relationship (Varian, 2014). Neural networks have some good features for a forecasting task. First, neural networks are data-driven self-adaptive methods different from the traditional model-based methods. Second, neural networks learn from data sets where the desired output is provided in advance. Hence, the neural networks learn by adjusting itself to find the right answer, increasing the accuracy of prediction. Finally, neural networks are non-linear. In fact, the linear model only provides limited support for large data set (Varian, 2014). It is difficult to convert a nonlinear model into a particular data set. Therefore, neural networks are nonlinear data-driven methods in comparison to the traditional model-based methods. They are able to accomplish nonlinear modeling without a priori knowledge about relationships between input and output (Zhang et.al, 1998). However, the performance of neural networks has been an important issue over the past few decades. Neural networks generally require tuning lots of parameters and consume the bulk of the processing time, especially the multi-layer neural networks.
With the advent of powerful hardware and deep learning frameworks, the learning speed performance of neural networks is improving a lot.
1.2 Motivation
Nowadays, people become more reliant on social media for accessing and sharing information. News, social media and forum have become a main source of information for many people when making decisions. Those data could provide invaluable forecasting insight in financial markets. Therefore, academic people and practitioners are wondering if the information and the discussions on the social media affect the daily movement of exchange rates. To this end, this study wants to explore the hypothesis that the relevant information in the news, the posts in forums and the discussions on the social
‧
media can really affect the movement of exchange rates. On the other hand, forecasting the movement of exchange rates has been a difficult task all the time that most econometric models are not capable of forecasting exchange rates with outstandingly higher accuracy than a naive random walk model. However, the variety, the volume, and the velocity of data on social media become bigger and bigger. Furthermore, the veracity of data on social media is an issue. Therefore, to justify the aforementioned hypothesis needs to apply the techniques of big data analytic and artificial intelligence (AI) (Walker , 2014) and the challenge of such study is huge.
Recently, concepts of big data analytic and artificial intelligence (AI) have been widely utilized in almost every field, such as defense, education, business, medicine, transport system and so on. On the other hand, the exchange rate is the value of one nation’s currency in comparison to another. With the increasing international trades, the fluctuations in exchange rates have a tremendous impact on the economy. The exchange rate is influenced by many factors, such as inflation rates, interest rates, political stability and so on (Levinson, 2014).
The time series model has been widely discussed and valued by many experts and scholars in forecasting the exchange rate. The time series data is an ordered sequence of observations in chronological order. Many financial data, such as unemployment rate can be thought of as time series data. The popular time series model is known as the autoregressive integrated moving average (ARIMA) model (Zhang, 2003).
Owing to the data is getting bigger and bigger, we may need some powerful tools such as big data analytics, artificial neural networks (ANN), deep learning and so on to deal with the more complex relationships. As a result, more and more economists are interested in the applications of big data analysis with machine learning (Varian, 2014)
‧
since they may provide effective ways to deal with more complex relationship (Zhang, 2003).
Big data used to mean a large of structured and unstructured data, which is very large and difficult to process using traditional database and software techniques. In general, big data is often characterized by its volume, variety, velocity and veracity.
Volume is how much data we have - the data measured in petabytes is now in zettabytes or even more. We can collect data from all kinds of sources, including social media, scientific instruments, mobile devices, Internet of Things and so on. Variety means that there are so many different types of data from textual to numerical to video. Velocity refers to the increasing speed at which data are generated and must be coped with in a timely manner. Veracity is about ensuring that the data is accurate, which requires a process to prevent wrong data from accumulating in system. The data is worthless if it is not accurate. Different from the conventional approach, big data analytics have already changed the way we process and manage data (Walker , 2014).
Over the coming decades, AI is the future trend around the world. Machine learning is a particular approach to AI. Machine learning techniques have increased the ability for computer to recognize patterns in big data. It also has the ability to learn from data and make prediction based on information. However, for instance, the performance of ANN has been an important issue over the past few decades. ANN generally consumes the bulk of the processing time and requires tuning lots of parameters. With the advent of more powerful hardware and deep learning frameworks recently, the performance of ANN has being improved.
In 2015, Google released its 2nd generation machine learning software, TensorFlow. It made easier for everyone to participate in machine learning. Not only
‧
algorithms, but also hardware makes a great progress. Besides the CPU, one of the most important parts in machine learning is the graphics processing unit (GPU). Both CPUs and GPUs can handle graphical operation, but GPUs have additional advantages over CPUs. First, a single GPU might have thousands of cores while a CPU usually has no more than multiple cores. Second, GPU designed for handling multiple tasks simultaneously. Finally, GPU is really fast at doing certain types of math calculations, particularly vector and matrix operations. To reinforce the performance of machine learning, Google even successfully developed a custom machine learning accelerator chip, the tensor processing unit (TPU) (Osborne, 2016).