nach oben

2007 | Buch

Kapitel lesen Erstes Kapitel lesen

Foreign-Exchange-Rate Forecasting With Artificial Neural Networks

verfasst von: Lean Yu, Shouyang Wang, Kin Keung Lai

Verlag: Springer US

Buchreihe : International Series in Operations Research & Management Science

Enthalten in: Springer Professional "Wirtschaft+Technik" , Springer Professional "Technik" , Springer Professional "Wirtschaft"

Einloggen, um Zugang zu erhalten

Über dieses Buch

The foreign exchange market is one of the most complex dynamic markets with the characteristics of high volatility, nonlinearity and irregularity. Since the Bretton Woods System collapsed in 1970s, the fluctuations in the foreign exchange market are more volatile than ever. Furthermore, some important factors, such as economic growth, trade development, interest rates and inflation rates, have significant impacts on the exchange rate fluctuation. Meantime, these characteristics also make it extremely difficult to predict foreign exchange rates. Therefore, exchange rates forecasting has become a very important and challenge research issue for both academic and ind- trial communities. In this monograph, the authors try to apply artificial neural networks (ANNs) to exchange rates forecasting. Selection of the ANN approach for - change rates forecasting is because of ANNs’ unique features and powerful pattern recognition capability. Unlike most of the traditional model-based forecasting techniques, ANNs are a class of data-driven, self-adaptive, and nonlinear methods that do not require specific assumptions on the und- lying data generating process. These features are particularly appealing for practical forecasting situations where data are abundant or easily available, even though the theoretical model or the underlying relationship is - known. Furthermore, ANNs have been successfully applied to a wide range of forecasting problems in almost all areas of business, industry and engineering. In addition, ANNs have been proved to be a universal fu- tional approximator that can capture any type of complex relationships.

Inhaltsverzeichnis

Frontmatter

Forecasting Foreign Exchange Rates with Artificial Neural Networks: An Analytical Survey

Frontmatter

1. Are Foreign Exchange Rates Predictable? — A Literature Review from Artificial Neural Networks Perspective

After more than two decades of research since Meese and Rogoff’s seminal work on exchange rates predictability (see Meese and Rogoff, 1983a, 1983b), the goal of exploiting foreign exchange rates forecasting model to beat naïve random walk forecasts remains as elusive as ever (Taylor, 1995) due to the fact that evidence supporting or refuting the exchange rate predictability seems plausible.

The remainder of the chapter is organized as follows. In next section, we explain how articles were selected. Section 1.3 examines and analyzes 45 articles in detail and investigates some main factors that affect the performance of the ANNs. Subsequently, some implications and future research directions are also pointed out in Section 1.4. Finally, Section 1.5 concludes the paper.

Basic Learning Principles of Artificial Neural Networks and Data Preparation

Frontmatter

2. Basic Learning Principles of Artificial Neural Networks

Artificial neural networks (ANNs), as an emerging discipline, studies or emulates the information processing capabilities of neurons of the human brain. It uses a distributed representation of the information stored in the network, and thus resulting in robustness against damage and corresponding fault tolerance (Shadbolt and Taylor, 2002). Usually, a neural network model takes an input vector X and produces output vector Y. The relationship between X and Y is determined by the network architecture. There are many forms of network architecture inspired by the neural architecture of the human brain.

In the neural network model, it is widely accepted that a three-layer back propagation neural network (BPNN) with an identity transfer function in the output unit and logistic functions in the middle-layer units can approximate any continuous function arbitrarily well given a sufficient amount of middle-layer units (White, 1990). Furthermore, in the practical applications, about 70 percent of all problems are usually trained on a three-layer back-propagation network, as revealed by Chapter 1. The backpropagation learning algorithm, designed to train a feed-forward network, is an effective learning technique used to exploit the regularities and exceptions in the training sample.

A major advantage of neural networks is their ability to provide flexible mapping between inputs and outputs. The arrangement of the simple units into a multilayer framework produces a map between inputs and outputs that is consistent with any underlying functional relationship regardless of its “true” functional form. Having a general map between the input and output vectors eliminates the need for unjustified priori restrictions that are needed in conventional statistical and econometric modeling. Therefore, a neural network is often viewed as a “universal approximator” i.e. a flexible functional form that can approximate any arbitrary function arbitrarily well, given sufficient middle-layer units and properly adjusted weights.

