DescriptionThere has been increasing interest in the application of neural networks to the field of finance. Several experiments have been carried out stating the success of neural networks for time series prediction.
Most of the existing systems recommend single neural network architecture to be used for a particular time series. Our experiments have shown that a fixed architecture may not be the best approach across different time horizons. The thesis proposes a new methodology where multiple NARX (nonlinear autoregressive network with exogenous inputs) networks with different architectures are generated and evaluated before the beginning of a new time horizon. A network is selected from this set and employed to make predictions. This selection is based on past datasets only -- making the system completely applicable to real world scenarios.
A framework of functions was built in MATLAB® to customize the Neural Network Toolbox ® for financial applications. This framework provides for all the basic functions required by a financial neural network system. An adaptive system that uses technical indicators and some external time series as inputs was built. Different rules were developed and tested for selecting the best performing neural networks.
The new approach was tested on 5 currencies and the gold series. Our results show that high realized values of returns in the past, along with generalization is the best parameter to select a network for the future. A system with adaptive approach performs better than one with a fixed architecture. Our adaptive system out performed not only the fixed architectures but also other benchmarks like technical indicators, linear regression and baseline buy or sell strategies.