by metacognitive algorithm, which is embodied in deleting samples similar to network knowledge, adding new rules to the network and updating network parameters according to the knowledge in the samples. The use of metacognitive learning algorithms improves the ability of reasoning systems to capture the underlying data distribution and avoids overtraining. Based on the predicted results, a concise trading strategy can be applied to carry out the trading, where some parameters of model can be adaptively adjusted according to the trading results.
The main contributions of this paper include: (1) the mechanism of gated recurrent unit is first involved into intuitionistic fuzzy neural network, where uncertain financial time series with the long-term time-dependence can be suitably modeled. Moreover, the interval type-2 fuzzy set can further improve the expression ability of fuzzy information. (2) EMD mechanism is used to capture the non-trivial trends of the financial data, so as to enhance the anti-noise ability of the system; (3) in the trading process, the parameters of the proposed model can be adaptively adjusted according to the trading results, which can effectively improve the model performance and reduce the time complexity of model updating.
The rest of this paper is arranged as follows: In the second part, we introduce the main techniques used in the system construction process. In the third part, we will introduce the data preprocessing process. The fourth part will introduce the main body of reasoning system. In the fifth part, the trading system is explained and the evaluation criteria are determined. The sixth part is verified. The seventh part provides the conclusions and discusses the future research direction.
Section snippets
Preliminaries
In this article, EMD technology and intuitionistic fuzzy sets are involved to construct the system. Here, we will briefly summarize the basic techniques used.
Data preprocessing
In the trading system proposed in this article, the flow of data processes can be shown in Fig. 1: Firstly, the financial data is preprocessed to extract the main trend features by using EMD. Secondly, the processed data are put into the prediction system and the coming data can be forecasted. Finally, by using the results obtained from prediction system, trading signals can be obtained.
Compared with wavelet transform (WT) and other time-frequency analysis techniques, EMD does not need to
The prediction system
In this section, the GRU based interval type-2 intuitionistic fuzzy system (IT2IFS-GRU) is established to implement the prediction of financial data. Due to the characteristics of high noises and large fluctuation, an online learning method using metacognitive algorithm is designed to process data, where metacognitive method can be seen as a kind of self-regulatory learning mechanism that controls the learning process, by deciding what-to-learn, when-to-learn and how-to-learn from sequential
Trading strategies
In this section, we will further explain how to apply the predicted results for the trading process. By comparing the forecasting value with the real value at time t, a forecasting spread is obtained, which is expressed as: where represents the predicted value and represents the true value at time t. When , the financial market will be a bullish trend, when , the financial market will be a bearish trend. In order to reduce worthless trading signals,
Experiments
In this section, the performance of the proposed IT2IFS-GRU is evaluated using real-world financial data. The data sets used are from the Chinese futures markets and foreign exchange markets.
Summary
This paper proposes an intelligent quantitative trading system, which is composed of EMD for the data preprocess, IT2IFS-GRU system for the prediction and a concise trading strategy. Since financial data has the characteristics of high noises and large fluctuations, the smoothness of the original data sequence is implemented by removing the high-frequency IMFs obtained by EMD, which can reduce the noises and outliers as well as retaining the main trends. The prediction system is based on the
CRediT authorship contribution statement
Yuan Wang: Conceptualization, Methodology, Software, Data curation, Writing - original draft. Chao Luo: Writing - review & editing, Supervision.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This research is supported by the National Natural Science Foundation of China (Nos: 61402267); Shandong Provincial Natural Science Foundation, China (ZR2019MF020).
References (40)
- et al.
Combining multiple feature selection methods for stock prediction: union, intersection, and multi-intersection approaches
Decis. Support Syst.
(2010) - et al.
A hybrid ARIMA and support vector machines model in stock price forecasting
Omega
(2005) - et al.
Estimating the impact of extreme events on crude oil price: An EMD-based event analysis method
Energy Econ.
(2009) - et al.
Empirical mode decomposition-based least square support vector regression for foreign exchange rate forecasting
Econ. Model.
(2012) - et al.
A novel time-series model based on empirical mode decomposition for forecasting TAIEX
Econ. Model.
(2014) Fuzzy sets
Inf. Control
(1965)The concept of a linguistic variable and its application to approximate reasoning
J. Inf. Sci.
(1975)- et al.
Some properties of fuzzy sets of type 2
Inf. Control
(1976) - et al.
Fuzzy sets and type 2 under algebraic product and algebraic sum
Fuzzy Sets and Systems
(1981) Intuitionistic fuzzy sets
Fuzzy Sets and Systems
(1986)
Interval valued intuitionistic fuzzy sets
Fuzzy Sets and Systems
Risk assessment based on novel intuitionistic fuzzy-hybrid-modified TOPSIS approach
Saf. Sci.
Quantcloud: Big data infrastructure for quantitative finance on the cloud
IEEE Trans. Big Data
Predict forex trend via convolutional neural networks
J. Intell. Syst.
Mastering R for Quantitative Finance
The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis
Proc. A
The Concept of a Linguistic Variable and Its Application To Approximate Reasoning. Learning Systems and Intelligent Robots
Fuzzy logic and its application to switching systems
IEEE Trans. Comput.
Application of fuzzy algorithms for control of simple dynamic plant
Proc. IEE
On the algebraic structure of fuzzy sets of type 2
Kybernetika
Cited by (23)
Development of technology predicting based on EEMD-GRU: An empirical study of aircraft assembly technology
2024, Expert Systems with ApplicationsA multi-model approach to the development of algorithmic trading systems for the Forex market
2024, Expert Systems with ApplicationsDCARS: Deep context-aware recommendation system based on session latent context
2023, Applied Soft ComputingFinancial trading decisions based on deep fuzzy self-organizing map
2023, Applied Soft ComputingCitation Excerpt :Niu et al. [7] propose a random data-time effective function and combine it with RBF neural network for financial time series forecasting. There are also neural networks such as RNN [8], LSTM [9], GRU [10], etc. These deep reinforcement learning based on recurrent neural networks can also extract useful features in complex financial data, and at the same time, their memory function can be used to learn the laws of data evolution in multidimensional financial derivatives.
Recurrent attention unit: A new gated recurrent unit for long-term memory of important parts in sequential data
2023, NeurocomputingCitation Excerpt :In this section, we mainly review the related work from two aspects: the GRU model and the attention mechanism. GRU is a powerful learning model for sequential data and has been widely applied in many areas, such as neural machine translation [20,3], traffic prediction [21,22], military [23] and financial [24]. Fig. 1 illustrates the architecture of the GRU cell, which includes two gates: the update gate and reset gate.