3. Data Preparation in Neural Network Data Analysis

Preparing data is an important and critical step in neural network data analysis and it has an immense impact on the success of a wide variety of complex data analysis, such as data mining and knowledge discovery (Hu, 2003). The main reason is that the quality of the input data into neural network models may strongly influence the results of the data analysis (Sattler and Schallehn, 2001). As Lou (1993) stated, the effect on the neural network’s performance can be significant if important input data are missing or distorted. In general, properly prepared data are easy to handle, which makes the data analysis task simple. On the other hand, improperly prepared data may make data analysis difficult, if not impossible. Furthermore, data from different sources and growing amounts of data produced by modern data acquisition techniques have made data preparation a time-consuming task. It has been claimed that 50–70 percent of the time and effort in data analysis projects is required for data preparation (Sattler and Schallehn, 2001; Pyle, 1999). Therefore, data preparation involves enhancing the data in an attempt to improve the performance of data analysis.

Individual Neural Network Models with Optimal Learning Rates and Adaptive Momentum Factors for Foreign Exchange Rates Prediction

Frontmatter

4. Forecasting Foreign Exchange Rates Using an Adaptive Back-Propagation Algorithm with Optimal Learning Rates and Momentum Factors

Foreign exchange rates prediction is regarded as a rather difficult and challenging task due to its high volatility and noisy market environment (Yu et al., 2005c). Usually, the difficulty in forecasting exchange rates is attributed to the limitation of many conventional forecasting models as many statistical and econometric models are unable to produce significantly better predictions than the random walk (RW) model (Ince and Trafalis, 2005; Chen and Leung, 2004), which has also encouraged academic researchers and business practitioners to develop more predictable forecasting models.

The rest of this chapter is organized as follows. In Section 4.2, an adaptive BP learning algorithm with optimal learning rate and momentum factor is first proposed in terms of the gradient descent rule and optimization techniques. For verification and illustration, Section 4.3 gives the experiments about three foreign exchange rates prediction and reports their results. Finally, conclusions and future directions are given in Section 4.4.

5. An Online BP Learning Algorithm with Adaptive Forgetting Factors for Foreign Exchange Rates Forecasting

The foreign exchange market is a complex, evolutionary, and nonlinear dynamical system. Foreign exchange rate series as a kind of financial time series are inherently noisy, non-stationary, and deterministically chaotic (Yaser and Atiya, 1996). This means that the distribution of foreign exchange rate series changes over time. Not only is a single data series non-stationary in the sense of the mean and variance of the series, but the relationship of the data series to other related data series may also be continuously changing (Yu et al., 2006b). Modeling such dynamical and non-stationary time series is a challenging task. Over the past few years, neural networks have been successfully applied to foreign exchange rates prediction and achieve promising results, as Chapters 1 and 4 indicated.

The rest of this chapter is organized as follows. In Section 5.2, the proposed online learning algorithm with adaptive forgetting factor is first presented in terms of the gradient descent algorithm and optimization techniques. For further illustration, an empirical analysis is then given in Section 5.3. Finally, some concluding remarks are drawn in Section 5.4.

6. An Improved BP Algorithm with Adaptive Smoothing Momentum Terms for Foreign Exchange Rates Prediction

Back-propagation neural network (BPNN) is one of the most popular neural networks which have been widely applied to many fields such as prediction and pattern recognition due to their strong capability to approximate any arbitrary function arbitrarily well and to provide flexible nonlinear mapping between inputs and outputs (Hornik et al., 1989; White, 1990). The basic learning rule of BPNN is based on the gradient descent optimization method and the chain rule (Widrow and Lehr, 1990). However, some typical drawbacks of the BPNN learning rule based on the gradient descent method are its slowness and its frequent confinement to local minima and over-fitting (Yu, 1992; Lawrence et al., 1997; Yu et al., 2006a). For these reasons, some global optimization algorithms, such as genetic algorithm (GA) (Jain et al., 1996) and simulated annealing (SA) (Karaboga and Pham, 2000), are proposed for escaping local minima.

In this chapter, we propose an improved BPNN learning algorithm with adaptive smoothing momentum. In this new algorithm, adaptive smoothing technique is used to adjust the momentum of weight updating formula automatically by tracking error signals in terms of “3σ limits theory”. For illustration and verification purposes, the proposed BPNN learning algorithm is applied to foreign exchange rates prediction.

