Exchange rate forecasting

Abstract

Forecasting of exchange rates between currencies is of utmost importance in the nancial world because its implication on imports-exports, trading and world economy in general. Since exhange rates data is a sequential data, modelling was traditionally done by time series analysis. However, forecasting using even non-linear time series models like the ARCH and the GARCH model did not give better forecasts in comparison to the simplest of the models , the random walk or the AR(1) model. However, with a growing interest in arti cial neural networks modelling since 1980s, they were also increasingly used for time series modelling. The neural network models gave a better forecast than the time series models used. Attempts were then made to combine the two modelling techniques to see if a hybrid model would perform better than either of the two models. With the advent of newer optimisation techniques like genetic algorithms in machine learning, these were incorporated as well to build new models. There was also an attempt to see how newer mathematical constructs like the fuzzy logic could be used for building an arti cial neural networks (Jang, 1993). By 2006, Hinton had proposed a new probabilistic model for data modelling which was called the deep belief network. Time series modelling done using these models gave better forecasts in comparison to even the xed geometry neural network models. In this thesis, I have attempted to study the theoretical basis behind such model and combine the forecasts of all the models used using information theoretic averaging. This is done to study whether an average of the forecasts gives a better forecast value than the individual models.