Forecasting in the Forex market with interval time series using a bayesian approach

Abstract

Exchange rates forecasting (ERF) is a key element in monetary policy decisionmaking. A standard benchmark in ERF, the random walk model, is considerably difficult to beat; this phenomenon is known as the Meese and Rogoff puzzle. An interval time series (ITS) assigns to each time period an interval covering the values taken by the observed variable. Each interval has four characteristic attributes, since it can be defined in terms of lower and upper boundaries, centre and radius. The analysis and forecasting of ITS is a very young research area, dating back less than 10 years, and it still presents a wide array of open issues such as the Bayesian approach to ITS forecasting. When dealing with FOREX time series one approach is to consider them as classic time series (CTS). The other one is to proceed with some kind of aggregation and get a symbolic data time series such as an ITS. This paper proposes to consider the use of Bayesian methods when forecasting in the FOREX market. In particular, to address the suitability of interval Bayesian neural networks to the forecasting of the EUR/USD exchange rate, and give an account as to their competitiveness compared to other neural network forecasting models. The relatively good performance of this framework will be compared to the random walk and to the non-Bayesian neural networks for ITS. Further research issues will be proposed.