The rest of this chapter is organized as follows. In Section 6.2, an improved BPNN learning algorithm with adaptive smoothing momentum terms is proposed in detail. In order to verify the effectiveness of the proposed algorithms, an exchange rate index prediction experiment of tradingweighted US dollar against currencies of major US trading partners is conducted and the corresponding results are reported in Section 6.3. In addition, Section 6.4 compares the proposed model with the two similar single neural network models proposed by Chapters 4 and 5. Finally, some concluding remarks are drawn in Section 6.5.

Hybridizing ANN with Other Forecasting Techniques for Foreign Exchange Rates Forecasting

Frontmatter

7. Hybridizing BPNN and Exponential Smoothing for Foreign Exchange Rate Prediction

A challenging task in financial market such as stock market and foreign exchange market is to predict the movement direction of financial markets so as to provide valuable decision information for investors (Lai et al., 2006b). Thus, many researchers and business practitioners have developed various kinds of forecasting methods. Of the various forecasting models, the exponential smoothing model has been found to be an effective forecasting method. Since Brown (1959) began to use simple exponential smoothing to forecast demand for inventories, the exponential smoothing models have been widely used in business and finance (Winters, 1960; Gardner, 1985; Alon, 1997; Leung, 2000). For example, Winters (1960) proposed exponentially weighed moving averages for sales forecasting and obtained good results. Gardner (1985) introduced exponential smoothing methods into supply chain management for forecasting demand, and achieved satisfactory results. Similarly, Alon (1997) found that Winters’ model forecasts aggregate retail sales more accurately than the simple exponential model. Leung et al. (2000) used an adaptive exponential smoothing model to predict stock indices, and achieved good forecasting performance for Nikkei 225.

In this chapter, a hybrid synergy model integrating both an exponential smoothing (ES) model and a BPNN model is proposed to take advantage of the unique strength of exponential smoothing and BPNN models in linear and nonlinear modeling. For testing purposes, two main exchange rates, EUR/USD and JPY/USD, are used. For comparison, individual exponential smoothing model and individual BPNN model are used as benchmark models.

The remainder of the chapter is organized as follows. Section 7.2 provides basic backgrounds about the exponential smoothing and neural network approaches to financial time series forecasting. Then the hybrid methodology combining the exponential smoothing and neural network model is introduced in Section 7.3. Subsequently, some experimental results are reported in Section 7.4. Finally, Section 7.5 concludes the study.

8. A Nonlinear Combined Model Hybridizing ANN and GLAR for Exchange Rates Forecasting

Foreign exchange rates modeling and forecasting has been a common research stream in the last few decades. Over this time, the research stream has gained momentum with the advancement of computer technologies, which have made many elaborate computation methods available and practical (Yu et al., 2005c). However, it is not easy to predict exchange rates due to their high volatility and noise. But the difficulty in forecasting exchange rates is usually attributed to the limitation of many conventional forecasting models; this has encouraged academic researchers and business practitioners to develop more predictable forecasting models. As a result models using artificial intelligence such as artificial neural network (ANN) techniques have been recognized as more useful than conventional statistical forecasting models. Literature documenting the research shows this is quite diverse and involves different architectural designs. Some examples are presented. De Matos (1994) compared the strength of a multilayer feed-forward neural network (MLFNN) with that of a recurrent network based on the forecasting of Japanese yen futures. Kuan and Liu (1995) provided a comparative evaluation of the performance of MLFNN and a recurrent network on the prediction of an array of commonly traded exchange rates. Hsu et al. (1995) developed a clustering neural network model to predict the direction of movements in the USD/DEM exchange rate. Their experimental results suggested that their proposed model achieved better forecasting performance relative to other indicators.

The rest of the chapter is organized as follows. The next section describes the model building process in detail. In order to verify the effectiveness and efficiency of the proposed model, empirical analysis of the three main currencies’ exchange rates is reported in Section 8.3. The conclusions are contained in Section 8.4.

9. A Hybrid GA-Based SVM Model for Foreign Exchange Market Tendency Exploration

Forecasting foreign exchange rates has been regards as one of the most challenging application of modern time series forecasting (Yu et al., 2005f ). Thus, numerous models have been developed to provide the investors with more precise predictions. Recently, artificial intelligence, such as artificial neural networks (ANN), has been widely applied to solve foreign exchange rates forecasting problems. Literature documenting the research shows the ANN has a powerful capability to predict foreign exchange rates. The previous chapters provide some literature reviews. Interested readers can be referred to the Chapter 1 and other chapters for more details about foreign exchange rates prediction. Two recently good surveys about foreign exchange rates prediction with ANN can be referred to Huang et al. (2004a, 2006) and Yu et al. (2005e) for more literature review.

