Contents
PATTERN RECOGNITION IN FOREX RATES USING NEURAL NETWORKS
by
Arijit R. Sarkar
U. Jayakrishnan
Abstract:
This presentation deals with the use of neural networks as a tool for pattern recognition, especially when applied to something as volatile as foreign exchange rates. We look at the nature of Artificial Neural Networks (ANN), and how they can be trained to recognize patterns, and interpolate/extrapolate data streams, more efficiently than conventional regression techniques. Specifically, we look at the kind of algorithms used while training a network, with focus on the Back-Propagation Algorithm. As an application of the pattern recognition capabilities of Artificial Neural Networks, we examine a neural network developed to forecast Foreign Exchange Rates. We look at the various factors affecting the Foreign Exchange Rates, and how these can be fed into a trained neural network to obtain predictions of future Foreign Exchange Rates. Finally, we examine presently available commercial software which use this principle, their drawbacks, and scope for further improvement in this field using wavelets.Introduction:
Artificial Neural Networks (ANN) are inspired by the high level of information processing capabilities observed in a human brain. Despite technological breakthroughs which have immensely increased the speed of conventional digital processing, as in computers, many tasks can be performed with a much greater speed and efficiency by a human brain. Taking motivation from this, neural networks are attempts to recreate these amazing properties of the brain, albeit on a smaller scale. Tasks such as image processing, pattern recognition and interpolation of data points are increasingly being entrusted to neural networks because of their robust, dynamic and efficient nature.Neural networks involve a paradigm shift in information processing ideas. They are basically models of the structural nature of a brain. Hence, their units, also called neuronal nodes, model actual neurons.
Neural networks always undergo a training period as such, when input data is fed to it, and adjustments made to improve the correlation between desired and obtained outputs. These adjustments are made to the synaptic weights, which are numbers associated with the interconnections between two neurons. Thus, the entire knowledge, or experience of a network, acquired during training, is stored in these interconnections (synapses). Neural networks, once trained, are assumed to give a minimum average error for the input data which was used to train it. However, trained neural networks have also been observed to be excellent at generalization, i.e. they yield fairly accurate results for interpolated and extrapolated data points. Hence, their outputs are reasonable even when faced with data inputs not previously encountered. This is one of the chief strengths of neural networks, i.e. their ability to handle cases for which they weren't programmed specifically.
A definition of Artificial Neural Networks which encompasses these details is as follows (ref. S. Haykin):
.An Artificial Neural Network (ANN) is a massively parallel distributed processor made up of simple processing units, which has a natural propensity for storing experiential knowledge and making it available for use. It resembles the brain in two respects:
1. Knowledge is acquired by the network from its environment through a learning process.
2. Inter neuron connection strengths, known as synaptic weights, are used to store the acquired knowledge.
Some advantages of neural networks are their non-linearity, adaptivity, implementability in both software and hardware, and especially their ease of implementation in VLSI. Coming to Foreign Exchange Rates (forex), we find that it is an area where application of neural networks is particularly apt. To be specific, we will be looking at the exchange rate between Indian Rupees and the U.S. Dollars. The major problem associated with forecasting the exchange rates is to decide exactly which factors affect it. In the present Indian context, this is an extremely tough decision. We can try and quantify some of these factors, such as interest rates, inflation rates, GDP, forex reserves, oil prices, etc. However, some of these cannot be quantified, such as fiscal and monetary policies, political scenarios, etc.
Making several assumptions about the constancy of most of these factors, we can arrive at a certain bouquet of critical variables. However, determining the output (tomorrow's forex rate) as a function of the current value of these variables is also a mammoth task. This is where neural networks enter the picture. With regards to accuracy of prediction, neural networks generally function much better than conventional regression techniques and numerical methods. This is chiefly due to the massively parallel nature of a neural network, with (typically) several layers of interconnections being used. As the number of synapses increases, the number of free parameters to be adjusted (synaptic weights) also increases. This yields better predictions and reduces errors. Now, let us look at the mathematical modeling of neural networks, and the kind of algorithms used to train these networks.