The main motivation of this chapter is to propose a new hybrid intelligent data mining approach integrating SVM with GA for exploring foreign exchange market tendency and to test the predictability of the proposed hybrid intelligent model by comparing it with statistical models and neural network models. The rest of the chapter is organized as follows. The next section will describe the formulation process of the proposed hybrid intelligent data mining model in detail. In Section 9.3, we give an experiment scheme and report the experimental results. For comparison, the similarities and difference of three hybrids model proposed by this part are presented in Section 9.4. And Section 9.5 concludes this chapter.

Neural Network Ensemble for Foreign Exchange Rates Forecasting

Frontmatter

10. Forecasting Foreign Exchange Rates with a Multistage Neural Network Ensemble Model

Some studies revealed that foreign exchange market is one of the most volatile markets. Due to its high volatility, foreign exchange rates forecasting is regarded as a rather challenging task (Yu et al., 2005c). For traditional statistical methods, it is hard to capture the volatility. In the last decades, many emerging artificial intelligent techniques, such as artificial neural networks (ANN), were widely used in foreign exchange rates forecasting and obtained good prediction performance.

The rest of this chapter is organized as follows. The next section explains the reasons of motivating neural network ensemble for prediction problem. In Section 10.3, we describe the building process of the multistage neural network ensemble forecasting model in detail. For further illustration, two foreign exchange rate series are used for testing in Section 10.4. Finally, some concluding remarks are drawn in Section 10.5.

11. Neural Networks Meta-Learning for Foreign Exchange Rate Ensemble Forecasting

Artificial neural network (ANN), first introduced by McCulloch and Pitts (1943), is a system derived through neuropsychology models (Hertz, 1989). It attempts to emulate the biological system of the human brain in learning and identifying patterns. Moreover, ANNs can more aptly recognize poorly defined patterns. Instead of extracting explicit rules from sample data, the ANN employs a learning algorithm to automatically: (a) extract the functional relationship between input and output, which is embedded in a set of historical data (called training exemplars or learning samples), and (b) encode it in connection weights. Training exemplars that are readily available allow neural networks to capture a large volume of information in a rather short period of time and to continuously learn throughout their lifespan. Furthermore, neural networks have the ability to not only deal with noisy, incomplete, or previously unseen input patterns, but to also generate a reasonable response (Tsaih et al., 1998). However, ANN is far from being optimal learner. For example, the existing studies (e.g., Breiman (1999)) have found that the ways neural networks have of getting to the global minima vary and some networks just settle into local minima instead of global minima through the analysis of error distributions. In this case, it is hard to justify which neural network’s error reaches the global minima if the error rate is not zero. Thus, it is not wise choice that only selecting a single neural network model with the best generalization from a limited number of neural networks if the error is larger than zero.

The main motivation of this study is to take full advantage of the inherent learning capability of meta-learning technique and to design a powerful neural network ensemble learning model. The rest of this chapter is organized as follows. Section 11.2 gives a brief introduction of neural network learning paradigm. In Section 11.3, a neural-network-based meta-learning process is provided in detail. To verify the effectiveness of the proposed meta-learning technique, an exchange rate prediction experiment is performed in Section 11.4. Finally, Section 11.5 concludes this chapter.

12. Predicting Foreign Exchange Market Movement Direction Using a Confidence-Based Neural Network Ensemble Model

Neural network ensemble has been turned out to be an efficient strategy for achieving high classification performance, especially in fields where the development of a powerful single classifier system requires considerable efforts. Usually, neural network ensemble model outperforms the individual neural network models, whose performance is limited by the imperfection of feature extraction, learning/classification algorithms, and the inadequacy of training data. Due to these reasons, there is a growing research stream about neural network ensemble learning methods (Perrone and Cooper, 1993; Krogh and Vedelsby, 1995; Rosen, 1996; Tumer and Ghosh, 1996; Yang and Browne, 2004; Yu et al., 2005c, 2006d; Lai et al., 2006c, 2006d). For example, performance improvement can result from training the individual networks to be decorrelated with each other (Rosen, 1996) with respect to their errors. To achieve high classification performance, there are some essential requirements to the ensemble members and the ensemble strategy. First of all, a basic condition is that the individual neural network classifiers must have enough training data. Secondly, the ensemble members are diverse or complementary, i.e., classifiers show different classification properties. Thirdly, to obtain high performance, a wise ensemble strategy is also required on a set of complementary classifiers.

The motivation of this chapter is to formulate a multistage confidencebased neural network ensemble model for predicting foreign exchange market movement direction and compare its performance with other existing approaches. The rest of the chapter is organized as follows. The next section presents a formulation process of the multistage neural network ensemble model in detail. To verify the effectiveness of the proposed method, two real examples are performed in Section 12.3. For comparison, Section 12.4 reports the comparisons of three ensemble neural network models. Finally, some conclusions are drawn in Section 12.5.

13. Foreign Exchange Rates Forecasting with Multiple Candidate Models: Selecting or Combining? A Further Discussion

From Chapter 4 to Chapter 12, nine typical foreign exchange rates forecasting models, including three single neural network models, three neural network hybrid models and three neural network ensemble models, are proposed from the perspectives of level estimation and direction exploration. However, in the development process of forecasting models, we often have to come up against two important dilemmas (Yu et al., 2005g): (1) Whether should we select an appropriate modeling approach for prediction purposes or should combine these different individual approaches into an ensemble forecast for the different/dissimilar models? (2) Whether should we select the best candidate model for forecasting or to mix the various candidate models with different parameters into a new forecast for the same/similar modeling approaches?

The motivation of this chapter is how to deal with the two dilemmas of selecting and combining for time series forecasting (in particular, foreign exchange rates forecasting problem) and meantime propose a solution to the two dilemmas. The rest of the chapter is organized as follows. The next section describes the procedure for dealing with the two dilemmas of selecting and combining in detail. To verify the effectiveness of the proposed procedures, a typical foreign exchange rate experiment is performed in Section 13.3. And Section 13.4 concludes this chapter and points out some future research directions.

Developing an Intelligent Foreign Exchange Rates Forecasting and Trading Decision Support System

Frontmatter

14. Developing an Intelligent Forex Rolling Forecasting and Trading Decision Support System I: Conceptual Framework, Modeling Techniques and System Implementations

Decision support system (DSS) is a powerful tool which can be used to support decision makers in making strategic decisions (Quah et al., 1996). One key objective of DSS is to improve management judgment (Kean and Morton, 1978). From the viewpoint of management, all forecasts are considered as a prerequisite for effective decisions that are based upon planning (Armstrong, 1983). Effective foreign exchange (forex) trading decision is usually dependent upon effective forex forecasting. Tsoi et al. (1993) have shown that forex forecasting on direction is more important than the actual forecast itself in terms of determining the profitability of a forecasting and trading system. A more comprehensive survey on forex forecasting can refer to Yu et al. (2005e, Huang et al., 2004a, 2006) for more details. Financial market (e.g., stock market or foreign exchange market) is a rather complicated environment. Traders must predict market price movements in order to sell at high points and to buy at low points. Therefore, forecasting often plays an important role in the process of decision-making in financial market.

The reminder of this chapter is organized as follows. Section 14.2 describes the general framework architecture and main functions of the advanced intelligent system. Modeling approach and quantitative measurements used in the system are presented in Section 14.3. Subsequently, the development, implementation and operation of the intelligent system are discussed in Section 14.4. Section 14.5 concludes this chapter.

15. Developing an Intelligent Forex Rolling Forecasting and Trading Decision Support System II: An Empirical and Comprehensive Assessment

In the previous chapter, the system framework of the intelligent forex rolling forecasting and trading decision support system (FRFTDSS) was presented, and the FRFTDSS integrating a BPNN-based forex rolling forecasting system (BPNNFRFS) and a web-based forex trading decision support system (WFTDSS) was developed and implemented (Yu et al., 2006f). The main goals of this intelligent integrated system are to present short-term forex forecasting and make corresponding e-trading decisions for forex traders and various investors.

The reminder of this chapter is organized as follows. Section 15.2 describes assessment methods and presents assessment results. Sections 15.3 and 15.4 compare the performance of FRFTDSS with that of some classical forecasting models and other existing forex forecasting and trading decision support systems, respectively. Section 15.5 concludes the chapter.

Backmatter

Titel: Foreign-Exchange-Rate Forecasting With Artificial Neural Networks
verfasst von: Lean Yu
Shouyang Wang
Kin Keung Lai
Verlag: Springer US
Electronic ISBN: 978-0-387-71720-3
Print ISBN: 978-0-387-71719-7
DOI: https://doi.org/10.1007/978-0-387-71720-3